content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
In a seminal paper, ~\citet{BL1} introduced a class of inequalities (\emph{Brascamp-Lieb inequalities}) that generalizes many well-known inequalities, including H{\"o}lder's inequality, the sharp Young convolution inequality, and the Loomis-Whitney inequality. As such, the class of Brascamp-Lieb inequalities provides a valuable tool of use to many areas of mathematics~\citep{tao-paper}. Brascamp-Lieb inequalities have also been applied in Machine Learning~\citep{dvir2016rank,pmlr-v30-Hardt13} and Information Theory~\citep{carlen2009subadditivity,liu2016smoothing}.
Each Brascamp-Lieb (BL) inequality is characterized by a \emph{BL-datum} $(\mathcal{A},w)$, where $\mathcal{A} = \big( A_1, \dots, A_m \big)$ is a tuple of linear transformations and $w=\lbrace w_j \rbrace$ is a set of exponents. Of central importance is the feasibility of a BL-datum, which is characterized by the following quantity:
\begin{defn}[Brascamp-Lieb Constant]
Let $\mathcal{A} = \big( A_1, \dots, A_m \big)$ be a tuple of surjective linear transformations $A_j: \mathbb{R}^d \rightarrow \mathbb{R}^{d'}$ and $\bm{w}=(w_1,\ldots,w_m)$ a real vector with non-negative entries. For each BL-datum $(\mathcal{A},\bm{w})$ there exists a constant $C(\mathcal{A},\bm{w})$ (finite or infinite) such that for any set of real-valued, non-negative, Lebesgue-measurable functions $f_j: \mathbb{R}^{d'} \rightarrow \mathbb{R}$, the following (\emph{Brascamp-Lieb}) inequality holds
\begin{equation}\label{eq:BL-inequ}
\int_{x \in \mathbb{R}^d} \Bigl( \prod\nolimits_{j \in [m]} f_j (A_j x)^{w_j} \Bigr) dx
\leq C(\mathcal{A},\bm{w}) \prod\nolimits_{j \in [m]} \Bigl( \int_{x \in \mathbb{R}^{d'}} f_j (x) dx \Bigr)^{w_j} \; .
\end{equation}
The smallest such constant that satisfies~\eqref{eq:BL-inequ} for a BL-datum $(\mathcal{A},w)$ and for any choice of $\lbrace f_j \rbrace$, is called the \emph{Brascamp-Lieb constant} ${\rm BL}(\mathcal{A},\bm{w})$. We further say that the datum $(\mathcal{A},\bm{w})$ is \emph{feasible}, if ${\rm BL}(\mathcal{A},\bm{w}) < \infty$, and \emph{infeasible} otherwise.
\end{defn}
Our aim is to characterize Brascamp-Lieb constants of feasible data; in particular, we want to compute the value
\begin{align}\label{eq:BL-const}
\sup_{Z \in \; \prod_{j \in [m]} \mathbb{P}_{d'}} \biggl[{\rm BL}(\mathcal{A},\bm{w}; Z) := \biggl( \frac{\prod_{j \in [m]} \det(Z_j)^{w_j}}{\det \bigl( \sum_{j \in [m]} w_j A_j^* Z_j A_j\bigr)} \biggr)^{1/2} \biggr] \; .
\end{align}
Then, the key question is whether there exists a solution to the following problem:
\begin{prob}\label{prob:BL}
Given a datum $(\mathcal{A},\bm{w})$, can we efficiently compute an $\epsilon$-approximation to ${\rm BL}(\mathcal{A},\bm{w})$, in time that is polynomial in the number of bits required to represent the datum. That is, for any $\epsilon>0$, can we compute a constant $C$, such that
\begin{align*}
{\rm BL}(\mathcal{A},\bm{w}) \leq C \leq (1+\epsilon) {\rm BL}(\mathcal{A},\bm{w}),
\end{align*}
in time that is polynomial in the bit length of the datum $(\mathcal{A}, \bm{w})$ and $\log(1/\epsilon)$.
\end{prob}
\noindent We will see below that the computation of the supremum in~\eqref{eq:BL-const} is equivalent to solving
\begin{equation}\label{eq:F}
\min_{X \in \mathbb{P}_d} F(X) := \sum\nolimits_j w_j \log \det \big( A_j^\ast X A_j \big) - \log \det (X),
\end{equation}
where $\bm{w}=(w_1, \dots, w_n)$ with $w_j \in (0,1)$ and $\sum_j w_j =1$.
A seminal result by~\citet{lieb} ensures that the optimum is positive definite, which allows for solving~\eqref{eq:F} over the positive definite matrices ($\posdef_d$).
Importantly, problem~\eqref{eq:F} turns out to be a geodesically convex optimization problem on the manifold of Hermitian positive definite matrices (under a suitable Riemannian metric~\citep{vishnoi}).
Given that $F(X)$ in~\eqref{eq:F} is geodesically convex, it is natural to consider Riemannian optimization approaches. For using first-order methods such as Riemannian Gradient Descent~\citep{absil2009optimization,zhang}, geodesic convexity proves key to ensuring sublinear rates, such as $O \big(\frac{1}{\epsilon} \big)$, on \emph{function suboptimality} $F(X_k)-F(X^*) \le \epsilon$. We will see below that the optimum lies in a compact convex set; hence \eqref{eq:F} could also be viewed as a constrained optimization task with a geodesic ball constraint. We can exploit this constraint by using first-order methods such as Projected Riemannian Gradient Descent~\citep{absil2009optimization,zhang} or Riemannian Frank-Wolfe~\citep{weber2022riemannian,weber2021projection}, both of which again achieve only sublinear convergence rates on function suboptimality. Furthermore, for a different geodesically convex formulation of~(\ref{eq:BL-const}) due to~\citet{garg2018algorithmic}, it is known that second-order methods can attain a faster $O \bigl({\rm poly} \bigl( \log \frac{1}{\epsilon}\bigr)\bigr)$ convergence rate~\citep{allen-zhu}. Under additional spectral assumptions, even linear convergence at rate $O \bigl( \log \frac{1}{\epsilon}\bigr)$ is possible~\citep{kwok}.
\paragraph{Our approach.} We take a different route in this paper; our idea is to analyze a \emph{fixed-point iteration} that solves~(\ref{eq:F}). In particular, we study Picard iterations obtained by iterating a nonlinear map $G: \mathcal{X} \subseteq \mathbb{P}_d \rightarrow \mathcal{X}$, defined as
\begin{equation}\label{eq:map}
G: \quad X \mapsto \Big(\sum\nolimits_{j} w_j A_j \big( A_j^\ast X A_j \big)^{-1} A_j^\ast \Big)^{-1} \; .
\end{equation}
The simplicity of the map~(\ref{eq:map}) makes a fixed-point approach particularly attractive, since it avoids expensive Riemannian operations such as exponential maps and parallel transports, which are required by standard Riemannian optimization approaches. Empirically, iterating map~\ref{eq:map} exhibits linear convergence -- see Figure~\ref{fig:BL} for a numerical experiment. The faster convergence in comparison with Riemannian gradient descent illustrates the attractiveness of our approach.
\begin{figure}[t]
\centering
\includegraphics[scale=0.2]{BL-paper.png}
\caption{\small Empirical performance of iterating $G$ (\ref{eq:map}) in comparison with Riemannian Gradient Descent (using the Manopt toolbox~\citep{boumal_manopt_2014}). The Brascamp-Lieb datum consists of a set of $n$ linear transformations $A_j: \mathbb{R}^d \rightarrow \mathbb{R}^k$ and an $n$-dimensional vector that characterizes the exponents.}
\label{fig:BL}
\end{figure}
We will present a detailed non-asymptotic convergence analysis of a fixed-point approach based on iterating~\eqref{eq:map}. The key insight of our work is to analyze the nonlinear map $G$ through the lens of nonlinear Perron-Frobenius theory, which views $\posdef_d$ as a metric space endowed with the Thompson part metric.
Our initial convergence analysis hinges on establishing \emph{non-expansivity} of G with respect to the Thompson part metric, as well as \emph{asymptotic regularity} of G, i.e.,
$\delta_T \big( G(X_{k+1}), G(X_k) \big) \rightarrow 0$.
More precisely, we make use of the following result that adapts a classical result on the convergence of non-expansive maps in Banach spaces (see, e.g.,~\citet[Theorem 1.2]{baillon_asymptotic_1978}) to Thompson geometry:
\begin{theorem}\label{thm:s-h}
Let $G$ be non-expansive and asymptotically regular, and let
the set of its fixed points ${\rm Fix}(G) := \lbrace X^\ast \in \mathbb{P}_d: \; X^\ast = G(X^\ast) \rbrace$ be non-empty. Then the iterative procedure defined by $G$ converges to a point in ${\rm Fix}(G)$.
\end{theorem}
\noindent Our plan is as follows: We show that $G$ has a fixed point in a compact and convex set $D$.
Establishing the non-expansivity and asymptotic regularity of $G$ then ensures asymptotic convergence (following Theorem~\ref{thm:s-h}). Subsequently, we refine our convergence analysis to seek non-asymptotic guarantees that ensure an $O \big(\log \frac{1}{\epsilon} \big)$ convergence rate. To this end, we show that a regularized variant of $G$ is \emph{strictly} contractive with respect to the Thompson part metric. Finally, we describe how this analysis implies a convergence rate for solving the original problem~\eqref{eq:F}, completing the analysis.
\subsection{Summary of contributions}
The main contributions of this paper are as follows:
\begin{enumerate}
\setlength{\itemsep}{0pt}
\item We propose a fixed-point approach for computing Brascamp-Lieb constants that is obtained by analyzing the first-order optimality conditions of the geodesically convex problem~\eqref{eq:F}.
\item We analyze the proposed approach and derive non-asymptotic guarantees on its convergence. Specifically, we study a regularized variant of our fixed-point approach that solves the original problem~\eqref{eq:F} to $\epsilon$-accuracy with an iteration complexity of $O \Big( \log \big( \frac{1}{\epsilon},\frac{\Delta}{\delta^{3/2}} ,\sqrt{d}, \frac{1}{R}\big) \Big)$. Here, $d, \delta, \Delta, R$\footnote{Note that the constants $\delta, \Delta, R$ are not necessary independent of each other. In particular, to be more precise, we should write $R(\delta)$, to indicate that $R$ may depend on $\delta$. However, to keep notation simple, we write just $R$.} are constants, which depend on the input datum $(\mathcal{A},\bm{w})$ only.
\end{enumerate}
Our analysis leverages the Thompson part metric on the manifold of positive definite matrices to model convergence of the fixed-point iteration. To our knowledge, this is the first work that analyzes the computation of Brascamp-Lieb constants via Thompson geometry.
We note that a similar Finslerian lens can be employed to understand other Picard iterations arising from problem~\ref{eq:F}. We discuss this observation in more detail below.
Our proof techniques build on ideas from nonlinear Perron-Frobenius theory~\citep{lemmens_nussbaum_2012}, and may be of independent interest for related problems.
Moreover, the simplicity of our approach and its ease of implementation make it attractive for applications.
\section{Background and Notation}
\subsection{Thompson geometry on positive definite matrices}
We start by recalling some background and notation. Throughout, we consider optimization tasks on the manifold of positive definite matrices, i.e.,
\begin{align*}
\posdef_d := \lbrace X \in \mathbb{R}^{d \times d}: \; X \succ 0 \rbrace \; .
\end{align*}
Here, ``$\succ$'' denotes the Löwner partial order on $\posdef_d$, such that ``$X \succ Y$" (``$X \succeq Y$") indicates that $X-Y$ is positive definite (semidefinite); similarly, ``$X \succ 0$'' implies that $X$ has only positive eigenvalues. With a slight abuse of notation, we use $\posdef_d$ to denote \emph{Hermitian} positive definite matrices. Recall the \emph{Schatten $p$-norm}, which on $\mathbb{P}_d$ becomes $\norm{X}_p := \big( \trace X^p \big)^{1/p}$. Important cases are the \emph{trace norm} $\norm{X}_1 = \trace(X)$ ($p=1$) and the \emph{Frobenius norm} $\norm{X}_F=\big( \trace X^2\big)^{1/2}$ ($p=2$). We further use the usual operator norm $\norm{\cdot}$;
note that $\norm{X} \leq \norm{X}_F \leq \sqrt{d} \norm{X}$.
Most Riemannian optimization literature that studies problems on the manifold $\mathbb{P}_d$ considers the Riemannian geometry induced by the following metric:
\begin{equation}
\ip{A}{B}_X := \trace (X^{-1} A X^{-1} B), \quad X \in \posdef_d, A,B \in T_X(\posdef_d)= \mathbb{H}_d \; ,
\end{equation}
where the tangent space $\mathbb{H}_d$ is the space of Hermitian matrices. Equipped with this metric, $\posdef_d$ is a Cartan-Hadamard manifold. Its (global) non-positive sectional curvature gives rise to many desirable properties, including geodesic convexity.
While we use the Riemannian lens above for exploiting the geodesic convexity of~\eqref{eq:F}, our algorithm design and analysis views $\posdef_d$ through the lens of nonlinear Perron-Frobenius theory~\citep{lemmens_nussbaum_2012} instead.
Here, we view $\posdef_d$ as a convex cone that is endowed with the Thompson part metric, which induces a Finsler geometry on $\posdef_d$. Our analysis does not require any specific Finslerian properties; hence, we recall only basic definitions and some key properties below.
\begin{defn}[Thompson metric]\label{def:thompson}
The Thompson metric $\delta_T$ is defined as
\begin{equation}
\delta_T(X,Y) := \log \max \lbrace M(X/Y), M(Y/X) \rbrace,
\end{equation}
where $M(X/Y) := \inf \lbrace \lambda > 0 : X \preceq \lambda Y \rbrace$.
\end{defn}
\begin{lem}[Properties of the Thompson metric]\label{lem:prop-dT}
We assume $X,Y \in \posdef_d$.
\begin{enumerate}
\item $\delta_T (X^{-1},Y^{-1}) = \delta_T (X,Y)$;
\item $\delta_T (B^\ast X B, B^\ast Y B) = \delta_T (X,Y)$ for $B \in GL_n(\mathbb{C})$;
\item $\delta_T (A^\ast X A, A^\ast Y A) \leq \delta_T (X,Y) $ for $A \in \mathbb{C}^{d \times r}$, having full rank;
\item $\delta_T (X^t, Y^t) \leq \vert t \vert \delta_T (X,Y)$ for $t \in [-1,1]$;
\item $\delta_T (A+B, C+D) \leq \max \lbrace \delta_T(A,C), \delta_T(B,D) \rbrace$, which can be recursively generalized to sums with $m$ terms, i.e.
\begin{align*}
\delta_T \Big( \sum\nolimits_j X_j, \sum\nolimits_j Y_j \Big) \leq \max_{1 \leq j \leq m} \delta_T (X_j, Y_j) \; ;
\end{align*}
\item $\delta_T (X+A, Y+A) \leq \frac{\alpha}{\alpha + \beta} \delta_T (X,Y)$, where $A \succeq 0$, $\alpha = \norm{X} \vee \norm{Y}$, and $\beta = \lambda_{min}(A)$.
\end{enumerate}
\end{lem}
\noindent For a proofs of Lemma~\ref{lem:prop-dT} we refer the reader to~\citep{sra-hosseini}.\\
\noindent Furthermore, we will use the following notions of non-expansitivity and contractivity with respect to the Thompson metric.
\begin{defn}\label{def:contract}
We say that a map $G$ on $\cal X$ is
\begin{enumerate}
\item \emph{non-expansive}, if $\delta_T(G(X),G(Y)) \leq \delta_T (X,Y)$ for all $X,Y \in \mathcal{X}$;
\item \emph{contractive}, if $\delta_T(G(X),G(Y)) < \delta_T (X,Y)$ for all $X,Y \in \mathcal{X}$.
\end{enumerate}
\end{defn}
\noindent Lastly, recall that the \emph{Kronecker product} $X \otimes Y$ of matrices $X \in \mathbb{R}^{m \times n}, Y \in \mathbb{R}^{p \times q}$ is given as $(X \otimes Y)_{\alpha \beta} = a_{ij} b_{kl}$, where $\alpha = p(i-1)+k$, $\beta = q(j-1)+l$. The Kronecker product is bilinear and associative.
\subsection{Brascamp-Lieb inequalities}
\subsubsection{Basic notation}
The goal of the present paper is the computation of Brascamp-Lieb constants for feasible data. The following seminal result by~\citet{lieb} allows for phrasing this problem as an optimization task on $\posdef_d$.
\begin{theorem}[\citet{lieb}]
Let $(\mathcal{A},\bm{w})$ denote a BL-datum, where $A_j: \mathbb{R}^d \rightarrow \mathbb{R}^{d'}$ for all $A_j \in \mathcal{A}$ and $\bm{w}=\lbrace w_j \rbrace$ ($1\le j\le m$). Consider the function
\begin{equation}\label{eq:BL-lieb}
{\rm BL}(\mathcal{A},\bm{w}; Z) := \left( \frac{\prod_{j \in [m]} \det(Z_j)^{w_j}}{\det \bigl( \sum_{j \in [m]} w_j A_j^* Z_j A_j\bigr)} \right)^{1/2},
\end{equation}
where $Z := \lbrace Z_j \rbrace \in \; \prod_{j \in [m]} \posdef_{d'}$. The \emph{Brascamp-Lieb constant} for $(\mathcal{A},\bm{w})$ can be computed as
\begin{equation}\label{eq:BL-sup}
\sup_{Z \in \; \prod_{j \in [m]} \posdef_{d'}} {\rm BL} \big(\mathcal{A},\bm{w}; Z \big) \; .
\end{equation}
\end{theorem}
\noindent \cite{tao-paper} characterized the feasibility of $(\mathcal{A},\bm{w})$ as follows:
\begin{theorem}[\citet{tao-paper}]\label{thm:feasible}
$(\mathcal{A},\bm{w})$ is feasible if and only if
\begin{enumerate}
\item $d=\sum_{j \in [m]} w_j d'$,
\item $\dim(H) \leq \sum_{j \in [m]} w_j \dim(A_j H)$ for any subspace $H \subseteq \mathbb{R}^d$.
\end{enumerate}
\end{theorem}
\noindent This result renders the feasibility of Brascamp-Lieb data into a constrained optimization problem: It induces linear constraints on $\bm{w}$ as $H$ varies over different subspaces. Importantly, since $\dim(A_j H) \leq d$, the set of constraints is finite.
Therefore, the feasible set is a convex polytope $\mathcal{X}_{\mathcal{A}}$ (also known as \emph{Brascamp-Lieb polytope}). If $\bm{w}$ lies on the boundary of $\mathcal{X}_{\mathcal{A}}$, then there exists a non-trivial subspace $H$ for which the inequality constraints in Theorem~\ref{thm:feasible} are tight.
\begin{defn}[Simple Brascamp-Lieb data]\label{def:simple}
Let $(\mathcal{A},\bm{w})$ feasible. A subspace $H \subseteq \mathbb{R}^d$ is \emph{critical}, if
\begin{equation}
\dim(H) = \sum_{j \in [m]} w_j \dim(A_j H) \; .
\end{equation}
We say that $(\mathcal{A},\bm{w})$ is \emph{simple}, if there exists no non-trivial, critical subspace.
\end{defn}
Geometrically, simple Brascamp-Lieb data lie in the interior of $\mathcal{X}_{\mathcal{A}}$. Importantly, Problem~\ref{eq:BL-const} has a unique global solution for simple data.
\begin{theorem}\label{thm:maximizer}
Let $(\mathcal{A},\bm{w})$ as before and $X := \sum_{j \in [m]} w_j A_j^* Z_j A_j$. Then, the following statements are equivalent:
\begin{enumerate}
\item $Z$ is a global maximizer for ${\rm BL}(\mathcal{A},\bm{w},Z)$;
\item $Z$ is a local maximizer for ${\rm BL}(\mathcal{A},\bm{w},Z)$;
\item $X$ is invertible and $Z_j^{-1} = A_j X^{-1} A_j^*$.
\end{enumerate}
\end{theorem}
\noindent Theorem~\ref{thm:maximizer} implies that the feasibility of a BL datum can be characterized by maximizing ${\rm BL}(\mathcal{A},\bm{w},Z)$. Moreover, Theorem~\ref{thm:maximizer}(3) allows us to rewrite~\eqref{eq:BL-sup} with respect to $X$, which results in the formulation that we will analyze in this work (see~\citep[Definition~7]{vishnoi}):
\begin{equation}
\min_{X \in \mathbb{P}_d} F(X) := \sum\nolimits_j w_j \log \det \big( A_j^\ast X A_j \big) - \log \det (X) \; .
\end{equation}
\subsubsection{Examples}
Many well-known inequalities in mathematics can be written as Brascamp-Lieb inequalities. In this section, we recall a few key examples. A more comprehensive list of examples can be found in~\citep{tao-paper}.
\begin{example}[H{\"o}lder]
Consider $A_j = {\rm id}_{\mathbb{R}^d}$ for all $j$, i.e., $\mathcal{A} = \lbrace {\rm id}_{\mathbb{R}^d} \rbrace_{1 \leq j \leq m}$. Then the multilinear \emph{H{\"o}lder inequality} asserts that ${\rm BL}(\mathcal{A},\bm{w})=1$, if $\sum_j w_j = 1$ and ${\rm BL}(\mathcal{A},\bm{w})=+\infty$ otherwise.
\end{example}
\begin{example}[Sharp Young inequality]
Consider $A_j: \mathbb{R}^d \times \mathbb{R}^d \rightarrow \mathbb{R}^d$ for $j=1,2,3$ with
\begin{equation*}
\begin{cases}
A_1(x,y) := x \\
A_2(x,y) := y \\
A_3(x,y) := x-y.
\end{cases}
\end{equation*}
Then the \emph{Sharp Young inequality} asserts that
\begin{equation*}
{\rm BL}(\mathcal{A},\bm{w}) = \Big( \prod_{j=1}^3 \frac{(1-w_j)^{1-w_j}}{w_j^{w_j}} \Big)^{d/2} \; ,
\end{equation*}
if $\sum_{j=1}^3 w_j =2$ and ${\rm BL}(\mathcal{A},\bm{w})=+\infty$ otherwise.
\end{example}
\citet{barthe} introduced a class of \emph{Reverse Brascamp-Lieb inequalities} that generalize several important inequalities that cannot be encoded with the original Brascamp-Lieb framework. A key example is the \emph{Gaussian Brunn-Minkowski inequality}~\citep{barthe2008gaussian}. Importantly, since Reverse Brascamp-Lieb inequalities have the same data as the standard Brascamp-Lieb inequalities, there exists an intrinsic relation between their constants: The optimal constants of standard and Reverse Brascamp-Lieb inequalities multiply to 1, if the datum is feasible. Infeasible data have Brascamp-Lieb constants of $+\infty$ and Reverse Brascamp-Lieb constants of $0$~\citep{barthe}. This connection ensures that the algorithmic results presented here and in the related work translate from standard to Reverse Brascamp-Lieb inequalities.
\section{Overview of our results}
We start with a derivation of the map~\eqref{eq:map} with the goal of establishing that its Picard iteration solves Problem~\eqref{eq:F}. We start by computing the (Euclidean) gradient of $F$:
\begin{align}
\nabla F(X) = \sum\nolimits_{j} w_j A_j ( A_j^\ast X A_j )^{-1} A_j^\ast - X^{-1} \; .
\end{align}
Since the constraint set $\posdef_d$ is open, $\nabla F(X)=0$ is necessary for $X$ to be a local minimum. Importantly, since $F$ is geodesically convex, this local minimum must be global. Thus, to solve~\eqref{eq:F}, it suffices to solve the equation
$\nabla F(X)=0$, or equivalently $\sum\nolimits_{j} w_j A_j ( A_j^\ast X A_j )^{-1} A_j^\ast - X^{-1} = 0$,
which can be rewritten as the nonlinear matrix equation
\begin{equation}
\label{eq:pre1}
X = \bigl( \sum\nolimits_{j} w_j A_j ( A_j^\ast X A_j )^{-1} A_j^\ast \bigr)^{-1}.
\end{equation}
Once written as~\eqref{eq:pre1}, it is natural to consider the following iterative map:
\begin{equation}
\label{eq:giter}
G: \; X_{k+1} \gets \bigl( \sum\nolimits_{j} w_j A_j ( A_j^\ast X_k A_j)^{-1} A_j^\ast \bigr)^{-1} \; .
\end{equation}
Remarkably, this na\"ive consideration turns out to be a meaningful algorithm. That is, we can show that the sequence of iterates $(X_k)_{k>0}$ produced by iterating~\eqref{eq:giter} has a fixed point under the usual assumptions. A crucial ingredient of our analysis is the observation that the solution of Problem~\eqref{eq:F} is contained in a compact convex set of the form
\begin{equation}\label{eq:def-D}
D_{\delta, \Delta} := \big\{\delta I_d \preceq X \preceq \Delta I_d \big\} \; ,
\end{equation}
where $\delta, \Delta >0$ are parameters that depend only on the input datum $(\mathcal{A}, \bm{w})$ only.
The upper and lower bounds on the solution are implied by~\citep[Proposition~5.2]{tao-paper}.
Formally, we will show the following convergence result:
\begin{theorem}\label{thm:conv-G1}
Let $X_0 \in D_{\delta, \Delta}$; the sequence $\{X_k\}$ generated by iterating $G$ converges to a solution of~\eqref{eq:F} as $k\to \infty$.
\end{theorem}
\noindent The proof of this key result relies on Theorem~\ref{thm:s-h}. We have to show the three conditions of the theorem, (1) that $G$ has a fixed point in $D_{\delta,\Delta}$, (2) that $G$ is non-expansive on $D_{\delta,\Delta}$, and (3) that $G$ is asymptotically regular.
\noindent Proofs for all three conditions will be given in Section~\ref{sec:convergence-G}.
The existence of a suitable set $D_{\delta, \Delta}$ and its dependency on the input datum $(\mathcal{A},\bm{w})$ via $\delta$ and $\Delta$ will be discussed in Section~\ref{sec:delta}.
However, to establish stronger, \emph{non-asymptotic} convergence guarantees, \emph{strict} contractivity (see Definition~\ref{def:contract}(2)) of the
iterative map is typically required. Unfortunately, $G$ only fulfills the weaker notion of non-expansivity (Definition~\ref{def:contract}(1)). Therefore, we cannot hope to develop non-asymptotic convergence guarantees with the tools proposed so far. To mitigate this problem, we introduce a regularization of $G$, which instead of minimizing $F$ directly, minimizes a perturbed objective. Consider thus the regularized problem
\begin{equation}\label{eq:F-regul}
\min_{X \succ 0} \; F_{\mu} (X) := F(X) + \mu \trace (X) \; ,
\end{equation}
for some $\mu > 0$. By analyzing the gradient of $F_{\mu}$, we can derive an iterative procedure for solving Problem~\eqref{eq:F-regul} as
\begin{align*}
\nabla F_\mu (X) &= \nabla F(X) + \mu I = 0, \\
X^{-1} &= \mu I + \sum\nolimits_j w_j A_j (A_j^\ast X A_j)^{-1} A_j^\ast \\
X &= \bigl(\mu I + \sum\nolimits_j w_j A_j (A_j^\ast X A_j)^{-1} A_j^\ast \bigr)^{-1} \; ,
\end{align*}
using which we can define the regularized map
\begin{align}\label{eq:G2}
G_{\mu}: \quad X_{k+1} \mapsto \bigl(\mu I + \sum\nolimits_j w_j A_j (A_j^\ast X_k A_j)^{-1} A_j^\ast \bigr)^{-1} \; .
\end{align}
We show now that the regularized map is indeed contractive.
\begin{lem}\label{lem:G4-contract}
For any $\mu>0$, the map $G_\mu$ is contractive on $D_{\delta,\Delta}$, i.e., $\delta_T(G_\mu(X), G_\mu(Y))< \delta_T(X,Y)$ for all $X,Y \in D_{\delta,\Delta}$.
\end{lem}
\noindent Finally, we analyze the convergence rate of the proposed fixed-point approach.
The central result of the paper gives the following non-asymptotic convergence guarantee:
\begin{theorem}\label{thm:conv-non-asymp}
For a suitable choice of $\mu$, $G_{\mu}$ computes an $\epsilon$-approximate solution to problem~\eqref{eq:F}, i.e., a constant $C$, such that ${\rm BL}(\mathcal{A},\bm{w}) \leq C \leq (1+\epsilon) {\rm BL}(\mathcal{A},\bm{w})$, with iteration complexity $O \Big( \log \big( \frac{1}{\epsilon},\frac{\Delta}{\delta^{3/2}} ,\sqrt{d}, \frac{1}{R}\big) \Big)$, where $d, \delta, \Delta, R$ depend on the input datum $(\mathcal{A},\bm{w})$ only.
\end{theorem}
We discuss the complexity of the algorithm in more detail in Section~\ref{sec:convergence-G-reg}. The proof relies on a convergence analysis of the regularized map $G_\mu$ and quantifies the approximation error of $G_\mu$ with respect to $G$. The latter relies on the smoothness of $F_\mu$, which is characterized by the following lemma:
\begin{lem}\label{lem:F-smooth}
${\rm Lip}(\nabla F_\mu) \leq \frac{2}{\delta}$, i.e., $\norm{\nabla F_\mu(X) - \nabla F_{\mu}(Y)} \leq \frac{2}{\delta} \norm{X-Y}$.
\end{lem}
Putting both results together yields the claim. The proofs of Theorem~\ref{thm:conv-non-asymp} and Lemma~\ref{lem:G4-contract} are given in Section~\ref{sec:convergence-G-reg}.
A practical implementation of our fixed-point approach requires an initialization in the set $D_{\delta,\Delta}$ and in turn an explicit computation of the parameters $\delta$ and $\Delta$ for a given input datum $(\mathcal{A},\bm{w})$. We discuss the difficulty of this problem, as well as its connection to constrained subdeterminant maximization in Section~\ref{sec:delta}.
\section{Related work}
Computing Brascamp-Lieb constants has been an actively studied problem for many years. Following seminal work by~\citet{BL1,BL2,lieb} and~\citet{tao-paper}, recent work has analyzed the problem through a geometric lens, exploiting geodesic convexity to utilize tools from Riemannian Optimization.
A growing body of literature has analyzed a formulation of the problem, which links to the more general \emph{operator scaling problem}~\citep{allen-zhu,garg,garg2018algorithmic,kwok,franks2018operator,burgisser2018efficient}. We briefly recall some of the key results in this line of work and comment on their relation to the current paper.
The \emph{operator scaling problem} introduced by~\citet{gurvits} asks to compute, for a given operator $\Phi$, left and right scaling matrices, such that the scaled operator $\Phi'$ is doubly stochastic. Formally, let $\mathcal{A} = \big(A_1, \dots, A_m \big)$ be a tuple of matrices with $A_j \in \mathbb{R}^{d \times d'}$ that define a postive operator
\begin{equation*}
\Phi_\mathcal{A}(X) = \sum_{j=1}^m A_j X A_j^* \; .
\end{equation*}
Furthermore, consider the notion of \emph{capacity} for an operator $\Phi_\mathcal{A}$, defined as
\begin{align}\label{eq:cap}
{\rm cap} (\Phi_\mathcal{A}(X)) := \inf_{X \succ 0, \det(X)=1} \det(\Phi_\mathcal{A}(X)) \; .
\end{align}
If an infimum $X^*$ can be attained in~(\ref{eq:cap}), then the scaling $A_j' = \big(\Phi_\mathcal{A}(X^*)^{-1} \big)^{1/2} A_j (X^*)^{1/2}$
gives $\sum_j A_j' (A_j')^* = 1$ and $\sum_j (A_j')^* A_j' = 1$. In consequence, the scaled operator
\begin{align}
\Phi_\mathcal{A}'(X) = \sum_{j=1}^m A_j' X (A_j')^*
\end{align}
is doubly stochastic as desired. Therefore, minimizing the capacity provides a means to computing a suitable scaling. A simple method for this task is Gurvits' algorithm~\citep{gurvits}, which can be viewed as a generalization of Sinkhorn's algorithm~\citep{sinkhorn}. The operator scaling problem itself can be viewed as a generalization of the \emph{matrix scaling problem} for which Sinkhorn's algorithm provides a simple, iterative solution.
A convergence analysis of Gurvits' algorithm was later presented by~\citet{garg,garg2018algorithmic}. They give a
$O \big( {\rm poly} \big( b,d, \frac{1}{\epsilon} \big) \big)$ guarantee, where $b$ denotes the bit complexity of the Brascamp-Lieb datum, which can be exponential in $m,d$ (\citet{garg2018algorithmic}, Theorem~1.5).
A second-order approach by~\citet{allen-zhu} reduces the dependency on $\frac{1}{\epsilon}$ to ${\rm poly}(\log \frac{1}{\epsilon})$ (\citet{allen-zhu}, Theorem M1).
Furthermore, under the additional assumption that $\mathcal{A}$ fulfills the \emph{$\lambda$-spectral gap condition} ($\sigma_2$ is the 2$^\text{nd}$ singular value):
\begin{align*}
\sigma_2 \Big( \sum\nolimits_{j=1}^m A_j \otimes A_j \Big) \leq (1 - \lambda) \frac{\sum_{j=1}^m \norm{A_j}_F^2}{\sqrt{d d'}} \; ,
\end{align*}
stronger results can be shown.
Indeed, \citet{kwok} show that
a gradient-based method achieves a factor $\log \frac{1}{\epsilon}$ (\citet{kwok}, Theorem~1.5).
The formulation of the Brascamp-Lieb problem utilized in these approaches differs from ours~(\ref{eq:F}). Notably, their formulation has an \emph{exponential} dependency on the bit complexity of the Brascamp-Lieb datum. In contrast, ~(\ref{eq:F}) is only \emph{polynomial} in the bit complexity of the datum~\citep{vishnoi}. In addition, while these approaches employ a more commonly used Riemannian structure on $\posdef_d$ induced by the Frobenius norm, our analysis uses the Thompson metric on $\posdef_d$.
\section{Proofs of the results}
\vspace*{-12pt}
\subsection{Convergence of the fixed-point iteration using map $G$}
\label{sec:convergence-G}
We begin by showing that the sequence of iterates $(X_k)_{k>0}$ produced by the map $G$ converges to a solution of Problem~\ref{eq:F}:
\begin{theorem}[Theorem~\ref{thm:conv-G1}]
Let $X_0 \in D_{\delta, \Delta}$; the sequence $\{X_k\}$ generated by iterating $G$ converges to a solution of problem~\ref{eq:F} as $k\to \infty$.
\end{theorem}
\noindent We first show the existence of a fixed point:
\begin{lem}\label{lem:fix-point}
$G$ has a fixed point in $D_{\delta, \Delta}$.
\end{lem}
\begin{proof}
From Theorem~\ref{thm:maximizer} we know that the minimum is attained in problem~\ref{eq:F}, i.e., there exists a point $X^*$, such that $F(X^*) = \inf_{X} F(X)$. $X^*$ must be a stationary point, i.e., $\nabla F(X^*)=0$. Hence, by construction $X^* = G(X^*)$. Moreover,~\citet[Prop. 5.2]{tao-paper} implies that the minimum lies in a set of the form $D_{\delta,\Delta}$ (see sec.~\ref{sec:delta} for more details). This implies that $G$ has a fixed point in $D_{\delta,\Delta}$.
\end{proof}
\noindent For proving Theorem~\ref{thm:conv-G1} we need a few additional auxiliary results that we now present. We start by showing that $G$ is homogeneous and order-preserving:
\begin{lem}\label{lem:G1-prop}
Let $G$ be as defined in~\eqref{eq:map}. Then:
\begin{enumerate}
\item For a scalar $\alpha \in \mathbb{R}$ we have $G(\alpha X)=\alpha G(X)$, i.e., $G$ is \emph{homogeneous}.
\item For $X,Y \in \mathbb{P}_d$ with $X \preceq Y$ ($X \prec Y$) we have $G(X) \preceq G(Y)$ ($G(X) \prec G(Y)$), i.e., $G$ is \emph{order-preserving}.
\end{enumerate}
\end{lem}
\begin{proof}
Property (1) is trivial. We therefore move directly to property (2). First, note that for $X \preceq Y$, we have $A_j^\ast X A_j \preceq A_j^\ast Y A_j$. Assuming that $A_j^\ast X A_j$ is invertible, this implies
\begin{align*}
(A_j^\ast X A_j)^{-1} \succeq (A_j^\ast Y A_j)^{-1} \; \Rightarrow \; \Big( \sum\nolimits_j w_j A_j (A_j^\ast X A_j)^{-1} A_j^\ast \Big)^{-1} \preceq \Big( \sum\nolimits_j w_j A_j (A_j^\ast Y A_j)^{-1} A_j^\ast \Big)^{-1} \; .
\end{align*}
The same arguments apply in the $\prec$ case.
\end{proof}
\noindent A direct consequence of this property is that $G$ is non-expansive.
\begin{lem}\label{lem:nonexp}
$G$ is non-expansive.
\end{lem}
\begin{proof}
We must prove that $\delta_T(G(X), G(Y)) \leq \delta_T(X,Y)$. By Definition~\ref{def:thompson}(2) we have
$M(Y/X)^{-1} Y \preceq X \preceq M(X/Y) Y$, and also
\begin{equation} \label{eq:G-bounds}
M(G(Y)/G(X))^{-1} G(Y) \preceq G(X) \preceq M(G(X)/G(Y)) G(Y) \; .
\end{equation}
Furthermore, since $G$ is order-preserving and homogeneous (Lemma~\ref{lem:G1-prop}), we have
\begin{align*}
G(X) &\preceq G(M(X/Y) Y) = M(X/Y) G(Y) \\
G(X) &\succeq G(M(Y/X)^{-1} Y) = M(Y/X)^{-1} G(Y) \; .
\end{align*}
Since $M(G(Y)/G(X)), M(G(X)/G(Y))$ are extremal in inequality~\eqref{eq:G-bounds}, this implies
\begin{align*}
M(X/Y) &\geq M(G(X)/G(Y)) \\
M(Y/X) &\geq M(G(Y)/G(X)) \; ,
\end{align*}
which gives
\begin{align*}
\max \lbrace M(X/Y), M(Y/X) \rbrace \geq \max \lbrace M(G(X)/G(Y)), M(G(Y)/G(X)) \rbrace
\end{align*}
and thereby the claim.
\end{proof}
\noindent Next, we establish asymptotic regularity of $G$.
\begin{lem}
G is asymptotically regular with respect to the Thompson metric, i.e.,
\begin{equation*}
\delta_T \big(G(X_{k+1}), G(X_k) \big) \rightarrow 0 \; .
\end{equation*}
\end{lem}
\begin{proof}
From Lemma~\ref{lem:fix-point} we know that ${\rm Fix(G) \neq \emptyset}$. Let $X^* \in {\rm Fix}(G)$. Due to the non-expansivity of $G$ (Lemma~\ref{lem:nonexp}), we have Fejér-monotonicity at $X^*$:
\begin{equation*}
\delta_T \big( X_{k+1}, X^* \big) = \delta_T \big( G(X_k), G(X^*) \big) \leq \delta_T \big( X_k, X^* \big) \; .
\end{equation*}
Thus, the iterates $\{G(X_k)\}_{k \geq 0}$ must remain bounded, which in turn implies asymptotic regularity of $G$.
\end{proof}
The proof further implies that $G$ is bounded, which we summarize in the following corollary:
\begin{cor}\label{lem:bounded}
$G$ is bounded on $D_{\delta, \Delta}$, i.e., there exists a $R>0$ depending on the input datum $(\mathcal{A},\bm{w})$ only, such that $\norm{G(X_k)} < R$ for all $k \geq 0$.
\end{cor}
\begin{rmk}\normalfont
We note that it is a well-known fact that non-expansive maps either do not have a fixed point or are bounded. Classical versions of this result can be found in~\citep{pazy_asymptotic_1971} (for Hilbert spaces) and in~\citep{reich_asymptotic_1973} (for Banach spaces).
\end{rmk}
\noindent In summary, we obtain our asymptotic convergence result (Theorem~\ref{thm:conv-G1}) as a direct consequence of Theorem~\ref{thm:s-h}:
With initialization at any $X \in D_{\delta, \Delta}$ the sequence produced by the map
$G$ converges to a solution of problem~\eqref{eq:F}. \\
\noindent Finally, we want to give a lower bound on the sequence of iterates produced by the map $G$, which is needed for the subsequent analysis. Recall the definition of \emph{operator convexity}. Consider a function $f$ defined on a real interval. We say that $f$ is operator convex, if for all $\lambda \in [0,1]$ and Hermitian matrices $A,B$, we have
\begin{align}\label{eq:def-op-convex}
f((1-\lambda)A + \lambda B) \leq (1-\lambda)f(A) + \lambda f(B) \; .
\end{align}
Furthermore, recall the following facts on the convexity of matrix operators (see, e.g.,~\cite{nordstroem,bhatia}):
\begin{lem}[Convex matrix operators]\hspace{5em}
\label{lem:convex-op}
\begin{enumerate}
\item $(\lambda X + \bar{\lambda} Y)^{-1} \preceq \lambda X^{-1} + \bar{\lambda} Y^{-1}$ for $\lambda \in (0,1)$, $\bar{\lambda}=1-\lambda$, $X,Y$ HPD.
\item If $f$ is operator convex, then $f(X^* L X) \preceq X^* f(L) X$ for $L$ PSD.
\end{enumerate}
\end{lem}
\noindent Finally, we need the following classical result on eigenvalue maps:
\begin{lem}[Problem III.6.14,~\citep{bhatia}]
\label{lem.eval}
Let $X, Y$ be HPD matrices. Then
\begin{equation*}
\lambda^\downarrow(X) \cdot \lambda^\uparrow(Y) \prec_w \lambda (XY) \prec_w \lambda^\downarrow(X) \cdot \lambda^\downarrow(Y),
\end{equation*}
which implies in particular that
\begin{equation*}
\langle \lambda^\downarrow (X), \lambda^\uparrow (Y) \rangle \leq \trace (XY) \leq \langle \lambda^\downarrow (X), \lambda^\downarrow (Y) \rangle \; ,
\end{equation*}
where $\lambda^\downarrow(X)$ ($\lambda^\uparrow(X)$) denotes eigenvalues of $X$ arranged in decreasing (increasing) order, $\prec_w$ denotes the weak-majorization order, and $\cdot$ denotes elementwise product. In the second inequality, we have equality, if $XY$ is symmetric.
\end{lem}
\noindent Recall that for $x \in \mathbb{R}^d$, we denote with $x^\downarrow = \big( x_1^\downarrow, \dots, x_d^\downarrow \big)$ the vector with entries of $x$ in decreasing order, i.e., $x_1^\downarrow \geq ... \geq x_d^\downarrow$. For $x, y \in \mathbb{R}^d$ we say that $x$ is \emph{majorized} by $y$ ($x \prec y$), if
\begin{align}
\sum_{i=1}^k x_i^\downarrow &\leq \sum_{i=1}^k y_i^\downarrow \quad {\rm for} \; 1 \leq k \leq d \\
\sum_{i=1}^d x_i^\downarrow &= \sum_{i=1}^d y_i^\downarrow \; .
\end{align}
If only the first condition holds, we say that $x$ is \emph{weakly majorized} by $y$ ($x \prec_w y$).
We show the following bound:
\begin{lem}\label{lem:G-low-bound}
Let $X_0 \in D_{\delta,\Delta}$. The sequence $\lbrace X_k \rbrace$ generated by iterating $G$ is bounded from below, i.e., $G(X_k) \succeq \delta I$.
\end{lem}
\begin{proof}
Let $X \in D_{\delta,\Delta}$ be arbitrary; we want to show that $G(X) \succeq \delta I$.
From Lemma~\ref{lem:G1-prop}(2) we know that $G$ is order-preserving, i.e.
\begin{align*}
X \succeq\delta I \quad \Rightarrow \quad G(X) \succeq G(\delta I) \;.
\end{align*}
Therefore, it remains to be shown that $G(\delta I) \succeq \delta I$. Recall the following lemma:
\begin{lem}[\citet{sra-hosseini}]\label{prop:4-12}
Let $A \in \mathbb{C}^{d \times d'}$ ($d' \leq d$) be of full rank. Then the function $f: \mathbb{P}_d \rightarrow \mathbb{R}$ with $f(Z) = \log \det (A^* Z A) - \log \det (Z)$ is operator decreasing.
\end{lem}
\noindent This lemma implies in particular that
\begin{equation*}
A(A^* Z A)^{-1} A^* \preceq Z^{-1} \; .
\end{equation*}
For $Z = \delta I$, this gives (together with the min-max theorem)
\begin{align*}
\lambda_{\max}\bigl(\big(G(\delta I)\big)^{-1}\bigr) = \lambda_{\max} \bigl( A_j^* \bigl( A_j^* (\delta I) A_j \bigr)^{-1} A_j \bigr) \leq \lambda_{\max} \bigl( \delta^{-1} I \bigr) \; .
\end{align*}
With $\sum_j w_j =1$, this implies $\lambda_{\max}\bigl(\bigl(G(\delta I)\bigr)^{-1}\bigr) \leq \delta^{-1}$, i.e. $(G(\delta I))^{-1} \preceq \delta^{-1} I$. This gives $G(\delta I) \succeq \delta I$.
\end{proof}
\subsection{Non-asymptotic convergence analysis of $G_\mu$}
\label{sec:convergence-G-reg}
We start by showing that the regularized map $G_\mu$ has the crucial contractivity property.
\begin{lem}[Lemma~\ref{lem:G4-contract}]
$G_\mu$ is contractive.
\end{lem}
\begin{proof}
\begin{align*}
\delta_T (G_\mu(X), G_\mu(Y)) &= \delta_T \big( \mu I + G^{-1}(X), \mu I + G^{-1}(Y) \big) \\
&\overset{(1)}{\leq} \frac{\gamma}{\gamma + \mu} \delta_T \big( G^{-1}(X), G^{-1}(Y) \big) \\
&\overset{(2)}{=} \frac{\gamma}{\gamma + \mu} \delta_T \big( G(X), G(Y) \big) \\
&\overset{(3)}{\leq}\frac{\gamma}{\gamma + \mu} \delta_T \big( X, Y \big) \; ,
\end{align*}
where ($1$) follows from Lemma~\ref{lem:prop-dT}(6) with $\gamma = \max \lbrace \norm{G^{-1}(X)}, \norm{G^{-1}(Y)} \rbrace \leq \frac{1}{\sqrt{\delta}}$ by Lemma~\ref{lem:G-low-bound}. ($2$) follows from Lemma~\ref{lem:prop-dT}(1) and ($3$) from $G$ being log-nonexpansive (Lemma~\ref{lem:nonexp}). Since $\mu>0$, this implies $\delta_T (G_\mu(X), G_\mu(Y)) < \delta_T (X,Y)$.
\end{proof}
The contractivity of the regularized map furthermore allows us to quantify its convergence rate non-asymptotically.
\begin{theorem}
For a suitable choice of $\mu$, $G_{\mu}$ computes an $\epsilon$-approximate solution to problem (Eq.~\ref{eq:F}) with iteration complexity $O \Big( \log \big( \frac{1}{\epsilon},\frac{\Delta}{\delta^{3/2}} ,\sqrt{d}, \frac{1}{R}\big) \Big)$, where $d, \delta, \Delta, R$ depend on the input daturm $(\mathcal{A},\bm{w})$ only.
\end{theorem}
\begin{proof}
\noindent Our plan is as follows: We start by analyzing the choice of the parameter $\mu$ and the convergence rate of the regularized approach. This will allow us to develop a non-asymptotic convergence result for problem~\ref{eq:F}.
The optimality gap is given by
\begin{align*}
\vert F(X_k) - F(X^\ast) \vert \leq \vert F(X_k) - F_\mu (X_k) \vert + \vert F_\mu (X_k) - F_\mu (X^*) \vert + \vert F_\mu (X^*)- F(X^*) \vert \; ,
\end{align*}
where the first and third term represent the \emph{approximation error} of $F_\mu$ and the second term the \emph{convergence rate} of $G_\mu$ in terms of the function values. Therefore, we have to perform the following steps:
\begin{enumerate}
\item bound the approximation error of $F_\mu$, i.e., bound $\vert F(X) - F_\mu (X) \vert$;
\item characterize the convergence rate of $G_\mu$, i.e., bound $\vert F_\mu(X) - F_\mu (X^\ast) \vert$.
\end{enumerate}
\paragraph{Approximation error of $F_\mu$.}
By construction we have $F_\mu (X) - F(X) = \mu \trace (X)$. Note that at a fixed point $X$, we have
\begin{align*}
\trace (X) = \trace (G_\mu (X)) = \trace \Big( \big( \mu I + \underbrace{\sum\nolimits_j w_j A_j (A_j^\ast X A_j)^{-1} A_j^\ast }_{=:G^{-1}(X)} \big)^{-1} \Big) \; .
\end{align*}
Let $\lbrace \nu_1, \dots, \nu_n \rbrace$ denote the eigenvalues of $G^{-1}(X)$. Then the eigenvalues of $G^{-1}(X) + \mu I$ are $\lbrace \nu_1 + \mu, \dots, \nu_n + \mu \rbrace$, which gives (for a fixed point $X$)
\begin{equation*}
\trace X = \trace \big[ \big( \mu I + G^{-1}(X) \big)^{-1} \big] = \sum_i \frac{1}{\lambda_i \big( \mu I + G^{-1}(X) \big)} = \sum_i \frac{1}{\mu + \nu_i} \; .
\end{equation*}
Let $\nu_{\rm min}$ be the minimal eigenvalue of $G^{-1}(X)$. Then
\begin{equation*}
\trace X = \sum_i \frac{1}{\mu + \nu_i} \leq \sum_i \frac{1}{\mu + \nu_{\rm min}} = \frac{d}{\mu + \nu_{\rm min}} \; .
\end{equation*}
Due to the boundedness of $G$ (Lemma~\ref{lem:bounded}), we have that $\nu_{\min} \geq \frac{1}{R}$.
Inserting this above, we obtain the bound
\begin{equation*}
\trace (X) = \frac{d}{\mu + \nu_{\rm min}} \leq \frac{d}{\mu + \frac{1}{R}} \; ,
\end{equation*}
which in turn implies the bound
\begin{equation}\label{eq:choose-mu}
F_\mu (X) - F(X) = \mu \trace (X) \leq \frac{\mu d}{\mu + \frac{1}{R}} \; .
\end{equation}
This bound gives a means for choosing $\mu$ ($d, \delta, R$ are determined by the input $(\mathcal{A},\bm{w})$), such that a solution to the regularized problem gives an $\frac{\epsilon}{4}$-accurate solution to the original one. We defer the further discussion to the end of the proof.
\paragraph{Convergence rate of $F_\mu$.}
From Lemma~\ref{lem:G4-contract} we know that $G_\mu$ is log-contractive, i.e., we have
\begin{align*}
\delta_T (\underbrace{G_\mu(X_k)}_{=X_{k+1}},\underbrace{ G_\mu(X^*)}_{=X^*}) \leq \frac{\gamma}{\gamma + \mu} \delta_T (X_k, X^*) \\
\Rightarrow \delta_T(X_{k+1}, X^\ast) \leq \frac{\gamma}{\gamma + \mu} \delta_T (X_k, X^*) \; ,
\end{align*}
where $\gamma := \max \lbrace \norm{G^{-1}(X_k)}, \norm{G^{-1}(X^*)} \rbrace$. Recursively, this gives
\begin{align}\label{eq:G4-bound}
\delta_T (X_k, X^*) \leq \Big( \frac{\tilde{\gamma}}{\tilde{\gamma} + \mu} \Big)^k \delta_T (X_0, X^*) \; ,
\end{align}
where we set $\tilde{\gamma} := \max \lbrace \max_{1 \leq i \leq k} \norm{G^{-1}(X_i)}, \norm{G^{-1}(X^*)} \rbrace$, which is bounded as $\frac{1}{R} \leq \tilde{\gamma} \leq \frac{1}{\sqrt{\delta}}$ by construction of the set $D_{\delta,\Delta}$ and the boundedness of $G$.
If $F_\mu$ is $L$-smooth, then the following bound holds:
\begin{align*}
F_\mu (X^\ast) &\leq F_\mu (X_k) + \ip{\nabla F_\mu (X_k)}{X_k - X^\ast}_F + \frac{L}{2} \norm{X^* - X_k}_F^2 \\
\Rightarrow \left\vert F_\mu (X^\ast) - F_\mu (X_k) \right\vert &\leq \left\vert \ip{\nabla F_\mu (X_k)}{X_k - X^\ast}_F \right\vert + \frac{L}{2} \norm{X^* - X_k}_F^2;
\end{align*}
here, $\ip{\cdot}{\cdot}_F$ denotes the usual Frobenius product $\ip{X}{Y}_F = \trace \big( X^* Y \big)$. In order to use this bound in our analysis, we must show that $F_\mu$ is indeed $L$-smooth. To that end, recall that
\begin{align*}
\nabla F_\mu(X) = \sum\nolimits_j w_j A_j (A_j^* X A_j)^{-1} A_j^* - X^{-1}+ \mu I \;
\end{align*}
and let $H_1(X) := \sum_j w_j A_j \big( A_j^* X A_j\big)^{-1} A_j^* $ and $H_2(X) := - X^{-1}$.
Then ${\rm Lip}(\nabla F_\mu) \leq L$, if ${\rm Lip}(H_1) + {\rm Lip}(H_2) \leq L$ for some $L$. For ${\rm Lip}(H_1)$ note that
\begin{align*}
\nabla \Big( \sum\nolimits_j w_j A_j (A_j^* X A_j)^{-1} A_j^* \Big)
&= - \sum\nolimits_j w_j \big[ \big(A_j (A_j^* X A_j)^{-1} A_j^* \big) \otimes \big(A_j (A_j^* X A_j)^{-1} A_j^* \big) \big] \\
&= - \Big( \sum\nolimits_j w_j A_j\big(A_j^* X A_j \big)^{-1} A_j^* \Big) \otimes \Big( \sum\nolimits_j w_j A_j\big(A_j^* X A_j \big)^{-1} A_j^* \Big)\\
&= - G^{-1}(X) \otimes G^{-1}(X) \; ,
\end{align*}
where the second equality follows from the distributive law. Then the construction of $D_{\delta,\Delta}$ and the fact that $G$ is a self-map imply
\begin{align*}
\norm{\nabla H_1(X)} = \norm{G^{-1}(X)}^2 \leq \frac{1}{\delta} \; .
\end{align*}
Furthermore, we have
\begin{align*}
\nabla \big(-X^{-1} \big) = (X^{-1}) \otimes (X^{-1}) \; ,
\end{align*}
which implies that $\norm{\nabla H_2(X)} = \norm{X^{-1}}^2 \leq \frac{1}{\delta}$, again by contruction of $D_{\delta,\Delta}$.
This implies $L=\frac{2}{\delta}$ and therefore Lemma~\ref{lem:F-smooth}.
With that, the Lipschitz bound above implies
\begin{align*}
\vert F_\mu (X^\ast) - F_\mu (X_k) \vert &\leq \vert \ip{\nabla F_\mu (X_k)}{X_k - X^\ast}_F \vert + \frac{1}{\delta} \norm{X^* - X_k}_F^2 \; .
\end{align*}
Hence, for the first term on the right hand side, the
Cauchy-Schwarz inequality implies
\begin{equation*}
\vert \ip{\nabla F_\mu (X_k)}{X_k - X^\ast}_F \vert \leq \norm{\nabla F_\mu (X_k)}_F \norm{X^* - X_k}_F
\leq \sqrt{d} \norm{\nabla F_\mu (X_k)} \norm{X^* - X_k}_F \; .
\end{equation*}
It remains to derive bounds on (a) $\norm{\nabla F_\mu (X_k)}$ and (b) the convergence rate of $\lbrace \norm{X^* - X_k}_F \rbrace_k$. For (a), we have
\begin{equation*}
\norm{\nabla F_\mu (X)}
= \norm{\nabla \big( F(X) + \mu \trace(X) \big)}
= \norm{ \underbrace{\big(\sum_j w_j A_j (A_j^* X A_j)^{-1} A_j \big)}_{=G^{-1}(X)} - X^{-1} + \mu I} \; .
\end{equation*}
Recall that for $X,Y$ Hermitian, Weyl's Perturbation Theorem (see, e.g.,~\citep{bhatia}, III.2.7) gives
\begin{equation*}
\norm{X - Y} \leq \max_i \left\vert \lambda_i (X) - \lambda_{n-i+1} (Y) \right\vert \; .
\end{equation*}
With that, we have
\begin{align*}
\norm{ ( G^{-1}(X) + \mu I ) - X^{-1} }
&\leq \max_i \vert \lambda_i ( G^{-1}(X) + \mu I) - \lambda_{n-i+1} (X^{-1}) \vert \\
&\leq \max \big\{ \lambda_{\max} ( G^{-1}(X) + \mu I ) , \lambda_{\max} (X^{-1}) \big\} \; .
\end{align*}
Note that $G(X), X \succeq \delta I$ by Lemma~\ref{lem:G-low-bound} and the construction of the set $D_{\delta,\Delta}$. This implies $\lambda_{\min}(G(X)), \lambda_{\min}(X) \geq \delta$ and therefore $\lambda_{\max}(G^{-1}(X)), \lambda_{\max}(X^{-1}) \leq \frac{1}{\delta}$. Inserting this above, we get
\begin{align}\label{eq:bound-grad-F}
\norm{\nabla F_\mu (X)}
&\leq \max \big\{ \lambda_{\max} \big( G^{-1}(X) + \mu I \big) , \lambda_{\max} (X^{-1}) \big\} \\
&\leq \max \Big\{ \frac{1}{\delta} + \mu, \frac{1}{\delta} \Big\} =\Big(\frac{1}{\delta} + \mu \Big) =: C_1 \; .
\end{align}
\noindent For (b), quantifying the convergence rate of $\big\{ \norm{X^* - X_k}_F \big\}_k$, we
recall an inequality of Snyder that relates the Thompson metric to Schatten norms.
\begin{theorem}[\citet{snyder}]
Let $X,Y$ HPD and $\norm{\cdot}_p$ denotes the Schatten $p$-norm. Then:
\begin{align*}
\norm{X-Y}_p &\leq 2^{\frac{1}{p}} \frac{e^{\delta_T(X,Y)} - 1}{e^{\delta_T(X,Y)}} \max \big\{ \norm{X}_p, \norm{Y}_p \big\} \; .
\end{align*}
\end{theorem}
\noindent Since $\norm{\cdot}_F$ is Schatten $2$-norm, we therefore get
\begin{align}\label{eq:snyder-bound}
\norm{X_k-X^*}_F
&\leq \sqrt{2} \max \big\{ \norm{X_k}_F, \norm{X^*}_F \big\} \frac{e^{\delta_T(X_k,X^*)} - 1}{e^{\delta_T(X_k,X^*)}} \\
&\le \sqrt{2d} \max \big\{ \norm{X_k}, \norm{X^*} \big\} \frac{e^{\delta_T(X_k,X^*)} - 1}{e^{\delta_T(X_k,X^*)}} \; .
\end{align}
For ease of notation, we introduce the shorthand $\delta_{T,k}:=\delta_T(X_k, X^*)$. We have seen above that $G_\mu$ is strictly contractive and that in particular $\delta_{T,k} \leq \big( \frac{\tilde{\gamma}}{\tilde{\gamma} + \mu} \big)^k \delta_{T,0}$.
If we initialize $X_0 \in D_{\delta,\Delta}$, then the initial distance from the optimum must be bounded due to the boundedness of $G$. Hence, there is some constant $C_2>0$, such that $\delta_{T,0} \leq C_2$. With this, we can analyze the exponential terms in~\eqref{eq:snyder-bound}. From the above inequality and the monotonicity of the exponential function, we have $e^{\delta_{T,k}} \leq e^{\big( \frac{\tilde{\gamma}}{\tilde{\gamma} + \mu} \big)^k \delta_{T,0}}
= e^{C_2 \alpha^k}$, where we set $\Big( \frac{\tilde{\gamma}}{\tilde{\gamma} + \mu} \Big)^k =: \alpha^k$.
Inserting this into the Lipschitz bound above, we then obtain
\begin{align*}
\left\vert F_\mu (X^\ast) - F_\mu (X_k) \right\vert
&\leq \sqrt{d}\big\|\nabla F_\mu (X_k) \big\| \big\|X^* - X_k\big\|_F + \frac{1}{\delta} \big\|X^* - X_k\big\|_F^2 \\
&\overset{(1)}{\leq} \Big(\sqrt{2} C_1 + \frac{1}{\delta} \Big) \big\|X^* - X_k \big\|_F \\
&\overset{(2)}{\leq} \underbrace{ \Big(\sqrt{2} C_1 + \frac{1}{\delta} \Big) \sqrt{\frac{2d}{\delta}} }_{=: \widehat{C}_1} \frac{e^{\delta_{T,k}} -1}{e^{\delta_{T,k}} }\; ,
\end{align*}
where (1) follows since $\Big\{ \Big(\frac{e^{\delta_{T,k}} -1}{e^{\delta_{T,k}} }\Big)^2 \Big\}_k$ decays faster than $\Big\{ \frac{e^{\delta_{T,k}} -1}{e^{\delta_{T,k}} } \Big\}_k$ and (2) follows from Snyder's theorem, and upon noting that
\begin{equation*}
\max \Big\{ \max_{1 \leq i \leq k} \big\|X_i^{-1} \big\|, \big\|\big(X^*\big)^{-1}\big\| \Big\} \leq \frac{1}{\sqrt{\delta}},
\end{equation*}
by construction of $D_{\delta,\Delta}$. The constant $\widehat{C}_1$ summarizes the dependency on the input datum $(\mathcal{A},\bm{w})$ and the regularization parameter $\mu$.
We want to bound the iteration complexity of an $\epsilon'$-accurate solution ($\epsilon' := \frac{\epsilon}{2}$) of $F_\mu$, so let $\epsilon'=\left\vert F_\mu (X^\ast) - F_\mu (X_k) \right\vert$. Then with the above, we have
\begin{equation*}
\epsilon' \leq \widehat{C}_1 \frac{e^{\delta_{T,k}} - 1}{e^{\delta_{T,k}}} \;.
\end{equation*}
Furthermore, with the constants $\alpha$ and $C_2$ as introduced above, we get
\begin{equation*}
\epsilon' \leq \widehat{C}_1 \frac{e^{C_2 \alpha^k} - 1}{e^{C_2 \alpha^k}} = \widehat{C}_1 \Big( 1 - \frac{1}{e^{C_2 \alpha^k}} \Big)\;.
\end{equation*}
This can be rewritten as follows:
\begin{align*}
\frac{\epsilon'}{\widehat{C}_1} \leq 1 - \frac{1}{e^{C_2 \alpha^k}}
&\Leftrightarrow e^{C_2 \alpha^k} \geq \frac{1}{1 - \frac{\epsilon'}{\widehat{C}_1}}
\Leftrightarrow \alpha^k \geq \frac{1}{C_2} \log \Bigg( \frac{1}{1 - \frac{\epsilon'}{\widehat{C}_1}} \Bigg) \\
&\Leftrightarrow \Bigg( \frac{1}{\alpha} \Bigg)^{-k} \geq \frac{1}{C_2} \log \Bigg( \frac{1}{1 - \frac{\epsilon'}{\widehat{C}_1}} \Bigg)
\Leftrightarrow -k \geq \frac{\log \Bigg( \frac{1}{C_2} \log \Big( \frac{1}{1 - \frac{\epsilon'}{\widehat{C}_1}} \Big) \Bigg)}{\log \Big( \frac{1}{\alpha} \Big)} \; .
\end{align*}
In summary, we have the following bound on the number of iterations needed to achieve an $\epsilon$-accurate solution.
\begin{equation}
k \leq -\frac{\log \Bigg( \frac{1}{C_2} \log \Big( \frac{1}{1 - \frac{\epsilon'}{\widehat{C}_1}} \Big) \Bigg)}{\log \Big( \frac{1}{\alpha} \Big)} \; .
\end{equation}
We first evaluate the dependency on $\frac{1}{\epsilon}$. Assuming $\epsilon$ is small, we get for the inner term
\begin{equation*}
- \log \Big( 1 - \frac{\epsilon'}{\widehat{C}_1} \Big) \geq \frac{\epsilon'}{\widehat{C}_1} \; .
\end{equation*}
This implies a $\log \left(\frac{1}{\epsilon} \right)$ dependency as
\begin{equation}
\boxed{
k \leq \frac{\log \left( \frac{\widehat{C}_1 C_2}{\epsilon'} \right)}{\log \left( \frac{1}{\alpha} \right)}
}
\end{equation}
\paragraph{Constants.} The result above depends on three constants $\widehat{C}_1, C_2,\alpha$, which we relate to the complexity of the input $(\mathcal{A},\bm{w})$, characterized by the constants $d, \delta, \Delta, R$.
\begin{rmk}\normalfont
As stated earlier, those constants are not necessary mutually independent. In particular, we expect that $R$ may dependent on $\delta$, i.e., a more precise notation would read $R(\delta)$. However, to keep notation simply, we just write $R$. For feasible BL data, we expect $\delta$ to be small and $R \geq 1$ (?).
\end{rmk}
For $\widehat{C}_1$, note that
\begin{align*}
\widehat{C}_1 &= \Big(\sqrt{2} C_1 + \frac{1}{\delta} \Big) \sqrt{\frac{2d}{\delta}}
\leq 2 \sqrt{\frac{d}{\delta}} \Big( \frac{1}{\delta} + \mu \Big) + \frac{\sqrt{2d}}{\delta^{3/2}} \overset{(1)}{=} O \Big( \sqrt{d}, \frac{1}{\delta^{3/2}},\frac{1}{R} \Big)
\; .
\end{align*}
To see (1), recall that to control the approximation error (Eq.~\ref{eq:choose-mu}), we set $\mu:= \frac{\epsilon' }{2R(d-\epsilon'/2)}$. Since $\widehat{C}_1$ enters logarithmically, this introduces a $O \Big( \log \big( \frac{1}{\delta^{3/2}}, \sqrt{d} ,\frac{1}{R}\big) \Big)$ dependency.
Furthermore, we have that
\begin{equation*}
C_2 = \delta_T (X_0, X^*) \leq {\rm diam}(D_{\delta,\Delta}) \leq \log \frac{\Delta}{\delta} \; ,
\end{equation*}
since by construction
\begin{equation*}
{\rm diam}(D_{\delta,\Delta}) = \max_{X,Y \in D_{\delta,\Delta}} \delta_T (X,Y) = \log M \big( \Delta I_d / \delta I_d \big) = \log \frac{\Delta}{\delta} \; ,
\end{equation*}
which only depends on the parameters $\delta, \Delta$.
Finally, we have
\begin{align*}
\frac{1}{\alpha} =\frac{\tilde{\gamma} + \mu}{\tilde{\gamma}} \geq \frac{\frac{1}{R} + \mu}{\frac{1}{\sqrt{\delta}}}
\geq \sqrt{\delta}\Big( \frac{1}{R} + \frac{\epsilon'}{2Rd} \Big) = O \Big(\frac{\sqrt{\delta}}{R} \Big) \; .
\end{align*}
This implies that the contribution of the factor $\big(\log \big( \frac{1}{\alpha}\big)\big)^{-1}$ is dominated by the other terms.
Overall, the regularized approach has a complexity of $O \Big( \log \big( \frac{1}{\epsilon},\frac{\Delta}{\delta^{3/2}} ,\sqrt{d}, \frac{1}{R}\big) \Big)$ with respect to the accuracy $\epsilon$ and $d, \delta,\Delta,R$, which depend on the input datum $(\mathcal{A},\bm{w})$ only. We will discuss the dependency on the parameter $\delta,\Delta$ in more detail in the following section.
\end{proof}
\subsection{Dependency on the input datum $(\mathcal{A},w)$}\label{sec:delta}
Our convergence guarantees rely on the existence of suitable parameters $\Delta, \delta>0$ for a given Brascamp-Lieb datum $(\mathcal{A},\bm{w})$, such that the global optimum $X^*$ of Eq.~\ref{eq:F} is contained in a set of the form $D_{\delta,\Delta}$. The existence of such a $\delta$ follows from a result in~\citep{tao-paper}, which ensures that the $X^*$ is bounded from above and below. In this section, we recall the argument by~\citet{tao-paper} and highlight its connection to our parameters $\delta,\Delta$. Furthermore, we discuss a conjectured connection to the \emph{subdeterminant maximization} problem. The development of an efficient approach for computing $\delta$ (and hence determining $D_{\delta,\Delta}$ explicitely) is left for future work.\\
\noindent We recall some crucial properties of the optimum $X^*$ from~\citet{tao-paper}. The existence of suitable upper and lower bounds on $X^*$ is implied by~\citep[Proposition~5.2]{tao-paper}, which relies on the following lemma:
\begin{lem}[\cite{tao-paper}, Lemma~5.1]\label{prop:5.1}
Assume that $(\mathcal{A},\bm{w})$ is feasible, i.e., the conditions in Theorem~\ref{thm:feasible} hold. Then there exists a $c>0$ and an ONB $\lbrace e_1, \dots, e_d \rbrace$ of $\mathbb{R}^d$, such that there exists an index set $I_j$ with $\vert I_j \vert = \dim(H_j) = \dim (A_j H)$, such that $\sum_j w_j \vert I_j \cap \lbrace 1, \dots, k \rbrace \vert \leq k \quad \forall \; 0 \leq k \leq d$
and
\begin{equation}\label{eq:36}
\Big\| \bigwedge_{i \in I_j} A_j e_i \Big\| \geq c \quad \forall \; 1 \leq j \leq m \; .
\end{equation}
Furthermore,
\begin{equation}\label{eq:37}
\sum_j w_j \left\vert I_j \cap \lbrace k+1, \dots, d \rbrace \right\vert \leq d-k \quad \forall \; 0 \leq k \leq d \; .
\end{equation}
\end{lem}
The connection to our parameters $\delta,\Delta$ follows from a crucial property, which is implicitly established in the proof of~\citep[Proposition~5.2]{tao-paper}:
\begin{cor}\label{cor:delta}
The optimizer $X^*$ for Problem~\ref{eq:F} is bounded from above and below, with both bounds depending on the input datum $\big( \mathcal{A}, w \big)$ only.
\end{cor}
In short, the argument is as follows: We again consider Gaussian inputs $(Z_j)_{j \in [m]}$ and set $\sum_j w_j \big( A_j^* Z_j A_j \big) =: X$.\footnote{Note that here, again, we perform a variable change as in Theorem~\ref{thm:maximizer} to be consistent with our geodesically convex formulation (problem~\ref{eq:F}).} By Lemma~\ref{prop:5.1}, this transformation is self-adjoint, since the matrices $\big( A_j \big)_{j \in [m]}$ are non-degenerate and $w_j >0$. Therefore, we can diagonalize $X= P^{-1} {\rm diag}\big(\lambda_1, \dots, \lambda_d \big) P$, where $\lambda_d >0$. As a consequence, there must exists a $\delta>0$, such that $\lambda_d \geq \delta$.
Furthermore, from Eq.~\ref{eq:36}, we see that $\lbrace A_j e_i \rbrace_{i \in I_j}$ forms a basis of $H_j$ with a lower bound on degeneracy. Therefore
\begin{equation*}
\det (Z_j) \leq C \prod_{i \in I_j} \lambda_i
\end{equation*}
for some $C>0$, which depends only on the input datum. This further implies
\begin{align*}
\prod_{j=1}^m \big( \det Z_j \big)^{w_j}
&\leq C \prod_{i=1}^d \lambda_i^{\sum_{j=1}^m w_j \left\vert I_j \cap \lbrace i \rbrace \right\vert} \\
&\overset{(1)}{\leq} C \lambda_1^d \prod_{0 \leq k \leq d-1} \Big( \frac{\lambda_{k+1}}{\lambda_k} \Big)^{\sum_{j=1}^m w_j \left\vert I_j \cap \lbrace k+1, \dots, d \rbrace \right\vert} \\
&\overset{(2)}{\leq} C \lambda_1^d \prod_{0 \leq k \leq d-1} \Big( \frac{\lambda_{k+1}}{\lambda_k} \Big)^{d-k} \\
&\overset{(3)}{\leq} C \lambda_1 \cdot \dots \cdot \lambda_k \; ,
\end{align*}
where (1) follows from telescoping and applying the scaling condition (Theorem~\ref{thm:feasible}(2)), (2) from Eq.~\ref{eq:37} and (3) from reverse telescoping.
Since we have restricted ourselves to simple BL input data (see Theorem~\ref{thm:maximizer}), Eq.~\ref{eq:37} holds with inequality (see also Definition~\ref{def:simple}). A refinement of the above argument then implies that
\begin{equation}\label{eq:38}
\prod_{j=1}^m \big( \det (Z_j) \big)^{w_j} \leq C \det(X) \prod_{1 \leq k \leq d-1} \Big( \frac{\lambda_{k+1}}{\lambda_k} \Big)^{c} \; ,
\end{equation}
for $c>0$ as in Lemma~\ref{prop:5.1}, which again depends only on the input datum. This implies that Eq.~\ref{eq:BL-lieb} goes to zero as $\frac{\lambda_d}{\lambda_1}$ goes to zero. Hence, the supremum (Eq.~\ref{eq:BL-sup}) must lie in a region with $\lambda_1 \leq C \lambda_d$. This implies upper and lower bounds on $X^*$ and thus the existence of a compact convex set of the form $D_{\delta,\Delta}$ that contains $X^*$.
To implement our proposed algorithm, we need to find a good initialization $X_0 \in D_\delta$, which requires the explicit computation of $\delta$ and $\Delta$. We leave this as an open problem.\\
\noindent We conjecture that there is a close relation between our $\delta$ and the parameter $c$ in Theorem~\ref{prop:5.1} (via Eq.~\ref{eq:36}), which is characterized by the bit complexity of the input datum $(\mathcal{A}, w)$. In contrast to $\delta$, $c$ can be directly computed by solving a constrained subdeterminant maximization problem over the set of linear transforms $\big(A_j\big)_{j \in [m]}$. This observation may serve as a starting point for future investigation into the computation of the parameter $\delta$. Given $\delta$ and $c$, the parameter $\Delta$ could then be inferred from Eq.~\ref{eq:38}.
To see that $c$ can be computed via subdeterminant maximization, note that with an ONB and the index sets $I_j$ in Theorem~\ref{prop:5.1}, we have
\begin{align}
\ip{Z_j A_j e_i}{A_j e_i}_{H_j} = \ip{e_i}{A_j^* Z_j A_j e_i}_H \leq \frac{1}{w_j} \ip{e_i}{X e_i}_H = \frac{\lambda_i}{w_j} \; .
\end{align}
Furthermore, we have with Eq.~\ref{eq:36} that $\big( A_j e_i \big)_{i \in I_j}$ forms a basis of $H_j$ with a lower bound on degeneracy, characterized by $c$.
Recall that $c$ is defined, such that we can find for each $j \in [m]$ an index set $I_j \subseteq \lbrace 1, \dots, d \rbrace$ with $\vert I_j \vert = d'$, such that
\begin{equation*}
\Big\| \bigwedge_{i \in I_j} A_j e_i \Big\|_{d'} > c \; .
\end{equation*}
Consider first the case where the $A_j$ are vectors $\big( a_1^j, \dots, a_d^j \big)$, i.e. $\vert I_j \vert =1$. Then $\norm{a_j e_i} \geq \vert a_i^j \vert$
and therefore $c = \min_{j \in [m]} \vert a_i^j \vert = \min_{j \in [m]} \norm{a^j}_\infty$, i.e., each $A_j$ has an entry with absolute value at least $c$. In the general case $A_j: \mathbb{R}^d \rightarrow \mathbb{R}^{d'}$ we have (${\rm row}_k A$ indicating the $k$th row of $A$)
\begin{align*}
\Big\| \bigwedge_{i \in I_j} A_j e_i \Big\|_{d'} &= \Big\|A_j e_{i_1} \wedge \dots \wedge A_j e_{i_{d'}}\Big\|_{d'} \\
&= \Big\| {\rm row}_{i_1} A_j \wedge \dots \wedge {\rm row}_{d'} A_j \Big\|_{d'} \\
&= \Big\| \det \big( ({\rm row}_{i_1} A_j)^T, \dots, ({\rm row}_{d'} A_j)^T \big) e_{i_1} \wedge \dots \wedge e_{i_{d'}} \Big\|_{d'} \\
&= \left\vert \det \big( ({\rm row}_{i_1} A_j)^T, \dots, ({\rm row}_{d'} A_j)^T \big) \right\vert \; .
\end{align*}
For each $j$, this translates to an optimization problem of the form
\begin{equation}\label{eq:subdet}
\max_{\vert I \vert = d'} \left\vert \det \big( (A_j)_I \big) \right\vert \; ,
\end{equation}
which corresponds to a constrained \emph{subdeterminant maximization}. For computing $c$, we simply take the minimum of all subdeterminant maximizers, i.e.
\begin{align*}
c := \min_{j \in [m]} \max_{\vert I \vert = d'} \left\vert \det \big( (A_j)_I \big) \right\vert \; .
\end{align*}
Problem~\ref{eq:subdet} is known to be NP-hard. An approximation algorithm, which achieves accuracy up to a factor of $O(e^{d'})$ was given by~\citet{nikolov}:
\begin{theorem}[\citet{nikolov}, Theorem~3]
There exists a deterministic polynomial time algorithm, which approximates problem~\ref{eq:subdet} within a factor of $e^{d' + o(d')}$.
\end{theorem}
\noindent For a detailed analysis of the algorithmic aspects of this problem and Nikolov's approach, see also~\citep{vishnoi_subdet}.
\section{Other Picard iterations for computing Brascamp-Lieb constants}\label{sec:other-picard}
This paper analyzes a specific Picard iteration, generated by the map $G$, for computing Brascamp-Lieb constants and analyzes its convergence via the Thompson part metric. We note that neither the choice of the Picard iteration, nor the specific Finslerian lens employed in the analysis is unique. Our particular choice is motivated by the observation that one can establish \emph{strict contractivity} of $G$ with respect to the Thompson part metric, which allows for developing \emph{non-asymptotic} convergence guarantees. In this section we discuss one other Picard iteration that arises from problem~\ref{eq:F} and establish asymptotic convergence.
Consider the map
\begin{equation}
\tilde{G} := \frac{G}{\norm{G}} \; ,
\end{equation}
where $\norm{\cdot}$ may denote any matrix norm. We want to analyze $\tilde{G}$ with respect to the \emph{Hilbert projective metric}, which is given by
\begin{equation}
\delta_H(X,Y) := \log \big( M(X/Y) M(Y/X) \big),
\end{equation}
where again $M(X/Y) := \inf \lbrace \lambda > 0 : X \preceq \lambda Y \rbrace$. We note the close relationship between the Thompson part metric $\delta_T$ (see Definition~\ref{def:thompson}) and the Hilbert projective metric $\delta_H$. Both metrics induce a Finslerian geometry on manifold of positive definite matrices $\posdef_d$.
One can show that $\tilde{G}$ is non-expansive and asymptptically regular with respect to $\delta_H$ and that it has a (not necessary unqiue) fixed point in the compact set $\tilde{D}:= \lbrace X: \frac{X}{\norm{X}}$. Asymptotic convergence of $\tilde{G}$ is then guaranteed by a Hilbert geometry version of Theorem~\ref{thm:s-h}. Such a result can be obtained by specializing the more general Banach space version of the theorem~\citep[Theorem 1.2]{baillon_asymptotic_1978} to Hilbert geometry.
We can show that the Picard iteration defined by the map $\tilde{G}$ convergences to the same fixed point as that of the map $G$: Let $X^*$ denote a fixed point of $\tilde{G}$. Then $G(X^*)=\lambda X^*$ for some $\lambda >0$. Corollary~\ref{cor:delta} implies that $\lambda=1$, which gives the claim.
It remains an open question to show, whether $\tilde{G}$ is strictly contractive with respect to the Hilbert projective norm. This would allow to adapt our non-asymptotic convergence analysis to the map $\tilde{G}$ as well.
\section{Discussion}
In this paper, we introduced a novel fixed-point approach for computing Brascamp-Lieb constants, which is grounded in nonlinear Perron-Frobenius theory. In contrast to much of the prior literature, which has analyzed the problem through a Riemannian lens, our approach utilizes a Finslerian geometry on the manifold of Hermitian, positive definite matrices, which arises from the Thompson part metric. To the best of our knowledge, this introduces a novel geometric perspective on the problem of computing Brascamp-Lieb constants.
We establish convergence of our proposed fixed-point approach and present a full convergence analysis for a regularized variant. Importantly, we show that the regularized approach attains a non-asymptotic convergence rate of $O \Big( \log \big( \frac{1}{\epsilon},\frac{\Delta}{\delta^{3/2}} ,\sqrt{d}, \frac{1}{R}\big) \Big)$, where $d, \delta,\Delta,R$ depend on the size and structure of the input datum only.
A shortcoming of the present approach lies in the difficulty of characterizing the structural dependency on the input datum more explicitly. In particular, while
a result by~\citet{tao-paper} guarantees the existence of suitable $\delta, \Delta>0$ for each (feasible) Brascamp-Lieb datum, we were unable to develop an approach for explicitly computing their values. Access to suitable $\delta,\Delta$ would allow us to find a good initialization for our proposed algorithm and therefore render our approach into a practical method. In addition, we expect the parameter $R$ to depend on $\delta$, but were unable to derive a precise relation. The mutual dependency of the parameters $d, \delta, \Delta, R$ is left for future work.
In future work, we hope to further investigate approaches for computing $\delta, \Delta$ explicitly. In addition, we hope to apply the techniques presented here to geodesically convex optimization problems similar to ~\eqref{eq:F}, such as the operator scaling problem~\cite{garg2018algorithmic} and difference of convex functions~\citep{sra-weber-2022}.
Other avenues for future investigation include practical applications of our fixed-point approach. Due to the strong connections of Brascamp-Lieb inequalities to important questions in machine learning and information theory, this approach could be of wider interest.
More generally, we hope that the Finslerian lens on fixed point iterations provides a new perspective on the problem of computing Brascamp-Lieb constants. We believe that the tools developed in this work can be applied to a wider class of Picard iterations that arise in the context of Brascamp-Lieb constants (such as those described in sec.~\ref{sec:other-picard}).
\subsection*{Acknowledgments}
This project was started during a visit of MW to MIT, supported by an Amazon Research Award. Part of this work was done while MW visited the Simons Institute for the Theory of Computing in Berkeley, CA, supported by a Simons-Berkeley Research Fellowship. SS acknowledges support from an NSF-CAREER award (1846088).\\
\noindent The authors thank Brian Lins for helpful comments.
\setlength{\bibsep}{3pt}
\bibliographystyle{plainnat}
|
\section{Introduction}\label{sec:intro}
Infrared divergences \cite{Bloch1937, yennie1961,weinberg1965} due to a
massless real scalar field in four spacetime dimensions can be, for the most part, structurally identical to those of the massless photons of quantum electrodynamics or the massless gravitons in the perturbative effective field theory of quantum gravity near a flat background. However, there is a big difference between scalar fields and photons or gravitons in that massless scalar fields can in principle occur in quantum field theories which do not exhibit any gauge invariance or even continuous or discrete global symmetry and therefore do not exhibit the asymptotic symmetries which seem intimately tied to the infrared problem of photons and gravitons \cite{Strominger}. In this paper, we shall study the infrared divergences and their cure in such a scalar field theory and attempt to answer the question as to what is different in that case.
What we will find, in the context of a very specific example, where there are no obvious internal symmetries at all, is one big difference. In this model, one can construct a dressed state of a hard particle where the hard particle is accompanied by a cloud of soft scalars and the soft scalar content of the cloud is fine-tuned in such a way that scattering amplitudes for these dressed particles are free of infrared divergences. Of course such a dressing by soft photons or soft gravitons is already well known, and is already proposed as a solution of the infrared problem of quantum electrodynamics \cite{Chung1965,Kibble1968,Kulish1971} and perturbative quantum gravity . In those cases, once electrically and gravitationally charged particles are dressed, the dressing can be further fine-tuned in such a way that the soft photons and soft gravitons decouple completely. The $S$ matrix factorizes into a hard sector and a soft sector. Corrections to this factorization are suppressed by powers of an infrared cutoff, $E_{res}$.
For the soft scalar fields in our model, this is not the case. There is a residual coupling of the scalar-dressed hard particles to soft scalars which is generically as important as the coupling of the dressed hard particles with each other. There is no way to adjust the dressing in order to remove this interaction. This means that even once the hard particles are dressed with soft scalars, an interaction of the hard particles will still produce more soft scalars.
We find that what does decouple is information. Like the case of photons and gravitons, the interaction of undressed hard particles produces infinite numbers of soft scalars.
Moreover, the $S$ matrix for scattering processes involving hard particles is infrared divergent. We will confirm these facts in the context of our scalar model. We will also confirm that, like photons and gravitons, the soft scalars which escape detection in a scattering experiment carry away copious amounts of information to the point that their entanglement entropy with the hard particles left behind is itself infrared divergent. The result of this entanglement, like for photons and gravitons, is decoherence of the final state of the hard particles and suppression of interference phenomena \cite{Carney2017,Carney2018,Carney2018a}.
When the hard particles are dressed with soft scalars, on the other hand, even if they don't decouple and their interaction still produces some soft scalars, the entanglement between those scalars and the dressed hard particles is negligible, suppressed by powers of the infrared cutoff $E_{res}$. Even when the soft scalars fly away from a scattering event undetected, the loss of quantum information, which would be manifest in decoherence, is negligible. In this sense it is information that decouples.
The rest of this paper is an exposition of the results described in the paragraphs above. It is organized as follows. In section \ref{sec: soft_theorem} we will discuss the details of the scalar field theory model that we will use and we will give a derivation of the soft scalar theorem. The soft scalar theorem will be used to study scattering amplitudes in later sections. In section \ref{canc} we will confirm that, for undressed states, the structure of infrared divergences is practically identical to those of photons or gravitons. We will discuss how the infrared problem is addressed using the Bloch-Nordseick scheme of using inclusive probabilities and we will demonstrate that infrared divergences indeed cancel for the questions that are traditionally asked by particle physicists. Then we will show that some other questions about certain interference phenomena or decoherence are still severely affected by the infrared divergences.
In section \ref{sec: dressing} we will construct dressed states of hard particles. There, we insist that the dressing is added by a formally unitary transformation. This is important since, in the infrared cutoff theory, this is a unitary transform, a simple change of basis in the Hilbert space. When the fundamental infrared cutoff is removed, it becomes an improper unitary transformation. However, even in that case, if it is formally unitary, it implements a canonical transformation between inequivalent representations of the operator algebra of the quantum field theory. We also discuss the Faddeev-Kulish modification of the $S$ matrix that is to be computed to find the scattering amplitudes for dressed particles. In section \ref{cancellation} we demonstrate that the infrared singularities indeed cancel from those amplitudes.
In section \ref{sec: non-decoupling} we establish that soft scalars do not decouple from the dressed states. In section \ref{conclusion} we discuss the implications. We also argue that, even though soft scalars do not decouple, they do not drive appreciable decoherence.
\section{Soft Scalar Theorem} \label{sec: soft_theorem}
Consider a Majorana fermion of mass $m$ coupled to a massless real scalar field $\phi$ via a Yukawa coupling in four spacetime dimensions. The Lagrangian density is given by
\begin{equation}\label{lagrangian_density}
\mathcal{L}=-i\bar{\psi}\left[\slashed{\partial}+m-g\phi\right]\psi-\frac{1}{2}\partial_\mu \phi\;\partial ^\mu \phi - V(\phi)
\end{equation}
where $\psi$ is the Majorana spinor field, $g$ is the Yukawa coupling constant and $\phi$ is the scalar field. Since the Yukawa coupling is marginal in four dimensions, $g$ is dimensionless. We will assume that we can always choose counter-terms so that the scalar field tadpole $\sim\phi$, the scalar field mass term $\sim \phi^2$ and the scalar field trivalent coupling $\sim \phi^3$ are all canceled exactly at each order of perturbation theory. To make the Lehmann-Symanzik-Zimmermann (LSZ) reduction formulae simpler, we will also assume a scalar that the subtraction scheme can be chosen so that the pole in the scalar field propagator has unit residue, $\sim \frac{-i}{q^2-i\epsilon}+\ldots$.
We note that this quantum field theory has no obvious internal symmetries at all. We will study the infrared divergences which occur in the analysis of scattering experiments involving the asymptotic particles of this quantum field theory. We will assume that the coupling is weak so that the particle spectrum resembles the tree level one, containing one Majorana fermion with mass $m$ and one massless real scalar field. There has already been some discussion of soft scalar theorems \cite{Cheung2022}, infrared divergences and
the possibility of asymptotic symmetries playing a role for scalar fields \cite{Henneaux2019,Campiglia2018,Campiglia2018a}. The latter has been discussed in the context of a dual antisymmetric tensor gauge field representation which, for a Yukawa coupling like we use here, is not related to this theory by a local transformation. This is not the direction that we will not pursue here. Instead we will examine the behariour of the scattering matrix in the quantized theory corresponding to (\ref{lagrangian_density}).
A scattering experiment has an incoming state. We will use the notation $|\{p\}\{k\}\{q\}\rangle$ for such a state, with fermions having momenta and helicity in a set $\{p\}$, hard scalars with momenta $\{k\}$ and soft scalars with momenta $\{q\}$. In the course of the scattering, an incoming state evolves to an outgoing state which is a quantum superposition of the incoming states,
\begin{align}
& |\{p\}\{k\}\{q\}\rangle\implies
\nonumber \\
&\sum_{\{p'\}\{k'\}\{q'\}} |\{p'\}\{k'\}\{q'\}\rangle~S^\dagger( \{p'\}\{k'\}\{q'\} ; \{p\}\{k\}\{q\} )
\label{scattering}
\end{align}
The coefficients in the superposition, $~S^\dagger( \{p'\}\{k'\}\{q'\} ; \{p\}\{k\}\{q\} )$, are elements of the $S$ matrix (the dagger in the above formula is there to match conventions),
\begin{align}\label{S}
&S^\lambda( \{p\}\{k\}\{q\}; \{p'\}\{k'\}\{q'\} ) \nonumber\\
&=\langle\{p\}\{k\}\{q\}|
~S^\lambda~|\{p'\}\{k'\}\{q'\} \rangle
\end{align}
We have added a superscript $\lambda$ to the $S$ matrix to remind ourselves that its definition requires a fundamental infrared cutoff.
We will denote this fundamental cutoff by $\lambda$.\footnote{An easy, Lorentz invariant way of introducing such a cutoff is to simply allow the scalar field to have a small mass. }
With the cutoff taken into account, the operator $S^\lambda $ whose matrix elements are discussed above is the Dyson $S$ matrix which is computed using the usual LSZ reduction formulae, time dependent perturbation theory and Feynman diagrams. In the rest of this paper, what we mean by an infrared finite quantity is that said quantity remains finite as we take $\lambda\to 0$. The matrix elements of $S^\lambda$ in (\ref{S}) are generally not such a quantity.
Beyond the fundamental infrared cutoff $\lambda$, we shall also require a distinction between hard and soft particles. Any particle which has energy above a threshold $E_{res}$ will be called a hard particle. Any particle which has energy less than $E_{res}$ will be called a soft particle. We shall call $E_{res}$ the ``detector resolution'' . We will require a hierarchy of scales
\begin{align}\label{hierarchy}
\lambda\ll E_{res},\Lambda
\ll m,~{\rm energies~ of~ hard~ particles}
\end{align}
where $\Lambda$ is a third infrared cutoff, distinct from $\lambda$ and $E_{res}$ that we will introduce shortly.
Fermions are always hard particles. The massless scalar, on the other hand can either be hard or soft, depending on its energy. The validity of the arguments in the rest of this paper will need the inequalities in equation (\ref{hierarchy}).
Let us consider an amplitude for the scattering of some fermions and hard scalar particles as well as the production of a soft scalar particle. Whenever this occurs, there are some contributions to the amplitude which are singular as the momentum of the soft scalar approaches zero.
In direct analogy with the same phenomenon in quantum electrodynamics, which is outlined in beautiful detail in Weinberg's book \cite{Weinberg2005}, the most singular parts come from the emission of the soft scalar from external lines of the amplitude. This singular part, $\sim\frac{1}{q}$ for soft scalar with momentum $q$,
and the next-to leading behaviour
due to a scalar emitted from an outgoing fermion line with momentum $p'$, is gotten from the amplitude without the scalar emission by making the replacement
\begin{align}\label{outgoing_soft}
&\bar {u}_r(p') ~~\to~~ \lim_{q\rightarrow 0}\;\bar {u}_r(p')(-g)\frac{-1}{i\slashed{p}'+i\slashed{q}+m-i\epsilon}\nonumber\\
& =\left(\frac{g m}{p'\cdot q-i\epsilon}-\frac{g}{2m}\right) \bar{u}_r(p') +\mathcal O(q)
\end{align}
where $ \bar {u}_r(p)$ is the momentum space spinor wave-function of the outgoing fermion and
we have used the fact that the spinor satisfies the
Dirac equation $\bar{u}_r(p')(i\slashed{p}'+m)=0 $. Here, we note that the contribution to the next-to leading behaviour from the external fermion line has a very simple form for a scalar field. Corrections to this formula go to zero as $q$ goes to zero.
Similarly, when the soft scalar emission is from an incoming fermion line, the singular part and next-to-leading contribution is
\begin{align}\label{incoming_soft}
u_r(p)~\to~ u_r(p)~ \left( \frac{g m}{-p\cdot q-i\epsilon} - \frac{g}{2m}\right)+\mathcal O(q)
\end{align}
Of course, soft emissions can also take place from internal lines in the Feynman diagrams which contribute to the amplitude. The contribution of such processes is not singular at small $q$, but it does compete with the next-to-leading terms in (\ref{outgoing_soft}) and (\ref{incoming_soft}). The effect of one soft emission from an internal line is the same as adding a vertex, together with a vertex counter-term to each of the fermion propagators inside the amputated correlation function that is used to form the $S$ matrix element $S^\lambda(\{p\}\{k\}; \{p'\}\{k'\} ) $. We will denote this process by the symbol $-g\hat\partial_m$. We note that the operation $\hat\partial_m$ is related to but not exactly the same as taking the derivative of amputated correlation function by the renormalized fermion mass $m$. The discrepancy is due to the fact that we need a subtraction scheme where some counter-terms are $m$-dependent in a way that does not preserve the $m-g\phi$ structure that appears in the Lagrangian (\ref{lagrangian_density}). It is easy to see that even though the $m-g\phi$ structure can be maintained at the tree level, it is already violated at the one-loop level.
Putting this together, and a similar one for soft scalar absorption, we have the leading and next-to-leading soft scalar theorem
\begin{align}
S^\lambda(&\{p\},\{k\}; \{p'\}\{k'\},q' ) \nonumber\\
&= \left\{\sum_{p_n\in\{p\}\{p'\} }\left(\frac{\eta_n g m }{p_n\cdot q'-i\eta_n\epsilon}- \frac{ g}{m}\right)- g\hat\partial_m\right\}\nonumber\\
&\times
S^\lambda(\{p\}\{k\}; \{p'\}\{k'\} )
+\mathcal O(q')
\label{sub_soft_theorem}
\end{align}
\begin{align}
S^\lambda(&\{p\},\{k\},q; \{p'\}\{k'\})
\nonumber\\
&= \left\{\sum_{p_n\in\{p\}\{p'\} }\left(\frac{\eta_n g m }{-p_n\cdot q-i\eta_n\epsilon}- \frac{ g}{m}\right)- g\hat\partial_m\right\}\nonumber\\
&\times S^\lambda(\{p\}\{k\}; \{p'\}\{k'\} )
+\mathcal O(q)
\label{sub_soft_theorem_1}
\end{align}
where $\eta_n=+1(-1)$ if $p_n$ is the momentum of an outgoing (incoming) line.
Now let us consider the amplitude for a process where an incoming state of hard particles and $M$ additional soft scalars $|\{p\}\{k\},q_1,\ldots,q_M\rangle$ evolves to another state of hard particles but with $N$ additional soft particles $|\{p'\}\{k'\},q_1',\ldots,q_N'\rangle$. Using the leading parts of the soft scalar theorem in equations (\ref{sub_soft_theorem}) and (\ref{sub_soft_theorem_1}), it follows that the most singular part of the $S$ matrix element is given by\footnote{We have not attached momentum space wave-functions for the soft scalars. We will assume take the convention of including them as the appropriate factors in the photon states. }
\begin{align}
& S^\lambda(\{p\}\{k\},q_1,\ldots,q_M; \{p'\}\{k'\}, q_1',...q_N')\nonumber \\ &= S^\lambda(\{p\}\{k\}; \{p'\}\{k'\} )
\prod_{j=1}^M\left( \sum_{p_n\in\{p\}\{p'\}}\frac {\eta_n g m }{-p_n \cdot q_j-i\eta_n\epsilon}\right)\nonumber\\
& \times
\prod_{r=1}^N\left( \sum_{p_n\in\{p\}\{p'\}}\frac {\eta_n g m }{p_n \cdot q_r'-i\eta_n\epsilon}\right) +\ldots
\label{soft1}
\end{align}
where the ellipses denote terms less singular at small $q$ or $q'$ than $\frac{1}{q^M}\frac{1}{{q'}^N}$.
The remarkable fact about this soft theorem is that it gives us the most important part of the soft scalar production amplitude for any process if we know the amplitude of the process without the soft scalar production. What is more, it is practically identical to the one for soft photons or gravitons where only the numerators in the singular factors ( here it is $gm$) are slightly different.
\section{Infrared divergence cancellation in the Bloch-Nordseick Scheme}\label{canc}
It is easy to see that, in direct parallel with quantum electrodynamics and perturbative quantum gravity, the infrared divergences in the $S$ matrix itself come from loop integrals where both ends of a single scalar propagator in the loop end on external fermion lines. Introduction of the fundamental infrared cutoff $\lambda$ renders these loop integrals finite and of order $\sim \ln(\lambda)$.
We imagine that, in these loop integrals, the infrared cutoff $\lambda$ could be replaced by a more convenient one which we shall call $\Lambda$. This is a third infrared cutoff, distinct from $E_{res}$ and $\lambda$. with the assumption that it is much larger than the fundamental cutoff, $\Lambda\gg\lambda$, but it would still work as an infrared cutoff in that $\Lambda\ll m$ and that $\Lambda$ is much smaller than the momentum scales of any of the hard particles. It is in the same interval of the hierarchy (\ref{hierarchy}) as the detector resolution $E_{res}$. Then with the new cutoff the logarithmically divergent integral goes as $\sim \ln(\Lambda)$ and the original integral with cutoff $\lambda$ goes like $\sim\ln(\lambda)= \ln (\Lambda) + \ln (\lambda/\Lambda)$, the first logarithm being produced by the integration over loop momenta from $\Lambda$ to infinity and the second being produced by the integration over loop momenta between $\lambda$ and $\Lambda$.
Moreover, it is well known how to separate and sum up the latter, $ \ln (\lambda/\Lambda)$ contributions. We refer the reader to Weinberg's book \cite{Weinberg2005} for the details in the case of photons and we note that, for massless scalars, the argument is practically identical. The result is a relationship between the $S$ matrix defined with the two different infrared cutoffs
\begin{widetext}
\begin{align}
S^\lambda( &\{p\}\{k\}; \{p'\}\{k'\} )
\nonumber\\
&=S^\Lambda( \{p\}\{k\}; \{p'\}\{k'\} )e^{-i \bar\Phi(\{p\})-i\bar \Phi(\{p'\})} \exp\biggl\{- \frac{1}{2}\int_\lambda^\Lambda\frac{d^3q}{(2\pi)^32|\vec q|}
\sum_{p_n\in\{p\}\{p'\}}\frac{gm\eta_n}{p_n\cdot q}\sum_{p_m\in\{p\}\{p'\}}\frac{gm\eta_m}{p_m\cdot q} \biggr\}
\label{infrared_cutoff_s_matrix} \\
&=S^\Lambda( \{p\}\{k\}; \{p'\}\{k'\} ) \left(\frac{\lambda}{\Lambda}\right)^{\frac{1}{2}\bar A(\{p\},\{p'\})}e^{-i\bar \Phi(\{p\})-i\bar \Phi(\{p'\})}\, ;
\label{infrared_cutoff_s_matrix_1}\\
&\bar A(\{p\},\{p'\})=-\frac{1}{8\pi^2}\sum_{p_np_m\in\{p\}\{p'\}}\frac{g^2m^2\eta_n\eta_m}{\gamma_{nm}}\ln\left(\frac{1+\xi_{nm}}{1-\xi_{nm}}\right), ~~\bar \Phi(\{p\})=\frac{1}{8\pi}\sum_{\substack{p_mp_n\in\{p\}\\mu\neq n}}\frac{g^2m^2}{\gamma_{nm}} ~\ln\frac{\Lambda}{\lambda}\,,\\ &\xi_{nm}=\sqrt{1-\frac{m^4}{(p_n \cdot p_m)^2}}~,~~
\gamma_{nm}=(p_n \cdot p_m)\sqrt{1-\frac{m^4}{(p_n \cdot p_m)^2}}\;.
\label{phase}
\end{align}
\end{widetext}
To be clear, the beautiful formula (\ref{infrared_cutoff_s_matrix}) does not remove the fundamental infrared cutoff. It simply gives us a relationship between $S$ matrices computed with different infrared cutoffs. Furthermore, it is strictly valid only when
$
\lambda \ll \Lambda \ll m
$.
We note that the exponent of the ratio of cutoffs contains data about the incoming and outgoing fermions only.
It is independent of the hard incoming or outgoing scalar particles. The phases are separated into two functions, one of incoming and one of outgoing fermion momenta. Notice that the two functions have the same sign.
For reasons which will become clear shortly, it is useful to express the evolution from the initial to final state in equation (\ref{scattering})
in the language of density matrices where a more general incoming state composed entirely of hard particles would be
\begin{align*}
\sum_{ \substack{ \{p\}\{k\}\\ \{\tilde p\}\{\tilde k\} } }~~ |\{p\}\{k\}\rangle\rho_{ \{p\}\{k\}\{\tilde p\}\{\tilde k\} }\langle\{\tilde p\}\{\tilde k\}|
\end{align*}
with $\rho_{\{p\}\{k\}\{\tilde p\}\{\tilde k\} }$ an incoming density matrix. In the course of a scattering experiment, this
incoming density matrix evolves to an outgoing one where the evolution is governed by the $S$ matrix,
\begin{widetext}
\begin{align}
\sum_{ \substack{ \{p\}\{k\}\\ \{\tilde p\}\{\tilde k\} } }~ |\{p\}\{k\}\rangle\rho_{\{p\}\{k\};\{\tilde p\}\{\tilde k\}}\langle\{\tilde p\}\{\tilde k\}|&\implies
\sum_{ \substack{ \{p\}\{k\}\\ \{\tilde p\}\{\tilde k\} } } \rho_{\{p\}\{k\}\{\tilde p\}\{\tilde k\}} \sum_{\substack{ \{p'\}\{k'\}\{q'\} \\ \{\tilde p'\} \{\tilde k'\}\{\tilde q'\} } } |\{p'\}\{k'\}\{q'\}\rangle\langle\{\tilde p'\}\{\tilde k'\}\{\tilde q'\}|
\nonumber \\ &
~~~~~~~~ \times S^{\lambda\dagger}( \{p'\}\{ k'\}\{q'\} ; \{p\}\{k\})~~
S^\lambda( \{\tilde p\}\{\tilde k\} ; \{\tilde p'\}\{\tilde k'\}\{\tilde q'\} ) \,.
\label{scattering_1}
\end{align}
\end{widetext}
We have assumed that there are no soft particles in the incoming states. However, soft particles are produced when the hard particles interact and they must appear in the final state density matrix.
We remember that the $S$ matrix elements in the expression above are infrared divergent and they are defined with a fundamental infrared cutoff, $\lambda$.
We will implement the Bloch-Nordsieck mechanism where we make the assumption that, due to the limitations of detector resolution, the soft particles which are produced by the scattering are unobservable. They fly away from the scattering experiment undetected. What is left behind are
the hard particles. All of the experimentally accessible properties of the quantum state of the hard particles which remain are embedded in the reduced density matrix that is gotten from the final state density matrix in (\ref{scattering_1}) by taking a trace over all of the soft scalar states.
The reduced density matrix of the final state is thus
\begin{widetext}
\begin{align}
&\rho_{\rm final}= \sum_{ \substack{ \{p\}\{k\}\\ \{\tilde p\}\{\tilde k\} } }~\rho_{\{p\}\{k\};\{\tilde p\}\{\tilde k\}}~~ \sum_{\substack{ \{p'\}\{k'\} \\ \{\tilde p'\} \{\tilde k'\} } } |\{p'\}\{k'\}\rangle\langle\{\tilde p'\}\{\tilde k'\}|
\sum_{\{q\}} S^{\lambda\dagger}( \{p'\}\{ k'\}\{q\} ; \{p\}\{k\})
S^\lambda( \{\tilde p\}\{\tilde k\} ; \{\tilde p'\}\{\tilde k'\}\{ q\} )
\label{scattering_2}
\end{align}
where $\sum_{\{q\}} $ denotes integration and summation over all possible soft scalar states.
We can use the soft scalar theorem (\ref{soft1}) to simplify equation (\ref{scattering_2}). To take the trace, we identify pairs of ingoing and outgoing $q$'s and we integrate each identified pair over all values of $\vec q$ with $\lambda<|\vec q|<E_{res}$. Using the soft scalar theorem yields the expression\footnote{Since, once the scalar momenta $q$ are on-shell, $p_n\cdot q>0$ and we can drop the $i\epsilon$'s from the denominators.}
\begin{align}
& \sum_{\{q\}} S^{\lambda\dagger}( \{p'\}\{ k'\}\{q\} ; \{p\}\{k\})~~
S^\lambda( \{\tilde p\}\{\tilde k\} ; \{\tilde p'\}\{\tilde k'\}\{ q\} ) \nonumber \\
&= \sum_{N=0}^\infty\frac{1}{N!}\int_\lambda^{E_{res}}\frac{d^3\vec{q_1}}{(2\pi)^32|\vec{q_1}|}\ldots\frac{d^3\vec{q_N}}{(2\pi)^32 |\vec{q_N}|}
\prod_{r=1}^N\left( \sum_{ p_n\in\{p\}\{p'\}}\frac{gm\eta_n}{p_n\cdot q}
\sum_{ p_n\in\{p\}\{p'\}} \frac{gm\eta_m}{-p_m\cdot q} \right)
\nonumber \\ &~~~~~~~~~~~~~~~ \times S^{\lambda\dagger}( \{p'\}\{ k'\} ; \{p\}\{k\})~~
S^\lambda( \{\tilde p\}\{\tilde k\} ; \{\tilde p'\}\{\tilde k'\} )\,. \nonumber
\end{align}
\end{widetext}
Notice that, in each factor in the product of integrals in the equation above, the three dimensional integration over $\vec q$ is over a narrow shell with $\lambda<|\vec q|<E_{res}$ and the result would be small, $\sim E_{res}^2$, if it were not for the singular terms due to soft photon emission. It is those singular terms which allow the integrals to be appreciable, in fact logarithmically infrared divergent $\sim \ln E_{res}/\lambda$. Corrections to the above formula due to the non-singular next-to-leading contributions to the soft scalar theorem would be relatively suppressed by positive powers of $E_{res}$.
The summation in the equation above exponentiates and we find the expression for the reduced final state density matrix
\begin{widetext}
\begin{align}
\rho_{\rm final}&= \sum_{ \substack{ \{p\}\{k\}\\ \{\tilde p\}\{\tilde k\} } }~ \rho_{\{p\}\{k\};\{\tilde p\}\{\tilde k\}} \sum_{\substack{ \{p'\}\{k'\} \\ \{\tilde p'\}\{\tilde k'\} } } |\{p'\}\{k'\}\rangle\langle\{\tilde p'\}\{\tilde k'\}|~ S^{\lambda\dagger}( \{p'\}\{ k'\} ; \{p\}\{k\})
S^\lambda( \{\tilde p\}\{\tilde k\} ; \{\tilde p'\}\{\tilde k'\} )
\nonumber \\ &~~
\times \exp\biggl\{ \int_\lambda^{E_{res}}\frac{d^3\vec{q}}{(2\pi)^32|\vec{q}|}
\sum_{ p_n\in\{p\}\{p'\} } \frac{gm\eta_n}{p_n\cdot q}
\sum_{ p_m\in\{\tilde p\}\{\tilde p'\} }
\frac{ g m \eta_{m} }{-p_{m}\cdot q}
\biggr\}\,.
\label{sscattering_3}
\end{align}
Now, we must examine the infrared cutoff dependence of the $S$ matrix elements on the right-hand-side of equation (\ref{sscattering_3}). For this, we must use the $\lambda$-dependence of the $S$ matrix elements that is summarized in equation (\ref{infrared_cutoff_s_matrix}). We then get
\begin{align}
&\rho_{\rm final}= \sum_{ \substack{ \{p\}\{k\}\\ \{\tilde p\}\{\tilde k\} } }~\rho_{\{p\}\{k\}\{\tilde p\}\{\tilde k\}}~ \sum_{\substack{ \{p'\}\{k'\} \\ \{\tilde p'\} \{\tilde k'\} } } |\{p'\}\{k'\}\rangle\langle\{\tilde p'\}\{\tilde k'\}| \nonumber \\ & \times
S^{\Lambda\dagger}( \{p'\}\{ k'\} ; \{p\}\{k\})
S^\Lambda( \{\tilde p\}\{\tilde k\} ; \{\tilde p'\}\{\tilde k'\} )
e^{i \bar\Phi(\{p\}) + i\bar \Phi(\{p'\}) - i \bar\Phi(\{\tilde p\}) - i\bar \Phi(\{\tilde p'\})} \nonumber \\ &
\times \exp\biggl\{ \int_\lambda^{E_{res}}\frac{d^3\vec{q}}{(2\pi)^32|\vec{q}|}
\sum_{ p_n\in\{p\}\{p'\} } \frac{gm\eta_n}{p_n\cdot q}
\sum_{ p_m\in\{\tilde p\}\{\tilde p'\} }
\frac{ g m \eta_{m} }{-p_{m}\cdot q}
\biggr\}
\nonumber \\ &
\times \exp\biggl\{- \int_\lambda^\Lambda\frac{d^3q}{(2\pi)^32|\vec q|}\biggl[ \frac{1}{2}
\biggl[ \sum_{ p_n\in\{ p\}\{ p'\}} \frac{gm\eta_n}{p_n\cdot q} \biggr]^2 +\frac{1}{2}
\bigg[ \sum_{ p_n\in\{\tilde p\}\{\tilde p'\}} \frac{gm\eta_n}{p_n\cdot q} \biggr]^2
\biggr] \biggr\}\,.
\label{scattering_5}
\end{align}
Upon combining the exponentials in the last two lines, we can write the above expression as
\begin{align}
&\rho_{\rm final}=\sum_{ \substack{ \{p\}\{k\}\\ \{\tilde p\}\{\tilde k\} } }~\rho_{\{p\}\{k\}\{\tilde p\}\{\tilde k\}}~~ \sum_{\substack{ \{p'\}\{k'\} \\ \{\tilde p'\} \{\tilde k'\} } } |\{p'\}\{k'\}\rangle\langle\{\tilde p'\}\{\tilde k'\}| \nonumber \\ & \times
S^{\Lambda\dagger}( \{p'\}\{ k'\} ; \{p\}\{k\})
S^\Lambda( \{\tilde p\}\{\tilde k\} ; \{\tilde p'\}\{\tilde k'\} )
e^{i \bar\Phi(\{p\}) + i\bar \Phi(\{p'\}) - i \bar\Phi(\{\tilde p\}) - i\bar \Phi(\{\tilde p'\})} \nonumber \\ &
\times \exp\biggl\{ \int_\Lambda^{E_{res}}\frac{d^3\vec{q}}{(2\pi)^32|\vec{q}|}
\sum_{ p_n\in\{p\}\{p'\} } \frac{gm\eta_n}{p_n\cdot q}
\sum_{ p_m\in\{\tilde p\}\{\tilde p'\} }
\frac{ g m \eta_{m} }{-p_{m}\cdot q}
\biggr\}
\nonumber \\ &
\times \exp\biggl\{-\frac{1}{2} \int_\lambda^\Lambda\frac{d^3q}{(2\pi)^32|\vec q|}\biggl[ \sum_{ p_n\in\{ p\}\{ p'\}} \frac{gm\eta_n}{p_n\cdot q}- \sum_{p_m \in\{ \tilde p\}\{ \tilde p'\}} \frac{gm\eta_m}{p_m\cdot q} \biggr]^2\biggr\}\,.
\label{scattering_7}
\end{align}
\end{widetext}
Now we want to examine the right-hand-side of equation (\ref{scattering_7}) as the fundamental infrared cutoff $\lambda\to 0$.
The second line contains phases which are separately infrared divergent and we reserve comment on them for later.
The third line is $\lambda$-independent and infrared finite.
The fourth (last) line has a negative semi-definite exponent which can be written as
\begin{align*}
&\exp\biggl\{-\frac{1}{2} \int_\lambda^\Lambda\frac{d^3q}{(2\pi)^32|\vec q|}\biggl[ \sum_{ p_n\in\{ p\}\{ p'\}} \frac{gm\eta_n}{p_n\cdot q}\nonumber\\
&~~~~~~~~~- \sum_{p_m \in\{ \tilde p\}\{ \tilde p'\}} \frac{gm\eta_m}{p_m\cdot q} \biggr]^2\biggr\}
\\ &
=\exp\biggl\{-\frac{1}{32\pi^3} \ln\frac{\Lambda}{\lambda}\cdot\int d\hat q \biggl[ \sum_{ p_n\in\{ p\}\{ p'\}} \frac{gm\eta_n}{p_n\cdot v}\nonumber\\&~~~~~~~~- \sum_{p_m \in\{ \tilde p\}\{ \tilde p'\}} \frac{gm\eta_m}{p_m\cdot v} \biggr]^2\biggr\}
\;;\; v^\mu=(1,\hat q)~,~~\hat q\equiv\vec q/|\vec q|\,.
\end{align*}
This exponent is either negative and logarithmically divergent or it vanishes. It can vanish only if the integrand in the integration over unit vectors vanishes, that is, if
\begin{align}
\sum_{ p_n\in \{p\} \{\tilde p'\} }\frac{gm}{p_n\cdot v} =
\sum_{ p_m\in \{p'\} \{\tilde p\} }\frac{gm}{p_m\cdot v}\,.
\label{condition}
\end{align}
It is only in this case where the $|\{p'\}\{k'\}\rangle\langle\{\tilde p'\}\{\tilde q'\}|$ matrix element of the reduced outgoing density matrix can be nonzero when the matrix
element $|\{p\}\{k\}\rangle\langle\{\tilde p\}\{\tilde q\}|$ of the incoming density matrix was nonzero.
Remember that the sums in (\ref{condition}) are over terms containing hard fermion momenta only. Hard scalar momenta do not enter in these expressions.
The above equation must be so for all values of the null four-vector $v^\mu = \left(1,\hat q\right)$. If we Taylor expand the above in powers off $\frac{\vec p_n}{\sqrt{\vec p_n^2+m^2}}<1$ and equate each order we see that
\begin{align*}
\sum_{p_n\in \{p\}\{\tilde p'\}} \frac{(\hat q\cdot \vec p_n)^\ell }{(\sqrt{\vec p_n^2+m^2})^{\ell+1}}
&=
\sum_{p_n\in \{\tilde p\}\{p'\} } \frac{(\hat q \cdot \vec p_n)^\ell }{(\sqrt{\vec p_n^2+m^2})^{\ell+1}},\nonumber\\
&~~~~~~~~~~~~~~~~~~~~~~~~\forall\ell,~\forall \hat q\,. \end{align*}
which implies that all multipole moments of the set of fermion momentum vectors $\{p\}\cup \{\tilde p'\}$ are equal to all moments of the set $\{p'\}\cup \{\tilde p\}$ which can only be so if
the two sets of vectors are identical
\begin{align}
\{p\}\cup \{\tilde p'\}=\{p'\}\cup \{\tilde p\}\,.
\label{constraint}
\end{align}
Only those elements of the density matrix for which this criterion is satisfied survive the limit $\lambda\to 0$.
Now, notice that, when the constraint (\ref{constraint}) is obeyed, the phases in the third line of equation (\ref{scattering_7}) also cancel exactly.
The result is, for evolutions from the initial to the final state for which (\ref{constraint}) holds, the final density matrix is free of infrared divergences.
Let us examine some of the consequences of this result.
Since, by unitarity, both the incoming and reduced outgoing density matrices must have unit trace, they both must have nonzero diagonal matrix elements.
Moreover, a question about the evolution of a diagonal element of the density matrix element to another diagonal element of the density matrix is unaffected by the constraint (\ref{constraint}). This is due to the fact that, in such a case, $\{p\}=\{p'\}$ and $\{\tilde p\}=\{\tilde p'\}$ and (\ref{constraint}) is automatically satisfied.
This is the question that is usually asked in particle physics: what is the probability that the state $|\{p\}\{k\}\rangle\langle\{p\}\{k\}|$ will evolve to the state $|\{p'\}\{k'\}\rangle\langle\{p'\}\{k'\}|$? This quantity is finite in the limit $\lambda\to 0$ and it is unconstrained by the condition (\ref{constraint}). We have nothing new to say about it.
On the other hand, if we ask the questions which probe the off-diagonal elements of the density matrix, the $\lambda\to0$ limit can have drastic consequences.
If we ask what is the probability of the process
\begin{align*}
& \frac{1}{\sqrt{2}}\left(|\{p_1\}\{k_1\}\rangle+|\{p_2\}\{k_2\}\rangle\right)\times~\nonumber\\
&\frac{1}{\sqrt{2}}\left(\langle\{p_1\}\{k_1\} | + \langle\{p_2\}\{k_2\} |\right)\implies |\{p'\}\{k'\}\rangle\langle\{p'\}\{k'\}|\,,
\end{align*}
we find that, unless $\{p_1\}=\{p_2\}$, it is
\begin{align*}
\frac{1}{2}~{\rm Prob.~of~} |\{p_1\}\{k_1\}\rangle\langle\{p_1\}\{k_1\}|& \implies |\{p'\}\{k'\}\rangle\langle\{p'\}\{k'\}|\nonumber\\
&+\nonumber\\
\frac{1}{2}~{\rm Prob.~of~} |\{p_2\}\{k_2\}\rangle\langle\{p_2\}\{k_2\} |& \implies |\{p'\}\{k'\}\rangle\langle\{p'\}\{k'\}|\,.
\end{align*}
There is no interference between the incoming states.
The probability for evolving to a superposition, on the other hand
\begin{align*}
|\{p\}\{k\}\rangle\langle\{p\}\{k\}|&\implies \frac{1}{\sqrt{2}}\left(|\{ p_1'\}\{ k_1'\}\rangle+|\{p_2'\}\{k_2'\}\rangle\right)\nonumber\\
&\times~\frac{1}{\sqrt{2}}\left( \langle\{p_1'\}\{k_1'\} | + \langle\{p_2'\}\{k_2'\} |\right)
\end{align*}
unless $\{p_1'\}=\{p_2'\}$, is simply the sum
\begin{align*}
\frac{1}{2}~{\rm Prob.~of~} |\{p\}\{k\}\rangle\langle\{p\}\{k\}| & \implies |\{p_1'\}\{k_1'\}\rangle\langle\{p_1'\}\{k_1'\}|\nonumber\\
&+\nonumber\\
\frac{1}{2}~{\rm Prob.~of~} |\{p\}\{k\}\rangle\langle\{p\}\{k\}| &\implies |\{p_2'\}\{k_2'\}\rangle\langle\{p_2'\}\{k_2'\}|\,.
\end{align*}
The state experiences complete decoherence.
In order to get the results outlined above, we reduced the final state density matrix by tracing it over the Fock space states
of the soft scalar fields. Of course, since we are tracing over this entire subspace of the total Hilbert space,
tracing over any redefinition of the Fock basis for soft scalars by a unitary transformation
must give the same answer. To find a result that is any different than what we have obtained,
one would have to use a redefinition of the basis that is not implemented by a proper unitary transformation.
Of course, it is easy to find such improper unitary transformations in a Fock space.
Here, the relevant one is the basis constructed around certain coherent states which become improper coherent states when the
fundamental infrared cutoff is removed. Indeed,
this redefinition of the basis for soft scalar states is easily implemented and, when one subsequently traces in such a basis, the reduced final state density matrix differs in ways that are physically consequential. In that basis the evolution does not exhibit the severe
decoherence or suppression of interference that we found for the Fock state basis.
However, then we would be in a situation where the $S$ matrix evolves an initial soft scalar Fock vacuum to soft scalar coherent states which live in a different Hilbert space and the $S$ matrix itself is therefore not a proper unitary operator. The only way to preserve unitarity of $S$ in this context is to also use the coherent states as incoming states, so that $S$ evolves coherent states to coherent states in such a way that it is unitary. This is the gist of what is done in the dressed state formalism which we will discuss in the context of soft scalar fields in the next section. We expect that the dressed states, in the limit where the cutoff is removed, will have the same problems with violations of Lorentz invariance as photon and graviton dressed states \cite{Froehlich1979,Froehlich1979a,Balachandran2013}.
\section{ Soft Scalar Dressing} \label{sec: dressing}
In quantum electrodynamics, a dressed state \cite{Chung1965,Kibble1968} is a modification of the quantum state of a charged hard particle which attaches a coherent state of soft photons to it. As well, it must be accompanied by a singular redefinition of the phases of the $S$ matrix \cite{Kulish1971}. A similar idea can be used to obtain dressed states of gravitationally charged particles in quantum gravity when that theory written as an effective field theory for perturbations of flat spacetime \cite{Ware2013,Choi2019}. In both electrodynamics and gravity, the dressing of states can be done in such a way that the $S$ matrix that describes the scattering of hard dressed particles is infrared finite. In both cases, it has been argued that the soft photons or gravitons simply decouple from the dressed states \cite{Kulish1971,Mirbabayi2016}
in that their interactions are suppressed by factors of $E_{res}/$(hard particle scales) which can be very small.
In the following, we will argue that the same dressing procedure can be implemented for hard particles which interact with soft scalar fields. In our simple model (\ref{lagrangian_density}) it is a close parallel to the construction for photons or gravitons. We will reserve the discussoin of decoupling or non-decoupling for a later section.
Consider an incoming Fock space state of hard particles $ |\{p\}\{k\}\rangle$. We will consider our quantum field theory with a fundamental infrared cutoff $\lambda$.
Following Chung and Faddeev and Kulish \cite{Chung1965,Kulish1971}, we define the dressed state, which we denote by $ |\{p\}\{k\}\rangle\rangle$, as
\begin{align}\label{dressing}
|\{p\}\{k\}\rangle\rangle~\equiv~ W(\{p\}) ~ |\{p\}\{k\}\rangle\,
\end{align}
where $W(\{p\})$ is the unitary operation
implemented on an incoming state with fermion quantum numbers $\{p\}$ as
\begin{align}
& W(\{p\})=\exp\left( R(\{p\} ) \right) \label{dressing1} \\
&R(\{p\})= \int_\lambda^{E_{res}}\frac{d^3k}{\sqrt{(2\pi)^32|\vec k|}}\ \biggl\{
\left[\sum_{p_n\in\{p\}} f(k,p_n)\right] a^\dagger(k)\nonumber\\
&~~~~~~~~~~~~~~~~~~~~~~~~~~~-\left[ \sum_{p_n\in\{p\}} f^*(k,p_n)\right] a(k)\biggr\} \label{dressing2} \\
&f(k,p)=\frac{gm}{k\cdot p}\,. \label{dressing3}\end{align}
Note that $R(\{p\})$ is anti-Hermitian and $W(\{p\})$ is unitary.
In this expression, $\lambda$ is a fundamental infrared cutoff, and $E_{res}$ is a second cutoff analogous to the detector resolution
of the previous section, and we use the same symbol for it here \footnote{Note that $f$ defined in \eqref{dressing3} is real. Thus $f^*$ in \eqref{dressing2} appears to over-complicate the expression. However, we prefer to write it this way to allow for any (possibly complex) subleading term in the definition of the dressed states, should we require it.}.
Our normalization of the creation and annihilation operators is such that
free field is
\begin{align}\label{phi}
\phi_{\rm in}(x)=\int \frac{d^3k}{\sqrt{(2\pi)^32|\vec k|}}\left( e^{ikx} a(k)+e^{-ikx}a^\dagger(k)\right)
\end{align}
and $$\left[a(k),a^\dagger(k')\right]=\delta^3(\vec k-\vec k')\,.$$ We have omitted the wave-function, $\frac{1}{\sqrt{(2\pi)^32|\vec k|}}$, for the scalar fields from our expressions for the $S$ matrix. This means that we must compensate by taking the normalization integral for the states to be the Lorentz invariant measure
$\int \frac{d^3k}{(2\pi)^32|\vec k|}$ when we finally sum over scalar field states.
In addition to the dressing of states described in equations (\ref{dressing1})-(\ref{dressing3}), the $S$ matrix must be modified. As an operator on Fock space, the modified $S$ matrix is similar to the Dyson $S$ matrix that is computed in Feynman-Dyson-Wick perturbation theory and which we used in the previous sections, the only difference is that it should be multiplied by some phases which take into account the infinite range of interactions.
If we consider the transition between a dressed state $|\{p\}\{k\}\rangle\rangle$ and a dressed state $|\{p'\}\{k'\}\rangle\rangle$, modified $S$ matrix, which we shall denote by the symbol ${\bf S}$ is defined by
\begin{align}
& {\bf S} ( \{p\}\{k\};\{p'\}\{k'\})~=~ \langle\langle \{p'\}\{k'\}| ~{\bf S}~|\{p\}\{k\}\rangle\rangle\\
& ~~\equiv~~ e^{i\Phi(\{p\})}~ \langle\langle \{p'\}\{k'\}|~ { S^\lambda}~|\{p\}\{k\}\rangle\rangle~e^{i\Phi(\{p'\})} \label{defn_bf_S}\,;\\
&\Phi(\{p\}) =-\frac{1}{8\pi}\sum_{\substack{p_n,p_m\in \{p\}\\mu\neq n}}\frac{g^2m^2}{ \sqrt{(p_n \cdot p_m)^2-m^4}} \ln\frac{E_{res}}{\lambda}\,.
\end{align}
As we shall see, the infrared diverging phases serve to cancel the infrared divergent parts of the phases due to infrared divergent loop integrals encountered in the computation of $S^\lambda$.
\section{Infrared Finiteness of the Dressed $S$ Matrix}\label{cancellation}
In this section we shall show that matrix elements of the $\bf S$ matrix between dressed states are free of infrared divergences.
The proof is very similar to the analogous one for quantum electrodynamics and for perturbative quantum gravity \cite{Chung1965,Ware2013}. Consider the Dyson $S$ matrix operator $S^\lambda$ computed in renormalized perturbation theory and the dressed states defined in equations (\ref{dressing})-(\ref{dressing3}).
The matrix element of the $\bf S$ matrix between dressed states is given in equation (\ref{defn_bf_S}) which we recopy here for the reader's convenience:
\begin{align}\label{mathcal_S}
&{\bf S}(\{p\}\{k\};\{p'\}\{k'\})\nonumber\\
&= e^{i\Phi(\{p\})}~ \langle\langle \{p'\}\{k'\}|~ { S^\lambda}~|\{p\}\{k\}\rangle\rangle~e^{i\Phi(\{p'\})}\,.
\end{align}
The superscript $\lambda$ on $S^\lambda$ indicates that the matrix elements on the right-hand-side of the above equation are computed while using $\lambda$ as a fundamental
infrared cutoff. The dressed states $|\{p\}\{k\} \rangle\rangle$ and the phases $\bar\Phi(\{p\})$ are also defined with this cutoff. Cancellation of this singular dependence
on $\lambda$ and finiteness as $\lambda$ is put to zero on the right-hand-side of equation (\ref{mathcal_S}) is the ``infrared finiteness'' that we are seeking in this section.
We can use the Baker-Campbell-Hausdorff formula\footnote{For operators $A$ and $B$ with the properties
$[A,[A,B]]=0$ and $[B,[A,B]]=0$,
the Baker-Campbell-Hausdorff formula is
$[A,[A,B]]=0$ and $[B,[A,B]]=0$,
$$
e^A e^B =e^{\frac{1}{2}[A,B]} e^{A+B} = e^{[A,B]} e^B e^A
$$
} to rewrite the exponential operators in the dressed states defined in equations (\ref{dressing})-(\ref{dressing3}) as
\begin{align}
&|\{p'\}\{k'\}\rangle\rangle\nonumber\\
&=
\exp\biggl\{ \int_\lambda^{E_{res}}\frac{d^3\ell}{ {(2\pi)^32|\vec \ell|} } \biggl[ -\frac{1}{2}
\biggl|\sum_{p_n'\in\{p'\}} f( \ell,p_n ')\biggr|^2\nonumber\\
&+ \sum_{p_n'\in\{p'\}}
f(\ell,p_n')a^\dagger(\ell) \biggr] \biggr\}
|\{p'\}\{k'\}\rangle\,,
\end{align}
\begin{align}
&\langle\langle\{p\}\{k\}|\nonumber\\
&=
\langle\{p\}\{k\}|
\exp\biggl\{ \int_\lambda^{E_{res}}\frac{d^3\ell}{ {(2\pi)^32|\vec \ell|} }\nonumber\\
&\times \biggl[ -\frac{1}{2}\biggl|
\sum_{p_n\in\{p\}} f( \ell,p_n )\biggr|^2-
\sum_{p_n\in\{p\}} f^*( \ell,p_n )a(\ell) \biggr] \biggr\}\,.
\end{align}
Then, using these states, equation (\ref{mathcal_S}) becomes
\begin{widetext}
\begin{align}
&{\bf S} (\{p\}\{k\};\{p'\}\{k'\})= \nonumber \\ &\exp\biggl\{ i\Phi(\{p\})+i\Phi(\{p'\})
+ \int_\lambda^{E_{res}}\frac{d^3\ell}{ {(2\pi)^32|\vec \ell|}} \biggl[ -\frac{1}{2}
\left|\sum_{p_n\in\{p\}}f(\ell,p_n)\right|^2 -\frac{1}{2} \left|\sum_{p_n'\in\{p'\}} f(\ell,p_n')\right|^2\ \biggr] \bigg\} \times \nonumber
\\
&
\langle\{p\}\{k\}|\exp\biggl\{ - \int_\lambda^{E_{res}}\frac{d^3\ell}{\sqrt{(2\pi)^32|\vec \ell|}}\sum_{p_n\in \{p\}}
f^*(\ell,p_n)a(\ell) \biggr\} S^\lambda
\exp\biggl\{ \int_\lambda^{E_{res}}\frac{d^3\ell}{\sqrt{(2\pi)^32|\vec \ell|}}\sum_{p_m'\in\{p'\}}
f(\ell,p_m')a^\dagger(\ell) \biggr\}|\{p'\}\{k'\}\rangle\,.
\end{align}
A compact form for the LSZ formula for the scalar field part of the Dyson $S$ matrix is obtained using a generating functional
\begin{align*}
\left. S^\lambda = ~:e^{ \int dz\phi_{\rm in}(z)(-\partial^2)\frac{\delta}{\delta J(z) } }:~\ldots \mathcal T \ldots e^{ i\int J\phi} \ldots
\ldots \right|_{J=0}
\end{align*}
where $\phi_{\rm in}(x)$ is the asymptotic free field as in equation (\ref{phi}) and $:\ldots:$ denotes the normal ordering.
With this formula, we can find the action of the dressing operator on $S^\lambda$ as
\begin{align*}
&\langle\{p\}\{k\}| e^{ - \int_\lambda^{E_{res}}\frac{d^3\ell}{\sqrt{(2\pi)^32|\vec \ell|}}\sum_{p_n\in \{p\}}
f^*(\ell,p_n)a(\ell) }
S^\lambda
e^{ \int_\lambda^{E_{res}}\frac{d^3\ell}{\sqrt{(2\pi)^32|\vec \ell|}}\sum_{p_m'\in\{p'\}}
f(\ell,p_m')a^\dagger(\ell) } |\{p'\}\{k'\}\rangle
\\
&=
\exp\biggl\{ - \int_\lambda^{E_{res}}\frac{d^3\ell}{{(2\pi)^32|\vec \ell|}}\sum_{ p_n\in\{p\}}f^*(\ell,p_n) \sum_{p_m'\in\{p'\} } f(\ell,p_m') \biggr\}
\exp\left(\int dzf(z)(-\partial^2)\frac{\delta}{\delta J(z) } \right) \\ & \times \left.
\langle\{p\}\{k\}|~:e^{\int dz\phi_{\text{in}}(z)(-\partial^2)\frac{\delta}{\delta J(z) } }:~\ldots \mathcal T \ldots e^{ i\int J\phi} \ldots
\ldots ~|\{p'\}\{k'\}\rangle~\right|_{J=0}
\end{align*}
In the above formula, we have used the equations
\begin{align}
& e^{ - \int_\lambda^{E_{res}}\frac{d^3\ell}{\sqrt{(2\pi)^32|\vec \ell|}}\sum_{p_n\in \{p\}}
f^*(\ell,p_n)a(\ell) }a^\dagger(k) =
\left( a^\dagger(k)-\sum_{p_n\in\{p\}}f^*(p_n,k)\right) e^{ - \int_\lambda^{E_{res}}\frac{d^3\ell}{\sqrt{(2\pi)^32|\vec \ell|}}\sum_{p_n\in \{p\}}
f^*(\ell,p_n)a(\ell) }\,,
\label{coherent_identity_1}\\
&a(k) e^{ \int_\lambda^{E_{res}}\frac{d^3\ell}{\sqrt{(2\pi)^32|\vec \ell|}}\sum_{p_m'\in\{p'\}}
f(\ell,p_m')a^\dagger(\ell) }
=
e^{ \int_\lambda^{E_{res}}\frac{d^3\ell}{\sqrt{(2\pi)^32|\vec \ell|}}\sum_{p_m'\in\{p'\}}
f(\ell,p_m')a^\dagger(\ell)}\left( a(k)+\sum_{p_n'\in\{p'\}}f(k,p_n')\right)
\label{coherent_identity_2}\,,
\end{align}
and
the result is the classical scalar field,
$$
f(z)=\int_\lambda^{E_{res}} \frac{d^3\ell }{(2\pi)^32|\vec \ell|}\biggl(\sum_{p_n'\in\{p'\}} f(\ell,p_n')e^{i\ell z} - \sum_{p_m\in\{p\}} f^*(\ell,p_m)e^{-i\ell z}\biggr)
$$
occurring in the exponential functional derivative operator in the second line,
$$
\exp\left(\int dz f(z)(-\partial^2)\frac{\delta}{\delta J(z) }\right)\,.$$
This operation inserts ingoing and outgoing soft scalars with wave-functions $\sim f,-f^*$ into the $S$ matrix element
that it operates on. We can use the soft scalar theorem (\ref{soft1}) to re-write these as the $S$ matrix element without the soft scalars and with an exponential factor,
$$
\exp\biggl\{ \int_\lambda^{E_{res}} \frac{d^3\ell}{(2\pi)^32|\vec \ell |} \left[ \sum_{p_n\in\{p\}} \frac{gm}{p_n\cdot \ell} - \sum_{p_n'\in\{p'\} } \frac{gm}{p_n\cdot \ell}\right]
\left[ \sum_{p_m\in\{p\} } f(\ell,p_m)-
\sum_{p_m'\in\{p'\} } f^*(-\ell,p_m') \right] \biggr\}
$$
which then leads us to
\begin{align}
&{\bf S} (\{p\}\{k\};\{p'\}\{k'\})= { S}^\lambda (\{p\}\{k\};\{p'\}\{k'\})~e^{i\Phi(\{p\})+i\Phi(\{p'\})} \nonumber \\ &\times
\exp\biggl\{ \int_\lambda^{E_{res}}\frac{d^3\ell}{ {(2\pi)^32|\vec \ell|}}\biggl[ -\frac{1}{2} \left|\sum_{p_n'\in\{p\}}
f(\ell,p_n)\right|^2 -\frac{1}{2}\left| \sum_{p_n'\in\{p'\}}
f(\ell,p_n')\right|^2-\sum_{p_n\in\{p\}}f^*(\ell,p_n) \sum_{p_m'\in\{p'\}} f(\ell,p_m') \nonumber \\ &
+ \biggl[ \sum_{p_n\in\{p\}} \frac{gm}{p_n\cdot \ell} - \sum_{p_n'\in\{p'\} } \frac{gm}{p_n\cdot \ell}\biggr]
\biggl[ \sum_{p_m\in\{p\} } f(\ell,p_m)-
\sum_{p_m'\in\{p'\} } f^*(-\ell,p_m') \biggr] \biggr] \biggr\}
\end{align}
Then, we can use equations (\ref{infrared_cutoff_s_matrix}) to write the above equation as
\begin{align}
&{\bf S}^\lambda (\{p\}\{k\};\{p'\}\{k'\})= { S}^\Lambda (\{p\}\{k\};\{p'\}\{k'\})~e^{i\Phi(\{p\})-i\bar \Phi(\{p\})+i \Phi(\{p'\})-i\bar \Phi(\{p'\})} \nonumber \\
&\times
\exp\biggl\{ \int_\lambda^{E_{res}}\frac{d^3\ell}{ {(2\pi)^32|\vec \ell|}}\biggl[ -\frac{1}{2} \left|\sum_{p_n'\in\{p\}}
f(\ell,p_n)\right|^2 -\frac{1}{2} \left|\sum_{p_n'\in\{p'\}}
f(\ell,p_n')\right|^2 -\sum_{p_n\in\{p\}}f^*(\ell,p_n) \sum_{p_m'\in\{p'\}} f(\ell,p_m') \nonumber \\ &
+ \biggl[ \sum_{p_n\in\{p\}} \frac{gm}{p_n\cdot \ell} - \sum_{p_n'\in\{p'\} } \frac{gm}{p_n'\cdot \ell}\biggr]
\biggl[ \sum_{p_m\in\{p\} } f(\ell,p_m)-
\sum_{p_m'\in\{p'\} } f^*(-\ell,p_m') \biggr]
\nonumber \\ &
+ \int_\lambda^{\Lambda}\frac{d^3\ell}{ {(2\pi)^32|\vec \ell|}}\biggl[ -\frac{1}{2} \left( \sum_{p_n\in\{p\}} \frac{gm}{p_n\cdot \ell } \right)^2
-\frac{1}{2}\left( \sum_{p_n'\in\{p'\}}\frac{gm}{p_n'\cdot \ell} \right)^2
+ \sum_{p_n\in\{p\}}\frac{gm}{p_n\cdot \ell} \sum_{p_m'\in\{p'\}} \frac{gm}{p_m'\cdot \ell }
\biggr]\biggr\}
\label{almost_final}\end{align}
\end{widetext}
The first line in equation (\ref{almost_final}) contains the phases that come from the Faddeev-Kulish prescription plus from the internal loop contributions to the infrared cutoff $S^\lambda$. The logarithmic infrared divergences cancel in the combinations in which the phases appear there, leaving behind finite parts which we will display shortly. The second and third lines in equation (\ref{almost_final}) comes from the normalizations and overlaps of the coherent states. The fourth line in equation (\ref{almost_final}) contains the result of using the soft scalar theorem to take into account the soft photons coming from the coherent states. The last line in equation (\ref{almost_final}) is the contribution of internal loops encountered in the computation of $S^\lambda$. It is easy to see that the $\lambda$-dependence of the sum of all of these terms cancels when we put $f(p,\ell)=\frac{gm}{p\cdot\ell}$. The latter exponential factors simplify as
\begin{align*}
&\exp\biggl\{
\int_{E_{res}}^{\Lambda}\frac{d^3\ell}{ {(2\pi)^32|\vec \ell|}}\biggl[ -\frac{1}{2} \left( \sum_{p_n\in\{p\}} \frac{gm}{p_n\cdot \ell } \right)^2 \nonumber\\
&
-\frac{1}{2}\left( \sum_{p_n'\in\{p'\}}\frac{gm}{p_n'\cdot \ell} \right)^2
+ \sum_{p_n\in\{p\}}\frac{gm}{p_n\cdot \ell} \sum_{p_m'\in\{p'\}} \frac{gm}{p_m'\cdot \ell }
\biggr]\biggr\}
\nonumber \\
&=\left( \frac{E_{\rm res}}{\Lambda}\right)^{A(\{p\}\{p'\})}
\end{align*}
where
\begin{align*}
A(\{p\}\{p'\})& = \frac{1}{8\pi^2}\sum_{nm} \frac{g^2m^2\eta_m\eta_n}{\sqrt{(p_m\cdot p_n)^2-m^4}}
\biggl[i\biggl(\frac{1+\eta_m\eta_n}{2}\biggr)
\nonumber\\
&+ \ln\frac{1+\sqrt{ 1 - \frac{m^4}{ (p_m\cdot p_n)^2} } }{1-\sqrt{ 1 - \frac{m^4}{ (p_m\cdot p_n)^2} } } \biggr]\,.
\end{align*}
The final result for the element of the $S$ matrix in dressed states is simply
\begin{align}\label{final_formula_for_dressed_S_matrix}
{\bf S} (&\{p\}\{k\};\{p'\}\{k'\})\nonumber\\
&=\left( \frac{E_{\rm res}}{\Lambda}\right)^{A(\{p\}\{p'\})}~ { S}^\Lambda (\{p\}\{k\};\{p'\}\{k'\})\,.
\end{align}
The left-hand-side of the above formula (\ref{final_formula_for_dressed_S_matrix}) is the matrix element of the modified $S$ matrix, $\bf S$, computed with dressed states.
It is the amplitude for the transition from dressed state $|\{p\}\{k\}\rangle\rangle$ to dressed state $|\{p'\}\{k'\}\rangle\rangle$. The right-hand-side simply contains the Dyson $S$ matrix, $S^\lambda$, its matrix elements $ S^\Lambda (\{p\}\{k\};\{p'\}\{k'\})$ computed with undressed Fock space states and with an infrared cutoff $\Lambda$. It is multiplied the factor $\left( \frac{E_{\rm res}}{\Lambda}\right)^{A(\{p\}\{p'\})}$ made from the ratio of cutoffs raised to a complex, momentum-dependent exponent. The right-hand-side does not depend on $\Lambda$ in that the $\Lambda$-dependence of $S^\Lambda$ is compensated by the $\Lambda$ dependence of the factor. However, the right-hand-side does depend on the cutoff, $E_{res}$ which is now a parameter of the theory.
\section{Non-decoupling of Soft Scalar Emission}\label{sec: non-decoupling}
Having constructed the infrared finite $S$ matrix for dressed states, we are now interested in computing the amplitude of emission of additional soft scalars beyond those in
the dressing. The vanishing of such amplitudes and the factorization of the $S$ matrix into hard and soft sectors for the scattering amplitudes of dressed states is
an already well-known feature of quantum electrodynamics and perturbative quantum gravity \cite{Kulish1971,Mirbabayi2016,Carney2018,Choi2019}.
There, the factorization has to do with the fact that one could correct the dressing factors to take into account the next-to-leading contributions to the soft theorem. Then amplitude for the interaction of suitably dressed states to emit or absorb an additional soft photon or soft graviton is indeed suppressed by powers of the detector resolution cutoff $E_{res}$. This decoupling also has sound physical reasoning. Photons or gravitons with wavelengths the size of the solar system should have nothing to do with electrodynamic or gravitational physics at a subatomic scale.
As we shall see, the case of the massless scalar field is a little different. We will find that soft scalars do not decouple.
A scattering event for hard dressed particles can produce soft scalars with an amplitude of the same order as other radiative corrections to the amplitude and at the same order as the interactions between the hard particles themselves. We will outline the argument for this in the following. Later, in the next section, we will examine the entanglement of the dressed hard particles and the soft particles that are produced in the scattering of hard particles.
Let us consider an outgoing dressed state which contains an additional soft scalar
\begin{align}
|\{p'\}\{k'\}q'\rangle\rangle
=e^{R(\{p'\})} a^\dagger(q') |\{p'\} \{k'\} \rangle
\end{align}
We would like to compute the dressed $S$ matrix element ${\bf S}( \{p\}\{k\};\{p'\}\{k'\}q')$ which is defined
as the quantity
\begin{align}
&{\bf S}( \{p\}\{k\};\{p'\}\{k'\}q')=
e^{i \Phi(\{p\}) }\times \nonumber\\
& \langle\{p\}\{k\}|W^\dagger (\{p\})S^\lambda W(\{p'\}) a^\dagger(q')|\{p'\}\{k'\}\rangle e^{i \Phi(\{p'\})}
\end{align}
We can move the scalar creation operator past the operator $W(\{p'\})$ by using the identity (\ref{coherent_identity_1}) to get
\begin{align}
&{\bf S}( \{p\}\{k\};\{p'\}\{k'\}q')
\nonumber \\ &
=e^{i \Phi(\{p\}) + i \Phi(\{p'\}) } \langle\{p\}\{k\}|W^\dagger (\{p\})S^\lambda\nonumber\\
&\times\left( a^\dagger (q')- \sum_{p_n'\in\{p'\} }f(p'_n,q')\right)W(\{p'\}) )|\{p'\}\{k'\}\rangle
\end{align}
Then, the $a^\dagger(q')$ can either act on $W^\dagger(\{p\})$, for which we use equation (\ref{coherent_identity_2}) or it could be absorbed into the $S$ matrix, in which case we can use the soft scalar theorem (\ref{sub_soft_theorem}), which we will keep to next-to-leading order. The result is a cancellation between the singular factors and the $f$'s, leaving only the contributions from the next-to-leading soft scalar theorem,
\begin{align}
&{\bf S}( \{p\}\{k\};\{p'\}\{k'\}q')
= ~\left( \frac{E_{\rm res}}{\Lambda}\right)^{A(\{p\}\{p'\})}\nonumber\\
&\times \left\{-\sum_{p_n\in\{p\}\{p'\} }\frac{ g}{m}- g\hat\partial_m\right\}{ S^\Lambda}( \{p\}\{k\};\{p'\}\{k'\})
\end{align}
where quantities on the right-had-side are defined in the discussions around equation (\ref{sub_soft_theorem})-(\ref{sub_soft_theorem_1}) and (\ref{final_formula_for_dressed_S_matrix}).
This equation is one of our central results. The right-hand-side, unlike what occurs for photons or gravitons, cannot be written in a form that depends only on the initial and final states in a way that its effect can be absorbed into $W(\{p'\})$ or $W^\dagger(\{p\})$. Even absorbing the first term with $-\frac{g}{m}$ for each external fermion would modify $W$ in such a way that it is no longer unitary. It is easy to confirm by a simple tree level computation that $-g\hat\partial_m$ operating on an amputated correlation function simply cannot in general be written as something that depends only on $\{p\}\{k\}$ plus something that depends only on $\{p'\}\{k'\}$ and its action therefore cannot be absorbed into the $W$'s either. Our conclusion is that the soft scalars couple to the dressed $S$ matrix at order $g$ in the coupling constant, which makes them just as coupled as the other particles.
\section{Conclusions}\label{conclusion}
We have shown that, in the context of our admittedly rather specialized model, the infrared problem due to massless scalar fields is practically identical to that for photons in quantum electrodynamics and for gravitons in perturbative quantum gravity.
This is in spite of the fact that there are no apparent internal symmetries, either continuous or discrete, whatsoever.
This means that there are no conserved Noether currents beyond the energy-momentum tensor, no Ward-Takahashi identities and no apparent asymptotic symmetries.
We have found what we conjecture is different in the soft scalar theory. The difference lies in the next-to-leading soft scalar theorem. Unlike the case of photons or gravitons where Ward-Takahashi identities help to write those terms as referring only to the initial and final states \cite{Choi2019}, their scalar analog cannot be written that way. Then, unlike for photons and gravitons, the next-to-leading behaviour cannot be absorbed by modifying the hard particle dressing. Soft scalars are still produced when dressed hard particles interact.
Now that we have demonstrated that soft scalars do not decouple from the interactions of dressed states, we can revisit the question as to whether they carry any significant amount of information. We still expect the scenario where the interactions of hard particles in a scattering event also produces a cloud of soft particles and those soft particles are undetectable. Our experimental resources only have access to the dressed hard particles. To proceed, we could simply ask the same question that we did for scattering of undressed hard particles. We consider an incoming density matrix state
\begin{align} \rho_{\{p\}\{k\};\{\tilde p\}\{\tilde k\}}~~ \sum_{\substack{ \{p\}\{k\} \\ \{\tilde p\} \{\tilde k\} } } |\{p\}\{k\}\rangle\rangle \rho_{\{p\}\{k\};\{\tilde p\}\{\tilde k\}}~\langle\langle\{\tilde p\}\{\tilde k\}| \,
\end{align}
which should evolve to an outgoing density matrix
\begin{align}
& ~ \sum_{\substack{ \{p\}\{k\}\ \\ \{\tilde p\} \{\tilde k\} } } ~ \sum_{\substack{ \{p'\}\{k'\}\{q'\} \\ \{\tilde p'\} \{\tilde k'\} \{\tilde q'\} } } |\{p'\}\{k'\}\{q'\}\rangle\rangle \langle\langle\{\tilde p'\}\{\tilde k'\}\{\tilde q'\} |
\nonumber \\ &
~~~~~~~ \times {\bf S}^{ \dagger}( \{p'\}\{ k'\}\{q\} ; \{p\}\{k\})~ \rho_{\{p\}\{k\};\{\tilde p\}\{\tilde k\}}~\nonumber\\ &
~~~~~~~ \times {\bf S}( \{\tilde p\}\{\tilde k\} ; \{\tilde p'\}\{\tilde k'\}\{ q'\} )\,
\end{align}
which we now trace over the outgoing soft scalars to obtain a reduced density matrix which describes all of the accessible physics of the hard particles in the outgoing state,
\begin{align}
&\rho_{\rm final}= ~ \sum_{\substack{ \{p\}\{k\}\ \\ \{\tilde p\} \{\tilde k\} } } ~ \sum_{\substack{ \{p'\}\{k'\} \\ \{\tilde p'\} \{\tilde k'\} } } |\{p'\}\{k'\}\rangle\rangle \langle\langle\{\tilde p'\}\{\tilde k'\} |
\nonumber \\ &
~~~~~~~ \times\sum_{\{q\}} {\bf S}^{ \dagger}( \{p'\}\{ k'\}\{q\} ; \{p\}\{k\})~ \rho_{\{p\}\{k\};\{\tilde p\}\{\tilde k\}}~\nonumber \\ &
~~~~~~~
{\bf S}( \{\tilde p\}\{\tilde k\} ; \{\tilde p'\}\{\tilde k'\}\{ q\} ) \,.
\label{soft_info_decoupling_1}\end{align}
Then, we observe that the trace over the soft scalars involves
\begin{widetext}
\begin{align}
& \sum_{\{q\}} {\bf S}^{ \dagger}( \{p'\}\{ k'\}\{q\} ; \{p\}\{k\})~ \rho_{\{p\}\{k\};\{\tilde p\}\{\tilde k\}}~
{\bf S}( \{\tilde p\}\{\tilde k\} ; \{\tilde p'\}\{\tilde k'\}\{ q\} ) \nonumber \\ &
= {\bf S}^{ \dagger}( \{p'\}\{ k'\} ; \{p\}\{k\})~ \rho_{\{p\}\{k\};\{\tilde p\}\{\tilde k\}}~
{\bf S}( \{\tilde p\}\{\tilde k\} ; \{\tilde p'\}\{\tilde k'\} )
\nonumber \\
& +\int_\lambda^{E_{res}}\frac{d^3q'}{(2\pi)^32|\vec q'|}
{\bf S}^{ \dagger}( \{p'\}\{ k'\}q'; \{p\}\{k\})~ \rho_{\{p\}\{k\};\{\tilde p\}\{\tilde k\}}~
{\bf S}( \{\tilde p\}\{\tilde k\} ; \{\tilde p'\}\{\tilde k'\}q')
+\ldots
\end{align}
\end{widetext}
Since the combination
\begin{align*}
&{\bf S}^{ \dagger}( \{p'\}\{ k'\}q'; \{p\}\{k\}) \rho_{\{p\}\{k\};\{\tilde p\}\{\tilde k\}}
{\bf S}( \{\tilde p\}\{\tilde k\} ; \{\tilde p'\}\{\tilde k'\}q') \nonumber\\
&~\sim~ ({q'})^0
\end{align*}
is not singular as $q\to 0$, it
goes like a constant, the integration in the last line of the above equation produces a factor of $E_{res}^2$ and the higher order terms represented by the ellipses there have multiple volume integrals which produce higher orders of $E_{res}$. All of these are suppressed. In the approximation where we neglect contributions with positive powers of $E_{res}$, we neglect such terms. Then equation (\ref{soft_info_decoupling_1})
becomes
\begin{align}
&\rho_{\rm final}= ~ \sum_{\substack{ \{p\}\{k\}\ \\ \{\tilde p\} \{\tilde k\} } } ~ \sum_{\substack{ \{p'\}\{k'\} \\ \{\tilde p'\} \{\tilde k'\} } } |\{p'\}\{k'\}\rangle\rangle \langle\langle\{\tilde p'\}\{\tilde k'\} |
\nonumber \\ &
~~~~~~~ \times {\bf S}^{ \dagger}( \{p'\}\{ k'\} ; \{p\}\{k\})~ \rho_{\{p\}\{k\};\{\tilde p\}\{\tilde k\}}\nonumber \\ &
~~~~~~~ \times
{\bf S}( \{\tilde p\}\{\tilde k\} ; \{\tilde p'\}\{\tilde k'\} ) +\mathcal O(E_{res}^2)
\label{soft_info_decoupling_2}
\end{align}
with no reference to soft particles at all.
This is the sense in which the information contained in the soft particles decouples. We could produce soft scalars in a scattering experiment. However, if we do not have the detector resolution to see them directly, we have no way of knowing that they are there. It will also be interesting to understand the implications of our result in the context of Hawking, Perry, and Strominger's proposal \cite{Hawking2016,Strominger2019} of the resolution of the black hole information paradox which has been heavily criticized based on the decoupling of soft gravitons in the dressed formalism \cite{Mirbabayi2016}.
\begin{acknowledgments}
This work is supported in part by the Natural Sciences and Engineering Research Council of Canada (NSERC).
\end{acknowledgments}
|
\section*{I. INTRODUCTION}
Quantum information science is a new frontier subject combining quantum mechanics and information science. The quantum nature of particle superposition, entanglement, and measurement are applicable for more efficient information processing, computation, transmission and storage. The development of quantum information industry promises more powerful computing power, high security information communication and a deeper understanding of nature. It is expected to resolve numerous scientific problems that are difficult to be effectively solved by existing classical techniques. For example, absolutely secure information transmission \cite{Bennett1984,Gisin2002}, exponentially or polynomial accelerating the resolving of hard scientific problems \cite{Shor1994,Grover1997}, efficient simulation of molecular structures \cite{Feynman1982} and precision measurement beyond the standard quantum limit \cite{Giovannetti2011}.
In the last decades, quantum fundamental science is rapidly transformed into quantum technologies with huge resources invested by global academia, research centers and industry. Quantum scientific research is moving from the stage of principle verification of quantum rules to the stage of practical device research and development governed by these rules. We have seen that quantum computational advantage realized in different systems \cite{Arute2019,Zhong2020,Wu2021}, quantum key distribution via thousands of kilometer level satellite-to-ground quantum network \cite{Liao2017} and quantum-enhanced sensing for gravitational waves \cite{Abbott2016}, magnetic field \cite{Boss2017} and mass of nanoparticle \cite{Zheng2020}.
Photon, as one ideal information carrier, has been widely used in quantum information processing and shows several unique advantages such as fast transmission speed, low noise, multiple degrees of freedom for information encoding and high capacity. Besides, light has a wide range of applications in energy, communication, computation, medical care etc. These increasingly mature industrial applications provide favorable supports for photonic quantum technologies. Currently, fibre optical communication technology has become the pillar for information transmission, and photonic integration and optoelectronic integration for information processing has developed rapidly in recent years for the future all-optical network. Therefore, optical quantum systems are becoming a promising platform for quantum information processing \cite{O'brien2009,Flamini2018,Slussarenko2019}.
To enhance the complexity of quantum optical experiments, optical quantum systems tend to use photonic integrated circuits \cite{Wang2020}. Compared to systems that use discrete optical components on an optical table, the integrated photonic devices enable localize and manipulate photons
at micro/nano scales, and thus greatly improve the stability and scalability of quantum optical experiments and provide a complex and compact quantum photonics approach for quantum communication, sensing and computing applications \cite{Lu2021}. Therefore, the integrated techniques will lead quantum applications moving out of the laboratory and into large-scale and practical. Multiple optical waveguide materials, such as silica \cite{Politi2008,Jiang2017,Zhang20192,Li2022}, silicon (Si) \cite{Silverstone2016}, silicon nitride (SiN) \cite{Zhang2018,Lu2019,Taballione2021,Ren2022}, lithium niobate \cite{Jin2014,Wang20172,Xu2022}, and techniques, such as lithography \cite{Bojko2011}, laser-writing \cite{Sugioka2014}, have been developed. Among those materials, silicon photonics is a good candidate because of its easy preparation, high integration density and excellent optical properties \cite{Rudolph2017}. Also thanks to the leadership of classical silicon photonics in large-scale photonic integrated circuits, the performance of silicon-based devices has been rapidly improved \cite{Thomson2016}.
Currently, photonic devices on the silicon chip for various quantum applications have been developed, sucn as high-efficiency chip-fiber coupler \cite{Son2018}, large-scale programmable quantum photonic circuits \cite{Harris2016} and single-photon detectors \cite{Pernice2012}. However, scalable quantum applications put forward new and higher requirements for integrated photonic devices. For example, high isolation on-chip filter is needed to separate pump and single-photon level signals, and modulator should also operate at cryogenic condition. In this paper, we review the relevant research results and state-of-the-art technologies on the silicon photonic chip for scalable quantum applications. Since silicon optical chips can be easy prepared, their applications are very wide and diverse. Here we just review common complementary metal oxide semiconductor (CMOS) compatible silicon-on-insulator (SOI) chips with telecom optical band O + C. We're glad to see that properties of some integrated components have already met the requirements for further quantum application and expansion. Furthermore, we point out the challenges ahead and further research directions. This review mainly focus on advances of silicon photonic devices, and for more information about related novel quantum experiments, we recommend some related good reviews, such as preparation of photon-pair sources with silicon photonics \cite{Feng2020,Chen2021}, advances of quantum experiments on the silicon chip \cite{Wang2020,Adcock2020}, and hybrid integrated techniques for quantum applications \cite{Elshaari2020,Kim2020}.
The review is structured as follows: Section II reviews the state-of-the-art technologies of basic components on the silicon chip, such as on-chip photon sources, detectors, chip interconnects and so on. In Section III, we introduce some key and fundamental quantum information processing in silicon photonic circuits and review research works on multiphoton and high dimensional applications, quantum error correction, quantum key distribution as well as quantum state teleportation between chips. These excellent works leads to scalable quantum computing and communication. Finally, in Section IV, we outlook the future of silicon photonics for scalable quantum information applications and provide challenges for further scaling.
\section*{II. Scalable techniques on silicon photonic chips}
Scalable quantum information processing needs to improve the complexity on the single chip as much as possible. Different degrees of freedom of photons should achieve universal quantum operation, and quantum photonic source, detector, logic operation and other core functions should be improved to high quality and be integrated on the same chip. Considering the limited chip size and difficulty in hybrid integration with different materials, multifunction chip can also be achieved
by high efficiency optical interconnection for quantum communication and distributed quantum computing and metrology. Accordingly, we divide the state-of-the-art techniques in silicon photonics into two aspects to review, single-chip and interconnection technologies. The contents of this section include single-photon sources, photon detection, wavelength and mode division multiplexing and cryogenic techniques on the single chip and chip interconnects techniques.
\subsection{A. Single-photon sources}
\begin{figure*}[t]
\centering
\includegraphics[width=14cm]{1.jpg}
\caption {\textbf{Solid-state quantum emitters in silicon photonics.} a. Position grown InAs/InP quantum dots on a silicon photonic chip by the pick-and-place technique. Adapted from Ref. \cite{Kim2017}. b. Integrate heterogeneous optical components with transfer-printing-based
approach. Adapted from Ref. \cite{Katsumi2019}. c. Atomic structure of the G-centre. Adapted from Ref. \cite{Redjem2020}. d. Si nanopillar including the G-centre. Adapted from Ref. \cite{Hollenbach2021}. e. Bullseye structures used to enhance vertical coupling of G-centers. Adapted from Ref. \cite{Prabhu2022}. }
\label{fig1}
\end{figure*}
\begin{table*}
\caption{\textbf{State-of-the-art photonic sources at telecommunication wavelengths.} For herald single-photon sources, pump power normalized pair generation rate are given. While for solid-state emitters, saturated emission rate are given. PGR: pair generation rate; SER: saturated emission rate: CAR: coincidence-to-accident ratio. $\ast$ indicates the use of pulsed laser pump.}
\begin{tabular}{ |p{1cm}|p{3.2cm}|p{2cm}|p{2.5cm}|p{1cm}|p{1.5cm}|p{4cm}| }
\hline
\makecell[c]{Ref.}& \makecell[c]{Structure} &\makecell[c]{Bandwidth}&\makecell[c]{PGR/SER}&\makecell[c]{CAR}&\makecell[c]{$g^2(0)$}&\makecell[c]{Wavelength}\\
\hline
\makecell[c]{\cite{Li2017}} &\makecell[c]{Single-mode waveguide} & \makecell[c]{100$\,$GHz}& \makecell[c]{0.7$\,$MHz.mW$^{-2}$}&
\makecell[c]{80}&\makecell[c]{-- --}&\makecell[c]{1538.2$\,$\&$\,$1562.2$\,$nm}\\
\hline
\makecell[c]{\cite{Ma2017}}& \makecell[c]{Micro-ring resonator} &\makecell[c]{2.1$\,$GHz}& \makecell[c]{149$\,$MHz.mW$^{-2}$}&\makecell[c]{12105}&\makecell[c]{0.00533}&\makecell[c]{1535.5$\,$\&$\,$1574.7$\,$nm} \\
\hline
\makecell[c]{\cite{Paesani2020} $\ast$}& \makecell[c]{Multimode waveguide} & \makecell[c]{4$\,$nm} &\makecell[c]{18.6$\,$MHz.mW$^{-2}$}&\makecell[c]{-- --}&\makecell[c]{0.053}&\makecell[c]{1516$\,$\&$\,$1588$\,$nm}\\
\hline
\makecell[c]{\cite{Zhou2018}} &\makecell[c]{GaN Defect} & \makecell[c]{3--50$\,$nm}& \makecell[c]{1.5$\,$MHz}&\makecell[c]{-- --}&\makecell[c]{0.05}&\makecell[c]{1085--1340$\,$nm}\\
\hline
\makecell[c]{\cite{Zhao2021} $\ast$}& \makecell[c]{2D MoTe$_2$} & \makecell[c]{8.5--37$\,$nm}&\makecell[c]{-- --}&\makecell[c]{-- --}& \makecell[c]{0.058}&\makecell[c]{1080--1550$\,$nm}\\
\hline
\makecell[c]{\cite{Hollenbach2020}}& \makecell[c]{G center} & \makecell[c]{0.5$\,$nm} &\makecell[c]{99$\,$kHz}&\makecell[c]{-- --}&\makecell[c]{0.07}&\makecell[c]{1278$\,$nm}\\
\hline
\makecell[c]{\cite{Higginbottom2022}}& \makecell[c]{T center} & \makecell[c]{255$\,$MHz} &\makecell[c]{-- --}&\makecell[c]{-- --}&\makecell[c]{0.2}&\makecell[c]{1326$\,$nm}\\
\hline
\end{tabular}
\end{table*}
Scalable quantum photonic information processing requires multiple high-quality single photon sources. Today, there are two promising approaches to achieving a near-deterministic single photon source, one is based on multiplexing an array of probabilistic parametric photon pair sources, and the other is based on solid-state single-photon emitters.
With strong third-order nonlinear response, silicon waveguides can be directly used to prepare photon pair sources via spontaneous four-wave mixing process (SFWM) \cite{Li2017}. Many techniques were developed to improve photonic source quality, such as using micro-ring resonators \cite{Ma2017,Liu2020,Burridge2020} and introducing special phase-matching condition \cite{Paesani2020}. More details are given in Table I. Upon pair generation, one of the photons is detected to herald the presence of its partner. Although each photon pair source is probabilistic, multiple heralded single photons can be dynamically switched to one single output mode, and thus increasing the output probability. Any degree of freedom of the photon pairs can be used for multiplexing, such as path, frequency and time bins. Also, these can be combined to multiplicatively increase the number of multiplexed modes. In ref. \cite{Collins2013}, a 63.1\% efficiency increase in the heralded single-photon output has been demonstrated through multiplexing photons generated in two silicon waveguides. In ref. \cite{Zhang2015}, the enhancement was increased to 90$\pm$5\% by using a silicon waveguide pumped by time and wavelength division multiplexed pulses. In ref. \cite{Xiong2016}, 100\% enhancement was achieved by multiplexing photons from four temporal modes. Recently, although not within silicon photonics, a factor of 9.7 enhancement in efficiency has been realized using optical delay to multiplex 40 heralded single-photon sources \cite{Kaneda2019}.
Despite these developments, on-chip integration of multiplexing systems is still an outstanding challenge, and the main obstacle is the device loss. For example, assuming that 1 ns delay line corresponds to 10 cm long silicon waveguide, and the waveguide has the state-of-the-art low loss of 0.08 dB/m \cite{Lee2012}, 100 ns delay line will introduce 0.8 dB loss, which is still much higher than that in commercial fiber. The integrated high-speed switch, one necessary component in multiplexing systems, usually needs ion doping or hybrid integration with other electro-optic materials such as lithium niobate \cite{He2019}, and these post-processing processes always introduce extra losses. Significant reduction in these losses will enable multiple near-deterministic single photon sources achieved on the silicon chip with probabilistic nonlinear parametric processes.
Another approach, the solid-state emitters generate deterministic single photons and can be integrated or transferred on the silicon photonic circuits. Since silicon is opaque below around 1100 nm, single-photon emitters integrated with silicon would need to radiate photons beyond this wavelength limit. Multiple quantum dots at telecommunication wavelengths have been demonstrated \cite{Haffouz2018,Kim2017,Muller2018,Katsumi2018,Katsumi2019,Zhou2018,Redjem2020,Zhao2021}. Among them, the most common are semiconductor quantum dots including InAsP \cite{Haffouz2018}, InAs/InP \cite{Kim2017,Muller2018} and InAs/GaAs \cite{Katsumi2018,Katsumi2019}, etc. By using emission-wavelength-optimized waveguides, ref. \cite{Haffouz2018} demonstrated photon emission from single InAsP quantum dot with large tuning range from 880 to 1550 nm.
With hybrid integration techniques, transfer semiconductor quantum dots on the silicon waveguides (Fig. 1a and 1b) has been implemented with $g^2(0)$ values around 0.3 \cite{Kim2017,Katsumi2018,Katsumi2019}. By using localized defects in the gallium nitride crystal, high-quality solid-state quantum emitters in the telecom range has been achieved \cite{Zhou2018}. Even at room temperatures, $g^2(0) = 0.05$ was obtained with continuous wave (CW) laser excitation. Two-dimensional (2D) materials show many amazing properties and have also been used to produce quantum emitters, for example, telecom-wavelength single-photon emitters via coupling 2D molybdenum ditelluride (MoTe$_2$) to nano-pillar arrays have been reported with $g^2(0) = 0.058$ and $g^2(0) = 0.181$ under pulsed and CW laser excitation, respectively \cite{Zhao2021}. Colour centers, known as G centers and T centers, which originate from carbon-related defect in silicon, are becoming another kind of candidates with telecom-O band radiation \cite{Redjem2020,Hollenbach2020,Bergeron2020,Higginbottom2022}. Different from other methods,
colour centers can be directly integrated into silicon waveguides without hybrid integration for large-scale quantum photonic information applications \cite{Hollenbach2021,Prabhu2022}. For example, tens of thousands of individually addressed photon-spin qubits have been demonstrated with T centers, which will provide photonic links between spin qubits and greatly advance quantum information networks \cite{Higginbottom2022}.
To integrate solid-state emitters with photonic waveguides, many approaches are proposed such as nanoscale positioning approaches \cite{Liu20212}. For the important coupling efficiency,
ref. \cite{Katsumi2018} showed total single-photon coupling efficiency of 63\% experimentally and in theory higher than 99\%.
In ref. \cite{Prabhu2022}, emitters were directly integrated in silicon waveguides and showed 40\% coupling efficiency in simulation. In addition, the use of optical micro-cavities deserves further consideration to improve the extraction efficiency \cite{Peng2017,Liu20213,Wei2022}. To further scaling up, simultaneous operation of multiple emitters is necessary. However, no two solid-state quantum emitters are alike when being produced experimentally. To keep coherence among them, wavelength tunability is necessary. Electronic integration or induced strain have been utilised to tune the wavelength and quantum interference between photons generated from independent quantum emitters have been demonstrated \cite{Patel2010,Flagg2010}. In nanophotonic devices, the wavelength tuning of the selected integrated quantum emitter have been realized \cite{Elshaari2018,Machielse2019,Katsumi2020}. The rapid development of processing and tuning techniques has enhanced the indistinguishability of photons between different emitters and increased the number of emitters on-chip. Though not in the telecom band, quantum interference of photons from two remote quantum dots with a visibility of 93.0\% \cite{Zhai2022} and large-scale integration of artificial atoms in hybrid AlN photonic circuits \cite{Wan2020} have been demonstrated recently.
\subsection{B. Photon Detection}
\begin{figure*}[t]
\centering
\includegraphics[width=15cm]{2.jpg}
\caption {\textbf{Integrated SNSPDs for single-photon detection.} a. Principle of travelling wave coupling. Adapted from Ref. \cite{Pernice2012}. b. SNSPD within a high-quality factor microcavity. Adapted from Ref. \cite{Akhlaghi2015}. c. Cavity-integrated SNSPD. Adapted from Ref. \cite{Vetter2016}. d. SNSPD implemented in a two-dimensional photonic crystal cavity. Adapted from Ref. \cite{Munzberg2018}. e. A typical chain of single-photon detector segments for signal multiplexing and number resolution. Adapted from Ref. \cite{Zhu2018}. }
\label{fig2}
\end{figure*}
\begin{table*}
\caption{\textbf{State-of-the-art techniques for integrated single-photon detection.} Except ref. \cite{Llin2020}, whose operating wavelength is 1310$\,$nm, all other operating wavelengths are around 1550$\,$nm. $\ast$ represents the system detection efficiency.}
\begin{tabular}{ |p{1cm}|p{2.8cm}|p{1.2cm}|p{2.5cm}|p{1.5cm}|p{2cm}|p{2.5cm}| }
\hline
\makecell[c]{Ref.}& \makecell[c]{Detection Efficiency} &\makecell[c]{Jitter}&\makecell[c]{Dark Count Rate}&\makecell[c]{Reset Time}&\makecell[c]{Temperature}&\makecell[c]{Number Resolving}\\
\hline
\makecell[c]{\cite{Pernice2012}}& \makecell[c]{91\%} &\makecell[c]{18$\,$ps}& \makecell[c]{50$\,$Hz}&\makecell[c]{505$\,$ps}&\makecell[c]{1.7$\,$K}&\makecell[c]{-- --} \\
\hline
\makecell[c]{\cite{Akhlaghi2015}}& \makecell[c]{100\%} & \makecell[c]{55$\,$ps} &\makecell[c]{0.1$\,$Hz}&\makecell[c]{7$\,$ns}&\makecell[c]{2.05$\,$K}&\makecell[c]{-- --}\\
\hline
\makecell[c]{\cite{Vetter2016}} &\makecell[c]{30\%} & \makecell[c]{32$\,$ps}&\makecell[c]{1$\,$Hz}&\makecell[c]{510$\,$ps}&
\makecell[c]{1.7$\,$K}&\makecell[c]{-- --}\\
\hline
\makecell[c]{\cite{Munzberg2018}} &\makecell[c]{70\%} & \makecell[c]{480$\,$ps}& \makecell[c]{0.1$\,$mHz}&\makecell[c]{480$\,$ps}&\makecell[c]{1.6$\,$K}&\makecell[c]{-- --}\\
\hline
\makecell[c]{\cite{Zhu2020}}& \makecell[c]{5.6\% $\ast$} & \makecell[c]{16.1$\,$ps}&\makecell[c]{2$\,$Hz}&\makecell[c]{85.8$\,$ns}&\makecell[c]{1.0$\,$K}& \makecell[c]{4}\\
\hline
\makecell[c]{\cite{Llin2020}}& \makecell[c]{29.4\%} & \makecell[c]{134$\,$ps} &\makecell[c]{100$\,$kHz}&\makecell[c]{-- --}&\makecell[c]{125$\,$K}&\makecell[c]{-- --}\\
\hline
\makecell[c]{\cite{Fang2020}}& \makecell[c]{60.1\%} & \makecell[c]{-- --} &\makecell[c]{340$\,$kHz}&\makecell[c]{88$\,$ns}&\makecell[c]{300$\,$K}&\makecell[c]{-- --}\\
\hline
\end{tabular}
\end{table*}
Photon detector converts optical signals into electrical signals and is an essential part for photonic quantum information applications. For squeezed light measurement, germanium (Ge) photodetectors integrated on silicon were developed \cite{Benedikovic2019} and homodyne detectors have been realized based on them \cite{Rafaelli2018}. Furthermore, by interfacing silicon photonics with integrated amplification electronics, complete integrated homodyne detectors have been demonstrated \cite{Tasker2021} and a shot-noise-limited bandwidth of more than 20 GHz has been achieved \cite{Bruynsteen2021}.
For single-photon level detection, superconducting nanowire single-photon detectors (SNSPDs) show excellent performance, such as near unity system detection efficiency \cite{Reddy2020} GHz maximum count rate \cite{Pernice2012} and picosecond level temporal resolution \cite{Korzh2020}. SNSPD uses superconducting material that works below its critical temperature. Even if only one photon hits, the energy is enough to excite the material, and generate voltage pulses subsequently. A range of superconducting materials have been developed, including NbN \cite{Pernice2012}, NbTiN \cite{Akhlaghi2015}, MoSi \cite{Li2016} and WSi \cite{Buckley2017}, etc. In vertical optical coupling between fibers and SNSPDs, fiber end faces are parallel to the SNSPDs' photonsensitive surfaces, and photons are vertically incident to the nanowires. By optimizing the device's vertical optical stack design, as well as the coupling of the guided fiber mode to the active detection area of the device, 98\% system detection efficiency has been achieved \cite{Reddy2020}. For coupling between waveguides and SNSPDs, travelling wave coupling is used, as shown in Fig. 2a, which can achieve efficient detection of photons in the optical waveguide. This hybrid integration method directly places SNSPDs into silicon photonics, thus greatly enhancing the scalability of quantum photonic integrated circuits. For example, in ref. \cite{Pernice2012}, ballistic photon transport in silicon ring resonators has been demonstrated by exploiting high temporal resolution detectors. Besides, SNSPDs are integrated with high-quality factor microcavities to increase the coupling efficiency (Fig. 2b) \cite{Akhlaghi2015} and decrease the dark count rate (Fig. 2c and 2d) \cite{Vetter2016,Munzberg2018}.
Typically, single-photon detectors operate in Geiger mode, that is, the electrical signal generated by the read-out circuit indicating the detection of one photon. However, in many experiments, it is desirable to use a detector that can resolve the number of photons. By analyzing the relationship between photon numbers and resistance of the nanowire, multi-photon detection was achieved with one convertioanl SNSPD \cite{Cahall2017}. SNSPDs with photon number resolving ability has also been achieved on the AlN platform through exploring the pulse shapes of the detector output (Fig. 2e) \cite{Zhu2018}. Besides, with an integrated impedance-matching taper on the superconducting nanowire, the detector’s output amplitude turned into more sensitive to the number of photons, and detector with photon number resolving up to four, 16.1 ps timing jitter and \textless$\,$2$\,$Hz dark count rate has been demonstrated \cite{Zhu2020}.
The high-quality of SNSPDs boosts various novel applications, such as quantum computational advantage using photons\cite{Zhong2020}, imaging and spectroscopy \cite{Zhao2017,Wollman2019,Cheng2019}, and integrated quantum key distribution \cite{Beutel2021,Zheng2021}.
Despite great progress achieved, SNSPDs cannot work without bulky and expensive cryogenic systems (\textless$\,$4$\,$K). Other types, such as germanium-on-silicon \cite{Vines2019,Llin2020} and InGaAs/InP single-photon detectors \cite{Zhangzhang2015,Fang2020}, are potential alternatives. For detection efficiency, germanium-on-silicon detectors have recently been demonstrated to be 38\% at 1310$\,$nm at 125$\,$K \cite{Vines2019}, and InGaAs/InP detectors to be 60.1\% at 1550$\,$nm at 300$\,$K. Further reduction of dark count rate will broaden their appeal in many quantum applications. More details about integrated single-photon detectors are summarized in Table II. In addition to these conventional devices, single-photon detection based on low-dimensional materials is emerging and has shown superior performance \cite{Wang2022}.
\subsection{C. Wavelength and mode division multiplexing}
\begin{figure}[t]
\centering
\includegraphics[width=8cm]{3.jpg}
\caption {\textbf{Wavelength division multiplexing techniques in silicon photonics.} a. Cascaded Mach-Zehnder demultiplexer. Adapted from Ref. \cite{Horst2013}. b. The wavelength division multiplexing receiver chip with an integrated arrayed waveguide grating. Adapted from Ref. \cite{Liuliu2019}. c. Coupled 5-ring silicon filter. Adapted from Ref. \cite{Ong2013}. d. Waveguide Bragg grating add-drop filter. Adapted from Ref. \cite{Qiu2017}.}
\label{fig3}
\end{figure}
\begin{table*}
\caption{\textbf{State-of-the-art techniques for filting in silicon photonics.} MZI: Mach-Zehnder interferometer; WBG: waveguide Bragg grating; IL: insertion loss; FSR: free spectral range.}
\begin{tabular}{ |p{1cm}|p{3cm}|p{1.5cm}|p{1.5cm}|p{1cm}|p{1cm}|p{4cm}| }
\hline
\makecell[c]{Ref.}& \makecell[c]{Structure} &\makecell[c]{Size}&\makecell[c]{Contrast}&\makecell[c]{IL}&\makecell[c]{FSR}&\makecell[c]{Bandwidth}\\
\hline
\makecell[c]{\cite{Xia2007}}& \makecell[c]{High-order micro-ring} &\makecell[c]{700$\,\upmu$m$^2$}& \makecell[c]{40$\,$dB}&\makecell[c]{1.8$\,$dB}&\makecell[c]{18$\,$nm}&\makecell[c]{310$\,$GHz (1$\,$dB)} \\
\hline
\makecell[c]{\cite{Ong2013}}& \makecell[c]{High-order micro-ring} & \makecell[c]{3000$\,\upmu$m$^2$} &\makecell[c]{50$\,$dB}&\makecell[c]{3$\,$dB}&\makecell[c]{7.3$\,$nm}&\makecell[c]{11.6--125$\,$GHz (3$\,$dB)}\\
\hline
\makecell[c]{\cite{Liao2014}} &\makecell[c]{Unbalanced MZIs} & \makecell[c]{2$\,$mm$^2$}& \makecell[c]{15$\,$dB}&
\makecell[c]{9$\,$dB}&\makecell[c]{0.8$\,$nm}&\makecell[c]{0.61, 0.34 and 0.21$\,$nm (3$\,$dB)}\\
\hline
\makecell[c]{\cite{Qiu2017}} &\makecell[c]{WBG} & \makecell[c]{600$\,\upmu$m$^2$}& \makecell[c]{35$\,$dB}&\makecell[c]{0.6$\,$dB}&\makecell[c]{-- --}&\makecell[c]{3$\,$nm (3$\,$dB)}\\
\hline
\makecell[c]{\cite{Harris2014}}& \makecell[c]{Cascaded WBG} & \makecell[c]{1280$\,\upmu$m$^2$}&\makecell[c]{65$\,$dB}&\makecell[c]{3$\,$dB}& \makecell[c]{-- --}&\makecell[c]{1--2$\,$nm (3$\,$dB)}\\
\hline
\makecell[c]{\cite{Oser2020}}& \makecell[c]{Cascaded WBG} & \makecell[c]{3105$\,\upmu$m$^2$} &\makecell[c]{60$\,$dB}&\makecell[c]{2$\,$dB}&\makecell[c]{-- --}&\makecell[c]{5.5$\,$nm (3$\,$dB)}\\
\hline
\end{tabular}
\end{table*}
Due to the significant enhancement in demand for information processing and communication capability, much effort has been done to encode more information on one single chip. One way is to increase the integration density. Devices are developed to be more compact and their spacing to be narrower. For example, by using the concept of symmetry breaking, high-density waveguide superlattices with low crosstalk were demonstrated \cite{Song2015}.This concept has also been used to construct two-qubit quantum logic gate, and greatly reduced the footprint \cite{Zhang2021}. Recently, an ultradensely integrated multidimensional optical system with a footprint of 20$\,\times\,$30$\,\upmu$m$^2$ has been demonstrated based on inverse-designed dielectric metasurface network \cite{Zhou2022}. Another way is to introduce new degrees of freedom through multiplexing, including wavelength division multiplexing (WDM) and mode division multiplexing (MDM).
WDM is a technique that encodes information with different optical wavelengths as different channels. Structures like ring resonators, unbalanced MZIs, waveguide Bragg gratings (WBGs) and arrayed waveguide gratings (AWGs) are used in the WDM systems for the purpose of wavelength multiplexing and de-multiplexing \cite{Horst2013,Cheung2014,Tan2014,Chen2015,Munk2019,Liuliu2019}. Among them, 512 channel dense WDM \cite{Cheung2014}, channel spacing down to 17$\,$GHz \cite{Munk2019}, hybrid integration with detectors (Fig. 3b) \cite{Liuliu2019} have been demonstrated. Although not with silicon photonics, high-speed quantum key distribution with WDM on integrated photonic circuits has been reported \cite{Price2018}.
\begin{figure*}[t]
\centering
\includegraphics[width=16cm]{4.jpg}
\caption {\textbf{Mode division multiplexing techniques in silicon photonics.} a. Mode (de)multiplexer with adiabatic taper. Adapted from Ref. \cite{Guo2017}. b. 10-channel mode (de)multiplexer. Adapted from Ref. \cite{Dai2018}. c. A multiport multimode waveguide crossing using a metamaterial-based Maxwell's fisheye lens. Adapted from Ref. \cite{Xu2018}. d. The digital meta-structure-based multimode bending. Adapted from Ref. \cite{Liuy2019}. e. High-speed optical two-mode switch. Adapted from Ref. \cite{Xiong2017}. f. Reconfigurable optical add-drop multiplexer for hybrid wavelength/mode-division-multiplexing systems. Adapted from Ref. \cite{Wang2017}.}
\label{fig4}
\end{figure*}
In addition to increasing channel capacity, WDM technique is also used as filters for the spectrum manipulation in various optical systems \cite{Liu2021}. With unbalanced MZIs, an programmable filter has been demonstrated and showed the tunability of the filter central wavelength, bandwidth and variable passband shape \cite{Liao2014}. With high-order ring resonators, on-chip filters are becoming ultra-compact, ultra-high-contrast and show high flexibility \cite{Xia2007,Ong2013,Chen2014,Liu2019}. For example, 5$^{\rm{th}}$ order ring resonator optical filters showed out-of-band rejection ratio of 40$\,$dB, and insertion loss of only 1.8$\,$dB within a footprint of 700$\,\upmu$m$^2$ (Fig. 3c) \cite{Xia2007}. To increase the free spectral range (FSR), a resonator with a small radius of 0.8$\,\upmu$m was demonstrated and a record large FSR of 93$\,$nm was achieved recently \cite{Liu2019}. WBGs are FSR free and could also be employed to build filters. The add-drop structure showed high-contrast and the filter bandwidths changed with the waveguide widths (Fig. 3d) \cite{Qiu2017}. More details about filting in silicon photonics are summarized in Table III. For scalable quantum information applications, these filters will be employed to filter out the pump light before detection or separate single photons at different wavelengths, as reported in refs. \cite{Harris2014,Oser2020}.
MDM is an emerging technique that uses the high-order transverse waveguide modes of multimode waveguides to encode more information \cite{Li2019}. Since multimode waveguides still support multiple wavelengths, this technique is compatible with WDM to further increase the channel capacity \cite{Luo2014,Dai2015,Wang2017}. As shown in Fig. 4, various multimode silicon photonic devices have been developed for low-loss and low-crosstalk light manipulation with multiple mode-channels in MDM systems, such as mode (de)multiplexer \cite{Dai2013,Guo2017}, grating couplers \cite{Lai2018}, high-speed switches \cite{Xiong2017}, sharp waveguide bends and mode-independent crossings \cite{Li2018,Xu2018,Liuy2019}. At present, mode (de)multiplexer including six mode-channels of TE polarization and four mode-channels of TM polarization has been realized \cite{Dai2018}. In addition, by using multimode optical waveguides with higher-order modes, some special silicon photonic devices that cannot be realized only with fundamental modes have been demonstrated, such as add-drop optical filters based on multimode Bragg gratings \cite{Qiu2017,Oser2020}.
The utilization of higher-order modes for quantum photonics has stimulated many novel applications. For example, in ref. \cite{Feng2016}, the transverse waveguide-mode degree of freedom was introduced for quantum encoding, and on-chip coherent conversion of photonic entangled quantum states between path, polarization, and transverse waveguide-mode degrees of freedom was demonstrated. In ref. \cite{Mohanty2017}, quantum interference between the guided modes was demonstrated within a multimode optical waveguide. Recently, with the help of two newly developed multimode devices, transverse mode-encoded 2-qubit logic quantum gate was realized \cite{Feng2022}, and showed the potential for universal transverse mode-encoded quantum operations and large-scale multimode multi-degrees of freedom quantum systems. The special inter-modal phase matching conditions were used to prepare quantum photonic sources. In particular, an on-chip transverse-mode entangled photon pair source via SFWM in a multimode waveguide was demonstrated \cite{Feng2019}. Furthermore, by exploiting special excitation scheme in low-loss multimode waveguides, quantum photonic source was engineered to high spectral purity and mutual indistinguishability through inter-modal SFWM \cite{Paesani2020}.
\subsection{D. Cryogenic techniques}
\begin{figure}[t]
\centering
\includegraphics[width=7cm]{5.jpg}
\caption {\textbf{Silicon photonic modulators at cryogenic temperatures.} a. The plasma dispersion micro-disk modulator. Adapted from Ref. \cite{Gehl2017}. b. The BaTiO$_3$–Si racetrack resonator. Adapted from Ref. \cite{Eltes2020}. c. The integrated PIN junction modulator and unbalanced Mach–Zehnder interferometer composed of the modulator. Adapted from Ref. \cite{Chakraborty2020}.}
\label{fig5}
\end{figure}
The large-scale expansion of quantum photonic integrated circuits requires the integration of all functions on one single chip including quantum photonic sources, state manipulation and photon detection. Because common and efficient SNSPDs must work at cryogenic temperatures, for example, 2$\,$K, quantum photonic sources and state manipulation processes should also extend to the same temperature condition. Besides, in quantum networks, optical interfaces with other quantum systems such as semiconductor and superconducting quantum computing systems also need to operate the whole system at cryogenic temperatures.
While passive optical components can often be used directly at low temperatures, implementation of active components and nonlinear processes has been a prominent challenge. At present, quantum state manipulation structures used in quantum photonic chips at ambient conditions are generally based on thermo-optic effect. However, the thermo-optic coefficient of silicon decreases significantly
at low temperatures \cite{Komma2012}. In particular, when the temperature is several kelvin, the thermo-optic coefficient is four orders of magnitude lower than that at room temperatures, which makes thermo-optic modulators difficult to work. Another problem is that as temperatures drop, so does the cooling capacity available of the cooling system.
Therefore, modulators operating at cryogenic conditions should maintain very low power consumption and heat production.
To date, some works have been reported on cryogenic modulation \cite{Gehl2017,Eltes2020,Chakraborty2020} with silicon photonics. In ref. \cite{Gehl2017}, by doping at higher concentrations, one microdisk modulator (Fig. 5a) prepared by plasma dispersion effect has been achieved at temperature 4.8 K and transmitting data rates up to 10 Gb/s has been achieved. As the first success implementation, this work opens the door for using silicon photonics to interface other systems at cryogenic conditions. Besides, electro-optic modulation at temperature 4 K by using the Pockels effect of integrated barium titanate (BaTiO$_3$) devices has been reported (Fig. 5b)) \cite{Eltes2020}. This material showed an effective Pockels coefficient of 200 pm V$^{-1}$ at 4 K, and the fabricated devices showed high electro-optic bandwidth (30 GHz), ultralow-power consumption and high-speed data modulation (20 Gb/s).
Another implementation employed DC Kerr effect of the silicon waveguides and achieved phase modulation at a temperature of 5 K at GHz speeds (Fig. 5c) \cite{Chakraborty2020}. Despite these progresses, these modulation devices are still in the initial research stage, and the relevant applications in quantum information research have not been demonstrated. For large-scale quantum information applications, they should be more compact and have lower excess loss. More recently, programmable Mach–Zehnder meshes have been realized by using aluminium nitride (AlN) piezo-optomechanical actuators coupled to SiN waveguides \cite{Dong2022}, which can also operate at cryogenic conditions. Though the modulators operate in the visible band and on the SiN platform, the work demonstrates the possibility of on-chip large-scale cryogenic modulation.
In addition to cryogenic modulation, some works have also been reported on cryogenic nonlinear processes in silicon photonics \cite{Pernice2011,Sun2013,Sinclair2019}. In ref. \cite{Sinclair2019}, the temperature dependence of the two-photon absorption and optical Kerr nonlinearity of a silicon waveguide with temperatures from 5.5 to 300 K was measured, and the nonlinear figure of merit was found to be improved at cryogenic temperatures. Quantum applications, however, such as quantum photonic sources preparation by cryogenic SFWM, has not been demonstrated yet.
\subsection{E. Chip interconnects}
Chip interconnection plays the key role in building large-scale quantum networks. In photonic quantum technologies, chip interconnection needs efficient transfer of optical signals among different optical components. However, due to the mismatch in the effective mode sizes between different components such as fibers and silicon waveguides, special coupling structures are needed \cite{Son2018}.
\begin{figure*}[t]
\centering
\includegraphics[width=16cm]{6.jpg}
\caption {\textbf{Chip interconnection techniques in silicon photonics.} a. Diffraction grating-based coupling structure. Adapted from Ref. \cite{Son2018}. b. The focusing grating. Adapted from Ref. \cite{Van2007}. c. The double-etched apodized waveguide grating coupler. Adapted from Ref. \cite{Li2013}. d. The grating coupler with a single aluminum backside mirror. Adapted from Ref. \cite{Hoppe2020}. e. The mode-size converter as end coupler. Adapted from Ref. \cite{Fang2011}. f. 3D coupler structure. Adapted from Ref. \cite{Luo2020}. g. The 3D-printed optical probes on the fiber end faces. Adapted from Ref. \cite{Trappen2020}. h. Fiber cores and different silicon waveguides connected by photonic wire bonds. Adapted from Ref. \cite{Lindenmann2015}. i. In situ 3D nanoprinted free-form lenses and expanders. Adapted from Ref. \cite{Dietrich2018}.}
\label{fig6}
\end{figure*}
\begin{table*}
\caption{\textbf{State-of-the-art techniques for chip interconnection in silicon photonics.} For wire bonding and in situ 3D nanoprinting techniques, different coupling efficiencies were reported for coupling different structures. Here, the maximum values are given.}
\begin{tabular}{ |p{1.5cm}|p{4cm}|p{4cm}|p{4.5cm}| }
\hline
\makecell[c]{Ref.}&\makecell[c]{Technique}&\makecell[c]{Loss}&\makecell[c]{Bandwidth}\\
\hline
\makecell[c]{\cite{Ding2014}}&\makecell[c]{Grating coupling}&\makecell[c]{0.58$\,$dB (TE)}&\makecell[c]{71$\,$nm (3$\,$dB)}\\
\hline
\makecell[c]{\cite{Pu2010}}&\makecell[c]{End coupling (tapered fiber)}&\makecell[c]{0.36$\,$dB (TM); 0.66$\,$dB (TE)}&\makecell[c]{\textgreater 80$\,$nm (1$\,$dB)}\\
\hline
\makecell[c]{\cite{Fang2011}} &\makecell[c]{End coupling (SMF)}&\makecell[c]{2.0$\,$dB (TM); 1.2$\,$dB (TE)}&\makecell[c]{\textgreater 120$\,$nm (1$\,$dB)}\\
\hline
\makecell[c]{\cite{Jia2018}} &\makecell[c]{End coupling (SMF)}&\makecell[c]{1.3$\,$dB (TM); 0.95$\,$dB (TE)}&\makecell[c]{\textgreater 100$\,$nm (1$\,$dB)}\\
\hline
\makecell[c]{\cite{Luo2020}} &\makecell[c]{3D vertical coupling}&\makecell[c]{1$\,$dB (TE and TM)}&\makecell[c]{170$\,$nm (TE); 104$\,$nm (TM) (1$\,$dB)}\\
\hline
\makecell[c]{\cite{Trappen2020}} &\makecell[c]{3D-printing}&\makecell[c]{1.9$\,$dB} &\makecell[c]{-- --}\\
\hline
\makecell[c]{\cite{Lindenmann2012,Lindenmann2015,Billah2018}} &\makecell[c]{Wire bonding}&\makecell[c]{0.4$\,$dB} &\makecell[c]{-- --}\\
\hline
\makecell[c]{\cite{Dietrich2018}} &\makecell[c]{In situ 3D nanoprinting}&\makecell[c]{0.6$\,$dB} &\makecell[c]{-- --}\\
\hline
\end{tabular}
\end{table*}
Grating coupler is one common such structure, which enlarges the mode size of the waveguide by etching diffraction gratings on the waveguide surface \cite{Taillaert2002}. The cross-section view of the uniform surface-corrugated grating structure is shown in Fig. 6a, and the picture of one fabricated grating coupler is shown in Fig. 6b. Typically, the grating couplers have an alignment tolerance value of $\pm$2$\,\upmu$m for 1$\,$dB excess losses and over 20$\,$nm 1$\,$dB bandwidth. The relaxed spatial alignment tolerance facilitates multi-path alignment, which is one key part for scalable photonic information applications. Since quantum signals are difficult to be amplified, the losses should be reduced as much as possible. For the uniform grating structure, the grating couplers have a typical coupling efficiency of 5$\,$dB \cite{Van2007}. To improve the coupling efficiency, different principles and techniques were introduced. One method is to spatially vary fill factors and etch depths (Fig. 6c), which can achieve larger overlap integral values between the field profiles of light from the grating coupler structure and the optical fiber \cite{Tang2010,Li2013,Benedikovic2014,Li2014,Marchetti2017}. These apodized grating couplers achieved a typical coupling efficiency of 1$\,$dB. Considering that a substantial portion of the optical power input is lost in the substrate due to the diffraction downward from the grating structure (Fig. 6a), another rule is to improve the radiation directionality. Many techniques were developed, such as using overlay grating elements \cite{Vermeulen2010} and substrate metal mirrors (Fig. 6d) \cite{Hoppe2020}. Especially, one can improve both the overlap integral and directionality \cite{Taillaert2004}, and below 1$\,$dB coupling loss was achieved in this way \cite{Ding2014}. It is worth mentioning that such high-efficiency grating couplers have been employed and played an important role in multi-photon quantum information processing on the silicon photonic circuits \cite{Paesani2019,Llewellyn2020,Vigliar2021}. The grating couplers are typically polarization dependent. In order to utilize polarization degree of freedom to encode quantum information, 2D grating couplers are alway used, which are able to couple orthogonal polarization of light into separate waveguides \cite{Xue2019}.
End coupler is another common structure for chip-fiber coupling. Designs of the tapered waveguide have been introduced to enlarge the effective mode size of the integrated silicon waveguides (Fig. 6e). Similar to grating couplers, the end couplers have also achieved below 1$\,$dB coupling losses with commercial single mode fibers \cite{Pu2010,Fang2011,Jia2018}. In addition, the end couplers can achieve greater 1$\,$dB bandwidth and support the simultaneous and efficient coupling of two polarization modes (Table IV). The locating on the chip edges facilitates direct packaging with the fiber array, despite requiring additional dedicated fabrication steps, such as chip dicing and polishing. Multi-channel packaging with end coupling has been demonstrated for deep learning \cite{Shen2017} and quantum transport simulations \cite{Harris2017}. Similar to end coupling, a three-dimensional (3D) fabricated polymer coupler showing coupling loss of 1$\,$dB was demonstrated recently (Fig. 6f). \cite{Luo2020}.
In addition to fabricating complex coupling structures on chip, many other efficient approaches are also worth considering. For example, 3D-printed optical probes on the fiber end faces (Fig. 6g), which can realize the detection of vertical cutting edge devices, were recently demonstrated with 1.9$\,$dB coupling loss \cite{Trappen2020}. Photonic wire bonding and in situ 3D nanoprinting are other novel and promising techniques for chip-scale multi-platform interconnects (Fig. 6h and 6i). Using polymer waveguides with 3D geometry, photonic wire bonding can bridge photonic circuits on different chips \cite{Lindenmann2012,Lindenmann2015,Billah2018,Rhee2022}. With this technique, 1.6$\,$dB coupling loss between silicon chips \cite{Lindenmann2012}, 1.7$\,$dB coupling loss between a four-core fiber and a silicon chip \cite{Lindenmann2015} and 0.4$\,$dB coupling loss between an indium phosphide chip and a silicon photonic chip \cite{Billah2018} were demonstrated. In situ 3D nanoprinting technique exploits direct-write two-photon laser lithography to create ultracompact elements such as lenses and expanders, which can be directly integrated onto surfaces of optical integrated devices \cite{Dietrich2018}. These elements can be optimized according to the coupling objectives, thus increasing the flexibility. In ref. \cite{Dietrich2018}, coupling efficiencies of up to 0.6$\,$dB between edge-emitting lasers and single-mode fibres and 2.5$\,$dB between lasers and passive chips were achieved. Introducing these two new technologies into integrated quantum photonics will greatly improve the chip complexity and integration.
\section*{III. Scalable quantum information applications}
After decades of development, silicon photonics has made a series of significant advances in quantum computing, quantum simulation, quantum communication and metrology, etc. Here we give a review of some important advances in recent years, and they are fundamental and have irreplaceable values in the extended applications of quantum information processing, including multiphoton and high-dimensional applications and quantum error correction on one single chip, and quantum key distribution and state teleportation among chips.
\subsection{A. Multiphoton and high-dimensional applications}
\begin{figure*}[t]
\centering
\includegraphics[width=16cm]{7.jpg}
\caption {\textbf{Multiphoton and high dimensional applications with silicon photonic devices.} a. Silicon photonic chip for the generation and sampling of quantum states. Adapted from Ref. \cite{Paesani2019}. b. Coherent pumping of two sources and processing of the emitted photons. Adapted from Ref. \cite{Paesani2020}. c. Chip-to-chip high-dimensional quantum key distribution based on multicore fiber. Adapted from Ref. \cite{Ding2017}. d. Silicon device for multidimensional quantum entanglement. Adapted from Ref. \cite{Wang2018}. e. Programmable qudit-based quantum processor. Adapted from Ref. \cite{Chi2022}. }
\label{fig7}
\end{figure*}
Assuming that a system contains $n$ photons with $m$ dimension, the system capacity is $m^n$. Therefore, increasing the number of photons will improve the system capacity exponentially, and it is vitally important. Although multiple solid-state quantum emitters can be integrated on the same chip and multiple determinate single photons can be generated, controlling emitters to ensure indistinguishability between photons remains a major difficulty. Through multiplexing techniques, single emitter can also generate multiple identical photons and complex quantum states \cite{Istrati2020,Wangw2017}, however, there are still difficulties in on-chip integration of multiplexing components. Alternative approach, SFWM can generate multiple photons directly on the single chip and the relevant work have made great progress in recent years.
With strong third-order nonlinear response, silicon waveguides and micro-ring resonators have been used to generate entangled photon pairs with different degrees of freedom \cite{Feng2019,Takesue2007,Takesue2008,Matsuda2012,Silverstone2015,Li2017} and to demonstrate various optical quantum applications \cite{Wangwang2017,Paesani2017,Qiang2018,Qiang2021}. Furthermore, multiple photon pairs were multiplexed to generate quantum states consisted with more photons. In 2011, ref. \cite{Harada2011} has performed quantum interference experiment with two heralded photons generated in two independent silicon waveguides. A visibility of 73\% has been observed. Later, in 2018, this multi-photon interference process has been demonstrated on one single chip with heralded photons from two independent micro-ring resonator sources and interference fringe visibility as 72\% was measured \cite{Faruque2018}. Almost at the same time, silicon waveguides have been used to prepare complex four-photon states \cite{Zhang2019} and frequency-degenerate entangled four-photon states \cite{Feng20192}. These works stimulate more multi-photon applications in integrated optical chips, such as preparation of programmable four-photon graph states \cite{Adcock2019}, generation and sampling of quantum states of light (Fig. 7a) \cite{Paesani2019} and observing nonlocal quantum interference \cite{Feng2021}. So far, the number of photons on one single silicon chip has been increased to 8 \cite{Paesani2019} and more photons could be achieved by further reducing losses. In improving the quality of multi-photon interference, different strategies have been demonstrated to improve the spectral purity of photon pairs \cite{Paesani2020,Liu2020,Burridge2020}. In particular, ref. \cite{Paesani2020} has achieved on-chip heralded two-photon quantum interference with a visibility of 96\% (Fig. 7b). These high-quality integrated photonic sources are promising in practical quantum applications.
High-dimensional encoding is another feasible approach to a larger system capacity. Moreover, it shows many unique quantum properties and provides improvements in particular applications such as higher capacity and noise robustness in quantum communications \cite{Hu2018} and higher efficiency and flexibility in quantum computing \cite{Lanyon2009}. On integrated chips, many degrees of freedom can be used for high-dimensional encoding, such as path, transverse-mode, frequency and time-bins. Among them, path encoding is the most common one due to its ease of implementation. With on-chip beam splitter and Mach-Zehnder interferometers, photons can be routed and manipulated in multiple paths. Coupled waveguide array also can be used to construct the desired dynamic evolution Hamiltonian for research including quantum walk, boson sampling and quantum simulation.
For high-dimensional quantum applications with integrated devices, ref. \cite{Ding2017} has experimentally demonstrated high-dimensional chip-to-chip quantum interconnection for the first time in 2017 (Fig. 7c). High-dimensional quantum information generated in one silicon photonic chip has been distributed to another chip through one multicore fiber.
Combined with on-chip quantum photonic sources, programmable high-dimensional bipartite entangled systems have been realized \cite{Wang2018,Lu2020}.
Recently, a programmable qudit-based quantum processor including all key functions for initialisation, manipulation, and measurement of two quantum quart states and multi-value quantum-controlled logic gates has been demonstrated on a silicon chip (Fig. 7e) \cite{Chi2022}.
Despite path, other degrees of freedom also have advantages in increasing the system capacity. For example, integrated photonic sources always have multiple frequencies \cite{Feng2020} and waveguide transverse-mode shows a compact way for parallel encoding \cite{Dai2018}. Furthermore, it is effective to use multiple degrees of freedom of a quantum particle simultaneously \cite{Feng2016}. The development of on-chip conversion devices will greatly increase the application prospects of these degrees of freedom.
\subsection{B. Quantum error correction}
\begin{figure*}[t]
\centering
\includegraphics[width=13cm]{8.jpg}
\caption {\textbf{Quantum error correction with silicon photonic devices.} Error-protected qubits for quantum computation. Adapted from Ref. \cite{Vigliar2021}.}
\label{fig8}
\end{figure*}
Despite many natural advantages, quantum information processor faces one basic difficulty, that is, the errors. When the product of logical operations and error rates of qubits reaches a certain level, the logical operations will loss their reliability. Meanwhile, due to quantum no-clone theorem, qubits cannot be copied, and eliminate errors by repetition as in classical processors is impossible. Therefore, effective quantum error correction is a necessary step to realize one realistic quantum processor. In photonic quantum technologies,
photon loss, quality of quantum states, and imperfect logical operations are all sources of errors. Various schemes were developed to improving fault-tolerant capability with photons, such as making efforts for decreasing fabrication imperfections to obtain high-performance optical components. In ref. \cite{Wilkes2016}, 60 dB extinction ratio MZI can be achieved with improved design, which equivalently can implement single-qubit quantum gate with 99.9999\% fidelity.
Another famous method is the generation of large-scale cluster states for measurement-based quantum computation (MBQC) \cite{Yao2012,Larsen2019,Asavanant2019}, which depends on a sequence of local measurements on a special entangled state, called the cluster or graph state. MBQC is equivalent to the circuit-based quantum computation model, and since it is relatively easy to prepare this kind of entangled states with nonlinear photonics, this model is highly valued in photonic quantum information applications. Besides being used in quantum computation, cluster state also shows important applications in other fields such as quantum error correction, multi-partite quantum communication and quantum metrology, as well as in the study of fundamental problems such as non-locality and decoherence \cite{Hein2006,Shettell2020}.
So far, with silicon photonic circuits, all types of four-photon graph state have been programmably generated \cite{Adcock2019}, and
a range of quantum information processing tasks with and without error-correction encodings have been implemented (Fig. 8) \cite{Vigliar2021}. The success rate was increased from 62.5\% to 95.8\% when running a phase-estimation algorithm by using the error-correction program.
Similar to MBQC, fusion-based quantum computation was proposed and developed in recent years by PsiQuantum \cite{Bartolucci2021,Bombin2021}. With small-scale entangled photons as resource states and fusion-measurement, large-scale universal fault-tolerant quantum computing can be realized with on-chip components.
Other novel error-correction methods includes entanglement purification and topologically electromagnetic modes. Entanglement purification is a way to extract a subset of states of high entanglement and high purity from a large set of less entangled states, which can significantly increase the quality of logic operations between different qubits and relax the requirement for high-accuracy logic operations \cite{Pan2003}. The topologically electromagnetic modes are much less affected by nanophotonic fabrication-induced disorder and can be used to avoid errors. Up to now, topological quantum light sources \cite{Mittal2018,Wang2019,Dai2022} and quantum interference process \cite{Chenchen2021} on silicon photonic chips have been demonstrated.
\subsection{C. Quantum key distribution}
\begin{figure*}[t]
\centering
\includegraphics[width=13cm]{9.jpg}
\caption {\textbf{Quantum key distribution (QKD) with silicon photonic devices.} a. Integrated devices for time-bin encoded BB84. Adapted from Ref. \cite{Sibson2017}. b. Integrated devices for high-speed measurement-device-independent QKD. Adapted from Ref. \cite{Wei2020}. c. Silicon photonics encoder with high-speed electro-optic phase modulators. Adapted from Ref. \cite{Bunandar2017}. d. Detector-integrated on-chip QKD receiver. Adapted from Ref. \cite{Beutel2021}.}
\label{fig9}
\end{figure*}
Governed by the laws of quantum mechanics, quantum key distribution (QKD) aims to share information with absolute security between the transmitters and receivers. After decades of development, QKD is becoming the building block of quantum network, and many quantum encoding protocols have been developed, such as BB84, two-state and Einstein-Podolsky-Rosen protocol \cite{Gisin2002}. As the common choice of information carrier, photons are distributed over 803-km fiber \cite{Wangw2022} and from the satellite to the ground over a distance of up to 1,200 kilometres \cite{Liao2017}. For convenience and practicality purposes, transmitters and receivers have been integrated on the photonic chip to achieve dense integration, high stability and scalability \cite{Wang2021}, and are connected by single-mode or multicore fibers \cite{Ding2017,Sibson2017,Bacco2017}.
Through combining slow thermo-optic DC biases and fast (10 GHz bandwidth) carrier-depletion modulation, high-speed low-error QKD modulation has been achieved with silicon photonic devices \cite{Sibson2017}. Besides, QKD systems with integrated silicon photonics have been demonstrated in an intercity metropolitan test with a 43-km fiber \cite{Bunandar2017} and even full daylight \cite{Avesani2019}. The QKD system is becoming more compact and different degrees of freedom encoding have been implemented such as path \cite{Ding2017}, polarization \cite{Bunandar2017} and time-bins \cite{Beutel2021,Geng2019}. Based on hybrid techniques, photon lasers \cite{Agnesi2019} and photon detection processes \cite{Rafaelli2018,Beutel2021,Zheng2021} have been integrated directly on the silicon chip. The high-dimensional QKD with quantum states generated in the silicon photonic circuits enables surpassing the information efficiency limit of traditional quantum key distribution protocols \cite{Ding2017}. With these technological advances, different key distribution protocols have been achieved with integrated silicon photonic devices such as BB84 \cite{Sibson2017}, continuous-variable \cite{Zhangzhang2019} and measurement-device-independent \cite{Wei2020}, as shown in Fig. 9.
\subsection{D. Quantum state teleportation}
\begin{figure*}[t]
\centering
\includegraphics[width=13cm]{10.jpg}
\caption {\textbf{Quantum state teleportation with silicon photonic devices.} Chip-to-chip quantum teleportation. Adapted from Ref. \cite{Llewellyn2020}.}
\label{fig10}
\end{figure*}
Quantum state teleportation aims to transfer a particle's quantum state to another particle rather than itself. As a new way of communication, this technique transfers quantum information carried in the quantum state with the help of quantum entanglement, and forms the basis of scalable quantum network and distributed quantum computing. It involves three particles (A, B, C), and we assume particle B and C are Bell entangled. The quantum state of particle A can be transferred to particle B through Bell state measurement between A and C. With quantum state teleportation, linear optical quantum information processor can be efficiently scaled by using a large number of cascaded gates \cite{Kok2007}.
The integrated quantum teleportation process was first implemented with silica slab waveguides \cite{Metcalf2014},
which was fabricated by the direct UV-writing technique. The silica waveguides have large cross-section (typically 4.5$\upmu$m$\times$4.5$\upmu$m), and low chip-fiber coupling loss and transmission loss enable us to achieve single photons generated in free-space feeding into the chip directly through optical fibers while maintaining high brightness. In ref. \cite{Metcalf2014}, four photons which were generated in free-space nonlinear crystals were used. Three photons were input into the chip, and the left one was input into the detector as heralding. On-chip interferometers were used to realize the functions of entanglement generation and Bell state measurement. Based on state measurement results, classical communications and corresponding quantum state operations, quantum state of one photon would be transferred to another photon.
Of course, quantum teleportation process can be achieved with silicon waveguides. In particular, strong SFWM in silicon waveguides can be used directly to manipulate multi-photon entangled quantum states. In ref. \cite{Llewellyn2020}, quantum teleportation between two silicon chips have been demonstrated, as shown in Fig. 10. Four microresonators were used to generate high-quality entangled quantum states directly on the silicon chip. Before the validation of teleportation process, photons that obtained the quantum states were transmitted to another chip with quantum photonic interconnect by path-polarization interconversion \cite{Wang2016}, where quantum states were reconstructed via state tomography measurements. This achievement
lays the foundation for large-scale integrated photon quantum technology in communication and computing.
\section*{IV. Challenges and outlook}
Despite progress mentioned above, further improvements are needed in some areas for scalable quantum information applications. Below, we outline some challenges in silicon photonic quantum technologies.
\textbf{Low-Loss Components:} Loss is one huge challenge for optical quantum integrated systems, including loss of passive structures, delay lines, switches and chip interconnects. Because of the small effective mode area, loss in silicon is more prominent than in other materials, such as silica. Although the state-of-the-art low transmission loss of 0.08 dB/m has been demonstrated based on sidewall smoothing
techniques \cite{Lee2012}, it is difficult to apply to waveguide structures with common strip waveguides. Multimode waveguide technology and mixing with silica or silicon nitride could enable the next generation of ultra low-loss components.
\textbf{Photon Generation:} For parametric photonic sources, the silicon waveguide has strong two-photon absorption at 1550 nm, and it is difficult to enhance the brightness of the photonic sources just through increasing the pump power. Besides, the properties of the photonic sources are relatively poor, especially for multi-photon interference. Photonic source capable of large-scale expansion with high brightness is still on the way. Extending photons to the far infrared band, where silicon has low two-photon absorption, is one potential method. Another possible solution is to develop hybrid integrated silicon chips with other materials with better nonlinear properties, such as silicon nitride, in which preparation of tens of photons has been demonstrated \cite{Arrazola2021}, and lithium niobate, which has strong second-order nonlinear response. The stochastic character of the photon sources needs to be solved, and multiplexed photon sources, long-time delay and high-speed modulation should be integrated totally on the same chip. For deterministic single-photon sources, modulation to increase the indistinguishability between different sources should be developed.
\textbf{Deterministic Quantum Operation:} The negligible photon-photon interaction limits the applicability of many photon-based schemes. To overcome this limitation, multiplexed photon sources and feedforward capability based on heralded detection of auxiliary photons should be integrated totally on the same chip. Strong nonlinear media such as atoms need to be introduced into the chip to enhance the photon-photon interaction to build deterministic multi-photon gates.
\textbf{Frequency Conversion:} Efficient frequency conversion will link various quantum systems to build up quantum networks, such as photonic conversion between microwave and telecom C-band. Efficient conversion at the single-photon level will allow us to take advantage of different systems, even if they are far away.
We have witnessed tremendous advances in silicon photonic devices for quantum information processing, especially in recent years. Photon source, quantum state manipulation and detection can be integrated on a chip and hopefully on the same chip, and integrated programmable multi-photon and high-dimensional quantum information processors have been demonstrated. With the further upgrading of fabrication technology, silicon photonics will have a greater prospect in quantum information processing. Of course, silicon photonic devices still face many defects of the material itself, and the future quantum information processors are most likely to be hybrid with various materials elevated to the extreme. In any case, we believe that silicon photonics will play an important role.
\textbf{Funding.} This work was supported by the National Natural Science Foundation of China (NSFC) (Nos. 62061160487, 62005239, 12004373, 61974168, 62075243), the Innovation Program for Quantum Science and Technology (No. 2021ZD0303200), the Natural Science Foundation of Zhejiang Province (LQ21F050006), the National Key Research and Development Program (2017YFA0305200), the Key Research and Development Program of Guangdong Province of China (2018B030329001 and 2018B030325001), the Postdoctoral Science Foundation of China (No. 2021T140647) and the Fundamental Research Funds for the Central Universities. This work was partially carried out at the USTC Centre for Micro and Nanoscale Research and Fabrication.
\textbf{Disclosures.} The authors declare no conflicts of interest.
|
\section*{Abstract}
We are investigating who has the winning strategy in a game in which two players take turns drawing arrows trying to complete cycle cells in a graph. A cycle cell is a cycle with no chords. We examine game boards where the winning strategy was previously unknown. Starting with a $C_{5}$ sharing two consecutive edges with a $C_{7}$ we solve multiple classes of graphs involving "stacked" polygons. We then expand upon and improve previous theorems and conjectures, and offer some new directions of research related to the Game of Cycles. The original game was described by Francis Su in his book \textit{Mathematics for Human Flourishing} \cite{su2020mathematics}. The first results on the game were published in \textit{The Game of Cycles} \cite{alvarado2021game}.
\section*{Introduction}
The Game of Cycles, introduced by Dr. Francis Su in \textit{Mathematics for Human Flourishing} \cite{su2020mathematics} is a two-player game on a graph. The \textit{gameboard} is a simple, connected, undirected, planar graph. The two players take turns directing edges of the graph using arrows, and the goal of the game is to be either the last player to make a move, or be the first player to create a \textit{cycle cell}, which is a chordless cycle of the graph whose edges are all directed the same way. Figure \ref{fig:cycle-cell-defintion} provides an example of how a cycle is not the same as a cycle cell, as the image on the right has a directed cycle, but not a cycle cell.
\begin{figure}
\centering
\begin{tikzpicture}
\draw (-1,0)--(1,0);
\draw (-1,0)--(0,1);
\draw (1,0)--(0,1);
\draw (0.5,0)--(0.3,0.2);
\draw (0.5,0)--(0.3,-0.2);
\draw (-0.5,0)--(-0.7,0.2);
\draw (-0.5,0)--(-0.7,-0.2);
\draw (0.5,0.5)--(0.7828,0.5);
\draw (0.5,0.5)--(0.5,0.2172);
\draw (-0.5,0.5)--(-0.5,0.7828);
\draw (-0.5,0.5)--(-0.2172,0.5);
\draw (-1,0)--(-1,-1);
\draw (1,0)--(1,-1);
\draw (-1,-1)--(1,-1);
\draw (0,0)--(0,-1);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (0,1) circle (2.5pt);
\draw [fill=black] (-1,-1) circle (2.5pt);
\draw [fill=black] (1,-1) circle (2.5pt);
\draw [fill=black] (0,-1) circle (2.5pt);
\end{tikzpicture}
\hspace{0.5in}
\begin{tikzpicture}
\draw (-1,0)--(1,0);
\draw (-1,0)--(0,1);
\draw (1,0)--(0,1);
\draw (-0.5,0)--(-0.7,0.2);
\draw (-0.5,0)--(-0.7,-0.2);
\draw (0.5,0.5)--(0.7828,0.5);
\draw (0.5,0.5)--(0.5,0.2172);
\draw (-0.5,0.5)--(-0.5,0.7828);
\draw (-0.5,0.5)--(-0.2172,0.5);
\draw (0,-0.5)--(-0.2,-0.3);
\draw (0,-0.5)--(0.2,-0.3);
\draw (0.5,-1)--(0.3,-0.8);
\draw (0.5,-1)--(0.3,-1.2);
\draw (1,-0.5)--(0.8,-0.7);
\draw (1,-0.5)--(1.2,-0.7);
\draw (-1,0)--(-1,-1);
\draw (1,0)--(1,-1);
\draw (-1,-1)--(1,-1);
\draw (0,0)--(0,-1);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (0,1) circle (2.5pt);
\draw [fill=black] (-1,-1) circle (2.5pt);
\draw [fill=black] (1,-1) circle (2.5pt);
\draw [fill=black] (0,-1) circle (2.5pt);
\end{tikzpicture}
\caption{Two partially filled-in gameboards, the left with a cycle cell, and the right without a cycle cell}
\label{fig:cycle-cell-defintion}
\end{figure}
The rules of the game are as follows:
1) A player must make a move if it is possible
2) Neither player is allowed to create a \textit{sink} or a \textit{source}, which are vertices in which all edges incident to the vertex are directed inwards or outwards, respectively.
The winner of the game is the player who first creates a directed cycle cell, or the player who makes the last available move.
Rule 2 creates the possibility of an \textit{unmarkable edge}, which is an edge that cannot be directed, as doing so would create either a sink or a source. Figure \ref{fig:unmarkable-edge-definition} provides an example of an unmarkable edge.
\begin{figure}
\centering
\begin{tikzpicture}
\draw (-1,0)--(0,0);
\draw [black] (0,0)--(1,0);
\draw (-1,0)--(0,1);
\draw (1,0)--(0,1);
\draw (-1,0)--(-1,-1);
\draw (1,0)--(1,-1);
\draw (-1,-1)--(1,-1);
\draw (0.5,0.5)--(0.7828,0.5);
\draw (0.5,0.5)--(0.5,0.2172);
\draw (-0.5,0.5)--(-0.5,0.7828);
\draw (-0.5,0.5)--(-0.2172,0.5);
\draw (-0.5,0)--(-0.3,0.2);
\draw (-0.5,0)--(-0.3,-0.2);
\draw (1,-0.5)--(0.8,-0.3);
\draw (1,-0.5)--(1.2,-0.3);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (0,1) circle (2.5pt);
\draw [fill=black] (-1,-1) circle (2.5pt);
\draw [fill=black] (1,-1) circle (2.5pt);
\draw [fill=black] (0,-1) circle (2.5pt);
\end{tikzpicture}
\hspace{0.5in}
\begin{tikzpicture}
\draw (-1,0)--(0,0);
\draw [black] (0,0)--(1,0);
\draw (-1,0)--(0,1);
\draw (1,0)--(0,1);
\draw (-1,0)--(-1,-1);
\draw (1,0)--(1,-1);
\draw (-1,-1)--(1,-1);
\draw (0.5,0.5)--(0.7828,0.5);
\draw (0.5,0.5)--(0.5,0.2172);
\draw (-0.5,0.5)--(-0.5,0.7828);
\draw (-0.5,0.5)--(-0.2172,0.5);
\draw (-0.5,0)--(-0.3,0.2);
\draw (-0.5,0)--(-0.3,-0.2);
\draw (1,-0.5)--(0.8,-0.3);
\draw (1,-0.5)--(1.2,-0.3);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (0,1) circle (2.5pt);
\draw [fill=black] (-1,-1) circle (2.5pt);
\draw [fill=black] (1,-1) circle (2.5pt);
\draw [fill=black] (0,-1) circle (2.5pt);
\draw (0,0.2) node[] {*};
\draw (0.5,0)--(0.3,0.2);
\draw (0.5,0)--(0.3,-0.2);
\end{tikzpicture}
\hspace{0.5in}
\begin{tikzpicture}
\draw (-1,0)--(0,0);
\draw [black] (0,0)--(1,0);
\draw (-1,0)--(0,1);
\draw (1,0)--(0,1);
\draw (-1,0)--(-1,-1);
\draw (1,0)--(1,-1);
\draw (-1,-1)--(1,-1);
\draw (0.5,0.5)--(0.7828,0.5);
\draw (0.5,0.5)--(0.5,0.2172);
\draw (-0.5,0.5)--(-0.5,0.7828);
\draw (-0.5,0.5)--(-0.2172,0.5);
\draw (-0.5,0)--(-0.3,0.2);
\draw (-0.5,0)--(-0.3,-0.2);
\draw (1,-0.5)--(0.8,-0.3);
\draw (1,-0.5)--(1.2,-0.3);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (0,1) circle (2.5pt);
\draw [fill=black] (-1,-1) circle (2.5pt);
\draw [fill=black] (1,-1) circle (2.5pt);
\draw [fill=black] (0,-1) circle (2.5pt);
\draw (1,0.2) node[] {*};
\draw (0.5,0)--(0.7,0.2);
\draw (0.5,0)--(0.7,-0.2);
\end{tikzpicture}
\caption{Either direction would create a source at the starred vertices, so the undirected edge is unmarkable}
\label{fig:unmarkable-edge-definition}
\end{figure}
Let's play the Game of Cycles with the gameboard in Figure \ref{fig:sample-gameboard} below in order to better understand the game. We will call the two edges paired if they are represented by the same type of lines. We show that Player 2 has a winning strategy by responding to Player 1's first move by playing the paired edge in the same direction.
First notice that the representation of the graph matters in this game, so we will need to consider two cases for Player 1's first move.
Case 1: If Player 1 makes a move on the outer perimeter, say $b \rightarrow a$, then Player 2 should make the move on the paired edge. In our example, it is $a \rightarrow d$. From here, Player 1 must avoid the moves $c \rightarrow b$ and $d \rightarrow c$ since Player 2 could make the other of these moves to complete a cycle cell and win the game. So, there are three subcases for Player 1's next move.
Subcase a) If Player 1 makes the move $b \rightarrow c$, then Player 2 should make the move $c \rightarrow d$. From here, Player 1 cannot make either of the moves $b \rightarrow e$ or $e \rightarrow d$, as this would make a source at vertex $b$ or a sink at $d$. So, Player 1 is forced to make the move $d \rightarrow e$ or $e \rightarrow b$, which means Player 2 can make the other move and complete a cycle cell to win the game.
Subcase b) If Player 1 makes the move $b \rightarrow e$, then Player 2 should make the move $e \rightarrow d$. From here, Player 1 cannot make either of the moves $b \rightarrow c$ or $c \rightarrow d$, as this would make a source at vertex $b$ or a sink at $d$. So, Player 1 is forced to make the move $d \rightarrow c$ or $c \rightarrow b$, which means Player 2 can make the other move and complete a cycle cell to win the game.
Subcase c) If Player 1 makes the move $e \rightarrow b$, then Player 2 should make the move $d \rightarrow e$. From here, all of the moves $b \rightarrow c$, $c \rightarrow d$, $d \rightarrow c$, and $c \rightarrow b$ are losing moves for Player 1, as Player 2 can make the last remaining legal move to complete a cycle cell and win the game.
Case 2: If Player 1 makes the move $b \rightarrow c$, then Player 2 should make the move $c \rightarrow d$. From here, Player 1 must avoid the moves $a \rightarrow b$, $d \rightarrow a$, $e \rightarrow b$, and $d \rightarrow e$ since Player 2 could make a move to complete a cycle cell and win the game. From here, Player 1 must make the move $b \rightarrow e$, or some equivalent move. So, Player 2 should make the move $e \rightarrow d$. From here, Player 1 cannot make either of the moves $b \rightarrow a$ or $a \rightarrow d$, as this would make a source at vertex $b$ or a sink at $d$. So, Player 1 is forced to make the move $d \rightarrow d$ or $a \rightarrow b$, which means Player 2 can make the other move and complete a cycle cell to win the game.
Therefore, for this gameboard Player 2 has a winning strategy.
\begin{figure}
\centering
\begin{tikzpicture}
\draw [dashed] (-1,0)--(1,0);
\draw (-1,0)--(0,1);
\draw (1,0)--(0,1);
\draw [dotted] (-1,0)--(0,-1);
\draw [dotted] (1,0)--(0,-1);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (0,1) circle (2.5pt);
\draw [fill=black] (0,-1) circle (2.5pt);
\draw (0,1.3) node[] {$a$};
\draw (-1,0.3) node[] {$b$};
\draw (0,0.3) node[] {$c$};
\draw (1,0.3) node[] {$d$};
\draw (0,-0.7) node[] {$e$};
\end{tikzpicture}
\caption{Example gameboard, with solid edges on top, dashed edges in middle, dotted edges on bottom}
\label{fig:sample-gameboard}
\end{figure}
Previous research into the Game of Cycles comes from the analysis in the article \textit{The Game of Cycles} \cite{alvarado2021game}. In this article, Alvarado et al. found a number of results regarding unmarkable edges and winning strategies for specific gameboards including:
\begin{itemize}
\item Player 2 has a winning strategy for graph $K_4$
\item The number of unmarkable edges of a polygonal gameboard must be even
\item Player 1 always wins polygons of odd length, and Player 2 always wins polygons of even length
\item Winning strategies for a gameboard composed of two polygons that share exactly one edge
\item Winning strategies for select symmetric gameboards
\end{itemize}
\begin{figure}
\centering
\begin{tikzpicture}
\draw (0,0)--(0,1);
\draw (0,0)--(-0.7071,-0.7071);
\draw (0,0)--(0.7071,-0.7071);
\draw (0,1)--(-0.7071,-0.7071);
\draw (0,1)--(0.7071,-0.7071);
\draw (-0.7071,-0.7071)--(0.7071,-0.7071);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (0,1) circle (2.5pt);
\draw [fill=black] (-0.7071,-0.7071) circle (2.5pt);
\draw [fill=black] (0.7071,-0.7071) circle (2.5pt);
\end{tikzpicture}
\hspace{0.5in}
\begin{tikzpicture}
\draw (0,0)--(0,1);
\draw (0,0)--(1,0);
\draw (1,0)--(1,1);
\draw [dashed] (0,1)--(1,1);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (0,1) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (1,1) circle (2.5pt);
\end{tikzpicture}
\hspace{0.5in}
\begin{tikzpicture}
\draw (0,0)--(1,0);
\draw (0,0.75)--(0,-0.75);
\draw (1,0.75)--(1,-0.75);
\draw [dashed] (0,0.75)--(1,0.75);
\draw [dashed] (0,-0.75)--(1,-0.75);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (0,0.75) circle (2.5pt);
\draw [fill=black] (0,-0.75) circle (2.5pt);
\draw [fill=black] (1,0.75) circle (2.5pt);
\draw [fill=black] (1,-0.75) circle (2.5pt);
\end{tikzpicture}
\hspace{0.5in}
\begin{tikzpicture}
\draw (-1,0)--(1,0);
\draw (-1,0)--(-1,-0.75);
\draw (1,0)--(1,-0.75);
\draw (-1,0)--(-0.5,0.75);
\draw (1,0)--(0.5,0.75);
\draw (-0.5,0.75)--(0.5,0.75);
\draw (-1,-0.75)--(-0.5,-1.25);
\draw (1,-0.75)--(0.5,-1.25);
\draw (-0.5,-1.25)--(0.5,-1.25);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (-0.5,0.75) circle (2.5pt);
\draw [fill=black] (0.5,0.75) circle (2.5pt);
\draw [fill=black] (-1,-0.75) circle (2.5pt);
\draw [fill=black] (1,-0.75) circle (2.5pt);
\draw [fill=black] (-0.5,-1.25) circle (2.5pt);
\draw [fill=black] (0.5,-1.25) circle (2.5pt);
\end{tikzpicture}
\caption{Various gameboards, the first three solved by Alvarado et al.}
\label{fig:various-gameboards}
\end{figure}
In this paper, we will extend some of the solved classes, as well as attempt to solve some of the questions from Alvarado et al. One such question that has motivated our research is solving the gameboard that is a pentagon and a heptagon that share two consecutive edges, as pictured in Figure \ref{fig:various-gameboards}. We have then expanded this research to consider polygons that share a variable number of edges, as well as various means of "stacked" polygons.
\section*{Playing the Game of Cycles on a Line}
Before analyzing various classes and structures of gameboard, we are going to consider the structure of the most basic possible gameboard, a single line, or \textit{path} between two vertices. Although a path has no cycles, many structures can be generalized by considering paths of variable length. Therefore, it is valuable to know the number of unmarkable edges of a path with as little information as possible. Note that when we play this game on a single path, we ignore the sink-source rule for the end vertices.
The number of unmarkable edges of a path depends on the direction of the first and last edges of the path.
\begin{lemma}\label{thm:unmarkable-edges-line}
A path will have an even number of unmarkable edges when the two end edges are directed the same direction and an odd number of unmarkable edges when the two end edges are directed in opposite directions.
\end{lemma}
\begin{proof}
Imagine we add an edge to the path between the two end vertices. What we have is a polygon with one more edge than the original path. Take note of which edge was not part of the original path. We know that polygons must have an even number of unmarkable edges \cite{alvarado2021game}.
If the two end edges point in the same direction, then the new edge will not be unmarkable. So, since there must be an even number of unmarkable edges overall, there are an even number of unmarkable edges on the original path.
If the two end edges point in opposite directions, then the new edge will be unmarkable. So, since there must be an even number of unmarkable edges overall, there are an odd number of unmarkable edges on the original path.
\end{proof}
Let's consider playing the Game of Cycles on a path. Since playing the game on a path does not have any cycles, the winner is the player who plays the last legal move. We can use Lemma \ref{thm:unmarkable-edges-line} as part of our strategy, as knowing the parity of unmarkable edges will tell us which player can make the last move.
Throughout this paper we will often concentrate on edges incident to a particular vertex. For ease of discussion we introduce two new definitions.
We call a \textit{primary edge} an edge that is incident to a vertex of interest, and a \textit{secondary edge} an edge that is adjacent to a primary edge, but is not itself a primary edge.
For a path, the vertices of interest are the end vertices. Notice that directing a secondary edge has the same effect as directing its adjacent primary edge, as in order to follow the sink-source rule the adjacent primary edge may only be directed one way. So, for paths with five edges or more, the player who marks a primary or secondary edge first will lose the game, as the other player can follow up by playing the primary edge on the opposite side of the path in the direction that would guarantee the appropriate number of unmarkable edges. Consider an example of this in Figure \ref{fig:primary-secondary-edges}. Player 1 has made the move $b \rightarrow c$. However, the edge $bc$ is a secondary edge, so Player 2 can follow up by playing the move $f \rightarrow e$. This guarantees that the primary edges will point in opposite directions, meaning there will be odd number of unmarkable edges total. Since the path has five edges to start with, there will be four edges that are marked throughout the game, meaning Player 2 will make the final legal move and win. We extended this result to find the winner for a game of any length of path.
\begin{figure}
\centering
\begin{tikzpicture}
\draw (-3,0)--(2,0);
\draw (-1.5,0)--(-1.75,0.25);
\draw (-1.5,0)--(-1.75,-0.25);
\draw [fill=black] (-3,0) circle (2.5pt);
\draw [fill=black] (-2,0) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (2,0) circle (2.5pt);
\draw (-3,0.3) node[] {$a$};
\draw (-2,0.3) node[] {$b$};
\draw (-1,0.3) node[] {$c$};
\draw (0,0.3) node[] {$d$};
\draw (1,0.3) node[] {$e$};
\draw (2,0.3) node[] {$f$};
\end{tikzpicture}
\caption{Player 1 has made a losing move in this game}
\label{fig:primary-secondary-edges}
\end{figure}
\begin{theorem}\label{thm:winning-strategy-line}
Assume we are playing the Game of Cycles on a path, and say Player W is the player with the winning strategy. If the path has an odd number of edges Player 1 is Player W, and if the path has an even number of edges Player 2 is Player W.
\end{theorem}
\begin{proof}
For lines with one edge or three edges, Player 1 obviously has a winning strategy by playing the only edge or playing the center edge respectively.
For lines of two edges or four edges, Player 2 obviously has a winning strategy by playing the only available edge or by orienting the opposite primary edge in the same direction as the last made move respectively.
Assume that Player W has a winning strategy for a path of $\ell$ edges. Consider a path of length $\ell + 4$. We may imagine that such a path is actually the path of length $\ell$ with two additional edges on either side. Those four additional edges are the two primary edges and the two secondary edges.
In order to win, Player $W$ wants to play the last legal move on the inner $\ell$ edges, forcing the other player to make the first move on one of the primary or secondary edges. We assumed that Player W has a strategy to win the line of length $\ell$ which would mean they have a strategy to play the last legal move on $\ell$ edges. Thus, Player W can force the other player to play on a primary or secondary edge first, and then orient the opposite primary edge accordingly.
Of course, if the other player plays a primary or secondary edge earlier than this, Player $W$ must play the opposite primary edge accordingly.
Therefore, by induction, Player 1 is Player W for a path with and odd number edges and Player 2 is Player W for a path with an even number of edges.
\end{proof}
We will use this theorem about the game on a path to analyze other classes of gameboards. Make note that on a path of three edges, Player 1 is the first player to direct a primary or secondary edge, but still wins the game, which goes against our earlier intuition. This is because the center edge for paths of length three is \textit{doubly secondary}, as it is secondary to two different primary edges. This will cause later problems.
\section*{$j-2-k$ Graphs}
Starting with the heptagon-pentagon graph mentioned in the introduction, we define a class of graphs with three paths between two vertices of interest, such that two vertices have degree 3 and the rest of the vertices have degree 2. This could also be considered two polygons that share 2 sides. We structure the graph as it is structured in Figure \ref{fig:j-2-k-primary-edges}, and say the upper path has $j$ edges, the inner path has 2 edges, and the lower path has $k$ edges. Note that only one of $j \text{ and } k$ can be equal to 1, as otherwise this would violate the simple graph assumption for the gameboard. The example game we played earlier used $j = k = 2$. [See Figure \ref{fig:sample-gameboard}.] The unsolved graph of a heptagon and a pentagon that share two sides is the case when $j = 3$ and $k = 5$.
The vertices of interest in these $j-2-k$ graphs are the degree-3 vertices. Ultimately, the way the primary edges are directed will determine the parity of the number of unmarkable edges in the graph, so we are interested in ways we may control these primary edges, seen in Figure \ref{fig:j-2-k-primary-edges}. We will label these primary edges so we may reference them later on.
\begin{figure}
\centering
\begin{tikzpicture}
\draw (-2.,4.)--(-2.,0.);
\draw (-2.,2.)--(2.,2.);
\draw (2.,4.)--(2.,0.);
\draw [dashed] (-2, 4)--(2,4);
\draw [dashed] (-2, 0)--(2,0);
\draw [fill=black] (-2.,0.) circle (3.5pt);
\draw [fill=black] (2.,0.) circle (3.5pt);
\draw [fill=black] (-2.,4.) circle (3.5pt);
\draw [fill=black] (2.,4.) circle (3.5pt);
\draw [fill=black] (-2.,2.) circle (3.5pt);
\draw [fill=black] (2.,2.) circle (3.5pt);
\draw [fill=black] (0.,2.) circle (3.5pt);
\draw (-2.3,3) node[] {1};
\draw (-2.3,1) node[] {2};
\draw (-1,2.3) node[] {3};
\draw (1,2.3) node[] {4};
\draw (2.3,1) node[] {5};
\draw (2.3,3) node[] {6};
\end{tikzpicture}
\caption{Primary edges and structure of $j-2-k$ gameboard}
\label{fig:j-2-k-primary-edges}
\end{figure}
Our strategy consists of finding a group of configurations with the following properties: 1) Reaching the given configuration guarantees an even number of unmarkable edges. 2) It is not possible to complete both cycle cells in the graph. 3) Both players can reach at least one of the given configurations regardless of the other player's moves.
Once we found such a group of configurations we can conclude that Player 1 will win if the total number of edges is odd, and player 2 will win if the total number of edges is even.
After investigating the possible directions that these primary edges could be labelled (as well as the possibility that they become unmarkable edges) we identified a few cases that always could guarantee an even number of unmarkable edges. These cases are shown in Figure \ref{fig:guaranteed-even-unmarkable-edges}. There are more cases which also meet this requirement, but they are equivalent via rotational, horizontal, or vertical symmetry.
\begin{figure}
\centering
\begin{tikzpicture}
\draw (-1,-1)--(-1,1);
\draw (-1,0)--(1,0);
\draw (1,1)--(1,-1);
\draw [dashed] (-1,1)--(1,1);
\draw [dashed] (-1,-1)--(1,-1);
\draw [fill=black] (-1,1) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (-1,-1) circle (2.5pt);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,-1) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (1,1) circle (2.5pt);
\draw (0,-1.3) node[] {A};
\draw (-1,0.5)--(-1.25,0.75);
\draw (-1,0.5)--(-0.75,0.75);
\draw (-1,-0.5)--(-1.25,-0.75);
\draw (-1,-0.5)--(-0.75,-0.75);
\draw (1,0.5)--(0.75,0.25);
\draw (1,0.5)--(1.25,0.25);
\draw (1,-0.5)--(0.75,-0.25);
\draw (1,-0.5)--(1.25,-0.25);
\end{tikzpicture}
\hspace{0.25in}
\begin{tikzpicture}
\draw (-1,-1)--(-1,1);
\draw (-1,0)--(1,0);
\draw (1,1)--(1,-1);
\draw [dashed] (-1,1)--(1,1);
\draw [dashed] (-1,-1)--(1,-1);
\draw [fill=black] (-1,1) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (-1,-1) circle (2.5pt);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,-1) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (1,1) circle (2.5pt);
\draw (0,-1.3) node[] {B};
\draw (-1,0.5)--(-1.25,0.75);
\draw (-1,0.5)--(-0.75,0.75);
\draw (1,0.5)--(0.75,0.25);
\draw (1,0.5)--(1.25,0.25);
\draw (-0.5,0)--(-0.25,0.25);
\draw (-0.5,0)--(-0.25,-0.25);
\end{tikzpicture}
\hspace{0.25in}
\begin{tikzpicture}
\draw (-1,-1)--(-1,1);
\draw (-1,0)--(1,0);
\draw (1,1)--(1,-1);
\draw [dashed] (-1,1)--(1,1);
\draw [dashed] (-1,-1)--(1,-1);
\draw [fill=black] (-1,1) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (-1,-1) circle (2.5pt);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,-1) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (1,1) circle (2.5pt);
\draw (0,-1.3) node[] {C};
\draw (-1,0.5)--(-1.25,0.75);
\draw (-1,0.5)--(-0.75,0.75);
\draw (1,-0.5)--(0.75,-0.25);
\draw (1,-0.5)--(1.25,-0.25);
\draw (-0.5,0)--(-0.25,0.25);
\draw (-0.5,0)--(-0.25,-0.25);
\end{tikzpicture}
\hspace{0.25in}
\begin{tikzpicture}
\draw (-1,-1)--(-1,1);
\draw (-1,0)--(1,0);
\draw (1,1)--(1,-1);
\draw [dashed] (-1,1)--(1,1);
\draw [dashed] (-1,-1)--(1,-1);
\draw [fill=black] (-1,1) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (-1,-1) circle (2.5pt);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,-1) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (1,1) circle (2.5pt);
\draw (0,-1.3) node[] {D};
\draw (-1,0.5)--(-1.25,0.25);
\draw (-1,0.5)--(-0.75,0.25);
\draw (1,0.5)--(0.75,0.75);
\draw (1,0.5)--(1.25,0.75);
\draw (-0.4,0)--(-0.65,0.25);
\draw (-0.4,0)--(-0.65,-0.25);
\end{tikzpicture}
\hspace{0.25in}
\begin{tikzpicture}
\draw (-1,-1)--(-1,1);
\draw (-1,0)--(1,0);
\draw (1,1)--(1,-1);
\draw [dashed] (-1,1)--(1,1);
\draw [dashed] (-1,-1)--(1,-1);
\draw [fill=black] (-1,1) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (-1,-1) circle (2.5pt);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,-1) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (1,1) circle (2.5pt);
\draw (0,-1.3) node[] {E};
\draw (-1,0.5)--(-1.25,0.25);
\draw (-1,0.5)--(-0.75,0.25);
\draw (1,-0.5)--(0.75,-0.75);
\draw (1,-0.5)--(1.25,-0.75);
\draw (-0.4,0)--(-0.65,0.25);
\draw (-0.4,0)--(-0.65,-0.25);
\end{tikzpicture}
\caption{Cases of $j-2-k$ graphs that guarantee an even number of unmarkable edges}
\label{fig:guaranteed-even-unmarkable-edges}
\end{figure}
In order to fully understand why these five cases must have an even number of unmarkable edges, let's explain what impact the direction of an edge can have on a path, as well as the direction of edges that are adjacent to a path. Say an \textit{explicit signal} is the direction given by a primary edge and an \textit{implicit signal} is the direction that an unmarkable primary edge should be according to the other edges incident the vertex of interest. We will use -1 to represent a signal directed away from a degree 3 vertex, and use 1 to represent a signal directed toward a degree 3 vertex.
Consider Case B of Figure \ref{fig:guaranteed-even-unmarkable-edges}. Edge 2 of this case is not directed. However, if we determined the direction of edge 2 based on edges 1 and 3, we would have to direct it way from the degree 3 vertex. This is the implicit signal of this edge. So, for the left degree 3 vertex of this case there are two explicit 1 signals and one implicit -1 signal.
We define an \textit{interaction} of a path, $\{a,b\}$, the set of the two signals on either end of the path, where $a$ and $b$ are both signal values. There are three types of interactions: $\{1,1\}$, $\{-1,-1\}$, and $\{1,-1\}$. When both of the signals given are explicit signals, the number of unmarkable edges of the path is given by Lemma \ref{thm:unmarkable-edges-line}. That is, for explicit signals the $\{1,1\}$ and $\{-1,-1\}$ interactions will result in an odd number of unmarkable edges, and the $\{1,-1\}$ interaction results in an even number of unmarkable edges. We will show that the same result holds when one or both of the signals are implicit.
We will denote the parity of the number of unmarkable edges by $p$. We will set $p=0$ for even parity and $p=1$ for odd parity.
\begin{lemma}\label{thm:implicit-signals-same}
Implicit signals and explicit signals in an interaction of a path with length$ \ \geq 4$ result in the same parity of the number of unmarkable edges of the path.
\end{lemma}
\begin{proof}
We handle this proof with two cases regarding the number of implicit signals in the interaction. We need the the length requirement in order to assure that no edge sends signals to both end vertices.
\noindent \textbf{Case 1: One implicit signal}
\noindent Since there is one implicit signal, then there is already one unmarkable edge, the primary edge that is giving the implicit signal. Consider the smaller path which excludes the unmarkable primary edge. This smaller path has an interaction in which both of the signals are explicit, and the parity of the number of unmarkable edges is $p + 1 \mod 2$. This is because we know the secondary edge adjacent to the unmarkable primary edge has the opposite signal than that given by the primary edge, so the parity of the number of unmarkable edges will also be the opposite.
\noindent Hence, the parity of unmarkable edges for the whole path is $((p + 1 \mod 2) + 1) \mod 2 = p \mod 2$. Therefore, the parity of unmarkable edges is unchanged from that of an interaction with two explicit signals.
\noindent \textbf{Case 2: Two implicit signals}
\noindent Since there are two implicit signals, then there is already two unmarkable edges, the primary edges that are giving the implicit signals. Consider the smaller path which excludes the primary edges. This smaller path has an interaction in which both of the signals are explicit, and the parity of the number of unmarkable edges is $p \mod 2$. This is because we know the secondary edges adjacent to the unmarkable primary edges have the opposite signal than that given by the primary edges, but since the signal is switched on both sides, there will be the same parity of unmarkable edges on this smaller path.
\noindent So, the parity of unmarkable edges for the whole path is $((p \mod 2) + 2) \mod 2 = p \mod 2$. Therefore, the parity of unmarkable edges is unchanged from that of an interaction with two explicit signals.
\end{proof}
It is not difficult to show that this lemma also holds for paths of length one, two, and three. We will use Lemma \ref{thm:implicit-signals-same} to prove information about our cases of interest.
\begin{lemma}\label{thm:proof-of-even-cases}
Each of the five cases of Figure \ref{fig:guaranteed-even-unmarkable-edges} has an even number of unmarkable edges.
\end{lemma}
\begin{proof}
Because of Lemma \ref{thm:implicit-signals-same}, as long as we know the signal that an edge will have in an interaction, we can predict the parity of unmarkable edges in the given path. So, we need not worry in this proof if a signal is implicit or explicit. Throughout this proof we will use Lemma \ref{thm:unmarkable-edges-line} repeatedly without explicitly referencing it.
\noindent \textbf{Case A:} We know that the top and bottom paths must have an even number of unmarkable edges. Also, both of the edges on the middle path must be directed the same direction. So, overall there must be an even number of unmarkable edges.
\noindent \textbf{Case B:} We know that the top path must have an even number of unmarkable edges. We know that edge 2 must have a signal of -1. Then, we consider two different subcases. If edge 5 has a signal of 1, then the bottom path must have an even number of unmarkable edges. Then, edge 4 is not unmarkable. So, there are an even number of unmarkable edges overall. Instead, if edge 5 has a signal of -1, the number of unmarkable edges on the bottom path would be odd and edge 4 would become unmarkable, so, there would be an even number of unmarkable edges overall. Therefore, there must be an even number of unmarkable edges for graph B.
\noindent \textbf{Case C:} We know that edge 2 will have a signal of -1, and the bottom path has an odd number of unmarkable edges. Then, we consider two different subcases. If edge 6 has a signal of -1, then we know that the top path will have an even number of unmarkable edges. We also know that edge 4 would become unmarkable. So, there are an even number of unmarkable edges overall. Instead, if edge 6 has a signal of 1, we know the top path will have an odd number of unmarkable edges, and edge 4 would not be unmarkable. So, there would be an even number of unmarkable edges overall. So, in all cases there must be an even number of unmarkable edges.
\noindent \textbf{Case D:} We know that the top path will have an even number of unmarkable edges. We also know that edge 2 must have a a signal of 1. Then, we consider two different subcases. If edge 5 has a signal of -1, then the bottom path will have an even number of unmarkable edges, and edge 4 is not unmarkable. So, there are an even number of unmarkable edges overall. If edge 5 has a signal of 1, then the bottom path will have an odd number of unmarkable edges, and edge 4 would become unmarkable. So, there would be an even number of unmarkable edges overall. Therefore, there must be an even number of unmarkable edges in graph D.
\noindent \textbf{Case E:} We know that edge 2 will have a signal of 1, so the bottom path will have an odd number of unmarkable edges. Then, we consider two different subcases. If edge 6 has a signal of 1, then the top path will have an even number of unmarkable edges and edge 4 would become unmarkable. So, there are an even number of unmarkable edges overall. If edge 5 has a signal of -1, then the top path will have an odd number of unmarkable edges, and edge 4 would not be unmarkable. So, there would be an even number of unmarkable edges overall. Therefore, there must be an even number of unmarkable edges in graph E.
\end{proof}
By showing that these cases will always have an even number of unmarkable edges, we prepare ourselves for a winning strategy for Player 1 when there is an odd number of total edges, and a winning strategy for Player 2 when there is an even number of total edges. However, this is not sufficient if both the upper polygon and the lower polygon can make a cycle cell based on our knowledge of the primary edges. Specifically, guaranteeing that a player can force Case A of Figure \ref{fig:guaranteed-even-unmarkable-edges} does not mean they can guarantee a victory. Consider the partially completed game shown in Figure \ref{fig:case-a-fails}. Despite the fact that it is Player 2's turn, there is an even number of total edges, and Player 2 has guaranteed an even number of unmarkable edges with Case A, all legal moves are losing moves.
\begin{figure}
\centering
\begin{tikzpicture}
\draw (-3,2)--(3,2);
\draw (-3,2)--(-1,4);
\draw (3,2)--(1,4);
\draw (1,4)--(-1,4);
\draw (-3,2)--(-3,1);
\draw (-3,1)--(-2,0);
\draw (-2,0)--(2,0);
\draw (2,0)--(3,1);
\draw (3,1)--(3,2);
\draw (-2,3)--(-2.35,3);
\draw (-2,3)--(-2,2.65);
\draw (-3,1.5)--(-3.25,1.75);
\draw (-3,1.5)--(-2.75,1.75);
\draw (-2.5,0.5)--(-2.5,0.85);
\draw (-2.5,0.5)--(-2.85,0.5);
\draw (0,0)--(-0.25,0.25);
\draw (0,0)--(-0.25,-0.25);
\draw (3,1.5)--(3.25,1.25);
\draw (3,1.5)--(2.75,1.25);
\draw (-1.5,2)--(-1.25,2.25);
\draw (-1.5,2)--(-1.25,1.75);
\draw (2,3)--(1.65,3);
\draw (2,3)--(2,3.35);
\draw [fill=black] (-3.,2.) circle (3.5pt);
\draw [fill=black] (0.,2.) circle (3.5pt);
\draw [fill=black] (3.,2.) circle (3.5pt);
\draw [fill=black] (1.,4.) circle (3.5pt);
\draw [fill=black] (-1.,4.) circle (3.5pt);
\draw [fill=black] (-3.,1.) circle (3.5pt);
\draw [fill=black] (-2.,0.) circle (3.5pt);
\draw [fill=black] (2.,0.) circle (3.5pt);
\draw [fill=black] (3.,1.) circle (3.5pt);
\end{tikzpicture}
\caption{Despite reaching Case A and there being an even number of edges, Player 2 is in a losing position}
\label{fig:case-a-fails}
\end{figure}
Therefore, in order to show that Player 1 will have a winning strategy for $j-2-k$ graphs with odd edges and Player 2 will have a winning strategy for $j-2-k$ graphs with even edges, we will need to use exclusively Cases B through E. In these cases only one of the polygons could potentially become a cycle cell, so by guaranteeing that a player can make the last move, they also guarantee they will be the player to complete the cycle cell.
What is it about these structures that guarantees an even number of unmarkable edges? It is the fact that all of these structures have an \textit{almost-source} and an \textit{almost-sink} at the two degree-3 vertices. An almost-source is a vertex in which all but one of the edges incident with the vertex are directed outward, and an almost-sink is a vertex in which all but one of the edges incident with the vertex are directed inward. It appears that only one of these properties is true for the four cases B through E. However, the directed inner edge of these four cases actually has an impact on both of the degree-3 vertices, such that having both an almost-sink and almost-source is unavoidable.
We will provide results for a few different cases of these $j-2-k$ graphs. For the smaller cases $1-2-2$ and $2-2-2$, note that Player 1 wins for $1-2-2$ and Player 2 wins for $2-2-2$. The second of these was shown in the introduction, the first can be shown using a fairly similar strategy, but by having Player 1 start with the path of length 1.
\begin{theorem}\label{thm:easy-j-2-k-case}
For any $j-2-k$ graph with $j,k \geq 4$, Player 1 has a winning strategy when $j+k$ is odd and Player 2 has a winning strategy when $j+k$ is even.
\end{theorem}
\begin{proof}
When $j,k \geq 4$, the each of the primary edges of the paths of lengths $j$ and $k$ has a distinct adjacent secondary edge. Therefore, directing a secondary edge whose primary edge has not yet been directed is the same as directing that primary edge, and it does not have an effect on the other degree-3 vertex. We will repeatedly use this fact and show that either player can force one of the cases B through E. Use Figure \ref{fig:greater-than-4-case} as a reference. Therefore, if there are an odd number of edges, Player 1 has a winning strategy, otherwise Player 2 has a winning strategy.
\noindent If the graph has an odd number of edges one winning strategy for Player 1 is as follows: Start by directing an inner edge in either direction. We will do this with the move $y \rightarrow b$. Then, over their next two moves, Player 1 must do two things. (1) Direct either $a \rightarrow b$ or $c \rightarrow b$ and (2) direct either $g \rightarrow f$ or $g \rightarrow h$. Since there is no edge on the paths with $j$ edges or $k$ edges that can affect both of the degree-3 vertices, Player 2 cannot stop Player 1 from completing both of these tasks. So, Player 1 can force one of the cases $B$ through $E$. Note that the order Player 1 must complete these tasks could vary based on Player 2's moves, but reaching one of the four cases is always possible for Player 1. Therefore, when $j+k$ is odd, Player 1 can force an even number of unmarkable edges, and only one possible cycle. This means that Player 1 has a winning strategy.
\noindent Given a graph with an even number of edges Player 2 can force one of the cases $B$ through $E$. To show this we must consider a few different cases based on Player 1's first move.
\noindent \textbf{Case 1: Player 1 directs an inner edge}
\noindent Without loss of generality, suppose Player 1's first move is $y \rightarrow b$. Then, Player 2 must do two things. (1) Direct either $a \rightarrow b$ or $c \rightarrow b$ and (2) direct either $g \rightarrow f$ or $g \rightarrow h$. Since there is no edge on the paths with $j$ edges or $k$ edges that can affect both of the degree-3 vertices, Player 1 cannot stop Player 2 from completing both of these tasks. So, Player 2 can force one of the cases $B$ through $E$.
\noindent \textbf{Case 2: Player 1 directs a primary or secondary edge on the path with $j$ edges or on the path with $k$ edges}
\noindent Without loss of generality, assume Player 1's first move is $a \rightarrow b$, or a secondary move that has the same effect on the degree-3 vertex, such as $x \rightarrow a$. In this case, Player 2 should direct the inner edge associated with the same degree-3 vertex in the same way, i.e., in a way to create an almost sink or an almost source. Here, that move is $y \rightarrow b$. Then, on their following move Player 2 must make one of the two moves $g \rightarrow f$ or $g \rightarrow h$. Player 1 cannot stop both of these moves since they will not be able to affect both of these edges with a single move.
\noindent \textbf{Case 3: Player 1 directs any other edge}
\noindent In this case, Player 2 can use the winning strategy described for Player 1, since a different edge must be some non-primary, non-secondary edge on either the path with $j$ edges or the path with $k$ edges. So, Player 2 can direct an inner edge, follow the strategy and force one of the cases $B$ through $E$.
\noindent As shown above, both players can force one of the cases B through E when $j,k \geq 4$. Once one of these cases is forced, when $j+k$ is odd, Player 1 will always have a move available that is not a losing move, and when $j+k$ is even, Player 2 will always have a move available that is not a losing move. I.e., Player 1 has a winning strategy when $j+k$ is odd and Player 2 has a winning strategy when $j+k$ is even.
\end{proof}
\begin{figure}
\centering
\begin{tikzpicture}
\draw (-2,-1)--(-2,1);
\draw (-2,0)--(2,0);
\draw (2,1)--(2,-1);
\draw (-2,1)--(-1,1);
\draw (-2,-1)--(-1,-1);
\draw (2,1)--(1,1);
\draw (2,-1)--(1,-1);
\draw [dashed] (-1,1)--(1,1);
\draw [dashed] (-1,-1)--(1,-1);
\draw [fill=black] (-2,1) circle (2.5pt);
\draw [fill=black] (-2,0) circle (2.5pt);
\draw [fill=black] (-2,-1) circle (2.5pt);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (2,-1) circle (2.5pt);
\draw [fill=black] (2,0) circle (2.5pt);
\draw [fill=black] (2,1) circle (2.5pt);
\draw [fill=black] (-1,1) circle (2.5pt);
\draw [fill=black] (-1,-1) circle (2.5pt);
\draw [fill=black] (1,1) circle (2.5pt);
\draw [fill=black] (1,-1) circle (2.5pt);
\draw (-2.3,1) node[] {$a$};
\draw (-2.3,0) node[] {$b$};
\draw (-2.3,-1) node[] {$c$};
\draw (-1,-0.7) node[] {$d$};
\draw (1,-0.7) node[] {$e$};
\draw (2.3,-1) node[] {$f$};
\draw (2.3,0) node[] {$g$};
\draw (2.3,1) node[] {$h$};
\draw (1,1.3) node[] {$i$};
\draw (-1,1.3) node[] {$x$};
\draw (0,0.3) node[] {$y$};
\end{tikzpicture}
\caption{Board structure for Theorem \ref{thm:easy-j-2-k-case}}
\label{fig:greater-than-4-case}
\end{figure}
Shorter paths require separate treatment as in a path of length three, the secondary edge sends a signal to both primary edges at the same time. However, an easy pairing argument gives the winning strategy for these cases.
\begin{theorem}\label{thm:3-2-k-case}
Given a $3-2-k$ graph with $k \geq 2$ Player 1 has a winning strategy if $k$ is even, and Player 2 has a winning strategy if $k$ is odd.
\end{theorem}
\begin{proof}
Let's label the edges of the middle path $e_{1}$ and $e_{2}$, the edges of the $3-$path $e_{j_{1}}, e_{j_{2}}$, and $ e_{j_{3}} $ consecutively, the vertices of the $k-$path $e_{k_{1}}, e_{k_{2}}, ..., e_{k_{k}}$ consecutively.
Let's pair the edges as follows:
$e_{1}$ with $e_{2}$, $e_{j_{1}}$ with $ e_{j_{3}}$, and $e_{k_{i}}$ with $e_{k_{k-i+1}}$ for $1 \leq i \leq \lfloor {\frac{k}{2} \rfloor} $.
When $k$ is odd pair the middle edge of the $k-$ path with $e_{j_{2}}$. Otherwise $e_{j_{2}}$ will not be paired.
If the total number of edges is odd, i.e., when $k$ is even, Player 1 should direct the unpaired edge, $e_{j_{2}}$ as her first move. After each move made by Player 2 Player 1 should direct the edge paired with the last played edge in the same direction. After each move Player 1 makes in each pair either both edges are markable, or both edges are unmarkable. Also, if any cycle cell is completed it will be completed by playing the second edge of a pair. Therefore, Player 1 will win the game.
If the total number of edges is even, i.e., when $k$ is odd, Player 2 should direct the edge paired with the last played edge in the same direction, with the possible exception of one of the middle edges. If Player 1 directs a middle edge, Player 2 will be able to direct the other middle edge, but possibly in the opposite direction.
There is a small variation to the above strategy. It is possible that using the above strategy lets Player 1 direct a cycle cell by directing the middle edge of the 3-path or the k-path last on the cycle cell. Therefore, the modified strategy should take care of avoiding this situation. When there are only three undirected edges left in either cycle cell and it is Player 1's turn if all three of these edges are markable then it must be that these three edges contain a pair of edges and the middle edge of the path. Also, the middle edge of the other path must be undirected and markable at this time. If Player 1 directs one of the paired undirected edges then player 2 should direct the middle edge of the other path on their next turn. Player 2 should consider the last two remaining undirected edges of the cycle cell paired, and follow the original strategy for the rest of the game.
Just as above, this strategy will guarantee a win for Player 2.
\end{proof}
\begin{figure}
\centering
\begin{tikzpicture}
\draw (-1,0)--(1,0);
\draw (-1,0)--(-1,-1);
\draw (1,0)--(1,-1);
\draw (1,0) arc (0:180:1);
\draw [dashed] (-1,-1)--(1,-1);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (-1,-0.5) circle (2.5pt);
\draw [fill=black] (-1,-1) circle (2.5pt);
\draw [fill=black] (1,-0.5) circle (2.5pt);
\draw [fill=black] (1,-1) circle (2.5pt);
\draw (-1.3,0) node[] {$a$};
\draw (0,0.3) node[] {$b$};
\draw (1.3,0) node[] {$c$};
\draw (-1.3,-0.5) node[] {$d$};
\draw (1.3,-0.5) node[] {$e$};
\draw (-1.3,-1) node[] {$f$};
\draw (1.3,-1) node[] {$g$};
\end{tikzpicture}
\caption{Structure for Theorem \ref{thm:1-2-k-case}}
\label{fig:1-2-k-case}
\end{figure}
\begin{theorem}\label{thm:1-2-k-case}
Given a $1-2-k$ graph with $k \geq 4$, Player 1 has a winning strategy when $k$ is even and Player 2 has a winning strategy when $k$ is odd.
\end{theorem}
\begin{proof}
See Figure \ref{fig:1-2-k-case} for this proof.
Assume $k$ is even. Then, Player 1 should start by playing $\overrightarrow{ac}$. From here, Player 1 may force one of the cases A-D by directing one of $\overline{ab}$ or $\overline{bc}$ accordingly on their next move, or completing a cycle cell if it is possible. So, Player 1 may win for these cases.
Assume $k$ is odd.
If Player 1 begins by directing edge $\overline{ac}$. Then, Player 2 should direct either of $\overline{ab}$ or $\overline{bc}$ so that the triangle cycle cannot be completed. From this point, we have reached one of cases A-D of Figure \ref{fig:guaranteed-even-unmarkable-edges}, so there will be an even number of unmarkable edges and at most one cycle cell, so Player 2 has a winning strategy.
If Player 1 begins by directing one of $\overline{ab}$ or $\overline{bc}$, then Player 2 should direct $\overline{ac}$ so the triangle cycle cannot be completed. This is then identical to the case above.
If Player 1 begins by directing one of $\overline{ad}$ or $\overline{df}$, then direct $\overline{bc}$ so that both edges are directed toward vertex $a$ or both edges are directed away from vertex $a$. From here, Player 2 may force one of the cases A-D by directing $\overline{ac}$ accordingly on their next move, or completing a cycle cell if it is possible.
If Player 1 begins by directing one of $\overline{ce}$ or $\overline{eg}$, then direct $\overline{ab}$ so that both edges are directed toward vertex $c$ or both edges are directed away from vertex $c$. From here, Player 2 may force one of the cases A-D by directing $\overline{ac}$ accordingly on their next move, or completing a cycle cell if it is possible.
If Player 1 begins by directing any other edge, Player 2 should direct edge $\overline{ac}$. From here, Player 2 may force one of the cases A-D by directing one of $\overline{ab}$ or $\overline{bc}$ accordingly on their next move, or completing a cycle cell if it is possible.
\end{proof}
It can be shown by exhaustion that Player 1 has a winning strategy for $1-2-2$ and Player 2 has a winning strategy for $1-2-3$, and we found that Player 2 has a winning strategy for $2-2-2$ in our example. So, we arrive at the following result.
\begin{corollary}\label{cor:j-2-k-full}
Consider the graph $j-2-k$ with $j \geq 1$ and $k \geq 2$. If $j+k$ is odd, Player 1 has a winning strategy. If $j+k$ is even, Player 2 has a winning strategy.
\end{corollary}
Note that this result answers the question asked by Alvarado at al. [CITATION NEEDED] about the gameboard of a glued heptagon and pentagon that motivated our research.
\section*{Graphs of $p$ Paths}
\subsection*{Structure of the Graph}
Before going further into our main result, something should be said regarding the actual structure of this graph. First, it should be said that graphs do not actually have a defined geometric structure. Because of the topological properties of this game, the geometric structure can be important for the actual strategy.
In the graph we will build, the strategy works regardless of geometric structure. Therefore, by imposing a geometric structure on the graph, we can have an intuitive understanding regarding the graph. Our graph will be $p$ paths between two vertices $u$ and $v$. We will require that $u$ and $v$ are the leftmost and rightmost vertices respectively, as seen in Figure \ref{fig:p-paths-graph}. We will also require that the graph is symmetric across a vertical line of symmetry.
\begin{figure}
\centering
\begin{tikzpicture}
\draw (-3,1)--(3,1);
\draw (-3.5,1.5)--(-3,1);
\draw (3,1)--(3.5,1.5);
\draw (-3.5,1.5)--(3.5,1.5);
\draw (-3.5,1.5)--(-3,2);
\draw (-1,2)--(1,2);
\draw (-1,2)--(-3.5,1.5);
\draw (1,2)--(3.5,1.5);
\draw (-3,2)--(-3,2.5);
\draw (-3,2.5)--(3,2.5);
\draw (3,2.5)--(3,2);
\draw (3.5,1.5)--(3,2);
\draw [dashed] (0,3)--(0,0.5);
\draw [fill=black] (1,2) circle (2.5pt);
\draw [fill=black] (-3,2) circle (2.5pt);
\draw [fill=black] (-3,1) circle (2.5pt);
\draw [fill=black] (-1,1) circle (2.5pt);
\draw [fill=black] (1,1) circle (2.5pt);
\draw [fill=black] (3,1) circle (2.5pt);
\draw [fill=black] (-1,2) circle (2.5pt);
\draw [fill=black] (3,2) circle (2.5pt);
\draw [fill=black] (-3,2.5) circle (2.5pt);
\draw [fill=black] (-1,2.5) circle (2.5pt);
\draw [fill=black] (3,2.5) circle (2.5pt);
\draw [fill=black] (1,2.5) circle (2.5pt);
\draw [fill=black] (0.,1.5) circle (2.5pt);
\draw [fill=white] (-3.5,1.5) circle (2.5pt);
\draw [fill=white] (3.5,1.5) circle (2.5pt);
\draw (-4,1.5) node[] {$u$};
\draw (4,1.5) node[] {$v$};
\end{tikzpicture}
\caption{Example of a path of $p$ paths}
\label{fig:p-paths-graph}
\end{figure}
This second requirement allows us to define two different types of edges. An \textit{unmirrorable edge} is an edge that is self-reflective under the vertical line of symmetry. These are the edges in Figure \ref{fig:p-paths-graph} in which the vertical line of symmetry goes through the edge. A \textit{mirrorable edge} is an edge that is not self-reflective under the vertical line of symmetry. These are all of the other edges of Figure \ref{fig:p-paths-graph}.
\subsection*{Strategy on $p$ Paths}
There are a few more definitions needed in order to design the strategy for the graph on $p$ paths. These definitions only exist within the realm of a constructive proof of a strategy. So, these definitions require that a previous move is made, and then we may define the ``response" to that move.
The \textit{mirror-reverse move} is the move made by flipping the previous move over the vertical line of symmetry, and then reversing the direction of that arrow. An example of the mirror-reverse move is shown in Figure \ref{fig:mirror-reverse-move}. We note that the mirror-reverse move is only a possible response if the previous move was made on a mirrorable edge.
\begin{figure}
\centering
\begin{tikzpicture}
\draw (-2,0)--(3,0);
\draw (-1.5,0)--(-1.75,0.25);
\draw (-1.5,0)--(-1.75,-0.25);
\draw [fill=black] (-2,0) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (2,0) circle (2.5pt);
\draw [fill=black] (3,0) circle (2.5pt);
\draw [white] (0,-0.75) circle (2.5pt);
\end{tikzpicture}
\hspace{4in}
\begin{tikzpicture}
\draw (-2,0)--(3,0);
\draw (-1.5,0)--(-1.75,0.25);
\draw (-1.5,0)--(-1.75,-0.25);
\draw (2.5,0)--(2.75,0.25);
\draw (2.5,0)--(2.75,-0.25);
\draw [fill=black] (-2,0) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (2,0) circle (2.5pt);
\draw [fill=black] (3,0) circle (2.5pt);
\draw [white] (0,-0.75) circle (2.5pt);
\end{tikzpicture}
\hspace{4in}
\begin{tikzpicture}
\draw (-2,0)--(3,0);
\draw (-1.5,0)--(-1.75,0.25);
\draw (-1.5,0)--(-1.75,-0.25);
\draw (2.5,0)--(2.25,0.25);
\draw (2.5,0)--(2.25,-0.25);
\draw [fill=black] (-2,0) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (2,0) circle (2.5pt);
\draw [fill=black] (3,0) circle (2.5pt);
\end{tikzpicture}
\caption{A move, its mirror move, and the mirror-reverse move}
\label{fig:mirror-reverse-move}
\end{figure}
As explored in Alvarado et al, continually playing the mirror-reverse move guarantees that later mirror-reverse moves will always be available. That is, if the mirror-reverse move would be illegal for some reason, it must be that the previous move was also in fact illegal. By contrapositive, this means if the previous move was legal and a mirror-reverse move exists, then the mirror-reverse move must also be legal. This is because mirror-reverse moves guarantee a structure that replicates direction changes on a path.
We also define two more types of edges that exists purely within the realm of this specific proof. We say that two edges are \textit{paired} in the proof construction if they are the losing player's move and the winning player's specific response to that move. At the beginning of the game, two edges are paired when they are each others' reflection over the vertical line of symmetry, so all of the mirrorable edges are paired edges.
An edge is a \textit{loner edge} if the edge is not paired. That is, in the proof construction the response to a move on a loner edge is not an edge on the same path. At the beginning of the game, the loner edges are the unmirrorable edges.
It is possible that a loner edge may become a paired edge, or that a paired edge may become a loner edge, and this happens in one specific case. The preexisting conditions of the path must be that there are three undirected markable edges left on the path, and that one of these must be the loner edge. This implies that this will only happen on paths of odd length. Then, if our opponent directs one of the paired edges, we redefine what the edges are. We say that the two remaining unmarked edges are now paired edges, and the edge that was just marked was a loner edge. In Figure \ref{fig:paired-loner-switch-roles} we have the initial conditions of this case, the move that is made in this case, and the new set of paired edges.
\begin{figure}
\centering
\begin{tikzpicture}
\draw (-2,0)--(3,0);
\draw (-0.5,0)--(-0.75,0.25);
\draw (-0.5,0)--(-0.75,-0.25);
\draw (1.5,0)--(1.25,0.25);
\draw (1.5,0)--(1.25,-0.25);
\draw [fill=black] (-2,0) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (2,0) circle (2.5pt);
\draw [fill=black] (3,0) circle (2.5pt);
\draw [white] (0,-0.75) circle (2.5pt);
\end{tikzpicture}
\hspace{4in}
\begin{tikzpicture}
\draw (-2,0)--(3,0);
\draw (-1.5,0)--(-1.75,0.25);
\draw (-1.5,0)--(-1.75,-0.25);
\draw (-0.5,0)--(-0.75,0.25);
\draw (-0.5,0)--(-0.75,-0.25);
\draw (1.5,0)--(1.25,0.25);
\draw (1.5,0)--(1.25,-0.25);
\draw [fill=black] (-2,0) circle (2.5pt);
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (2,0) circle (2.5pt);
\draw [fill=black] (3,0) circle (2.5pt);
\draw (0.5,0.2) node[] {*};
\draw (2.5,0.2) node[] {*};
\end{tikzpicture}
\caption{Top image provides initial conditions, bottom image provides new move, and stars newly paired edges}
\label{fig:paired-loner-switch-roles}
\end{figure}
Now that we have made these definitions, we may proceed with our result.
\begin{theorem}\label{p-paths}
Consider a graph consisting of $p$ disjoint paths between two vertices $u$ and $v$ with the property that every path with an odd number of edges between $u$ and $v$ has at least 5 edges. For such a graph, Player 1 has a winning strategy when there is an odd number of total edges and Player 2 has a winning strategy when there is an even number of total edges.
\end{theorem}
\begin{proof}
This strategy will be completed in cases regarding the the total number of edges and the opponent's move. Note that while a game may go from Subcase A to Subcase B, the opposite may not happen.
\begin{enumerate}
\item [\textbf{Case 1)}] \textbf{Odd number of total edges}
Player 1 will start by playing a loner edge.
\begin{enumerate}
\item [\textbf{Subcase A)}] \textbf{There is at least one path whose primary edges have or could be directed to have an explicit signal from left to right and at least one path whose primary edges have or could be directed to have an explicit signal from right to left}
In this subcase, we will show that Player 1 may make the last move on every path. Therefore, if the winning player wins with a cycle cell, Player 1 will complete the cycle cell by completing the path, and if the winning player wins by making the last move, Player 1 will make the last move on every path, and therefore the graph overall.
\begin{enumerate}
\item [\textbf{Subsubcase i)}]\textbf{Player 2 plays a loner edge}
Because there are an odd number of edges overall, there must be an odd number of paths with an odd number of edges, and therefore an odd number of loner edges at the beginning of the game. Since Player 1 started by playing a loner edge, if they play a loner edge whenever Player 2 plays a loner edge, Player 1 will direct the last loner edge.
Note that this also includes whenever a loner edge is played when a paired edge becomes a loner edge.
\item [\textbf{Subsubcase ii)}]\textbf{Player 2 plays a paired edge}
\begin{enumerate}
\item [\textbf{Subsubsubcase a)}] \textbf{Edge is on a path with an even number of edges}
Play the mirror-reverse move. Assuming that we have played the mirror-reverse move in all previous moves on this path, we know we can continue to do so.
\item [\textbf{Subsubsubcase b)}] \textbf{Edge is on a path with an odd number of edges and the unmirrorable edge is not a paired edge}
Play the mirror-reverse move. Assuming that we have played the mirror-reverse move in all previous moves on this path, we know we can continue to do so.
\item [\textbf{Subsubsubcase c)}] \textbf{Edge is on a path with an odd number of edges and the unmirrorable edge is a paired edge}
Because of our analysis of this case, we know there is only one other move available, the edge paired to the edge just played. Because the mirror-reverse move was used for all other moves on this edge, we know that this move must be available.
\end{enumerate}
\end{enumerate}
\item [\textbf{Subcase B)}] \textbf{Any other case}
In this case, it is either impossible to direct the primary edges of at least one path from left to right, or it is impossible to direct the primary edges of at least one path from right to left. Without loss of generality, consider the former.
The key fact to note regarding this case is that since no pair of primary edges may be directed from left to right, this means that it is impossible to create a cycle cell. So, all that is left to show is that there must be an even number of unmarkable edges overall.
None of the primary edges may be directed from left to right, so for the ones that are directed, they are directed from right to left. Not all of them may be directed from right to left, as this would create a sink at vertex $u$ and a source at vertex $v$. Therefore, we know that there is at least one unmarkable edge incident to vertex $u$ and at least one unmarkable edge incident to vertex $v$. Notice that for any given vertex, there may be only one unmarkable edge incident to that vertex, as if there are two undirected edges incident to the vertex, one of them can be directed without creating a sink or a source, simply meeting the needs of the other vertex incident that edge. Therefore, there is exactly one unmarkable edge incident to vertex $u$ and exactly one unmarkable edge incident to vertex $v$.
We know that all but two of the primary edges are directed from right to left. Using Lemma \ref{thm:unmarkable-edges-line}, any path that has both primary edges directed from right to left has an even number of unmarkable edges.
Then, the two unmarkable primary edges must have an implicit signal from the vertices $u$ and $v$ directed them from left to right. If these two primary edges are on the same path then we may apply Lemma \ref{thm:unmarkable-edges-line} and Lemma \ref{thm:implicit-signals-same} to conclude there is an even number of unmarkable edges on that path. If these two primary edges are on different paths, then we may apply Lemma \ref{thm:unmarkable-edges-line} and Lemma \ref{thm:implicit-signals-same} to conclude there is an odd number of unmarkable edges on both paths. In either case, there is an even number of unmarkable edges overall.
Therefore, the winning player is determined by the parity of the total number of edges.
\end{enumerate}
\item [\textbf{Case 2)}] \textbf{Even number of total edges}
In this case, follow subcases A and B from Case 1. The only different argument needed is that because there are an even number of edges overall, there must be an even number of paths with an odd number of edges, and therefore an even number of loner edges. So, if Player 2 plays a loner edge whenever Player 1 plays a loner edge, they will direct the last loner edge.
\end{enumerate}
\end{proof}
\begin{corollary}\label{cor:p-paths-extended}
Make the same assumptions for the graph above, except that for graphs that have an odd number of total edges, case 1, there may be a single path of length 3 or a single path of length 1. In such a case, Player 1 has a winning strategy.
\end{corollary}
\begin{proof}
For Player 1's first move, they must move on the central, or only edge of the path of length 3 or 1, respectively. From this point, they may follow the strategy as described in the proof above.
\end{proof}
\subsection*{Issue of Short Odd Paths}
There is an issue with having many short odd paths because when Player 1 does not have a chance to play these unmirrorable edges early on, and uses the above strategy, Player 2 may exploit their plans.
For example, consider a path with an even number of edges that is adjacent to a path of one edge. If Player 1 does not have the chance to direct this path of length 1 beforehand, Player 2 may play the edges on the path with even edges. Since Player 1's response is to play the mirror-reverse of these moves, Player 1 will complete the path, and then Player 2 may direct the path with one edge to complete the cycle.
For paths of length three, there is again an issue if Player 1 cannot make sure they play the unmirrorable edge early on. If not, Player 2 may play one of the primary edges of the path of length three. This has two potentially negative outcomes. If Player 1 leaves this path the way it is, Player 2 could later direct the other primary edge such that there is an odd number of unmarkable edges on the path. If Player 1 plays the mirror-reverse move, there could be a similar situation as that described with the path with one edge.
\section*{Grids}
In Alvarado et al \cite{alvarado2021game}, Theorem 6 provides that graphs with an involution that left each cell nowhere-involutive or self-involutive, then Player 1 has a winning strategy when there is exactly one edge fixed by the involution, and Player 2 has a winning strategy otherwise.
Consider a graph that is a grid of $a \times b$ vertices. That is, the height of the grid is $a-1$ edges and the width of the grid is $b-1$ edges. For such a graph, we are interested in the involution of 180 degree rotation since any other involution would fix multiple edges.
Under Alvarado et al's Theorem 6, Player 2 will win when $a$ and $b$ have the same parity. However, when $a$ and $b$ have an opposite parity, there is exactly one self-involutive edge, but this makes two cells part-involutive, so the theorem does not apply.
So, the following work will be done for when $a$ and $b$ have opposite parity. Say that we consider the grid to be the first quadrant with vertices placed at the integer coordinates of the coordinate plane with the leftmost path as the $y$-axis and the bottommost path as the $x$-axis. Then the left bottom vertex may be labelled $(0,0)$, and the top right vertex may be labelled $(b,a)$. Then, the vertex $(i,j)$ after rotating 180 degrees is $(b-i,a-j)$, and the center of rotation is the center of an edge.
Player $W$ will use the rotate-reverse strategy. This strategy involves rotating the other player's most recent move 180 degrees and then reversing the direction. Consider the move $(i_1, j_1) \rightarrow (i_2, j_2)$. Rotating this move 180 degrees gives the move $(b - i_1, a - j_1) \rightarrow (b - i_2, a - j_2)$. Then, reversing this move gives $(b - i_2, a - j_2) \rightarrow (b - i_1, a - j_1)$.
As shorthand for this notation, we will say that vertex $v = (i,j)$ and vertex $v^{\prime} = (b - i, a - j)$. In Figure \ref{fig:rotate-reverse-move}, say the bottom left vertex is $(0,0)$ and the top right vertex is $(3,2)$. Then, say we have the move $x \rightarrow y$ has the rotate-reverse move $y^{\prime} \rightarrow x^{\prime}$. That is, the rotate verse of the move $(0,0) \rightarrow (1,0)$ is $(2,2) \rightarrow (3,2)$
\begin{figure}
\centering
\begin{tikzpicture}
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (0,1) circle (2.5pt);
\draw [fill=black] (0,2) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (1,1) circle (2.5pt);
\draw [fill=black] (1,2) circle (2.5pt);
\draw [fill=black] (2,0) circle (2.5pt);
\draw [fill=black] (2,1) circle (2.5pt);
\draw [fill=black] (2,2) circle (2.5pt);
\draw [fill=black] (3,0) circle (2.5pt);
\draw [fill=black] (3,1) circle (2.5pt);
\draw [fill=black] (3,2) circle (2.5pt);
\draw (0,0)--(0,2);
\draw (1,0)--(1,2);
\draw (2,0)--(2,2);
\draw (3,0)--(3,2);
\draw (0,0)--(3,0);
\draw (0,1)--(3,1);
\draw (0,2)--(3,2);
\draw (0,-0.3) node[] {$x$};
\draw (1,-0.3) node[] {$y$};
\draw (3,2.3) node[] {$x^{\prime}$};
\draw (2,2.3) node[] {$y^{\prime}$};
\draw (0.5,0)--(0.3,0.2);
\draw (0.5,0)--(0.3,-0.2);
\draw (2.5,2)--(2.3,2.2);
\draw (2.5,2)--(2.3,1.8);
\end{tikzpicture}
\caption{Example of Rotate-Reverse Move}
\label{fig:rotate-reverse-move}
\end{figure}
\begin{figure}
\centering
\begin{tikzpicture}
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (0,1) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\draw [fill=black] (1,1) circle (2.5pt);
\draw (0,-0.3) node[] {$y^{\prime}$};
\draw (1,-0.3) node[] {$z^{\prime}$};
\draw (0,1.3) node[] {$x^{\prime}$};
\draw (1,1.3) node[] {$w^{\prime}$};
\draw (0,0)--(1,0);
\draw (1,0)--(1,1);
\draw (1,1)--(0,1);
\draw (0,1)--(0,0);
\draw (0.5,1)--(0.3,1.2);
\draw (0.5,1)--(0.3,0.8);
\draw (1,0.5)--(1.2,0.7);
\draw (1,0.5)--(0.8,0.7);
\draw [-{implies},double] (1.5,0.5) -- (2.5,0.5);
\draw [fill=black] (3,0) circle (2.5pt);
\draw [fill=black] (4,0) circle (2.5pt);
\draw [fill=black] (3,1) circle (2.5pt);
\draw [fill=black] (4,1) circle (2.5pt);
\draw (3,-0.3) node[] {$w$};
\draw (4,-0.3) node[] {$x$};
\draw (3,1.3) node[] {$z$};
\draw (4,1.3) node[] {$y$};
\draw (3,0)--(4,0);
\draw (4,0)--(4,1);
\draw (4,1)--(3,1);
\draw (3,1)--(3,0);
\draw (3.5,0)--(3.3,0.2);
\draw (3.5,0)--(3.3,-0.2);
\draw (3,0.5)--(3.2,0.7);
\draw (3,0.5)--(2.8,0.7);
\end{tikzpicture}
\caption{Player 1 will only make a cycle available if Player 2 makes a cycle available}
\label{fig:grid-cycle-image}
\end{figure}
\begin{theorem}\label{thm:grids}
Consider a graph that is a grid of $a \times b$ vertices, where exactly one of $a$ or $b$ is odd. Player 1 has a winning strategy for this gameboard.
\end{theorem}
\begin{proof}
Because one of the dimensions of the grid is odd, there is exactly one self-involutive edge under the involution of 180 degree rotation. Player 1 starts by playing this self-involutive edge. That is, if the move is $x \rightarrow y$, then $x = y^{\prime}$ and $y = x^{\prime}$.
From here, whenever Player 2 plays $x \rightarrow y$, Player 1 should respond with $y^{\prime} \rightarrow x^{\prime}$, unless it is possible to complete a cycle, in which case Player 1 will do that instead.
Assume that up until this point in the game, Player 1 has used this strategy. Then, we will show that Player 1's move must always be available, legal, and will not create a winning move for Player 2.
Player 1's move must be available. If the move $y^{\prime} \rightarrow x^{\prime}$ had already been played, then there are two cases. Either Player 1 played it in response to $x \rightarrow y$ having been played earlier, in which case, Player 2 could not have played $x \rightarrow y$ on the most recent move. Or, if Player 2 played $y^{\prime} \rightarrow x^{\prime}$, then Player 1 would have played $x \rightarrow y$ in response, which means Player 2 couldn't have played $x \rightarrow y$ on the most recent move.
Player 1's move must be legal, as if the move $y^{\prime} \rightarrow x^{\prime}$ creates a sink or a source, then the previous move $x \rightarrow y$ by Player 2 must have created a sink or a source as well. To explain this more precisely:
Consider a fixed $x$ vertex. Then, whenever Player 2 plays $x \rightarrow y$ and Player 1 plays $y^{\prime} \rightarrow x^{\prime}$ in response, there is an arrow going out of $x$ and an arrow going into $x^{\prime}$. So, Player 1 can only create a sink by playing $y^{\prime} \rightarrow x^{\prime}$ if Player 2 has created a source by playing $x \rightarrow y$. Since Player 2 cannot legally create a source, Player 1 cannot create a sink with this strategy.
Consider a fixed $y$ vertex. Then, whenever Player 2 plays $x \rightarrow y$ and Player 1 plays $y^{\prime} \rightarrow x^{\prime}$ in response, there is an arrow going into $y$ and an arrow going out of $y^{\prime}$. So, Player 1 can only create a source by playing $y^{\prime} \rightarrow x^{\prime}$ if Player 2 has created a sink by playing $x \rightarrow y$. Since Player 1 cannot legally create a sink, Player 1 cannot create a source with this strategy.
Therefore, Player 1's move must be legal. Since Player 1's move must always be legal and available, Player 1 will always make the last move if there are no cycle cells made. We next show that Player 1 will always have a winning move available before Player 2 does.
Player 1's move will not create a winning move for Player 2, because if the rotate-reverse move would create a winning move for Player 2, then Player 2's most recent move has created a winning move for Player 1 which they will make instead. Consider Figure \ref{fig:grid-cycle-image} with vertices $w,x,y,$ and $z$, and its 180 degree rotation with the vertices $w^{\prime}, x^{\prime}, y^{\prime},$ and $z^{\prime}$.
Say that Player 2 has just played $x \rightarrow y$, so the rotate-reverse move is for Player 1 to play $y^{\prime} \rightarrow x^{\prime}$. Say that this would give Player 2 a winning move. Therefore, this means that the second grid on the right has two of the three directed edges: $w^{\prime} \rightarrow x^{\prime}, z^{\prime} \rightarrow w^{\prime}, y^{\prime} \rightarrow z^{\prime}$. Say without loss of generality that the two directed edges are $w^{\prime} \rightarrow x^{\prime}$ and $z^{\prime} \rightarrow w^{\prime}$ as seen below. If these moves have been played, then the moves $x \rightarrow w$ and $w \rightarrow z$ must have also been played, as Player 1 has been using rotate-reverse strategy for previous turns. Therefore, if Player 2 plays $x \rightarrow y$, then the move $y \rightarrow z$ is the winning move for Player 1.
\end{proof}
\section*{Conclusion}
There are many other directions to consider for future research with the Game of Cycles. One such direction would be to expand the class of $j-2-k$ graphs to $j-n-k$ graphs. Many of these graphs are covered by Theorem \ref{p-paths}. While expanding Theorem \ref{p-paths} directly would be more efficient, finding solutions for a generalized $j-n-k$ would be easier, and would finish the entire class of graphs of two polygons that share a variable number of edges. To complete this, one would need to find strategies for:
\begin{itemize}
\item $n = 3$
\begin{itemize}
\item $1-3-k$ with $k \geq 2$
\item $3-3-k$ with $k \geq 2$
\item $j-3-k$ with $j \geq 5$ odd and $k \geq 2$ even
\end{itemize}
\item $n \geq 4$ even
\begin{itemize}
\item $1-n-k$ with $k \geq 3$ odd
\item $3-n-k$ with $k \geq 3$ odd
\end{itemize}
\item $n \geq 5$ odd
\begin{itemize}
\item $1-n-k$ with $k \geq 2$ even
\item $3-n-k$ with $k \geq 2$ even
\end{itemize}
\end{itemize}
However, there are also much wider areas of research related to the Game of Cycles. One such topic is the parity conjecture made by Alvarado et al suggesting that Player 1 has a winning strategy when there is an odd number of total edges and Player 2 has a winning strategy when there is an even number of total edges. However, consider Figure \ref{fig:counterexample}. In this graph, there are four edges, but by playing $\overrightarrow{cd}$ to begin, Player 1 has won the game. This is because $\overline{ab}$ is automatically an unmarkable edge, since it must create either a sink or a source at vertex $a$. So, we may design a new conjecture.
\begin{conjecture}
For gameboards in which there is no vertex with degree 1, Player 1 has a winning strategy when there is an odd number of total edges, and Player 2 has a winning strategy when there is an even number of total edges.
\end{conjecture}
\begin{figure}
\centering
\begin{tikzpicture}
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (-1,-1) circle (2.5pt);
\draw [fill=black] (1,-1) circle (2.5pt);
\draw [fill=black] (0,1) circle (2.5pt);
\draw (0,-0.3) node[] {$b$};
\draw (-1.3,-1) node[] {$c$};
\draw (0,1.3) node[] {$a$};
\draw (1.3,-1) node[] {$d$};
\draw (0,0)--(1,-1);
\draw (0,0)--(-1,-1);
\draw (1,-1)--(-1,-1);
\draw (0,0)--(0,1);
\draw (0,-1)--(-0.25,-0.75);
\draw (0,-1)--(-0.25,-1.25);
\end{tikzpicture}
\caption{Player 1 has a winning strategy from this position}
\label{fig:counterexample}
\end{figure}
There are also new versions of the game that could be played. The most simple of these would to ignore the geometric drawing of the graph, and to make the object of the game to complete a cycle (as defined in graph theory), rather than a cycle cell. In fact, how much the geometric drawing of a gameboard can affect game play is a nontrivial question. Consider Figure \ref{fig:geometry-issue}. These two graphs are isomorphic. However, because of the cycle cell rules in the Game of Cycles, they must be played differently. The image on the left has two cycle cells of length four and two of length three, while the image on the right has three cycle cells of length four and one of length three. Knowing when a different board geometry affects the game could be important to understanding the Game of Cycles more generally.
\begin{figure}
\centering
\begin{tikzpicture}
\draw (0,0)--(2,0);
\draw (0,0)--(-0.5,1.9365);
\draw (0,0)--(-0.5,-1.9365);
\draw (2,0)--(2.5,1.9365);
\draw (2,0)--(2.5,-1.9365);
\draw (-0.5,1.9365)--(2.5,1.9365);
\draw (-0.5,-1.9365)--(2.5,-1.9365);
\draw (-0.5,1.9365)--(-0.5,-1.9365);
\draw (2.5,1.9365)--(2.5,-1.9365);
\draw [fill=black] (0,0) circle (2.5pt);
\draw [fill=black] (2,0) circle (2.5pt);
\draw [fill=black] (-0.5,1.9365) circle (2.5pt);
\draw [fill=black] (-0.5,-1.9365) circle (2.5pt);
\draw [fill=black] (2.5,1.9365) circle (2.5pt);
\draw [fill=black] (2.5,-1.9365) circle (2.5pt);
\draw (-0.3,2.1365) node[] {$a$};
\draw (0.2,0.2) node[] {$b$};
\draw (-0.3,-2.1365) node[] {$c$};
\draw (2.3,2.1365) node[] {$d$};
\draw (1.8,0.2) node[] {$e$};
\draw (2.3,-2.1365) node[] {$f$};
\end{tikzpicture}
\hspace{0.25in}
\begin{tikzpicture}
` \draw (0,0) circle (2.25cm);
\draw (0,0) circle (1cm);
\draw (0,1)--(0,2.25);
\draw (-.866,-0.5)--(-1.9486,-1.125);
\draw (.866,-0.5)--(1.9486,-1.125);
\draw [fill=black] (0,1) circle (2.5pt);
\draw [fill=black] (0,2.25) circle (2.5pt);
\draw [fill=black] (-.866,-0.5) circle (2.5pt);
\draw [fill=black] (-1.9486,-1.125) circle (2.5pt);
\draw [fill=black] (.866,-0.5) circle (2.5pt);
\draw [fill=black] (1.9486,-1.125) circle (2.5pt);
\draw (0.2,2.45) node[] {$a$};
\draw (2.1486,-1.325) node[] {$b$};
\draw (-2.1486,-1.325) node[] {$c$};
\draw (0.2,1.2) node[] {$d$};
\draw (1.166,-0.4) node[] {$e$};
\draw (-1.166,-0.4) node[] {$f$};
\end{tikzpicture}
\caption{These are the same graph, but they have different cycle cells}
\label{fig:geometry-issue}
\end{figure}
One other step would be to play the Game of Cycles on a tree, thus rendering the ``cycle" aspect worthless. In fact, Mathews \cite{mathews2021game} has already begun to consider this topic, naming it the Game of Arrows. Specifically, they have focused on 3-legged spider graphs. Here, Mathews makes the rule that leaves of the graph can be directed, allowing for some sources and sinks. The work in this paper regarding signals and interactions may be effective for this topic.
\bibliographystyle{plain}
|
\section{Introduction}
The quest for a complete understanding of nucleonic structure and the mechanism by which hadrons form out of constituent partons is one of the ultimate goals of nuclear physics. In deep inelastic scattering (DIS) an electron scatters off a nucleon with sufficient energy and momentum transfer that the process is well-described by incoherent scattering from individual partons (quarks or gluons). In semi-inclusive deep inelastic scattering (SIDIS), one or more hadrons are detected in coincidence with the scattered electron, providing information on the initial quark flavor, transverse momentum and spin~\cite{Anselmino:2020vlp}. The majority of SIDIS studies have focused on the analysis of hadron production in the current-fragmentation region (CFR), where the final-state hadrons are produced from the struck quark. The production of hadrons in the CFR can be described in a factorized framework by the convolution of Parton Distribution Functions (PDFs) and Fragmentation Functions (FFs)~\cite{Bacchetta:2006tn}. Here the PDFs describe the probability of finding a specific quark or gluon in a particular state inside the nucleon~\cite{Gao:2017yyd,Ethier:2020way} and the FFs dictate the formation of hadrons out of quarks and gluons~\cite{Metz:2016swz}. However, hadrons produced in the target-fragmentation region (TFR), formed under the participation of the spectator partons, are not described by this picture and have been largely unexplored until now.
This letter describes the first ever SIDIS detection of a hadron in the CFR (a $\pi^+$) in coincidence with a hadron in the TFR (a proton). The corresponding theoretical basis to study the TFR is based on the fracture function formalism and was established in Ref.~\cite{Trentadue:1993ka}
for the collinear case. This approach has been generalized to the spin- and transverse-momentum-dependent (STMD) case~\cite{Anselmino:2011ss}. Similar to the case of PDFs and FFs in the CFR, the fracture functions describe the conditional probability for the target remnant to form a specific final state hadron after the ejection of a particular quark. In electroproduction, the polarization state of the virtual photon depends on the longitudinal polarization of the lepton beam, which in turn selects preferentially one polarization state of the struck quark. The opposite polarization and transverse momentum of the remnant can introduce correlations between final-state hadrons produced in the TFR and hadrons produced in the CFR. The study of this dihadron production in SIDIS with a longitudinally polarized electron, where one hadron is produced in the CFR and another in the TFR, provides access to leading twist fracture functions~\cite{Anselmino:2011bb}.
In the valence-quark region, the
polarization transfer from the beam to the active quark is expected to be significant at the relatively low center-of-mass energies accessible at CLAS12~\cite{Burkert:2020akg}. Preliminary studies using CLAS data indicated that these target-current correlations may be significant~\cite{Avakian:2016zos}. The high luminosity and high polarization of the electron beam along with a wide acceptance for the detection of many final-state particles makes CLAS12 an ideal place for studies of correlations between the target- and current-fragmentation regions.
Sizable beam single-spin asymmetries (SSAs) for a longitudinally polarized electron beam on nucleon targets and one or two hadrons detected in the CFR have been observed at JLab
\cite{CLAS:2003qum,CLAS:2021opg,CLAS:2020igs,Hayward:2021psm},
HERMES~\cite{Airapetian:2006rx} and COMPASS~\cite{COMPASS:2014kcy,Parsamyan:2018ovx,Parsamyan:2018evv}. These results have been interpreted in terms of higher twist contributions, related to quark-gluon correlations. Here, higher twist refers to quantities that are suppressed by the hard scale of the process~\cite{Jaffe:1997vlv}.
When one of two hadrons is detected in the TFR and the other in the CFR, the beam SSAs measured here appear at leading twist~\cite{Anselmino:2012zz} without this suppression.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figures/b2b_fig.pdf}
\caption{The SIDIS kinematics of back-to-back dihadron production in the center-of-mass frame. The $x$-$z$ plane is defined by the incoming and outgoing lepton with positive $z$ in the direction of the virtual photon. $\phi_1$ and $\phi_2$ are defined from the scattering plane to $P_1$ and $P_2$ in an anti-clockwise direction.}
\label{fig:b2b_plane}
\end{figure}
In the target fragmentation region it is not possible to separate quark emission from hadron production. This prevents access to a chiral-odd quantity, such as the Collins function, to pair with any chiral-odd fracture functions in single hadron production and ultimately makes any chiral-odd fracture functions inaccessible in single hadron production. In contrast, in the double hadron production process $l(\ell) + N(P) \to l(\ell') + h_1(P_1) + h_2(P_2) + X$, at perturbative QCD leading order, the cross section expression includes all twist-2 fracture functions and quark fragmentation functions~\cite{Anselmino:2011ss,Anselmino:2011bb} by pairing the fracture function with a fragmentation function that dictates the production of a hadron in the CFR.
The process considered here is shown in Fig.~\ref{fig:b2b_plane}. We use the standard DIS variables: the momentum of the exchanged virtual photon, $q=l-l^\prime$, the scale of the process, $Q^2=-q^2$, the fractional longitudinal target momentum carried by the struck quark, $x=Q^2/2P \cdot q$, the fractional energy loss of the scattered electron $y=P \cdot q/P \cdot l$ and the hadronic mass of the system, $W^2=(P+q)^2$. The hadronic variables are defined below. For the case of a longitudinally polarized beam and unpolarized target after the integration over $\phi_2$ (the azimuthal angle of the TFR hadron) and keeping $\Delta\phi = \phi_2 - \phi_1$ fixed, there are two contributions to the cross section, $\sigma_{UU}$ and $\sigma_{LU}$ ~\cite{Anselmino:2011vkz},
\begin{align}
\sigma_{UU} &= F_0^{{\hat u_1} \cdot D_1},
\label{Sig_UU}
\end{align}
\begin{align}
\sigma_{LU} & = \frac{ P_{{T1}} P_{{T2}}}{m_N m_2} F_{{k1}}^{{\hat l_1}^{\perp h}\cdot D_1} \, \sin(\Delta\phi),
\label{Sig_LU}
\end{align}
where the structure functions $F_0^{\hat{u}_1 \cdot D_1}$ and $F_{k1}^{\hat{l}_1^{\perp h} \cdot D_1}$ are convolutions~\cite{Anselmino:2011bb} of the leading twist fracture functions $\hat{u}_1$ and $\hat{l}_1^{\perp h}$ with the unpolarized fragmentation function, $D_1$, which depend on the kinematic variables $x, Q^2, z_1, \zeta_2, P_{T1}^2, P_{T2}^2$, and ${\Vec P}_{T1} \cdot {\Vec P}_{T2}$. The masses of the nucleon target, forward and backward produced hadrons are denoted as $m_N$, $m_1$ and $m_2$. The hadron 1, $h_1$, produced in the CFR ($x_{F1}>0$)
\footnote{We use the standard definition for Feynman variables $x_{F1}$ and $x_{F2}$, see next section.}
is described by the standard scaled variable $z_1 = P{\cdot}P_1/P{\cdot}q$, describing the fraction of the virtual photon energy carried by the CFR hadron, and its transverse momentum ${\Vec P}_{T1}$ (defined relative to the $q$ vector in the target rest frame) with magnitude $P_{T1}$ and azimuthal angle
$\phi_1$. The hadron 2, $h_2$, produced in the TFR ($x_{F2}<0$) is described by similar variables: the fractional longitudinal target momentum carried by the TFR hadron, $\zeta_2 \simeq E_2/E$ where $E$ is the energy of the target,
and ${\Vec P}_{T2}$ ($P_{T2}$ and $\phi_2$) in the virtual photon and target nucleon center-of-mass frame. The usual hadronic scaling variable, $z$, is not used in the TFR because of ambiguities between soft hadron emission and target fragmentation~\cite{Anselmino:2011ss}.
In particular, the structure function in Eq.~\ref{Sig_LU} contains the fracture function $\hat{l}_{1}^{\perp h}(\zeta_2,P_{T2})$, describing the production of $h_2$ after the emission of a longitudinally polarized quark in an unpolarized nucleon and $D_1 (z_1, P_{T1})$, the unpolarized fragmentation function describing the formation of $h_1$. This structure function depends on the relative azimuthal angle of the two hadrons, indicating a long-range correlation between hadrons produced in the CFR and the TFR. The resulting beam-spin asymmetry contains the convolution of the fracture function and the fragmentation function modulated by $\sin\Delta\phi$,
\begin{align}
{\mathcal A}_{LU} &= - \sqrt{1-\epsilon^2} \frac{\vert \Vec P_{T1} \vert \vert
\Vec P_{T2} \vert}{m_N \, m_2}
\, \frac{{\mathcal C} [w_5 \, \hat{l}_1^{\perp h} D_1]}
{{\mathcal C} [\hat{u}_1 D_1]}
\sin \Delta \phi.
\label{ptdep}
\end{align}
The depolarization factor in the front of Eq.~\ref{ptdep}, governing the polarization transfer from the electron to virtual photon, is described by the variable
\begin{equation}
\epsilon = \frac{1-y-\frac{1}{4}\gamma^2 y^2}{1-y+\frac{1}{2}y^2+\frac{1}{4}\gamma^2y^2},
\end{equation}
with $\gamma = 2m_N x/Q$. The weight factor is given by
\begin{align}
w_{5} &= \frac{(\Vec k_{\perp} \cdot \Vec P_{T2}) (\Vec P_{T1}
\cdot \Vec P_{T2}) - (\Vec k_{\perp} \cdot \Vec P_{T1})
\Vec P_{T2}^2}{(\Vec P_{T1} \cdot \Vec P_{T2})^2
- \Vec P_{T1}^2 \Vec P_{T2}^2}, \>
\end{align}
and the following notation is used for the transverse momentum convolution
\begin{align}
{\mathcal C} \, [f (\Vec k_{\perp}, &\Vec k'_{\perp}, \ldots)] = \sum_a e_a^2 \, x_B \, \int d^2 \Vec k_{\perp} \, \int d^2 \Vec k_{\perp}' \times
\nonumber \\
& \delta^2 (\Vec k_{\perp} - \Vec k_{\perp}' - \Vec P_{T1}/z_1)
\, f (\Vec k_{\perp}, \Vec k'_{\perp}, \ldots) \>,
\end{align}
where $k_\perp$ is the transverse momentum of the initial quark with respect to the virtual photon, $k_\perp'$ is the transverse momentum after the interaction and the summation runs over the quark flavors.
\section{Experiment}
The data, corresponding to SIDIS events with a $\pi^+$ in the CFR and proton in the TFR, were taken in two run periods in the fall of 2018 and spring 2019 using 10.6 and 10.2 GeV longitudinally polarized electron beams delivered by the Continuous Electron Beam Accelerator Facility at Jefferson Lab~\cite{Leemann:2001dg}. The electron beam was incident on a liquid-hydrogen target and reactions were recorded using the CLAS12 spectrometer~\cite{Burkert:2020akg}. The beam polarization averaged to $85.7\pm 1.6\%$ and was flipped at a rate of 30~Hz to minimize systematic effects.
A tracking subsystem consisting of drift chambers in a toroidal magnetic field was used to identify and reconstruct particles scattered in the forward direction. A high-threshold Cherenkov counter was used to distinguish between electrons and final-state hadrons. Additional identification criteria for the electrons was also imposed using a series of electromagnetic calorimeters. The CLAS12 forward time-of-flight systems, composed of six arrays of plastic scintillation counters, were used to analyze the velocity vs. momentum relationship of positive tracks to distinguish between hadron species. The pions were limited to momenta $1.2~<~p~<~4.0$~GeV in order to avoid regions of low efficiency (lower limit) and minimize misidentification of kaons (upper limit). Protons were required to have a momentum greater than 0.5~GeV (no strict upper limit was enforced but the distribution of upper proton momenta dies down around 2.5~GeV, well before any significant contamination from lighter hadron species). The reconstructed electron and hadrons were required to have been identified in the so-called ``forward detector'' of CLAS12 and a requirement has been placed on the polar angle of each track, $\theta < 30^\circ$.
SIDIS events were selected with the usual requirement that $Q^2>1$~GeV$^2$ and the mass of the hadronic final-state, $W>2$ GeV. Events with a radiated photon were limited by requiring events to have $y~<~0.75$. At energies accessible by fixed target experiments there is no rapidity gap and the forward and backward regions were defined by the variable $x_F$, in the virtual photon-nucleon center-of-mass frame, with the requirement that ${x_F}_1~>~0$ and ${x_F}_2~<~0$. The Feynman-$x$ variable is defined as ${x_F}_{1(2)}=2{P_\parallel}_{1(2)} / W$, where $P_\parallel$ is the longitudinal momentum of the hadron and takes a positive value if the hadron moves in the same direction as the momentum transfer dictated by virtual photon and a negative value if it moves in the same direction of the target in the center-of-mass frame.
An additional requirement on the boost-invariant quantity $\Delta Y \equiv Y_1 - Y_2 > 0$, where $Y$ is the rapidity evaluated in the Breit frame defined as $2Y_{1(2)}=\ln(E_{1(2)}+P_{\parallel1(2)})/(E_{1(2)}-P_{\parallel1(2)})$, was required to enforce separation between the hadrons. The asymmetries were studied as a function of both of these variables to investigate the transition from one region to another. Contributions from $\Delta^{++}$ decays were minimized by requiring the invariant mass of the observed hadrons to have $M_{p\pi} > 1.5$~GeV.
Finally, the missing mass of the process $ep~\rightarrow~e'p \pi^+X$ was restricted to be greater than 0.95~GeV in order to avoid contributions from diffractive meson production.
\section{Results}
The beam-spin asymmetry can be accessed using the yields, $N^\pm$, of events with a proton in the backward region and a positive pion in the forward region, produced from the scattering of an electron with helicity $\pm$, written
\begin{align}
\label{eq:fit}
& \mathcal{A}_{LU}(\Delta\phi)=
\frac{1}{P_\text{beam}} \frac{N^+(\Delta\phi)- N^-(\Delta\phi)}{N^+(\Delta\phi)+ N^-(\Delta\phi)}=\\
& \mathcal{A}_{LU}^{\sin\left(\Delta\phi\right)}\sin(\Delta\phi)+\mathcal{A}_{LU}^{\sin\left(2\Delta\phi\right)}\sin(2\Delta\phi)\nonumber \nonumber,
\end{align}
with the dependence on $\sin(\Delta\phi)$ and $\sin(2\Delta\phi)$ described in Ref.~~\cite{Anselmino:2012zz}, and fitting for the resulting azimuthal modulation amplitudes. The beam polarization, $P_\text{beam}$, has been divided out of the asymmetries. The amplitudes in Eq.~\eqref{eq:fit} were extracted from the data using an unbinned maximum likelihood fit that includes both modulations of $\mathcal{A}_{LU}$.
A binned $\chi^2$-minimization fit with $9$ bins in $\Delta\phi$ was also performed and is in very good agreement with the unbinned fit with a mean reduced $\chi^2$ of 1.01. The count-rate asymmetry between positive and negative electron helicities as a function of $\Delta\phi$ is shown in Fig.~\ref{fig:sinusoid}. The count rate difference exhibits a clear $\sin(\Delta\phi)$ behavior with a much smaller $\sin(2\Delta\phi)$ contribution.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figures/ALU.pdf}
\caption{The beam spin asymmetry, $\mathcal{A}_{LU}$, as a function of $\Delta \phi$ and integrated over all other kinematics for the entire data set. A clear $\sin(\Delta\phi)$ dependence is observed with small $\sin(2\Delta\phi)$ contributions.}
\label{fig:sinusoid}
\end{figure}
The dependence of $\mathcal{A}_{LU}^{\sin\Delta\phi}$ on the product of transverse momenta of the proton and pion is shown in Fig.~\ref{fig:pTpT} and is consistent with a linear increase in magnitude and approaches zero as the transverse momentum goes to zero, following the kinematic dependence predicted by theory (see. Eq.~\ref{ptdep}). Additional multidimensional asymmetries as a function of the product of the transverse momenta in bins of $z_1$ are given in the supplementary material.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figures/pTpT.pdf}
\caption{The measured $\mathcal{A}_{LU}^{\sin\Delta\phi}$ asymmetry as a function of $P_{T1} P_{T2}$. Thin black bars indicate statistical uncertainties and wide gray bars represent systematic uncertainties.}
\label{fig:pTpT}
\end{figure}
Due to the correlation between different kinematic variables and the product of the transverse momenta of both hadrons, the asymmetries can be weighted by dividing out the depolarization and kinematic weighting factor in Eq.~\ref{ptdep}, $\sqrt{1-\epsilon^2}(P_{T1} P_{T2}) / (m_N m_2)$, (note that for this measurement $m_N = m_2 = 0.938$~GeV). The dependence of the resulting ratio, which should directly depend on the ratio of the convolutions of fracture and fragmentation functions, was studied for several different kinematic variables.
The $x$-dependence, shown in Fig.~\ref{fig:x}, has the general trend of increasing in magnitude as $x$ increases. This strong dependence implies that the correlation of final-states hadrons is most significant in the valence quark region. The dependence on $z_{1}$ of the pion, which reflects the fragmentation function dependence, is shown in Fig.~\ref{fig:z_pi}. At relatively small $z_1$, contributions from the initial quark transverse momentum can be neglected and the main contribution to the produced hadron transverse momentum comes from the struck quark hadronization process. Indeed, this dependence appears relatively flat, with a possible decrease at higher values of $z_1$ where effects from decreasing transverse momentum begin to dominate. This relatively weak dependence may also be a consequence of cancellation between the pion fragmentation functions in the numerator and denominator. The dependence on $\zeta_2$, shown in Fig.~\ref{fig:zeta}, is stronger and may be interpreted in terms of strong correlations with other variables such as $x$; typically the higher longitudinal momentum carried by the struck quark, the lower the longitudinal momentum available for the TFR hadron.
Additional kinematic dependences are included in the supplementary material. The asymmetries plotted versus ${x_F}_1$ and $\Delta Y$ (with the kinematic constraints on ${x_F}_2$, ${x_F}_1$ and $\Delta Y$ removed) show a relatively flat dependence in the ${x_F}_1 < 0$ region with a possible transition in the positive region where the proton is increasingly likely to have originated in the CFR and the back-to-back fracture function formalism no longer holds. Finally, the dependence on the missing mass, with the $M_x > 0.9$~GeV requirement removed, shows a relatively flat behavior above the contributions from diffractive $ep \rightarrow e'p\pi^+\pi^-$ and $ep \rightarrow e'p\pi^+\rho^-$ events.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figures/x.pdf}
\caption{The measured weighted $\mathcal{A}_{{LU}}^{\sin\Delta\phi}$ asymmetry as a function of $x$. Thin black bars indicate statistical uncertainties and wide gray bars represent systematic uncertainties.}
\label{fig:x}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figures/zpi.pdf}
\caption{The measured weighted $\mathcal{A}_{{LU}}^{\sin\Delta\phi}$ asymmetry as a function of $z_1$. Thin black bars indicate statistical uncertainties and wide gray bars represent systematic uncertainties.}
\label{fig:z_pi}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figures/zeta.pdf}
\caption{The measured weighted $\mathcal{A}_{{LU}}^{\sin\Delta\phi}$ asymmetry as a function of $\zeta_2$. Thin black bars indicate statistical uncertainties and wide gray bars represent systematic uncertainties.}
\label{fig:zeta}
\end{figure}
Relevant systematic uncertainties have been estimated using a number of methods. Monte Carlo simulations were performed using the PEPSI generator~\cite{Mankiewicz:1991dp} and a GEANT4-based simulation~\cite{Agostinelli:2002hh,Ungaro:2020xlc} of the detector for acceptance, efficiency and particle-ID studies. Good agreement for all underlying variables was observed. Systematic uncertainties due to bin migration and the scale uncertainty on the beam polarization can reach a few percent in each bin. Other effects due to particle misidentification, accidental coincidences, photoproduction of electrons and contamination from target-fragmentation pions (baryonic decays like $\Delta^{++} \rightarrow p\pi^+$) have all been estimated to be small. Contributions from radiative effects are avoided by limiting our kinematic range and are estimated to be small. However, more theory effort will be needed in future for detailed studies of radiative corrections in SIDIS in general, and back-to-back dihadron production in particular.
The largest systematic uncertainty comes from contributions from the unpolarized cross section. After integration over $\phi_2$, the only remaining structure functions are $F_{UU}$ and $F^{\sin(\Delta\phi)}_{LU}$. However, due to the non-perfect acceptance of CLAS12 and the potential resulting non-orthogonality of modulations, the other unpolarized structure functions may impact our extraction of the $F_{LU}$ amplitudes. Since there is no experimental data on the spin-independent ($UU$) modulations, the uncertainty due to not including these modulations in our fit has been evaluated using Monte Carlo simulations by injecting various values of the amplitudes and then performing fits not including the $UU$ modulations in order to measure the possible deviation in our $LU$ amplitudes. The effect is heavily bin-dependent, but can only reach magnitudes similar to the statistical uncertainty at edges of kinematic space. For example, in the lowest $x$-bin we assign an uncertainty to $A_{LU}^{\sin\Delta\phi}$ of 0.006 compared to 0.021 at the highest $x$ bin where the measurement is less constrained. This study is both dependent on our injected amplitudes and fairly conservative and so, therefore, may correspond to an overestimate.
\section{Conclusions}
In summary, the kinematic dependences of beam SSAs in the production of two hadrons in opposite hemispheres have been measured for the first time. The asymmetries may be interpreted in a framework described by TMD factorization into fracture functions and fragmentation functions, with the conditional probability of finding a proton originating from the target remnant after the emission of a quark which undergoes hadronization to form a final-state $\pi^+$. The $P_T$ dependence of the asymmetries is consistent with predictions of the factorized framework and can ultimately be used to test TDM factorization once extractions of the relevant functions are available and predictions have been made.
Our measurement of correlations between the target- and current-fragmentation regions develop a new methodology to quantify the relationship between the spin and transverse momenta of quarks in the nucleon and provides a new avenue for studies of the complex nucleonic structure in terms of quark and gluon degrees of freedom. The kinematics of the generated asymmetries are not in the perturbative regime, but instead, the asymmetries likely originate from correlations between the longitudinally polarized struck quark's azimuthal angle and the azimuthal angle of the proton produced in the TFR~\cite{Anselmino:2011ss}.
Future work will extend the analysis to other hadron species in both the TFR and CFR in order to test the universal nature of fracture functions. The flavor dependence of fracture functions can be extracted by comparing with deuterium targets, of which comparable statistics to the proton-target data shown here have already been collected by CLAS12. Finally, polarized-target measurements will enable access to the complete set of leading-twist fracture functions, which cannot be observed in single-hadron production but require an additional hadron to produced in the CFR like the measurement performed here.
\section{Acknowledgements}
We acknowledge the outstanding efforts of the staff of the Accelerator and the Physics Divisions at Jefferson Lab in making this experiment possible. This work was supported in part by the U.S. Department of Energy, the National Science Foundation (NSF), the Italian Istituto Nazionale di Fisica Nucleare (INFN), the French Centre National de la Recherche Scientifique (CNRS), the French Commissariat pour l$^{\prime}$Energie Atomique, the UK Science and Technology Facilities Council, the National Research Foundation (NRF) of Korea, the Helmholtz-Forschungsakademie Hessen für FAIR (HFHF), the Ministry of Science and Higher Education of the Russian Federation, the Chilean Agencia Nacional de Investigacion y Desarrollo (ANID) via grant PIA/APOYO AFB180002 and the European Union's
Horizon 2020 Research and Innovation program under Grant
Agreement N.824093 (STRONG2020). The Southeastern Universities Research Association (SURA) operates the Thomas Jefferson National Accelerator Facility for the U.S. Department of Energy under Contract No. DE-AC05-06OR23177. The work is also supported in part by DOE grant DE-FG02-04ER41309.
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
Global and partial actions of Hopf algebras on algebras are widely studied subjects, with various interests, applications and ramifications. However, calculating and classifying such actions is a difficult task. There exist several works that calculate (global) actions of certain Hopf algebras on different types of algebras, such as fields, domains, Weyl algebras, path algebras of quivers, among others \cite{CuadraEtingofWaltonI,CuadraEtingofWaltonII, etingofewaltonSemisimple, etingofewaltonI, etingofewaltonFinite, etingofewaltonFinite2, etingofewaltonII}.
In general, the works investigate actions of two classes of Hopf algebras: the semisimple and the pointed (nonsemisimple).
In particular, the Taft algebra is an important example to be considered in the pointed case.
Because of that, some works deal particularly with Taft actions on algebras \cite{Acoes_taft_Centrone, Acoes_taft_Chelsea, montgomery_Schneider_taft_actions}.
In order to understand the actions of pointed Hopf algebras, another interesting example to be considered is the Nichols Hopf algebra: it is the archetype of pointed Hopf algebras.
Indeed, finite dimensional pointed Hopf algebras (with abelian group of group-like elements $G$) are expected to be a lifting of the bosonization of a Nichols algebra by the group algebra $\Bbbk G$ \cite{Nicolas_ICM}. In particular, the Nichols Hopf algebra is just the bosonization of an exterior algebra by the group algebra of the cyclic group of order 2.
For partial Hopf actions, the situation is even more unexplored, with just a few concrete examples known.
In \cite{taft_corpo_revista}, the authors exhibit explicitly all partial actions of Taft and Nichols Hopf algebras on their base fields, obtaining that such partial actions depend only on the partial action of the group of group-like elements, a derivation and a compatibility between both, same as it happens in the global case \cite{Acoes_taft_Centrone}.
In this work, the authors continue the investigation of the partial actions of Taft and Nichols Hopf algebras on algebras, extending the results of \cite{taft_corpo_revista}.
In particular, Theorem \ref{taft_principal} shows that, with suitable conditions, the partial actions still depend only on the partial group action, a derivation and a compatibility between them.
This paper is organized as follows.
In Section \ref{sec:preliminaries}, we recall the definitions of partial (co)actions of a bialgebra on an algebra and also some definitions and results about \emph{q-combinatorics}, that will be used in the computations of the partial Taft actions.
In Section \ref{sec:Taft}, we investigate and characterize suitable partial actions of the Taft algebra on an algebra. Furthermore, we present the partial Taft coactions obtained dually from such partial actions. At last, in Section \ref{sec:Nichols}, we compute some partial (co)actions of a Nichols Hopf algebra on algebras.
\section{Preliminaries}\label{sec:preliminaries}
Throughout the work, \emph{algebra} means \emph{unital associative algebra}.
Besides, $\Bbbk^{\times} = \Bbbk \backslash \{0\}$ and unadorned $\otimes$ means $\otimes_{\Bbbk}$.
For a coalgebra $C$, $\varepsilon_C$ and $\Delta_C$ stands for the counit and the comultiplication maps of $C$, respectively, and will be written as $\varepsilon$ and $\Delta$ if there is no ambiguity.
Moreover, we use Sweedler-Heynemann notation with summation sign suppressed for the comultiplication map, namely $\Delta(c)=c_1 \otimes c_2 \in C \otimes C$, for all $c \in C$.
We write $G(C)= \{g \in C \backslash \{0\} \ : \ \Delta(g) = g \otimes g \}$ for the set of the \emph{group-like elements} of the coalgebra $C$.
Given $g,h \in G(C)$, an element $x \in C$ is called a \emph{$(g,h)$-primitive element} if $\Delta(x)= x \otimes g + h \otimes x$.
If no emphasis on the elements $g, h \in G(C)$ is needed, a $(g,h)$-primitive element will be called simply of a \emph{skew-primitive element}.
For a group $G$, we write $\Bbbk G$ for the group algebra with the canonical basis $\{g \ : \ g \in G \}$.
Moreover, given a basis $\{v_1, v_2, \cdots , v_n\}$ of a finite-dimensional vector space, we write $ \{v_1^*, v_2^*, \cdots , v_n^*\}$ for its dual basis.
We use $\mathbb{Z}$, $\mathbb{N}$ and $\mathbb{N}_0$ for the sets of integers, the positive integers and $\mathbb{N} \cup \{0\}$, respectively.
Let $ j, k \in \mathbb{Z}$ and $n \in \mathbb{N}$.
The symbols $\delta_{j,k}$ and $C_n$ stand for Kronecker's delta and the cyclic group of order $n$, respectively.
If $j \leq k$, then $\I_{j, k} =\{j, j+1, \cdots, k\}$ and $\I_n = \I_{1, n}$.
\subsection{Partial (co)actions}
Partial actions and partial coactions of bialgebras on algebras were introduced in \cite{caenepeel2008partial}.
Since this earlier definition does not deal with symmetric conditions, we will present as it appears in later works.
These conditions are important for some developments in the theory of partial (co)actions, such as partial (co)representations \cite{corepresentations, Partial_representations, dilations}.
A \emph{partial action of a bialgebra $H$ on an algebra $A$} is a linear map $\cdot: H \otimes A \longrightarrow A$, denoted by $\cdot (h \otimes a) = h \cdot a$, such that
\begin{itemize}
\item[(PA.1)]\label{PA1} $1_H \cdot a=a$,
\item[(PA.2)]\label{PA2} $h\cdot ab=(h_1\cdot a)(h_2\cdot b)$, and
\item[(PA.3)]\label{PA3} $h\cdot(k\cdot a)=(h_1\cdot 1_A)(h_2k\cdot a)$,
\end{itemize}
hold for all $h,k\in H$ and $a,b\in A$.
In this case, $A$ is called a \textit{partial $H$-module algebra}.
A partial action is \emph{symmetric} if in addition we have
\begin{itemize}
\item[(PA.S)]\label{PAS} $h \cdot ( k \cdot a)=(h_1k \cdot a)(h_2 \cdot 1_A)$.
\end{itemize}
Note that conditions (PA.2) and (PA.3) are equivalent to the following:
\begin{itemize}
\item[(PA.2$^\prime$)]\label{PA2e3} $h \cdot (a( k \cdot b))=(h_1 \cdot a)(h_2k \cdot b)$.
\end{itemize}
Every $H$-module algebra is a symmetric partial $H$-module algebra. Moreover, a partial $H$-module algebra is an $H$-module algebra if and only if $h \cdot 1_A = \varepsilon(h) 1_A$, for all $h \in H$.
A \emph{(symmetric) partial coaction of a bialgebra $H$ on an algebra $A$} is a linear map $\rho: A \longrightarrow A\otimes H$ satisfying dual conditions to the previous conditions (PA.$\_$).
In fact, the concepts of partial actions and partial coactions are dually related in \cite{dual_constructions} as follows.
Let $\langle -,- \rangle: K \otimes H \longrightarrow \Bbbk$ be a dual pairing between bialgebras and $A$ an algebra. If $A$ is a (symmetric)
partial $K$-comodule algebra, then $A$ is a (symmetric) partial $H$-module algebra.
Furthermore, if the dual pairing $\langle -,- \rangle$ is non-degenerate and $A$ is a rational (symmetric) partial $H$-module algebra, then $A$ is a (symmetric) partial $K$-comodule algebra.
Thus, in the particular case that $H$ is a finite-dimensional bialgebra with basis $\{h_i\}_{i \in I}$, we obtain the following characterization.
\begin{prop} \label{dual}
If $A$ is a (symmetric) partial $H$-module algebra via $h \cdot a$, then $A$ is a (symmetric) partial $H^{\ast}$-comodule algebra via $\rho: A \rightarrow A\otimes H^*$, given by $\rho(a)=\sum_{i \in I} h_i\cdot a\otimes h_i^{\ast}$. Reciprocally, if $A$ is a (symmetric) partial $H$-comodule algebra via $\rho: A \rightarrow A\otimes H$, where $\rho(a)=a^0\otimes a^1$, then $A$ is a (symmetric) partial $H^{\ast}$-module algebra via $h^{\ast} \cdot a=h^{\ast}(a^1)a^0$.
\end{prop}
\subsection{q-combinatorics}
Let $n, m \in \mathbb{Z}, n \geq 0$, and $q \in \Bbbk^{\times}$. The \emph{q-binomial coefficients} are defined recursively as follows.
First, set ${0 \choose 0}_q = 1$ and ${n \choose m}_q = 0$ if $m>n$ or $m<0$.
Then, for $n \geq 1$ and $0 \leq m \leq n$,
\begin{align}\label{def_q_binomial}
{n \choose m}_q = {{n-1} \choose {m-1}}_q + q^{m} {{n-1} \choose m}_q.
\end{align}
It follows that ${n \choose 0}_q = 1 = {n \choose n}_q$ and ${n \choose m}_q = {n \choose {n-m}}_q$.
The \emph{q-numbers} and the \emph{q-factorials} also are defined recursively: $(0)_q = 0$, $(n)_q = \sum_{\ell=0}^{n-1}q^\ell;$ and
$(0)_q ! = 1$, $(n)_q ! = (n)_q (n-1)_q !,$ respectively.
Note that, if $q \neq 1$, then $q$ is an $n^{th}$ root of unity if and only if $(n)_q = 0$.
Suppose that $n \geq 1$ and $q \in \Bbbk^{\times}$. If $(n-1)_q! \neq 0$, then
\begin{align}\label{def_q_bino_fatorial}
{n \choose m}_q = \dfrac{(n)_q!}{(n-m)_q!(m)_q!},
\end{align}
for all $0 < m < n$.
In particular, if $q \neq 1$ is a primitive $n^{th}$ root of unity, then ${n \choose m}_q = 0,$
for all $0 < m < n$.
Suppose $q$ is a primitive $N^{th}$ root of unity. For $k \geq 0$, let $k_D$ and $k_R$ be the integers determined by $k =k_D \ N + k_R$ where $0 \leq k_R <N$.
Then
\begin{align}\label{radford}
{n \choose m}_q = {n_R \choose m_R}_q {n_D \choose m_D},
\end{align}
for all $0 \leq m \leq n$.
Furthermore, the following equalities are well-know:
\begin{align}
\sum_{k=0}^{m} (-1)^k {m \choose k}_q q^{\frac{k(k+1)}{2} - km}=0 \quad (m \geq 1),
\end{align}
\begin{align}\label{id_q}
{n \choose m}_q = q^{m(n-m)} { n \choose m}_{q^{-1}},
\end{align}
\begin{align} \label{6.7}
{n \choose m}_q = {{n-1} \choose m}_q + q^{n-m} {{n-1} \choose {m-1}}_q.
\end{align}
These definitions and equalities are classic for \emph{q-computations}, and one can find more details about them, for instance, in \cite{radford}.
However, to compute partial actions of Taft algebras on algebras, we are going to need some more.
We present below \emph{q-identities} that will be used to compute such partial actions in the following section.
\begin{lem}\cite[Lemmas 3.2, 3.3 and 3.9]{taft_corpo_revista} \label{prodqbinom} \label{lema1} \label{lema4}
Let $q \in \Bbbk^{\times}$ and $i,j,k,s, t \in \mathbb{N}_0$, such that $k \leq j$.
Then,
\begin{align} \label{Lema2_2}
&{ j \choose k }_q { {j-k} \choose {i-k} }_q = { j \choose i }_q { i \choose k }_q;\\
\label{Lema2_6}
&\sum_{\ell=0}^{j} (-1)^\ell q^{s \ell +\frac{\ell(\ell+1)}{2}} {j \choose \ell}_q {{j+t-\ell} \choose {j+s}}_q ={t \choose s}_q;\\
\label{Lema2_10}
&q^{s(i-j)} \sum_{\ell=0}^{j} {j \choose \ell}_q {{j+t-\ell} \choose {i+s-\ell}}_q {\ell \choose i}_q (-1)^{i-\ell} q^{\frac{(i-\ell)(i-\ell+1)}{2}}= {j \choose i}_q {t \choose s}_q.
\end{align}
\end{lem}
\begin{lem}
Let $t,\ell \in \mathbb{N}_0$ and $q \in \Bbbk^{\times}$.
Then,
\begin{itemize}
\item[(i)] if $0 \leq \ell < t$ and $(t-1)_q! \neq 0$, then
\begin{equation}\label{igualdade_estrela_simetria}
\left(1-q^{t - \ell}\right) { t \choose \ell}_q +q^{\ell+1} { t \choose \ell + 1}_q = { t \choose \ell + 1}_q;
\end{equation}
\item[(ii)] for each $m \in N_0$, it holds
\begin{equation} \label{lema_acao_da_acao}
\sum_{s=0}^{m} (-1)^{m-s} q^{\frac{s(s+1)}{2}-sm} {m \choose s}_q {{s+t} \choose {\ell}}_q = q^{\frac{m(m+1)}{2} + m(t-\ell)} {t \choose \ell-m}_q.
\end{equation}
\end{itemize}
\end{lem}
\begin{proof}
\begin{itemize}
\item[(i)] It is straightforward using that $ { t \choose \ell + 1}_q = \frac{(t-\ell)_q}{(\ell + 1)_q} { t \choose \ell}_q$ and \eqref{def_q_bino_fatorial}.
\item[(ii)] The proof is done by induction on $m$. The statement is clear for $m=0$. For $m > 0$, note that
\begin{align*}
& \sum_{s=0}^{m+1} (-1)^{m+1-s} q^{\frac{s(s+1)}{2}-s(m+1)} {m+1 \choose s}_q {{s+t} \choose {\ell}}_q \\
\stackrel{\eqref{def_q_binomial}}{=} & \sum_{s=1}^{m+1} (-1)^{m+1-s} q^{\frac{s(s+1)}{2}-s(m+1)} {m \choose s-1}_q {{s+t} \choose {\ell}}_q \\
- & \sum_{s=0}^{m} (-1)^{m-s} q^{\frac{s(s+1)}{2}-sm} {m \choose s}_q {{s+t} \choose {\ell}}_q\\
= & q^{-m}\sum_{s=0}^{m} (-1)^{m-s} q^{\frac{s(s+1)}{2}-sm} {m \choose s}_q {{s+(t+1)} \choose {\ell}}_q \\
- & \sum_{s=0}^{m} (-1)^{m-s} q^{\frac{s(s+1)}{2}-sm} {m \choose s}_q {{s+t} \choose {\ell}}_q.
\end{align*}
So, we conclude the induction step applying the induction hypothesis in both summations, and using equality \eqref{6.7}.
\end{itemize}
\end{proof}
\begin{lem}
Let $q\in \Bbbk \setminus\{1\}$ be a $n^{th}$ root of unity. Then,
\begin{itemize}
\item[(i)] for each $0 \leq i < n$, it holds that
\begin{equation} \label{igualdade_q_n}
(-1)^i q^{\frac{i(i+1)}{2}} { n-1 \choose i}_q = 1;
\end{equation}
\item[(ii)] if $q$ is a primitive root of unity, and $\ell, t \in \mathbb{N}_0$ such that $0 \leq \ell \leq t < n$, then it holds that
\begin{equation}\label{identidade_lema_simetria_n}
(-1)^\ell q^{t \ell - \frac{\ell(\ell-1)}{2}} { n+\ell-t-1 \choose \ell}_q = { t \choose \ell}_q.
\end{equation}
\end{itemize}
\end{lem}
\begin{proof} Both items are proved by induction. For item (i) it can be used equality \eqref{def_q_binomial}, while for item (ii) it is a computation using \eqref{def_q_bino_fatorial}.
\end{proof}
\begin{lem}\label{Lema_R_L_simetria}
Let $q\in \Bbbk$ be a primitive $n^{th}$ root of unity and $k, \ell, t \in \mathbb{N}_0$ such that $ k, \ell, t < n$. Then,
\begin{itemize}
\item[(i)] if $n \leq t+\ell$, then, for each $0 \leq m < t$, it holds that \begin{equation}\label{identidade_R_L}
\begin{split}
& \sum_{s=0}^{m}(-1)^s q^{\frac{s(s+1)}{2}-sm} { t \choose m-s}_q { s \choose t+\ell-n}_q \\
= & \ (-1)^m q^{\frac{-m(m-1)}{2}-t(t+\ell-m)} { m - t + n \choose \ell }_q;
\end{split}
\end{equation}
\item[(ii)] if $\ell \leq k$, then it holds that
\begin{equation}
\sum_{s=0}^{\ell}(-1)^s q^{\frac{s(s+1)}{2}-s(\ell-k)} { n+ \ell - k \choose s}_q = { k \choose \ell }_q;
\end{equation}
\item[(iii)] if $\ell \leq k$, then, for each $\ell \leq m < n$, it holds that
\begin{equation}\label{identidade_R_L_item}
\sum_{s=0}^{m}(-1)^s q^{\frac{s(s+1)}{2}-s(\ell-k)} { n+ m - k \choose s}_q { n + m - s \choose n+\ell-s}_q = { k \choose \ell }_q.
\end{equation}
\end{itemize}
\end{lem}
\begin{proof} We prove the statements by induction on the upper index of the summation symbol.
\begin{itemize}
\item[(i)] To prove this item, we consider two cases: $t+ \ell > n$ and $t + \ell =n$.
First, assume $t+\ell>n$.
For $m=0$, it is clear.
For the induction step, note that
\begin{align*}
& \sum_{s=0}^{m+1}(-1)^s q^{\frac{s(s+1)}{2}-s(m+1)} { t \choose m + 1-s}_q { s \choose t+\ell-n}_q \\
= \ & \ \sum_{s=0}^{m}(-1)^{s+1} q^{\frac{(s+1)(s+2)}{2}-(s+1)(m+1)} { t \choose m -s}_q { s + 1 \choose t+\ell-n}_q \\
\stackrel{\eqref{def_q_binomial}}{=} & \ -q^{-m}\sum_{s=0}^{m}(-1)^{s} q^{\frac{s(s+1)}{2}-sm} { t \choose m -s}_q { s \choose t+\ell-n-1}_q \\
- &\ q^{t+\ell-m}\sum_{s=0}^{m}(-1)^{s} q^{\frac{s(s+1)}{2}-sm} { t \choose m -s}_q { s \choose t+\ell-n}_q.
\end{align*}
Then, one conclude the induction step using the induction hypothesis in both summations and applying \eqref{def_q_binomial}.
Now, assume $t+\ell=n$.
Thus, identity \eqref{identidade_R_L} is rewritten as
\begin{align*}
\sum_{s=0}^{m}(-1)^s q^{\frac{s(s+1)}{2}-sm} { t \choose m-s}_q
= (-1)^m q^{mt - \frac{m(m-1)}{2}} { m - t + n \choose m }_q .
\end{align*}
Again, the case $m=0$ is clear. For the induction step, assume $m \geq 0$ and note that
\begin{align*}
& \sum_{s=0}^{m+1}(-1)^s q^{\frac{s(s+1)}{2}-s(m+1)} { t \choose m+1-s}_q \\
= \ & \ { t \choose m+1} - q^{-m} \sum_{s=0}^{m}(-1)^{s} q^{\frac{s(s+1)}{2}-sm} { t \choose m-s}_q \\
= \ & \ { t \choose m+1} - q^{-m} \left((-1)^m q^{mt- \frac{m(m-1)}{2}} { m - t + n \choose m }_q \right),
\end{align*}
where the induction hypothesis was used to obtain the last equality. Then, the result follows using equalities \eqref{identidade_lema_simetria_n} and \eqref{6.7}.
\item[(ii)] The statement is clear for $\ell=0$. Assume $\ell > 0$ and $\ell+ 1 \leq k$. Then,
\begin{align*}
& \sum_{s=0}^{\ell+1}(-1)^s q^{\frac{s(s+1)}{2}-s(\ell+1-k)} { n+ \ell +1 - k \choose s}_q \\
\stackrel{\eqref{def_q_binomial}}{=} & 1+ \sum_{s=1}^{\ell+1}(-1)^s q^{\frac{s(s+1)}{2}-s(\ell+1-k)} { n+ \ell - k \choose s -1}_q \\
+ & \sum_{s=1}^{\ell+1}(-1)^s q^{\frac{s(s+1)}{2}-s(\ell-k)} { n+ \ell - k \choose s}_q \\
= & \sum_{s=0}^{\ell}(-1)^{s+1} q^{\frac{(s+1)(s+2)}{2}-(s+1)(\ell+1-k)} { n+ \ell - k \choose s}_q \\
+ & \sum_{s=0}^{\ell}(-1)^s q^{\frac{s(s+1)}{2}-s(\ell-k)} { n+ \ell - k \choose s}_q \\
+ & (-1)^{\ell+1} q^{\frac{(\ell+1)(\ell+2)}{2}-(\ell+1)(\ell-k)} { n+ \ell - k \choose \ell + 1}_q\\
=& - q^{k - \ell} { k \choose \ell}_q + { k \choose \ell}_q +q^{\ell+1} { k \choose \ell + 1}_q.
\end{align*}
where the last equality follows from the induction hypothesis and equality \eqref{identidade_lema_simetria_n}. Hence, \eqref{igualdade_estrela_simetria} concludes the induction step.
\item[(iii)] The statement holds for $m= \ell$ by the previous item.
Now, for the induction hypothesis, note that
\begin{align*}
& \sum_{s=0}^{m+1}(-1)^s q^{\frac{s(s+1)}{2}-s(\ell-k)} { n+ m + 1 - k \choose s}_q { n + m + 1 - s \choose n+\ell-s}_q \\
= & \sum_{s=0}^{m}(-1)^s q^{\frac{s(s+1)}{2}-s(\ell-k)} { n+ m + 1 - k \choose s}_q { n + m + 1 - s \choose n+\ell-s}_q,
\end{align*}
since ${ n \choose n + \ell - m - 1}_q=0$, because $0<n+\ell -m -1 <n$.
Then, by \eqref{def_q_binomial} we obtain the result.
\end{itemize}
\end{proof}
\section{Partial (co)actions of the Taft Algebra}\label{sec:Taft}
\noindent{\bf Taft Algebra.}
Let $n \geq 2$ be an integer and suppose that $\Bbbk$ contains a primitive $n^{th}$ root of unity $q$.
In particular $char(\Bbbk) \nmid n$.
The \emph{Taft algebra of order $n$}, or shortly \emph{Taft algebra}, here denoted by $T_n(q)$, has the following structure:
as algebra it is generated over $\Bbbk$ by two elements $g$ and $x$ with relations $g^n=1,$ $x^n =0$ and $xg = q gx$.
Thus, the set $\{g^ix^j \ : \ 0 \leq i, j <n \}$ is the canonical basis for $T_n(q)$ and so $dim_\Bbbk(T_n(q)) = n^{2}$.
The coalgebra structure of $T_n(q)$ is induced by setting $g$ a group-like element and $x$ an $(1,g)$-primitive element, that is, $\Delta (g) = g \otimes g$, $\Delta (x) = x \otimes 1 + g \otimes x$, $\varepsilon (g) = 1$ and $\varepsilon (x) = 0$.
In general, the comultiplication map is $\Delta (g^ix^j) = \sum_{\ell=0}^j {j \choose \ell}_q g^{i+\ell}x^{j-\ell} \otimes g^ix^\ell$, for $i, j \in \I_{0, n-1}$.
To complete the Hopf algebra structure of $T_n(q)$, the antipode map $S$ is defined by $S(g) =g^{n-1}$ and $S(x) = -g^{n-1}x.$
Furthermore, note that the group of group-like elements is $G( T_n(q) ) = \{ 1, g, \cdots, g^{n-1} \}$, \emph{i.e.}, $ G( T_n(q) ) = C_n.$
\subsection{Partial actions}
In this section, we characterize suitable partial actions of $T_n(q)$ on an algebra $A$. If $A$ has only trivial idempotent elements, these are all partial actions.
\begin{lem}\label{idempotente_comuta}
Let $ \cdot : H \otimes A \longrightarrow A$ be a partial action of $H$ on $A$. For each $h \in G(H)$ and $a \in A$:
\begin{itemize}
\item[(i)] $(h \cdot 1_A)a(h \cdot 1_A) = (h \cdot 1_A)a$;
\item[(ii)] if the partial action $\cdot|_{\Bbbk G(H)\otimes A}$ is symmetric, then $(h \cdot 1_A)a=a(h \cdot 1_A)$;
\item[(iii)] if $h \cdot 1_A =0$, then $(h^i \cdot 1_A)(h^{i+1} \cdot 1_A)=0$, for all $i \in \mathbb{Z}$.
\end{itemize}
\end{lem}
\begin{proof}
Assume $h \in G(H)$ and $a \in A$.
\begin{itemize}
\item[(i)] First,
$
(h \cdot 1_A)(h \cdot (h^{-1} \cdot a)) \stackrel{(PA.2)}{=} (h \cdot 1_A)(h \cdot 1_A)(1_H \cdot a) = (h \cdot 1_A)a.
$
Then, \begin{align*}
(h \cdot 1_A)a(h \cdot 1_A) & = (h \cdot 1_A)(h \cdot (h^{-1} \cdot a))(h \cdot 1_A) \\
&\!\!\!\!\! \stackrel{(PA.2)}{=} (h \cdot 1_A)(h \cdot ((h^{-1} \cdot a) 1_A)) = (h \cdot 1_A)a.
\end{align*}
\item[(ii)] Analogously to the previous item.
\item[(iii)] For each $i \in \mathbb{Z}$, $(h^i \cdot 1_A)(h^{i+1} \cdot 1_A) \stackrel{(PA.3)}{=} h^i \cdot (h \cdot 1_A) = h^i \cdot 0 = 0.$
\end{itemize}
\end{proof}
\begin{prop}\label{parte_global}
Let $\cdot : T_n(q) \otimes A \longrightarrow A$ be a partial action of $T_n(q)$ on $A$.
If $g \cdot 1_A = 1_A$, then $\cdot$ is a global action.
\end{prop}
\begin{proof}
Suppose that $g \cdot 1_A = 1_A$.
It is enough to prove that $g \cdot (h \cdot a) = gh \cdot a$ and $x \cdot (h \cdot a) = xh \cdot a$, for any $h \in T_n(q)$ and $a \in A$.
First, by (PA.2), we get
$$x \cdot 1_A = x \cdot 1_A^2 = (x \cdot 1_A)(1 \cdot 1_A) + (g \cdot 1_A)(x \cdot 1_A)= 2 (x \cdot 1_A),$$
that is, $x \cdot 1_A = 0$. Then, by (PA.3),
$$x \cdot (h \cdot a) = (x \cdot 1_A)(h \cdot a) + (g \cdot 1_A)(xh \cdot a) = xh \cdot a.$$
Also by (PA.3), it follows that
$g \cdot (h \cdot a) = (g \cdot 1_A)(gh \cdot a) = gh \cdot a.$
\end{proof}
\begin{prop}\label{formula_eh_acao_parcial}
Let $\cdot : T_n(q) \otimes A \longrightarrow A$ be a partial action of $T_n(q)$ on $A$.
If $g \cdot 1_A = 0$, then
\begin{equation}\label{formula_3_1}
g^ix^j \cdot a = q^{-ij} \sum_{k=0}^j (-1)^k q^{- \frac{k(k-1)}{2}} { j \choose k }_{q^{-1}} (x \cdot 1_A)^{j-k}(g^{i+k} \cdot a)(x \cdot 1_A)^k,
\end{equation}
for any $a \in A$ and $i,j \in\I_{0, n-1}.$
\end{prop}
\begin{proof}
First, we note that $g^{-1}x \cdot a = -q a (x \cdot 1_A)$.
Indeed, by (PA.1) and (PA.3), we get $g^{-1} \cdot (g \cdot 1_A) = (g^{-1} \cdot 1_A)(1 \cdot 1_A)$, that is, $g^{-1} \cdot 1_A =0$.
Again, by (PA.1) and (PA.3), we conclude that
$$ g^{-1}x \cdot (g \cdot 1_A) = (g^{-1}x \cdot 1_A)(1 \cdot 1_A) + (1 \cdot 1_A)(g^{-1}xg \cdot 1_A) = (g^{-1}x \cdot 1_A) + q (x \cdot 1_A),$$
that is, $g^{-1}x \cdot 1_A = -q (x \cdot 1_A).$
Then, by (PA.1) and (PA.2), we obtain
$$g^{-1}x \cdot a = (g^{-1}x \cdot a)(g^{-1} \cdot 1_A) + (1 \cdot a)(g^{-1}x \cdot 1_A) = -q a (x \cdot 1_A).$$
Now, we prove the desired equality by induction on $j$. It is immediate for $j=0$, and if one wants to check for $j=1$ it follows by (PA.3) in $g^{-1}x \cdot (g^{i+1} \cdot a)$. Assume $j \geq 0$.
Then, by (PA.3), we get
\begin{align*}
g^{-1}x \cdot (g^{i+1}x^j \cdot a) = (g^{-1}x \cdot 1_A)(g^{-1}g^{i+1}x^j \cdot a)+ (1 \cdot 1_A)(g^{-1}xg^{i+1}x^{j} \cdot a),
\end{align*}
that is, $-q (g^{i+1}x^j \cdot a)(x \cdot 1_A) = -q (x \cdot 1_A)(g^ix^j \cdot a) + q^{i+1} (g^ix^{j+1} \cdot a),$
and thus $q^i\left(g^ix^{j+1} \cdot a \right) = (x \cdot 1_A)(g^{i}x^j \cdot a)-( g^{i+1}x^j \cdot a)(x \cdot 1_A).$
Using the induction hypothesis, we obtain
\begin{align*}
& q^i (g^{i}x^{j+1} \cdot a ) \\
= & (x \cdot 1_A) \left( q^{-ij} \sum_{k=0}^j (-1)^k q^{- \frac{k(k-1)}{2}} { j \choose k }_{q^{-1}} (x \cdot 1_A)^{j-k}(g^{i+k} \cdot a)(x \cdot 1_A)^k \right) \\
- & \left(\!\! q^{-(i+1)j} \sum_{k=0}^j (-1)^k q^{- \frac{k(k-1)}{2}} { j \choose k }_{q^{-1}} \!\!\!\!\!\!\!\!\!(x \cdot 1_A)^{j-k}(g^{i+1+k} \cdot a)(x \cdot 1_A)^k \!\!\right)\!\!(x \cdot 1_A) \\
= & \ q^{-ij}\Big( (x \cdot 1_A)^{j+1}(g^i \cdot 1_A) \\
+ & \sum_{k=1}^j (-1)^k q^{- \frac{k(k-1)}{2}} { j \choose k }_{q^{-1}} (x \cdot 1_A)^{j+1-k}(g^{i+k} \cdot a)(x \cdot 1_A)^k \\
- & q^{-j} \sum_{k=0}^{j-1} (-1)^k q^{- \frac{k(k-1)}{2}} { j \choose k }_{q^{-1}} (x \cdot 1_A)^{j-k}(g^{i+1+k} \cdot a)(x \cdot 1_A)^{k+1} \\
- & q^{-j} (-1)^j q^{-\frac{j(j-1)}{2}} (g^{i+1+j} \cdot a) (x \cdot 1_A)^{j+1} \Big) \\
= & \ q^{-ij}\Big( (x \cdot 1_A)^{j+1}(g^i \cdot 1_A) \\
+ & \sum_{k=1}^j (-1)^k q^{- \frac{k(k-1)}{2}} \left({ j \choose k }_{q^{-1}} + q^{-j+(k-1)} { j \choose k-1 }_{q^{-1}} \right) \\
\times & (x \cdot 1_A)^{j+1-k}(g^{i+k} \cdot a)(x \cdot 1_A)^k + (-1)^{j+1} q^{-\frac{j(j+1)}{2}} (g^{i+1+j} \cdot a) (x \cdot 1_A)^{j+1} \Big) \\
\stackrel{\eqref{6.7}}{=} & q^{-ij}\Big( (x \cdot 1_A)^{j+1}(g^i \cdot 1_A) \\
+ & \sum_{k=1}^j (-1)^k q^{- \frac{k(k-1)}{2}} { j + 1 \choose k }_{q^{-1}} (x \cdot 1_A)^{j+1-k}(g^{i+k} \cdot a)(x \cdot 1_A)^k \\
+ & (-1)^{j+1} q^{-\frac{j(j+1)}{2}} (g^{i+1+j} \cdot a) (x \cdot 1_A)^{j+1} \Big)
\end{align*}
that is,
\begin{align*}
& g^{i}x^{j+1} \cdot 1_A \\
= & q^{-i(j+1)}\sum_{k=0}^{j+1} (-1)^k q^{- \frac{k(k-1)}{2}} { j + 1 \choose k }_{q^{-1}} (x \cdot 1_A)^{j+1-k}(g^{i+k} \cdot a)(x \cdot 1_A)^k.
\end{align*}
It concludes the induction step.
\end{proof}
\begin{cor}\label{corolario_ida}
Let $\cdot : T_n(q) \otimes A \longrightarrow A$ be a partial action of $T_n(q)$ on $A$.
If $g \cdot 1_A = 0$, then
\begin{itemize}
\item[(i)] $(x\cdot 1_A)^n\in Z(A)$;
\item[(ii)] $g^i \cdot (x\cdot 1_A) =q^{-i}(g^i \cdot 1_A)(x\cdot 1_A)$, for all $ i\in\I_{0, n-1}$;
\item[(iii)] If there is $ i\in\I_{2, n-1}$ such that $g^i\cdot a=a$, for all $a \in A$, then $x\cdot 1_A=0.$
\end{itemize}
\end{cor}
\begin{proof}
\begin{itemize}
\item[(i)] By definition of the partial action, we have
$$g^{n-1}x\cdot(gx^{n-1}\cdot a)=(g^{n-1}x\cdot 1_A)(x^{n-1}\cdot a)$$
As seen in the proof of Proposition \ref{formula_eh_acao_parcial}, it holds that $g^{n-1}x\cdot a=-qa(x\cdot 1_A)$. Then, the above equality can be rewritten as
$$(gx^{n-1}\cdot a)(x\cdot 1_A)=(x\cdot 1_A)(x^{n-1}\cdot a).$$
Then, by \eqref{formula_3_1}, we obtain
\begin{eqnarray*}
&& q\sum_{k=0}^{n-1} (-1)^k q^{- \frac{k(k-1)}{2}} { n-1 \choose k }_{q^{-1}} (x \cdot 1_A)^{n-1-k}(g^{k+1} \cdot a)(x \cdot 1_A)^{k+1}\\
&=& \sum_{k=0}^{n-1} (-1)^k q^{- \frac{k(k-1)}{2}} { n-1 \choose k }_{q^{-1}} (x \cdot 1_A)^{n-k}(g^{k} \cdot a)(x \cdot 1_A)^k.
\end{eqnarray*}
Now, since $g \cdot 1_A =0$, and so $g \cdot a= g^{-1} \cdot a=0$, for all $a \in A$,
\begin{eqnarray*}
&&\sum_{k=1}^{n-3} (-1)^k q^{- \frac{k(k-1)}{2}} q { n-1 \choose k }_{q^{-1}} (x \cdot 1_A)^{n-(k+1)}(g^{k+1} \cdot a)(x \cdot 1_A)^{k+1}\\
&+& (-1)^{n-1} q^{- \frac{(n-1)(n-2)}{2}} q { n-1 \choose n-1 }_{q^{-1}} a(x\cdot 1_A)^n\\
&=& (x\cdot 1_A)^n a+\sum_{k=2}^{n-2} (-1)^k q^{- \frac{k(k-1)}{2}} { n-1 \choose k }_{q^{-1}} (x \cdot 1_A)^{n-k}(g^{k} \cdot a)(x \cdot 1_A)^k.
\end{eqnarray*}
Then, with a suitable change in the summation index,
\begin{eqnarray*}
&&\sum_{k=2}^{n-2} -(-1)^k q^{- \frac{k(k-1)}{2}} q^k { n-1 \choose k-1 }_{q^{-1}} (x \cdot 1_A)^{n-k}(g^{k} \cdot a)(x \cdot 1_A)^{k}\\
&+& (-1)^{n-1} q^{- \frac{n(n-1)}{2}} { n-1 \choose n-1 }_{q^{-1}} a(x\cdot 1_A)^n\\
&=& (x\cdot 1_A)^n a+\sum_{k=2}^{n-2} (-1)^k q^{- \frac{k(k-1)}{2}} { n-1 \choose k }_{q^{-1}} (x \cdot 1_A)^{n-k}(g^{k} \cdot a)(x \cdot 1_A)^k.
\end{eqnarray*}
At last, using \eqref{igualdade_q_n} and ${ n-1 \choose k }_{q^{-1}}=-q^k{ n-1 \choose k-1 }_{q^{-1}}$, we conclude that
$a(x\cdot 1_A)^n=(x\cdot 1_A)^na$, for all $a\in A$.
\item[(ii)]
Since $g^i \in G(T_n(q))$ and $g \cdot 1_A=0$, we have
\begin{align*}
g^i \cdot (x\cdot 1_A) = & \ (g^i\cdot 1_A)(g^ix\cdot 1_A) \\ \stackrel{\eqref{formula_3_1}}{=} & \ q^{-i}(g^i \cdot 1_A)((x\cdot 1_A)(g^i \cdot 1_A)-(g^{i+1} \cdot 1_A)(x\cdot 1_A)) \\
= & \ q^{-i}((g^i \cdot 1_A)(x\cdot 1_A)(g^i \cdot 1_A)-(g^i \cdot 1_A)(g^{i+1} \cdot 1_A)(x\cdot 1_A))\\
\stackrel{ \ref{idempotente_comuta} (iii)}{=} & \ q^{-i}(g^i \cdot 1_A)(x\cdot 1_A)(g^i \cdot 1_A) \\
\stackrel{ \ref{idempotente_comuta}(i)}{=} & q^{-i}(g^i \cdot 1_A)(x\cdot 1_A).
\end{align*}
\item[(iii)] In that case, by the previous item, we conclude that $(x\cdot 1_A)=q^{-i}(x\cdot 1_A)$, for some $i\in\I_{2, n-1}$, and therefore $x\cdot 1_A=0$.
\end{itemize}
\end{proof}
\begin{lem}\label{lema_ind}
Let $\cdot : \Bbbk C_n \otimes A \longrightarrow A$ be a partial action of $\Bbbk C_n$ on $A$.
Then, for each $w \in A$ such that
$g^i \cdot w =q^{-i}(g^i \cdot 1_A)w$,
for all $ i\in\I_{0, n-1}$, the identity
$$ ( g^i \cdot w)^\ell = q^{-i\ell}(g^i \cdot 1_A)w^\ell$$
holds for all $\ell \geq 1$. Moreover, if the partial action is symmetric,
then $$ ( g^i \cdot w)^\ell = q^{-i\ell}w^\ell(g^i \cdot 1_A).$$
\end{lem}
\begin{proof}
The result follows by induction on $\ell$ using Lemma \ref{idempotente_comuta} (i).
\end{proof}
From the above lemma we have the following consequence.
\begin{cor}\label{cor_condw}
Let $\cdot : \Bbbk C_n \otimes A \longrightarrow A$ be a partial action of $\Bbbk C_n$ on $A$. Then, for each $w \in A$ such that
$g^i \cdot w =q^{-i}(g^i \cdot 1_A)w$,
for all $i,j \in \I_{0,n-1}$, the identity
$$ g^i \cdot (w^\ell (g^j \cdot a) w^k) = q^{-i(\ell + k)}(g^i \cdot 1_A)w^\ell(g^{i+j} \cdot a) w^k$$
holds for all $a \in A$, $\ell, k \geq 0$.
In particular, $ g^i \cdot w^k = q^{-ik}(g^i \cdot 1) w^k$.
Moreover, if the partial action is symmetric, then $$ g^i \cdot (w^\ell (g^j \cdot a) w^k) = q^{-i(\ell + k)}w^\ell(g^{i+j} \cdot a) w^k(g^i \cdot 1_A).$$
\end{cor}
\begin{proof}
First, since $g^i$ is a group-like element, then $g^i \cdot a^{\ell} = (g^i \cdot a)^{\ell}$ for all $a \in A$. Now, notice that for each $w \in A$ such that $g^i \cdot w =q^{-i}(g^i \cdot 1_A)w$,
\begin{eqnarray*}
g^i \cdot (w^\ell (g^j \cdot a) w^k)&=& (g^i \cdot w)^{\ell} (g^i \cdot 1_A)(g^{i+j} \cdot a)(g^i \cdot w)^{k}\\
&\stackrel{\ref{lema_ind}}{=} & q^{-i\ell}(g^i \cdot 1_A) w^{\ell}(g^i \cdot 1_A)(g^{i+j} \cdot a) q^{-ik}(g^i \cdot 1_A) w^{k}\\
&\stackrel{\ref{idempotente_comuta}(i)}{=}& q^{-i(\ell + k)} (g^i \cdot 1_A) w^\ell(g^{i+j} \cdot a) w^k,
\end{eqnarray*}
for all $a \in A$, $\ell, k \geq 0$.
The symmetric part is analogous.
\end{proof}
\begin{prop}\label{prop_volta}
Let $\cdot : \Bbbk C_n \otimes A \longrightarrow A$ be a partial action of $\Bbbk C_n$ on $A$, and suppose $g \cdot 1_A = 0$.
Then, for each $w \in A$ such that $w^n \in Z(A)$ and $g^i \cdot w = q^{-i}(g^i \cdot 1_A)w$, for all $i \in \I_{0,n-1}$, the linear map $\cdot : T_n(q) \otimes A \longrightarrow A$, defined as
\begin{align}\label{formula}
g^ix^j \cdot a = q^{-ij} \sum_{k=0}^j (-1)^k q^{- \frac{k(k-1)}{2}} { j \choose k }_{q^{-1}} w^{j-k}(g^{i+k} \cdot a) w^k,
\end{align}
is a partial action of $T_n(q)$ on $A$.
\end{prop}
\begin{proof} Since (PA.1) clearly holds, it is sufficient to check condition (PA.2$^{\prime}$):
$$g^ix^j \cdot (a(g^t x^s \cdot b)) = ((g^ix^j)_1 \cdot a)((g^ix^j)_2 g^t x^s \cdot b).$$
We will proceed by fixing the vectors and checking the scalars. On one hand,
\begin{align*}
& g^ix^j \cdot (a(g^tx^s \cdot b)) = q^{-ij} \sum_{k=0}^j (-1)^{k} q^{\frac{-k(k-1)}{2}} {j \choose k}_{q^{-1}} w^{j-k} g^{i+k}\cdot(a(g^tx^s \cdot b)) w^{k}\\
& = q^{-ij-st} \sum_{k=0}^j\sum_{\ell=0}^s (-1)^{k+\ell} q^{- \frac{k(k-1)}{2}}q^{\frac{-\ell(\ell-1)}{2}} {j \choose k}_{q^{-1}}{s \choose \ell }_{q^{-1}} w^{j-k} (g^{i+k}\cdot a)\\
& \times (g^{i+k}\cdot(w^{s-\ell}(g^{t+\ell} \cdot b) w^\ell))w^{k}\\
& \stackrel{\ref{cor_condw}}{=} q^{-ij-s(t+i)} \sum_{k=0}^j\sum_{\ell=0}^s (-1)^{k+\ell} q^{-sk} q^{- \frac{k(k-1)}{2}} q^{\frac{-\ell(\ell-1)}{2}} {j \choose k}_{q^{-1}}{s \choose \ell }_{q^{-1}}\\
& \times w^{j-k} (g^{i+k}\cdot a) w^{s-\ell}(g^{i+t+k+\ell} \cdot b) w^{k+\ell}.
\end{align*}
Let $ k_0 \in\I_{0, j}$ and $ \ell_0\in\I_{0, s}$ be fixed elements, thus we obtain the vector
\begin{equation}\label{eq_vetor}
w^{j-k_0} (g^{i+k_0} \cdot a) w^{s-\ell_0} (g^{i+t+k_0+\ell_0} \cdot b) w^{k_0+\ell_0}
\end{equation} with scalar
\begin{equation}\label{eq_lado1}
q^{-ij-s(t+i)} \left( (-1)^{k_0+\ell_0} q^{-sk_0} q^{- \frac{k_0(k_0-1)}{2}} q^{\frac{-\ell_0(\ell_0-1)}{2}} {j \choose k_0}_{q^{-1}}{s \choose \ell_0}_{q^{-1}} \right).
\end{equation}
On the other hand,
\begin{align*}
& ((g^ix^j)_1 \cdot a)((g^ix^j)_2g^tx^s \cdot b)
= \ \sum_{\ell =0}^j {j \choose \ell}_q (g^{i+\ell}x^{j-\ell} \cdot a) ( g^ix^\ell g^t x^s \cdot b) \\
\stackrel{\ref{id_q}}{=} & \ \sum_{\ell=0}^j q^{\ell(j-\ell)} q^{\ell t} {j \choose \ell}_{q^{-1}} (g^{i+\ell}x^{j-\ell} \cdot a) ( g^{i+t}x^{\ell + s} \cdot b).
\end{align*}
Due to term $g^{i+t}x^{\ell+s} \cdot b$, we have to proceed by cases:
\underline{Case 1: $j+s < n$.}
\begin{align*}
& \sum_{\ell=0}^j q^{\ell(j-\ell)} q^{\ell t} {j \choose \ell}_{q^{-1}} (g^{i+\ell}x^{j-\ell} \cdot a) ( g^{i+t}x^{\ell + s} \cdot b)\\
= & \sum_{\ell=0}^j \sum_{r=0}^{j-l} \sum_{\theta=0}^{\ell+s} q^{-(i+\ell)(j-\ell)} q^{-(i+t)(\ell+s)}q^{\ell(j-\ell)} q^{\ell t} (-1)^{r+\theta} q^{\frac{-r(r-1)}{2}} q^{\frac{-\theta(\theta-1)}{2}} \\
\times & {j \choose \ell}_{q^{-1}} \!\!{j-\ell \choose r}_{q^{-1}} \!\!{\ell+s \choose \theta}_{q^{-1}} \!\!\!\!\!\!\! w^{j-(\ell+r)} (g^{i+\ell+r}\cdot a) w^{r+\ell+s-\theta} (g^{i+t+\theta}\cdot b)w^{\theta}\\
= & \ q^{-ij-s(t+i)} \sum_{\ell=0}^j \sum_{r=0}^{j-l} \sum_{\theta=0}^{\ell+s}(-1)^{r+\theta} q^{\frac{-r(r-1)}{2}} q^{\frac{-\theta(\theta-1)}{2}} {j \choose \ell}_{q^{-1}} {j-\ell \choose r}_{q^{-1}} \\
\times & {\ell+s \choose \theta}_{q^{-1}} w^{j-(\ell+r)} (g^{i+\ell+r}\cdot a) w^{r+\ell+s-\theta} (g^{i+t+\theta}\cdot b)w^{\theta}\\
\stackrel{k=\ell+r}{=} & q^{-ij-s(t+i)} \sum_{\ell=0}^j \sum_{k=\ell}^{j} \sum_{\theta=0}^{\ell+s}(-1)^{k-\ell+\theta} q^{\frac{-(k-\ell)(k-\ell-1)}{2}} q^{\frac{-\theta(\theta-1)}{2}} {j \choose \ell}_{q^{-1}} \\
\times & {j-\ell \choose k-\ell}_{q^{-1}} {\ell+s \choose \theta}_{q^{-1}} w^{j-k} (g^{i+k}\cdot a) w^{k+s-\theta} (g^{i+t+\theta}\cdot b)w^{\theta}.
\end{align*}
Setting $0\leq k_0\leq j$, we have that each $0 \leq \ell \leq j$ determines the vector in \eqref{eq_vetor}, considering $k=k_0$ and $\theta=k_0+\ell_0$, with scalar
\begin{align*}
& q^{-ij-s(t+i)} \left(\sum_{\ell=0}^{k_0} (-1)^{\ell_0-\ell} q^{\frac{-(k_0-\ell)(k_0-\ell-1)}{2}} q^{\frac{-(k_0+\ell_0)(k_0+\ell_0-1)}{2}} \right. \\
\times & \left. {j \choose \ell}_{q^{-1}} {j-\ell \choose k_0-\ell}_{q^{-1}} {\ell+s \choose k_0+\ell_0}_{q^{-1}} \right) \\
\stackrel{\ref{prodqbinom}}{=} & \ \ q^{-ij-s(t+i)} q^{\frac{-k_0(k_0-1)}{2}} q^{\frac{-\ell_0(\ell_0-1)}{2}} (-1)^{\ell_0} {j \choose k_0}_{q^{-1}}\\
\times & \left (\sum_{\ell=0}^{k_0} (-1)^{\ell} q^{\frac{-k_0(k_0-1)}{2}}q^{\frac{-\ell(\ell-1)}{2}} q^{\frac{-k_0(k_0-1)}{2}} q^{\ell k_0} q^{-\ell_0k_0} {k_0 \choose \ell}_{q^{-1}} {\ell+s \choose k_0+\ell_0}_{q^{-1}} \right).
\end{align*}
Therefore, looking to the highlighted scalar in \eqref{eq_lado1}, we just need to see that
\begin{align*}
& \sum_{\ell=0}^{k_0} (-1)^{k_0-\ell} q^{\frac{-\ell(\ell-1)}{2}+\ell k_0} q^{\frac{-k_0(k_0-1)}{2}} q^{-\ell_0k_0} {k_0 \choose \ell}_{q^{-1}} {\ell+s \choose k_0+\ell_0}_{q^{-1}} \\
=& \ q^{-sk_0} {j \choose k_0}_{q^{-1}}{s \choose \ell_0}_{q^{-1}},
\end{align*}
but it holds by Lemma \ref{lema_acao_da_acao}.
\underline{Case 2: $j+s \geq n$.}
\begin{align*}
& \sum_{\ell=0}^j q^{\ell(j-\ell)} q^{\ell t} {j \choose \ell}_{q^{-1}} (g^{i+\ell}x^{j-\ell} \cdot a) ( g^{i+t}x^{\ell + s} \cdot b)\\
= & \ q^{-ij-s(t+i)} \sum_{\ell =0}^{n-s-1} \sum_{r=\ell}^{j} \sum_{\theta=0}^{r+s} (-1)^{r - \ell+ \theta} q^{\frac{-(r - \ell)(r - \ell-1)}{2}-\frac{\theta(\theta-1)}{2}} {j \choose \ell}_{q^{-1}} {j-\ell \choose r - \ell}_{q^{-1}} \\
\times & {\ell+ s \choose \theta}_{q^{-1}} w^{j- r} (g^{i+r} \cdot a) w^{r+s-\theta} (g^{i+t+\theta} \cdot b) w^{\theta}.
\end{align*}
Note that, since $0 \leq \ell \leq n-s-1$, we can change the upper bound of the summation in $\theta$ for $n-1$ instead $r+s$, due to ${\ell+ s \choose \theta}_{q^{-1}}$:
\begin{align*}
& q^{-ij-s(t+i)} \sum_{\ell =0}^{n-s-1} \sum_{r=\ell}^{j} \sum_{\theta=0}^{n-1} (-1)^{r - \ell+ \theta} q^{\frac{-(r - \ell)(r - \ell-1)}{2}-\frac{\theta(\theta-1)}{2}} {j \choose \ell}_{q^{-1}} {j-\ell \choose r - \ell}_{q^{-1}} \\
\times & {\ell+ s \choose \theta}_{q^{-1}} w^{j- r} (g^{i+r} \cdot a) w^{r+s-\theta} (g^{i+t+\theta} \cdot b) w^{\theta}.
\end{align*}
Now, we will investigate when the vector highlighted in \eqref{eq_vetor} appears in the above expression. We can assume $k_0+\ell_0 \geq n$, since $k_0+\ell_0 < n$ is included in Case 1. Thus, $k_0+s \geq k_0 + \ell_0 \geq n$, what leads to $0 \leq n-s-1 < k_0$. So, to obtain the referred vector, there is a unique choice for $r$, and so only one for $\theta$ too, namely $r=k_0$ and $\theta= k_0+\ell_0-n$ (where the latter uses \eqref{radford}). Hence,
\begin{align*}
& q^{-ij-s(t+i)} \sum_{\ell =0}^{n-s-1} (-1)^{\ell_0-n-\ell} q^{\frac{-(k_0 - \ell)(k_0 - \ell-1)}{2}-\frac{(k_0 + \ell_0-n)(k_0 + \ell_0-n-1)}{2}} \\
\times & {j \choose \ell}_{q^{-1}} {j-\ell \choose k_0 - \ell}_{q^{-1}} {\ell+ s \choose k_0 + \ell_0-n}_{q^{-1}}\\
\times & \ w^{j- k_0} (g^{i+k_0} \cdot a) w^{s-\ell_0+n} (g^{i+t+k_0+\ell_0-n} \cdot b) w^{k_0+\ell_0-n}.
\end{align*}
The vector in \eqref{eq_vetor} and the one in the above expression are the same due the commutative of $w^n$. Therefore, comparing the above scalar with the one highlighted in \eqref{eq_lado1}, we want to see that the following equality is true
\begin{align*}
& (-1)^{k_0+\ell_0} q^{-sk_0} q^{\frac{-k_0(k_0-1)}{2}-\frac{\ell_0(\ell_0-1)}{2}} {j \choose k_0}_{q^{-1}} {s \choose \ell_0}_{q^{-1}} = \sum_{\ell =0}^{n-s-1} (-1)^{\ell_0-n-\ell}\\
\times & q^{\frac{-(k_0 - \ell)(k_0 - \ell-1)}{2}-\frac{(k_0 + \ell_0-n)(k_0 + \ell_0-n-1)}{2}} {j \choose \ell}_{q^{-1}} {j-\ell \choose k_0 - \ell}_{q^{-1}} {\ell+ s \choose k_0 + \ell_0-n}_{q^{-1}}.
\end{align*}
But, applying identities \eqref{Lema2_2}, \eqref{igualdade_q_n} and \eqref{lema_acao_da_acao} to the above expression, it is reduced to
\begin{equation}\label{eq_final}
q^{\frac{k_0(k_0-1)}{2}+k_0(\ell_0-s)}\! {s \choose \ell_
0}_{q^{-1}} \!\!\!\!=\!\!\!\sum_{\ell =n-s}^{k_0} \!\!\!(-1)^{k_0-\ell} q^{\frac{-\ell(\ell+1)}{2}+\ell k_0} {k_0 \choose \ell}_{q^{-1}} \!\!{\ell+ s-n \choose k_0 + \ell_0-n}_{q^{-1}}.
\end{equation}
Since $n-s \leq \ell \leq k_0$ and $0 \leq \ell-n+s \leq k_0-n+s$, the right side in above expression can be rewritten replacing the variable $\ell$ to $L+n-s$ as
\begin{align*}
\sum_{L =0}^{k_0-n+s} \!\! (-1)^{k_0-L-n+s} q^{\frac{-(L+n-s)(L+n-s+1-2k_0)}{2}} {k_0 \choose L+n-s}_{q^{-1}} {L \choose k_0 + \ell_0-n}_{q^{-1}}\\
\stackrel{\eqref{igualdade_q_n}}{=} \!\!\sum_{L =0}^{k_0-n+s} \!\! (-1)^{k_0-L+s+1} q^{\frac{-(L-s)(L-s+1-2k_0)}{2}} {k_0 \choose L+n-s}_{q^{-1}} {L \choose k_0 + \ell_0-n}_{q^{-1}}.
\end{align*}
Therefore, \eqref{eq_final} follows from identity \eqref{identidade_R_L}.
\end{proof}
\begin{rem}
Notice, in Proposition \ref{prop_volta}, that $x\cdot 1_A=w$, $x\cdot a=wa$, and $g^ix^j \cdot a$ is determined by the partial group action and this fixed element $w$. Hence, a partial action $T_n(q)$ on an algebra $A$ depends only on a partial group action, an element $w \in A$, and a compatibility relation between them.
\end{rem}
\begin{thm}\label{taft_principal}
Let $\cdot:T_n(q) \otimes A \longrightarrow A$ be a linear map, and suppose $g \cdot 1_A \in \{0, 1_A\}$.
Then, $\cdot$ is a partial action of $T_n(q)$ on $A$ if and only if $\cdot$ is a global action, or $\cdot$ is a linear map such that $\cdot|_{\Bbbk C_n \otimes A}$
is a partial group action such that $g \cdot 1_A = 0$ and
\begin{itemize}
\item[(i)] $ g^ix^j \cdot a = q^{-ij} \sum_{k=0}^j (-1)^k q^{- \frac{k(k-1)}{2}} { j \choose k }_{q^{-1}} (x \cdot 1_A)^{j-k}(g^{i+k} \cdot a)(x \cdot 1_A)^k;$
\item[(ii)] $(x \cdot 1_A)^n \in Z(A)$;
\item[(iii)] $g^i \cdot (x \cdot 1_A) = q^{-i}(g^i \cdot 1_A)(x \cdot 1_A)$,
\end{itemize}
for all $0 \leq i,j <n, a \in A$.
\end{thm}
\begin{proof}
One part follows by Propositions \ref{parte_global}, \ref{formula_eh_acao_parcial} and Corollary \ref{corolario_ida}, and the converse follows by Proposition \ref{prop_volta} .
\end{proof}
\begin{cor}\label{idempotente_trivial}
Suppose that $A$ is an algebra that has only the trivial idempotents. Then, Theorem \ref{taft_principal} characterizes all partial actions of $T_n(q)$ on A.
\end{cor}
From the above result we have a complete characterization for partial actions of $T_n(q)$ on some families of algebras, such as division algebras, the tensorial algebra, the Weyl algebra and the universal enveloping algebra of a Lie algebra.
All the partial actions of $T_n(q)$ on such algebras are described in Theorem \ref{taft_principal} .
\begin{cor}
Let $\cdot:T_n(q) \otimes A \longrightarrow A$ be a linear map, and suppose $1 \cdot 1_A=1_A$ and $g^i \cdot 1_A =0$, for all $1 \leq i <n$.
Then, $\cdot$ is a partial action of $T_n(q)$ on $A$ if and only if $(x \cdot 1_A)^n \in Z(A)$ and
\begin{eqnarray}\label{formula_corpo}
g^{n-i}x^j \cdot a = (-1)^i q^{ \frac{i(i+1)}{2}} { j \choose i }_{q} (x \cdot 1_A)^{j-i}a(x \cdot 1_A)^i,
\end{eqnarray}
for any $a \in A$ and $0 \leq i,j < n$.
\end{cor}
\begin{proof}
Since $g^i \cdot 1_A =0$, for all $1 \leq i <n$, then $g^{i+k}=0$ when $i+k\neq 0$ and $i+k\neq n$. Now, we analyze these two exceptional cases. By Theorem \ref{taft_principal}, we have
\begin{itemize}
\item If $i+k=0$, then $i=k=0$, \emph{i.e.}, $x^j=(x\cdot 1_A)^ja$, for all $0 \leq j <n$;
\item If $i+k=n$, then $g^{i+k}\cdot a=a$ and so
$$ g^{n-k}x^j \cdot a = q^{-(n-k)j} (-1)^{k} q^{- \frac{k(k-1)}{2}} { j \choose k }_{q^{-1}} (x \cdot 1_A)^{j-k}a(x \cdot 1_A)^{k}.$$
Then, by identity \eqref{id_q},
$ g^{n-k}x^j \cdot a = (-1)^k q^{ \frac{k(k+1)}{2}} { j \choose k }_{q} (x \cdot 1_A)^{j-k}a(x \cdot 1_A)^k,$ for all $0\leq k<n$.
\end{itemize}
\end{proof}
\noindent{\textit{Remark.}} If $A= \Bbbk$, the above corollary coincides with the partial action $\lambda_\alpha$, $\alpha\in\Bbbk$, of the result \cite[Theorem 3.1]{taft_corpo_revista}.
\begin{cor}\label{x_ponto_1_comuta}
Let $\cdot:T_n(q) \otimes A \longrightarrow A$ be a linear map, and suppose $g \cdot 1_A \in \{0, 1_A\}$ and $x \cdot 1_A \in Z(A)$.
Then, $\cdot$ is a partial action of $T_n(q)$ on $A$ if and only if $\cdot$ is a global action, or $\cdot$ is a linear map such that $\cdot|_{\Bbbk C_n \otimes A}$
is a partial group action such that $g \cdot 1_A = 0$ and, for all $i,j \in \I_{0,n-1}, a \in A$,
\begin{itemize}
\item[(i)] $ g^ix^j \cdot a = q^{-ij} (x \cdot 1_A)^j \sum_{k=0}^j (-1)^k q^{- \frac{k(k-1)}{2}} { j \choose k }_{q^{-1}}(g^{i+k} \cdot a);$
\item[(ii)] $g^i \cdot (x \cdot 1_A) = q^{-i}(g^i \cdot 1_A)(x \cdot 1_A)$.
\end{itemize}
\end{cor}
In particular, if $A$ is a commutative algebra, then Theorem \ref{taft_principal} and Corollary \ref{x_ponto_1_comuta} are equivalent.
\begin{cor}
Let $\cdot:T_n(q) \otimes A \longrightarrow A$ be a linear map, and suppose $g \cdot 1_A \in \{0, 1_A\}$ and that there exists $k \in \I_{2,n-1}$ such that $g^k \cdot a = a$, for all $a \in A$.
Then, $\cdot$ is a partial action of $T_n(q)$ on $A$ if and only if $\cdot$ is a global action, or $\cdot$ is a linear map where $\cdot|_{\Bbbk C_n \otimes A}$
is a partial group action such that $g \cdot 1_A = 0$ and $g^ix^j \cdot a = \delta_{j,0} (g^i \cdot a)$,
for all $i,j \in \I_{0,n-1}, a \in A$.
\end{cor}
\begin{proof}
By Theorem \ref{taft_principal} (iii), we get $x \cdot 1_A =0$.
\end{proof}
\begin{exa}\label{ex_taft}
Every partial action of Sweedler's $4$-dimensional Hopf algebra on $A$, namely $\cdot : \mathbb{H}_{4} \otimes A \longrightarrow A$, such that $g \cdot 1_A =0$, is given by
$$1\cdot a=a, \ \ \ g\cdot a=0, \ \ \ x\cdot a= w a \ \ \ \mbox{and} \ \ \ gx\cdot a=a w,$$
for all $a\in A$, where $w = x\cdot 1_A$ satisfies $w^2 \in Z(A)$.
In particular, if $w \in Z(A)$, then for all $a\in A$, $$1\cdot a=a, \ \ \ g\cdot a=0 \ \ \ \mbox{and} \ \ \ x\cdot a=gx\cdot a=w a.$$
When $A=\Bbbk$, every $w \in \Bbbk$ characterizes such a partial action by the linear map $\lambda_w: \mathbb{H}_{4}\longrightarrow\Bbbk$, defined as $\lambda_w (1)=1_{\Bbbk}, \lambda_w(g)=0, \lambda_w(x)=\lambda_w(gx)=w.$
\end{exa}
\begin{exa}
Every partial action $\cdot : T_3(q) \otimes A \longrightarrow A$,
such that $g \cdot 1_A =0$, is given by
$$1\cdot a=a, \ \ \ g\cdot a=g^2\cdot a=0, \ \ \ x\cdot a=wa, \ \ \ gx\cdot a=0,$$ $$g^2x\cdot a=-qaw, \ x^2\cdot a=w^2a, \ gx^2\cdot a=aw^2, \ g^2x^2\cdot a=waw,$$
for all $a\in A$, where $w = x\cdot 1_A$ satisfies $w^3 \in Z(A)$.
In particular, if $w \in Z(A)$, then, for all $a\in A$, we have
$$1\cdot a=a, \ \ \ g\cdot a=g^2\cdot a=0, \ \ \ x\cdot a=wa \ \ \ gx\cdot a=0,$$
$$g^2x\cdot a=-q wa, \ x^2\cdot a= gx^2\cdot a=g^2x^2\cdot a=w^2a.$$
When $A=\Bbbk$, every $w \in \Bbbk$ characterizes such a partial action by the linear map
$\lambda_w: T_3(q)\longrightarrow\Bbbk$, where $\lambda_w(g^i)=\delta_{i,0}$, $\lambda_w(x)=w$, $\lambda_w(gx)=0$, $\lambda_w(g^2x)= - q w$ and $\lambda_w(g^ix^2)=w^2$, for all $i \in \I_{0,2}$.
\end{exa}
\begin{exa}
Assume that $\cdot : T_4(q) \otimes A \longrightarrow A$ is a partial action such that $g \cdot 1_A =0$. Then,
$$1\cdot a=a, \ g\cdot a=g^3\cdot a=0, \ g^2\cdot a= \alpha_{g^2}(a), \ x\cdot a=wa,$$ $$gx\cdot a=q(g^2\cdot a)w, \ g^2x\cdot a=-w(g^2\cdot a), \ g^3x\cdot a=-qaw,$$ $$x^2\cdot a=w^2a-q(g^2\cdot a)w^2, \ gx^2\cdot a=(1-q)w(g^2\cdot a)w,$$ $$g^2x^2\cdot a= w^2(g^2\cdot a)-qaw^2, \ g^3x^2= (1-q)waw,$$ $$x^3\cdot a=w^3a-w(g^2\cdot a)w^2, \ gx^3\cdot a=-w^2(g^2\cdot a)w+aw^3,$$ $$g^2x^3\cdot a=-w^3(g^2\cdot a)+waw^2, \ g^3x^3\cdot a=w^2aw-(g^2\cdot a)w^3,$$
where $\alpha_{g^2}: D_{g^2} \longrightarrow D_{g^2}$, being $D_{g^2} = (g^2 \cdot 1_A) A$, is the linear map inherited from the partial group action $\Bbbk C_4$ on $A$, and any element $w \in A$ such that $w^4\in Z(A)$ and $g^2\cdot w=-(g^2\cdot 1_A)w$.
In particular, some interesting cases are:
\begin{enumerate}
\item If $g^2 \cdot 1_A = 0$, then
$$1\cdot a=a, \ \ g\cdot a=g^2\cdot a=g^3\cdot a=0, \ \ x\cdot a=wa, \ \ gx\cdot a=g^2x\cdot a=0,$$
$$ g^3x\cdot a=-qaw, \ x^2\cdot a=w^2a, \ gx^2\cdot a=0, \ g^2x^2\cdot a=-qaw^2, g^3x^2=(1-q)waw,$$
$x^3\cdot a=w^3a, \ \ gx^3\cdot a=aw^3, \ \ g^2x^3\cdot a=waw^2, \ \ g^3x^3\cdot a=w^2aw.$
\item If $\alpha_{g^2} = Id_{D_{g^2}}$, then for all $0\leq i\leq n-1$ and $1\leq j\leq n-1$,
$$1\cdot a=a, \ \ \ g\cdot a=g^3\cdot a=0, \ \ \ g^2\cdot a=a, \ \ \ g^ix^j\cdot a=0.$$
\item If $A=\Bbbk$, then the subcases (1) and (2) describe all the partial actions of $T_4(\omega)$ on $\Bbbk$, as in \cite[Example 3.8]{taft_corpo_revista}.
\end{enumerate}
\end{exa}
For a partial action to define a partial representation, it must be symmetric \cite[Example 3.5]{Partial_representations}.
Our next goal is to verify that the partial actions presented in Theorem \ref{taft_principal} are symmetric.
\begin{thm}\label{simetria_taft}
Every partial action $\cdot : T_n(q) \otimes A \longrightarrow A$ such that $g \cdot 1_A \in \{0, 1_A\}$ is symmetric.
\end{thm}
\begin{proof}
If $g \cdot 1_A = 1_A$, the partial action is, in fact, a global action, and so the result is clear.
Now, assume $g \cdot 1_A = 0$.
It is sufficient to check the symmetric condition (PA.S):
$g^ix^j \cdot ((g^tx^s \cdot a)b) = ((g^ix^j)_1 g^tx^s \cdot a)((g^ix^j)_2 \cdot b).$
As in the proof of Proposition \ref{prop_volta}, we proceed by fixing vectors and checking the scalars. On one hand,
\begin{align*}
g^i x^j & \cdot ((g^tx^s \cdot a)b) = q^{-ij} \sum_{k=0}^{j} q^{\frac{-k(k-1)}{2}} (-1)^k {j \choose k}_{q^{-1}} \!\!\!\! w^{j-k} \left( g^{i+k} \cdot (g^tx^s \cdot a)b \right) w^k\\
\stackrel{\ref{lema_ind}}{=} \ & q^{-ij-s(t+i)} \sum_{k=0}^{j} \sum_{\ell=0}^{s} q^{-sk -\frac{k(k-1)}{2} - \frac{\ell(\ell-1)}{2}} (-1)^{k+\ell} {j \choose k}_{q^{-1}} {s \choose \ell}_{q^{-1}}\\
\times \ & w^{j+s-(k+\ell)}(g^{i+t+k+\ell} \cdot a) w^{\ell}(g^{i+k}\cdot b)w^k.
\end{align*}
Considering $0 \leq k_0 \leq j$ and $0 \leq \ell_0 \leq s$ fixed, we obtain the vector
\begin{eqnarray}\label{eqsimetria_vetor}
w^{j+s-(k_0+\ell_0)}(g^{i+t+k_0+\ell_0} \cdot a) w^{\ell_0}(g^{i+k_0}\cdot b)w^{k_0}
\end{eqnarray}
with scalar
\begin{eqnarray}\label{eqsimetria}
q^{-ij-s(t+i)} q^{-sk_0 -\frac{k_0(k_0-1)}{2} - \frac{\ell_0(\ell_0-1)}{2}} (-1)^{k_0+\ell_0} {j \choose k_0}_{q^{-1}} {s \choose \ell_0}_{q^{-1}}.
\end{eqnarray}
On the other hand,
\begin{align*}
& ((g^ix^j)_1 g^t x^s \cdot a)((g^ix^j)_2 \cdot b) = \sum_{\ell=0}^{j} {j \choose \ell}_{q} (g^{i+\ell}x^{j-\ell}g^tx^s \cdot a)(g^ix^{\ell} \cdot b)\\
= \ & \sum_{\ell=0}^{j} q^{\ell(j-\ell) + t(j-\ell)}{j \choose \ell}_{q^{-1}} (g^{i+t+\ell}x^{j+s-\ell} \cdot a)(g^ix^{\ell} \cdot b).
\end{align*}
Due to the terms $g^{i+t+\ell}x^{j+s-\ell} \cdot a$, we have to proceed by cases.
\underline{Case 1: $j+s < n$.}
\begin{align*}
& \sum_{\ell=0}^{j} q^{\ell(j-\ell) + t(j-\ell)}{j \choose \ell}_{q^{-1}} (g^{i+t+\ell}x^{j+s-\ell} \cdot a)(g^ix^{\ell} \cdot b)\\
= \ & q^{-ij-s(i+t)}\sum_{\ell=0}^{j} \sum_{r=0}^{j+s-\ell}\sum_{\theta=0}^{\ell} q^{-\ell s -\frac{r(r-1)}{2} - \frac{\theta(\theta-1)}{2}} (-1)^{r + \theta} {j \choose \ell}_{q^{-1}} \\
\times \ & {j+s-\ell \choose r}_{q^{-1}} {\ell \choose \theta}_{q^{-1}} w^{j+s-(\ell+r)} (g^{i+t+\ell+r} \cdot a) w^{r+\ell-\theta} (g^{i+\theta} \cdot b) w^{\theta}\\
\stackrel{k = \ell + r}{=} & q^{-ij-s(i+t)}\sum_{\ell=0}^{j} \sum_{k=\ell}^{j+s}\sum_{\theta=0}^{\ell} q^{-\ell s - \frac{(k-\ell)(k-\ell-1)}{2} - \frac{\theta(\theta-1)}{2}} (-1)^{k+\theta -\ell}\\
\times \ & {j \choose \ell}_{q^{-1}} {j+s-\ell \choose k-\ell}_{q^{-1}} {\ell \choose \theta}_{q^{-1}} w^{j+s-k} (g^{i+t+k} \cdot a) w^{k-\theta} (g^{i+\theta} \cdot b) w^{\theta}.
\end{align*}
For $\theta = k_0$ and $k=k_0+\ell_0$, we obtain the following scalar
\begin{align*}
& \ q^{-ij-s(i+t)}\sum_{\ell=0}^{j} q^{-\ell s - \ell_0k_0+k_0\ell+\ell_0\ell -k_0(k_0-1) - \frac{\ell_0(\ell_0-1)}{2} - \frac{\ell(\ell+1)}{2} } (-1)^{\ell_0 -\ell}\\
& \times {j \choose \ell}_{q^{-1}} {j+s-\ell \choose k_0+\ell_0-\ell}_{q^{-1}} {\ell \choose k_0}_{q^{-1}}.
\end{align*}
Therefore, this case holds by \eqref{eqsimetria} and identity \eqref{Lema2_10}.
\underline{Case 2: $j+s \geq n$.}
\begin{align*}
& \sum_{\ell=0}^{j} q^{\ell(j-\ell) + t(j-\ell)}{j \choose \ell}_{q^{-1}} (g^{i+t+\ell}x^{j+s-\ell} \cdot a)(g^ix^{\ell} \cdot b)\\
= \ & q^{-ij - s(i+t)} \sum_{\ell=j+s-n+1}^{j} \sum_{k=\ell}^{j+s} \sum_{\theta=0}^{\ell} q^{-\ell s - \frac{(k-\ell)(k-\ell-1)}{2} - \frac{\theta(\theta-1)}{2}} (-1)^{k - \ell + \theta}\\
\times \ & {j \choose \ell}_{q^{-1}} {j+s-\ell \choose k-\ell}_{q^{-1}} {\ell \choose \theta}_{q^{-1}} w^{j+s-k} (g^{i+t+k} \cdot a ) w^{k-\theta}(g^{i+\theta} \cdot b)w^{\theta}.
\end{align*}
We will seek the vector highlighted in \eqref{eqsimetria_vetor} in the above expression and check its scalar.
If $j+s-(\ell_0 + k_0) < n$, one proceeds as in Case 1.
Thus, assume $j+s-(\ell_0 + k_0) \geq n$.
For $\theta = k_0$ fixed, $\ell \geq k_0$ and so $j+s-n+1 \geq k_0$.
Moreover, since $j+s-k < n$, for each $\ell \in \I_{j+s-n+1,j}$ there is a unique $k = \ell_0 + k_0+n$ such that the vector $w^{j+s-(k_0+\ell_0)-n} (g^{i+t+k_0+\ell_0} \cdot a ) w^{\ell_0+n}(g^{i+k_0} \cdot b)w^{k_0}$ appears in the above expression.
This vector is the same in \eqref{eqsimetria_vetor} due the commutative of $w^n$, but with scalar
\begin{align*}
& q^{-ij - s(i+t)} \sum_{\ell=j+s-n+1}^{j} q^{\ell s - \frac{(k + \ell_0+n-\ell)(k+\ell_0+n-\ell-1)}{2} - \frac{k_0(k_0-1)}{2}} (-1)^{\ell_0 - \ell + n} \\
& \times {j \choose \ell}_{q^{-1}} {j+s-\ell \choose k_0+\ell_0+n-\ell}_{q^{-1}} {\ell \choose k_0}_{q^{-1}}
\end{align*}
Now, we investigate this scalar. First, by identity \eqref{igualdade_q_n} it is equal to
\begin{align*}
& \ q^{-ij - s(i+t)} \sum_{\ell=j+s-n+1}^{j} q^{-\ell s - \ell_0k_0 + \ell(k_0+\ell_0) - k_0(k_0-1) - \frac{\ell_0(\ell_0-1)}{2} - \frac{\ell(\ell-1)}{2}} \\
& \times (-1)^{\ell_0 - \ell+1} {j \choose \ell}_{q^{-1}} {j+s-\ell \choose k_0+\ell_0+n-\ell}_{q^{-1}} {\ell \choose k_0}_{q^{-1}},
\end{align*}
and, comparing it with the scalar in \eqref{eqsimetria}, we only have to show that:
\begin{align*}
& \sum_{\ell=j+s-n+1}^{j} q^{-\ell s +sk_0 - \ell_0k_0 + \ell(k_0+\ell_0) -\frac{k_0(k_0-1)}{2} - \frac{\ell(\ell-1)}{2}} (-1)^{k_0- \ell+1}\\
& \times {j \choose \ell}_{q^{-1}} {j+s-\ell \choose k_0+\ell_0+n-\ell}_{q^{-1}} {\ell \choose k_0}_{q^{-1}} = {j \choose k_0}_{q^{-1}} {s \choose \ell_0}_{q^{-1}}.
\end{align*}
By identity \eqref{Lema2_10}, the expression above is equivalent to:
\begin{align*}
& \sum_{\ell=0}^{j+s-n} q^{-\ell s +sk_0 - \ell_0k_0 + \ell(k_0+\ell_0) -\frac{k_0(k_0-1)}{2} - \frac{\ell(\ell-1)}{2}} (-1)^{k_0- \ell}\\
& \times {j \choose \ell}_{q^{-1}} {j+s-\ell \choose k_0+\ell_0+n-\ell}_{q^{-1}} {\ell \choose k_0}_{q^{-1}} = {j \choose k_0}_{q^{-1}} {s \choose \ell_0}_{q^{-1}}.
\end{align*}
Note that, on the left side of the above equality, due to the term ${ \ell \choose k_0}_{q^{-1}}$, the lower index of the summation symbol can begin in $k_0$ instead $0$. Then, we replace $\ell$ by $L+k_0$ and use \eqref{Lema2_2}, to get the equivalent expression
\begin{align*}
& \sum_{L=0}^{j+s-n-k_0} q^{L(\ell_0-s) -\frac{L(L-1)}{2}} (-1)^{L} {j \choose k_0}_{q^{-1}} {j-k_0 \choose L}_{q^{-1}} {j+s-L-k_0 \choose \ell_0+n-L}_{q^{-1}} \\
= \ & {j \choose k_0}_{q^{-1}} {s \choose \ell_0}_{q^{-1}}.
\end{align*}
Finally, with a change of variable on the upper index of the summation, the above expression is true by identity \eqref{identidade_R_L_item}.
\end{proof}
\subsection{Partial Coactions}
\medbreak
In this section, we investigate the theory of the partial coactions of Taft algebras on algebras.
First, recall that $T_n(q)$ is a self-dual Hopf algebra.
\begin{lem}\label{inversa}
Consider the Taft algebra $T_n(q)$ and the linear maps:
\begin{itemize}
\item[(1)] $\psi : T_n(q) \longrightarrow (T_n(q))^*$ given by
$\psi(g^ix^j)=G^iX^j,$ where
\begin{align*}
G^iX^j = \sum_{k=0}^{n-1} (j)_q! q^{-i(k+j)-jk-\frac{j(j-1)}{2}} (g^kx^j)^*;
\end{align*}
\item[(2)] $\varphi : (T_n(q))^* \longrightarrow T_n(q)$ given by
\begin{align*}
\varphi ((g^ix^j)^*)=\frac{1}{n}((j)_q!)^{-1} q^{ij+\frac{j(j-1)}{2}} \sum_{k=0}^{n-1} q^{k(i+j)} g^kx^j.
\end{align*}
\end{itemize}
Then, both maps defined in (1) and (2) are isomorphisms of Hopf algebras.
Furthermore, $\varphi = \psi^{-1}$.
\end{lem}
For the next result, consider $\pi$ the canonic projection from $T_n(q)$ on $\Bbbk C_n$, \emph{i.e.}, $\pi : T_n(q) \longrightarrow \Bbbk C_n$ given by $\pi(g^ix^j)=\delta_{j,0}g^i$, for all $i,j \in \I_{0,n-1}$.
\begin{thm}\label{coacao_taft}
Let $\rho:A \longrightarrow A \otimes T_n(q)$ be a linear map, denoted by $\rho(a)=a^0\otimes a^1$, such that $A_1\in\{0,1\}$, where $A_i=\sum_{r=0}^{n-1}q^{-ir}(g^r)^{\ast}(1^1)1^0$, $0 \leq i <n$. Then, $\rho$ is a partial coaction if and only if $\rho$ is a global coaction, or $(I\otimes\pi)\rho$ is a partial coaction of $\Bbbk C_n$ on $A$ such that $A_1=0$ and
\begin{itemize}
\item[(i)] $\displaystyle \rho(a)= \sum\limits_{i,j = 0}^{n-1}\sum_{k=0}^j q^{-ij} (-1)^k q^{- \frac{k(k-1)}{2}} { j \choose k }_{q^{-1}} w^{j-k}\\\times\left(\sum_{t=0}^{n-1}q^{-(i+k)t}(g^t)^{\ast}(a^1)a^0\right) w^k
\otimes\frac{1}{n}((j)_q!)^{-1} q^{ij+\frac{j(j-1)}{2}} \sum_{\ell=0}^{n-1} q^{\ell(i+j)} g^\ell x^j$;
\item[(ii)] $w^n \in Z(A);$
\item[(iii)] $\sum_{s=0}^{n-1}q^{-is}(g^s)^{\ast}(w^1)w^0=q^{-i}A_iw,$
\end{itemize}
for all $a \in A$.
\end{thm}
\begin{proof}
By Proposition \ref{dual} and Lemma \ref{inversa}, $A$ is a partial $T_n(q)$-comodule algebra via $\cdot$
if and only if, $A$ is a partial $T_n(q)$-module algebra via $\rho$.
Notice that the maps $\cdot$ and $\rho$, when considered just as linear, are dual.
Therefore, considering $A_i=\sum_{r=0}^{n-1}q^{-ir}(g^r)^{\ast}(1^1)1^0$,
\begin{eqnarray*}
A_1=0&\Leftrightarrow & g \cdot 1_A=0 \ \ \mbox{and} \ \
A_1=1\Leftrightarrow g \cdot 1_A=1.
\end{eqnarray*}
By Proposition \ref{dual} and Theorem \ref{taft_principal},
\begin{eqnarray*}
\rho(a)&=& \sum\limits_{i,j} (g^ix^j)\cdot a \otimes \varphi((g^ix^j)^{\ast})\\
&=&\sum\limits_{i,j}\sum_{k=0}^j q^{-ij} (-1)^k q^{- \frac{k(k-1)}{2}} { j \choose k }_{q^{-1}} (x \cdot 1_A)^{j-k}(g^{i+k} \cdot a)(x \cdot 1_A)^k\\
&\otimes & \frac{1}{n}((j)_q!)^{-1} q^{ij+\frac{j(j-1)}{2}} \sum_{\ell=0}^{n-1} q^{\ell(i+j)} g^\ell x^j,
\end{eqnarray*}
for all $a\in A$ and $\varphi$ given by Lemma \ref{inversa}.
Moreover, $\cdot|_{\Bbbk C_n \otimes A}$
is a partial group action such that $g \cdot 1_A = 0$ if and only if $(I\otimes\pi)\rho$ is a partial coaction of $\Bbbk C_n$ on $A$ such that $A_1=0$.
Finally, considering $w=x\cdot 1_A$,
\begin{eqnarray*}
\sum_{s=0}^{n-1}q^{-is}(g^s)^{\ast}(w^1)w^0=q^{-i}A_iw \Leftrightarrow g^i \cdot (x \cdot 1_A) = q^{-i}(g^i \cdot 1_A)(x \cdot 1_A).
\end{eqnarray*}
\end{proof}
\noindent{\textit{Remark.}} Since the partial actions given by Theorem \ref{simetria_taft} are symmetric, it follows by Proposition \ref{dual} that the partial coactions of the above theorem are symmetric.
\begin{cor}\label{idempotente_trivial_coac}
Suppose that the algebra $A$ has only the trivial idempotents. Then, Theorem \ref{coacao_taft} characterizes all partial coactions of $T_n(q)$ on A.
\end{cor}
\begin{cor}\label{formula_coacao_corpo}
Let $\rho:A \longrightarrow A \otimes T_n(q)$ be a linear map, denoted by $\rho(a)=a^0\otimes a^1$, and suppose $A_0=1$ and $A_i =0$, for all $1 \leq i <n$, where $A_i=\sum_{r=0}^{n-1}q^{-ir}(g^r)^{\ast}(1^1)1^0$. Then, $\rho$ is a partial coaction of $T_n(q)$ on $A$ if and only if $w^n \in Z(A)$ and
\begin{equation*}
\displaystyle \rho(a)= \sum_{i,j =0}^{n-1} (-1)^i q^{ \frac{i(i+1)}{2}} { j \choose i }_{q} w^{j-i}a w^i \otimes \frac{1}{n}((j)_q!)^{-1} q^{-ij+\frac{j(j-1)}{2}} \sum_{k=0}^{n-1} q^{k(j-i)} g^kx^j,
\end{equation*}
for all $a \in A$.
\end{cor}
\noindent{\textit{Remark.}} If $A= \Bbbk$, this partial coaction coincides with the coaction $z_{\alpha}$, $\alpha\in\Bbbk$, of \cite[Theorem 3.13]{taft_corpo_revista}.
\begin{cor}\label{w_comuta}
Let $\rho:A \longrightarrow A \otimes T_n(q)$ be a linear map, denoted by $\rho(a)=a^0\otimes a^1$, such that $A_1\in\{0,1\}$, where $A_i=\sum_{r=0}^{n-1}q^{-ir}(g^r)^{\ast}(1^1)1^0$, $0 \leq i <n$, and suppose $w =\sum_{k=0}^{n-1} q^{-k}(g^kx)^{\ast}(1^1)1^0 \in Z(A)$.
Then, $\rho$ is a partial coaction of $T_n(q)$ on $A$ if and only if $\rho$ is a global coaction, or $(I\otimes\pi)\rho$ is a partial coaction of $\Bbbk C_n$ on $A$ such that $A_1=0$ and
\begin{itemize}
\item[(i)] $\displaystyle \rho(a)= \sum\limits_{i,j = 0}^{n-1} w^j \sum_{k=0}^j (-1)^k q^{- \frac{k(k-1)}{2}} { j \choose k }_{q^{-1}} \left(\sum_{t=0}^{n-1}q^{-(i+k)t}(g^t)^{\ast}(a^1)a^0\right)\\
\otimes \frac{1}{n}((j)_q!)^{-1} q^{\frac{j(j-1)}{2}} \sum_{\ell=0}^{n-1} q^{\ell(i+j)} g^\ell x^j$;
\item[(ii)] $ \sum_{s=0}^{n-1}q^{-is}(g^s)^{\ast}(w^1)w^0=q^{-i}A_iw,$
\end{itemize}
for all $a \in A$.
\end{cor}
In particular, if $A$ is a commutative algebra, then Theorem \ref{coacao_taft} and Corollary \ref{w_comuta} are equivalent.
\begin{cor}\label{coacao_identidade}
Let $\rho:A \longrightarrow A \otimes T_n(q)$ be a linear map, denoted by $\rho(a)=a^0\otimes a^1$, such that $A_1\in\{0,1\}$, where $A_i=\sum_{r=0}^{n-1}q^{-ir}(g^r)^{\ast}(1^1)1^0$, $i \in \I_{0,n-1}$, and suppose there exists $k \in I_{2,n-1}$, such that $\sum_{r=0}^{n-1}q^{-kr}(g^r)^{\ast}(a^1)a^0= a$, for all $a \in A$.
Then, $\rho$ is a partial coaction of $T_n(q)$ on $A$ if and only if $\rho$ is a global coaction, or $\rho$ is a linear map such that $(I \otimes \pi) \rho$
is a partial coaction of $\Bbbk C_n$ on $A$ such that $A_1=0$ and $$\rho(a)= \sum\limits_{i= 0}^{n-1} \left(\sum_{t=0}^{n-1}q^{-it}(g^t)^{\ast}(a^1)a^0\right)
\otimes \frac{1}{n} \sum_{\ell=0}^{n-1} q^{\ell i} g^\ell,$$
for all $a \in A$.
\end{cor}
\begin{proof}
By hypothesis it follows that $A_k=1$ and so, by Theorem \ref{coacao_taft} (iii), $w=0$.
\end{proof}
\begin{exa} \label{exa_coactiontaft}
Every partial coaction of Sweedler's $4$-dimensional Hopf algebra on $A$, namely $\rho : A \longrightarrow A \otimes \mathbb{H}_{4}$, such that $A_1 =0$, is given by
$$\rho(a)=a\otimes \left(\frac{1+g}{2}\right)+ w a\otimes \left(\frac{x-gx}{2}\right)-a w \otimes \left(\frac{x+gx}{2}\right),$$
for all $a\in A$, where $w^2 \in Z(A)$.
In particular, if $w \in Z(A)$, then for all $a\in A$, $$\rho(a)=a\otimes \left(\frac{1+g}{2}\right)- a w\otimes gx.$$
For instance, if $A=\Bbbk[z]$ is the polynomial algebra in the variable $z$, this example generalizes \cite[Example 3.14]{corepresentations}.
When $A=\Bbbk$, every $w \in \Bbbk$ characterizes such a partial coaction $\rho(1)=1\otimes z_w$, where $z_w= \frac{1 + g}{2} - w gx$.
\end{exa}
\begin{exa}
Every partial coaction $\rho : A \longrightarrow A \otimes T_3(q)$, such that $A_1 =0$, is given by
\begin{eqnarray*}
\rho(a)&=&a\otimes \left(\frac{1+g+g^2}{3}\right)+w a \otimes \left(\frac{x+qgx+q^2g^2x}{3}\right)\\
&-&aw\otimes \left(\frac{x+gx+g^2x}{3}\right)-w^2a\otimes \frac{q^2}{3}\left(x^2+q^2gx^2+qg^2x^2\right)\\
&-& aw^2\otimes \frac{q}{3}\left(x^2+gx^2+g^2x^2\right)-waw\otimes \frac{1}{3}\left(x^2+qgx^2+q^2g^2x^2\right),
\end{eqnarray*}
for all $a\in A$, where $w^3 \in Z(A)$.
In particular, if $w \in Z(A)$, then, for all $a\in A$,
\begin{eqnarray*}
\rho(a)&=&a\otimes \left(\frac{1+g+g^2}{3}\right)+ w a\otimes \frac{1}{3}\left( (q-1)gx + (q^2-1) g^2x\right)\\
&-&wa\otimes qgx^2.
\end{eqnarray*}
When $A=\Bbbk$, every $w \in \Bbbk$ characterizes such a partial coaction $\rho(1)=1\otimes z_w$, where $z_w = \frac{1 + g + g^2 }{3} + \frac{1}{3}\left( (q-1)w gx + (q^2-1) w g^2x - 3q w^2gx^2 \right).$
\end{exa}
\begin{exa}
The partial coaction of $T_4(\omega)$ on $A$ given by
$$\rho(a)=a\otimes \left(\frac{1+g^2}{2}\right),$$
for all $a \in A$, is an example that satisfies the conditions of Corollary \ref{coacao_identidade}.
\end{exa}
\section{Partial (co)actions of the Nichols Hopf algebra}\label{sec:Nichols}
First, we shall emphasize that Nichols Hopf algebras were introduced by Taft in \cite{Taft}, but were named after the work of Nichols \cite{nichols}.
Such Hopf algebras are the prototype of the theory of \emph{Nichols algebras}.
We will present this family of Hopf algebras as it appears in \cite[Section 2.2]{etingof}.
\medbreak
Let $n \geq 2$ be an integer and suppose $char(\Bbbk) \neq 2$.
The \emph{Nichols Hopf algebra of order $n$}, or shortly \emph{Nichols Hopf algebra}, here denoted by $\mathbb{H}_{2^n}$, has the structure as follows:
as algebra it is generated over $\Bbbk$ by the $n$ letters $g,x_1, \cdots, x_{n-1}$ with relations $g^2=1$, $x_i^2=0$, $x_i g = -g x_i$ and $x_ix_j = -x_jx_i$, for all $i, j \in \I_{n-1}$.
Thus, the set
$\mathcal{B} = \{g^{j_0}x_1^{j_1}x_2^{j_2}...x_{n-1}^{ j_{n-1} } : j_i \in \I_{0,1}, i \in \I_{0, n-1} \}$
is the canonical basis for $\mathbb{H}_{2^n}$ and consequently $dim_\Bbbk(\mathbb{H}_{2^n}) = 2^{n}$.
To complete the Hopf algebra structure of $\mathbb{H}_{2^n}$, we set
$\Delta (g) = g \otimes g$, $\varepsilon (g) = 1$, $S(g) =g^{-1}=g,$ and $\Delta (x_i) = x_i \otimes 1 + g \otimes x_i$, $\varepsilon (x_i) = 0$ and $S(x_i) = -gx_i,$ for all $i \in \I_{n-1}$.
Note that $G(\mathbb{H}_{2^n})= \{1, g \} = C_2$, for any $n$.
In particular, when $n=2$, the Nichols Hopf algebra $\mathbb{H}_{2^2}$ is exactly the Sweedler's $4$-dimensional Hopf algebra $\mathbb{H}_{4}$.
Similarly to global actions of Taft's algebra \cite[Proposition 4]{Acoes_taft_Centrone}, it is straightforward to determine global actions of Nichols Hopf algebra completely through an automorphism and skew-derivations.
\begin{prop}
Let A be an algebra, then an action of the Nichols Hopf algebra $H_{2^n}$ on $A$ is completely determinate by a choice of:
\begin{itemize}
\item [(i)] An automorphism $\alpha$ of $A$ of order $2$, \textit{i.e.}, $\alpha^2=\alpha$;
\item [(ii)] A family of $\alpha$-derivations $\{\delta_i\}_{i \in \mathbb{I}_{n-1}}$ of $A$ such that ${\delta_i}^2=0$ and $\alpha \delta_i = - \delta_i \alpha$, for all $i \in \mathbb{I}_{n-1}$.
\end{itemize}
Equivalently, the structure of $H_{2^n}$-module algebra on A is uniquely determined by a choice of:
\begin{itemize}
\item[(i)] A $\mathbb{Z}_2$-grading $A=A_{0}\oplus A_{1}$, defined by an automorphism $\alpha$;
\item [(ii)] A family of $\alpha$-derivations $\{\delta_i\}_{i \in \mathbb{I}_{n-1}}$ such that ${\delta_i}^2=0$, $\delta_i(A_0) \subset A_1$ and $\delta_i(A_1) \subset A_0$, for all $i \in \mathbb{I}_{n-1}$.
\end{itemize}
\end{prop}
\subsection{Partial Actions}
In this section, we compute a family of examples of partial actions of $\mathbb{H}_{2^n}$ on $A$.
First, we shall analyze when a partial action is global.
\begin{prop}\label{prop_nichols_global}
Let $\cdot : \mathbb{H}_{2^n} \otimes A \longrightarrow A$ be a partial action of $\mathbb{H}_{2^n}$ on $A$.
If $g \cdot 1_A = 1_A$, then $\cdot$ is global.
\end{prop}
\begin{proof}
Analogously to the proof of Proposition \ref{parte_global}.
\end{proof}
\begin{prop}\label{nichols}
Let $\cdot : \mathbb{H}_{2^n} \otimes A \longrightarrow A$ be a partial action of $\mathbb{H}_{2^n}$ on $A$.
If $g \cdot 1_A = 0$ and $x_i\cdot 1_A\in Z(A)$, $i \in \I_{n-1}$, then,
for all $a \in A$,
$$ gx_i\cdot a=x_i\cdot a=(x_i\cdot 1_A)a,$$
$$gx_{i_1}x_{i_2}...x_{i_s}\cdot a=x_{i_1}x_{i_2}...x_{i_s}\cdot a=0, \ \ \ s \geq 2, i_\ell \in \I_{n-1}, \ell \in \I_{s}.$$
\end{prop}
\begin{proof}
Let $\cdot : \mathbb{H}_{2^n} \otimes A \longrightarrow A$ be a partial action of $\mathbb{H}_{2^n}$ on $A$ with
$x_i\cdot 1_A\in Z(A)$, $i \in \I_{n-1}$, and $g \cdot 1_A = 0$.
First, note that $g\cdot a=0$, for $a\in A$. Then, by (PA.2), we have that $x_i\cdot a=(x_i\cdot 1_A)a$ and $gx_i\cdot a=a(gx_i\cdot 1_A)$. Moreover, considering $h=gx_i$ and $k=g$ in (PA.3), we obtain $gx_i\cdot 1_A=x_i\cdot 1_A$, and using that $x_i\cdot 1_A\in Z(A)$ we conclude $gx_i\cdot a=x_i\cdot a=(x_i\cdot 1_A)a$, for all $i \in \I_{n-1}$ and $a\in A$.
Now, for $h=gx_i$ and $k=x_j$ we get $x_ix_j\cdot a=0$, and finally taking $h=gx_i$ and $k=gx_j$ we deduce that $gx_ix_j\cdot a=0$, for all $i,j \in \I_{n-1}$, $i \neq j$.
By induction, one can prove that $x_{i_1}x_{i_2}...x_{i_s}\cdot a=gx_{i_1}x_{i_2}...x_{i_s}\cdot a=0$, for $s \geq 2$, $i_\ell \in \I_{n-1}$, $\ell \in \I_{s}$.
Indeed, we already done for $s=2$.
Assume $s \geq 2$ and consider $h=gx_{i_1}$, $k=x_{i_2}...x_{i_s}$ (resp. $h=gx_{i_1}$, $k=gx_{i_2}...x_{i_s}$) in (PA.3).
Then, using the induction hypothesis, it follows that $gx_{i_1}x_{i_2}...x_{i_s}\cdot a=0$ (resp. $x_{i_1}x_{i_2}...x_{i_s}\cdot a=0$).
\end{proof}
\begin{prop}\label{lda_alpha_eh_acao_nichols}
Let $\cdot : \mathbb{H}_{2^n} \otimes A \longrightarrow A$ be a linear map such that $1 \cdot 1_A = 1_A$, $g \cdot 1_A = 0$, $x_i\cdot 1_A\in Z(A)$, $i \in \I_{n-1}$, and for all $a \in A$,
$$ gx_i\cdot a=x_i\cdot a=(x_i\cdot 1_A)a,$$
$$gx_{i_1}x_{i_2}...x_{i_s}\cdot a=x_{i_1}x_{i_2}...x_{i_s}\cdot a=0, \ \ \ s \geq 2, i_\ell \in \I_{n-1}, \ell \in \I_{s}.$$
Then, $\cdot $ is a partial action of $\mathbb{H}_{2^n}$ on $A$.
\end{prop}
\begin{proof}
We need to check if condition (PA.2$^\prime$) holds.
We proceed by fixing $h \in \mathcal{B}$ and verifying (PA.2$^\prime$) for any $k \in \mathcal{B}$, where $\mathcal{B}$ is the canonical basis of $\mathbb{H}_{2^n}$.
First, if $h \in \{1, g, x_1, \cdots, x_{n-1} \}$, then (PA.2$^\prime$) holds trivially for any $k \in \mathcal{B}$.
If $h=gx_i$, $i \in I_{n-1}$, then (PA.2$^\prime$) means
$gx_i\cdot(a(k\cdot b)) =(gx_i\cdot a)(gk\cdot b) + a(gx_ik \cdot b),$
for any $k \in \mathcal{B}$.
Using that $x_i\cdot 1_A\in Z(A)$, $i \in \I_{n-1}$, and considering $k$ as $g^\ell, g^\ell x_j$ and $g^\ell x_{j_1}x_{j_2}...x_{j_s}$, where $\ell \in \I_{0,1},$ $j \in \I_{n-1}$ and $s \in \I_{2,n-1}, j_t \in \I_{n-1}, t \in \I_s$, it is a routine computation to get the equality.
Now, if $h=g^\ell x_{i_1}x_{i_2}$, $\ell \in \I_{0, 1}$, $i_1, i_2 \in \I_{n-1}$, $i_1 < i_2$, then
\begin{align*}
\Delta(g^\ell x_{i_1}x_{i_2}) = & \ g^\ell x_{i_1}x_{i_2} \otimes g^\ell - g^{\ell+1}x_{i_1} \otimes g^\ell x_{i_2} + g^{\ell +1} x_{i_2} \otimes g^{\ell}x_{i_1} \\
+ &\ g^\ell \otimes g^\ell x_{i_1}x_{i_2},
\end{align*}
and so (PA.2$^\prime$) becomes
\begin{align*}
g^\ell x_{i_1}x_{i_2}\cdot (a(k\cdot b)) = & \ (g^\ell x_{i_1}x_{i_2}\cdot a)(g^\ell k\cdot b) - (g^{\ell+1}x_{i_1}\cdot a) (g^\ell x_{i_2} k\cdot b) \\
+ & \ (g^{\ell+1} x_{i_2}\cdot a) (g^{\ell}x_{i_1} k\cdot b) + (g^\ell\cdot a) ( g^\ell x_{i_1}x_{i_2} k\cdot b),
\end{align*}
for any $k \in \mathcal{B}$.
Since $(gx_s\cdot a)=(x_s\cdot a)=(x_s\cdot 1_A)a$ and $(g^\ell x_{i_1}x_{i_2}\cdot a)=(g^\ell x_{i_1}x_{i_2}k\cdot a)=0$, for any $s \in \I_{n-1}$, $k \in \mathcal{B}$, (PA.2$^\prime$) is reduced to
$$(x_{i_1}\cdot 1_A)a(g^\ell x_{i_2}k\cdot b) = (x_{i_2}\cdot 1_A)a (g^\ell x_{i_1}k\cdot b),$$
for any $ k \in \mathcal{B}$.
One can verify this equality using $x_i\cdot 1_A\in Z(A)$ for $i \in \I_{n-1}$ and considering $k$ as $g^t$ and $g^t x_{j_1}...x_{j_s}$, where $t \in \I_{0,1}$ and $s \in \I_{n-1}$, $j_k \in \I_{n-1}$, $k \in \I_{s}$, $j_1 < \cdots < j_s$.
Finally, it remains to check (PA.2$^\prime$) for $h = g^\ell x_{i_1}x_{i_2}...x_{i_s}$, where $\ell \in \I_{0,1}$ and $s \geq 3$, and any $k \in \mathcal{B}$.
Write $h =g^\ell x_{i_1}x_{i_2}x_{i_3}w$, for some $w \in \mathbb{H}_{2^n}$.
Then,
\begin{align*}
\Delta(g^\ell x_{i_1}x_{i_2}x_{i_3}w) & = \Delta(g^\ell x_{i_1}x_{i_2})\Delta(x_{i_3}) \Delta(w) \\
& = g^\ell x_{i_1}x_{i_2}x_{i_3}w_1 \otimes g^\ell w_2 - g^{\ell +1}x_{i_1}x_{i_3}w_1 \otimes g^\ell x_{i_2}w_2 \\
& + g^{\ell +1} x_{i_2}x_{i_3}w_1 \otimes g^{\ell }x_{i_1}w_2 + g^\ell x_{i_3}w_1 \otimes g^\ell x_{i_1}x_{i_2}w_2 \\
& + g^{\ell +1}x_{i_1}x_{i_2}w_1 \otimes g^\ell x_{i_3} w_2 + g^{\ell +2}x_{i_1}w_1 \otimes g^\ell x_{i_2}x_{i_3} w_2 \\
& - g^{\ell +2} x_{i_2}w_1 \otimes g^{\ell }x_{i_1}x_{i_3} w_2 + g^{\ell +1} w_1 \otimes g^\ell x_{i_1}x_{i_2}x_{i_3} w_2,
\end{align*}
and for any $k \in \mathcal{B}$, condition (PA.2$^\prime$) means
\begin{align*}
g^\ell x_{i_1}x_{i_2}x_{i_3}w\cdot (a(k\cdot b)) & =( g^\ell x_{i_1}x_{i_2}x_{i_3}w_1\cdot a) (g^\ell w_2 k\cdot b) \\
& - ( g^{\ell +1}x_{i_1}x_{i_3}w_1\cdot a) (g^\ell x_{i_2}w_2 k\cdot b)\\
& + ( g^{\ell +1} x_{i_2}x_{i_3}w_1\cdot a) ( g^{\ell }x_{i_1}w_2 k\cdot b) \\
& + ( g^\ell x_{i_3}w_1\cdot a)( g^\ell x_{i_1}x_{i_2}w_2 k\cdot b) \\
& + ( g^{\ell +1}x_{i_1}x_{i_2}w_1 \cdot a) ( g^\ell x_{i_3} w_2 k\cdot b) \\
& + ( g^{\ell +2}x_{i_1}w_1\cdot a) ( g^\ell x_{i_2}x_{i_3} w_2 k\cdot b) \\
& - ( g^{\ell +2} x_{i_2}w_1 \cdot a) ( g^{\ell }x_{i_1}x_{i_3} w_2 k\cdot b) \\
& + ( g^{\ell +1} w_1 \cdot a) ( g^\ell x_{i_1}x_{i_2}x_{i_3} w_2 k\cdot b).
\end{align*}
Note that the above equality holds since the linear map $\cdot$ evaluated in a product that contains at least two skew-primitive elements as factors is equal to zero. Therefore, the map $\cdot$ as given is a partial action of $\mathbb{H}_{2^n}$ on $\Bbbk$.
\end{proof}
Considering the results presented in this section, we have characterized all partial actions $\cdot : \mathbb{H}_{2^n} \otimes A \longrightarrow A$ when $g\cdot 1_A = 0$ and $x_i\cdot 1_A\in Z(A)$ for all $i \in \I_{n-1}$.
\begin{thm}\label{simetria_nichols}
Every partial action $\cdot$ of the Nichols Hopf algebra $\mathbb{H}_{2^n}$ on $A$ such that $g\cdot 1_A = 1_A$, or $g\cdot 1_A = 0$ and $x_i\cdot 1_A\in Z(A)$ for all $i \in \I_{n-1}$ is symmetric.
\end{thm}
\begin{proof}
Let $\cdot$ be a partial action of the Nichols Hopf algebra $\mathbb{H}_{2^n}$ on $A$. If $g\cdot 1_A=1_A$, then, by Proposition \ref{prop_nichols_global}, $\cdot$ is a global action and so symmetric.
If $g\cdot 1_A=0$ and $x_i\cdot 1_A\in Z(A)$ for all $i \in \I_{n-1}$, the symmetric condition is verified similarly as the proof of Proposition \ref{lda_alpha_eh_acao_nichols}.
\end{proof}
\begin{exa} Every partial action of Nichols Hopf algebra of order $2$ on $A$, namely $\cdot : \mathbb{H}_{2^2} \otimes A \longrightarrow A$, such that $g \cdot 1_A =0$ and $x\cdot 1_A \in Z(A)$ is given by $$1\cdot a=a, \ \ \ g\cdot a=0 \ \ \ \mbox{and} \ \ \ x\cdot a=gx\cdot a=(x\cdot 1_A) a,$$
for all $a\in A$.
In fact, since the Nichols Hopf algebra of order $2$ is the Sweedler's Hopf algebra, it is expected by Example \ref{ex_taft}.
\end{exa}
\begin{exa}
Every partial action $\cdot : \mathbb{H}_{2^3} \otimes A \longrightarrow A$,
such that $g \cdot 1_A =0$ and $x_1\cdot 1_A, x_2\cdot 1_A\in Z(A)$, is given by
$$1\cdot a=a, \ \ \ g\cdot a=0, \ \ \ x_1\cdot a=gx_1\cdot a=(x_1\cdot 1_A)a,$$ $$ x_2\cdot a=gx_2\cdot a=(x_2\cdot 1_A)a, \ \ \ x_1x_2\cdot a=gx_1x_2\cdot a=0,$$
for all $a\in A$.
When $A=\Bbbk$, every $\alpha=(\alpha_1, \alpha_2) \in \Bbbk^2$ characterizes such a partial action by the linear map $\lambda_\alpha: \mathbb{H}_{2^3} \longrightarrow \Bbbk$, where $\lambda_\alpha(1)=1_\Bbbk$, $\lambda_\alpha(g)=0$, $\lambda_\alpha(x_1x_2)=0$, $\lambda_\alpha(gx_1x_2)=0$ and $\lambda_\alpha(x_i)=\lambda_\alpha(gx_i)=\alpha_i$, for all $i \in \I_{2}$.
\end{exa}
\subsection{Partial coactions}
In this section, we compute some partial coactions of Nichols Hopf algebra $\mathbb{H}_{2^n}$ on an algebra $A$.
The setting here is the same as for the Taft algebra, since both are self-dual Hopf algebras.
\begin{lem}\label{iso_nichols}
The linear map $\psi : \mathbb{H}_{2^n} \longrightarrow (\mathbb{H}_{2^n} )^*$
given by $\psi(g)=1^* - g^*$ and $\psi(x_i)=x_i^* - (gx_i)^*$, for all $i \in \I_{n-1}$, is an isomorphism of Hopf algebras.
\end{lem}
From the isomorphism $\psi$ in the above lemma, it follows that $$\psi\left(\frac{1+g}{2}\right) = 1^*, \psi\left(\frac{x_i-gx_i}{2}\right) = x_i^* \textrm{ and } \psi\left(\frac{-(x_i+gx_i)}{2}\right) = (gx_i)^*,$$
for all $i \in \I_{n-1}$.
Thus, we are able to prove the following result.
\begin{thm}\label{nichols_coac}
Let $\rho:A \longrightarrow A \otimes \mathbb{H}_{2^n}$ be a linear map, denoted by $\rho(a)=a^0\otimes a^1$, such that $A_1=(1^{\ast}-g^{\ast})(1^1)1^{0}=0$ and $w_i \in Z(A)$, where $w_i=(x_i^* - (gx_i)^*)(1^1)1^0$, for all $i \in \I_{n-1}$. Then, $\rho$ is a partial coaction if and only if
\begin{equation*}
\displaystyle \rho(a)= a\otimes \left(\frac{1+g}{2}\right)-\sum\limits_{i=1}^{n-1}w_ia\otimes gx_i,
\end{equation*}
for any $a \in A$.
Moreover, all these partial coactions are symmetric.
\end{thm}
\begin{proof}
Let $\mathcal{B} = \{g^{j_0}x_1^{j_1}x_2^{j_2}...x_{n-1}^{ j_{n-1} } : j_i \in \I_{0,1}, i \in \I_{0, n-1} \}$ be the canonical basis of $\mathbb{H}_{2^n}$ . By Proposition \ref{dual} and Proposition \ref{nichols}, $A_1=0\Leftrightarrow g \cdot 1_A=0$ and
\begin{eqnarray*}
\rho(a)&=& (g^{j_0}x_1^{j_1}x_2^{j_2}...x_{n-1}^{ j_{n-1} }\cdot a) \otimes \psi^{-1}((g^{j_0}x_1^{j_1}x_2^{j_2}...x_{n-1}^{ j_{n-1} })^{\ast})\\
&=& 1\cdot a \otimes \psi^{-1}(1^{\ast})+\sum_{i=1}^{n-1}x_i\cdot a \otimes \psi^{-1}(x_i^{\ast})+gx_i\cdot a \otimes \psi^{-1}(gx_i^{\ast})\\
&=&a\otimes \left(\frac{1+g}{2}\right)-\sum\limits_{i=1}^{n-1}w_ia\otimes gx_i,
\end{eqnarray*}
for all $a\in A$, where $\psi$ is given by Lemma \ref{iso_nichols}.
\end{proof}
\noindent{\textit{Remark.}} By the correspondence $A_1\Leftrightarrow g \cdot 1_A$, $A_1=1_A$ if only if $\rho$ is global coaction.
\begin{exa}
When $A=\Bbbk$, these partial coactions coincide with the partial coactions $\rho(a)=1\otimes z_\alpha$, where $z_\alpha= \frac{1+g}{2} - \sum_{i=1}^{n-1}\alpha_i \, gx_i$, presented in \cite[Theorem 4.6]{taft_corpo_revista}.
\end{exa}
\begin{exas}
If $n=2$ every partial coaction such that $A_1 =0$ and $w \in Z(A)$ is given by $\rho(a)=a\otimes \left(\frac{1+g}{2}\right)- wa\otimes gx,$
for all $a\in A$. See Example \ref{exa_coactiontaft}.
If $n=3$, every partial coaction such that $A_1 =0$ and $w_1, w_2\in Z(A)$ is given by $\rho(a)=a\otimes \left(\frac{1+g+g^2}{3}\right)- w_1 a\otimes gx_1-w_2a\otimes gx_2,$ for all $a\in A$.
\end{exas}
\bibliographystyle{abbrv}
|
\section{Introduction}
The Anti-de Sitter (AdS)/Conformal Field Theory (CFT) correspondence \cite{Maldacena:1997re} (for review see \cite{Aharony:1999ti, Casalderrey-Solana:2011dxg}) relates a Quantum Field Theory in $d$-dimensions to a gravitational theory defined on higher dimensions. In the appropriate limit, the gravitational theory is approximated by classical General Relativity, where metric of the latter describes the ``bulk". Different fields in the bulk theory correspond to different operators of the dual quantum field theory. The correspondence is further strengthened by the equality of the partition functions as described by the GKPW rule \cite{Gubser:1998bc,Witten:1998qj}. \\
After the initial discovery, there have been several extensions in different directions. Bulks which are not AdS everywhere have been studied. A particular class of example is asymptotically AdS space-time, as the name suggests the bulk asymptotes to AdS space-time near the boundary, though it can differ from AdS in the interior. It can have horizon, consequently, such black hole space-time carries a non-trivial Hawking temperature. On the dual side, the asymptotically AdS black hole corresponds to an equilibrium quantum field theory at a finite temperature. \\
To study non-equilibrium properties of quantum field theories, a natural tool is the Schwinger-Keldysh closed time path formalism \cite{calzetta_hu_2008, kamenev_2011}. Instead of defining the action on a positive real line for time, in the Schwinger-Keldysh formalism, the time runs forward and then backward in time forming a closed path. This is because of the fact that ket and bra in a density matrix evolves in opposite directions in time. In holography, the early attempts to capture the Schwinger-Keldysh physics was accomplished in \cite{Herzog:2002pc}, where the authors used Kruskal diagram to identify the field doubling as required by the Schwinger-Keldysh formalism. Recently, a prescription \cite{Glorioso:2018mmw} has been proposed to capture the Schwinger-Keldysh physics from the gravity dual. In this proposal the bulk geometry is complexified encircling the horizon. A cartoon of this geometry is shown in Fig. \ref{contour}.
\begin{figure}[t]
\centering
\begin{overpic}[scale=.4]{Figures/Contour}
\put(18,20){$r_+$}
\put(100,25){$\infty+i \epsilon$}
\put(100,9){$\infty-i \epsilon$}
\end{overpic}
\caption{Schwinger-Keldysh contour requires complexifying the bulk geometry.} \label{contour}
\end{figure}
Two segments of the bulk contour correspond to the two different segments of the Schwinger-Keldysh closed path, and we specify the sources at the boundaries corresponding to each segment. Using this prescription, several studies have been conducted for different kinds of systems \cite{Chakrabarty:2019aeu, Loganayagam:2020eue, Jana:2020vyx, Loganayagam:2020iol, Ghosh:2020lel, He:2021jna, He:2022jnc, He:2022deg, Bu:2020jfo, Bu:2021jlp, Bu:2022esd, Banerjee:2022aub}.
\\
Anomaly is an important concept in quantum field theory which is defined to be any classical symmetry of the Lagrangian that is broken by quantum mechanical effects \cite{Bertlmann:1996xk, Fujikawa:2004cx}. Although the classical Lagrangian is invariant under a symmetry, the measure of the path integral may not be invariant of the symmetry under consideration, and hence the divergencelessness of the corresponding current does not hold quantum mechanically. Different aspects of anomaly in holography have been extensively studied in the literature \cite{Fukushima:2008xe, Kharzeev:2013ffa, Landsteiner:2016led, Kharzeev:2020jxw, Ghosh:2021naw}. \\
The purpose of this paper is to understand anomaly in Schwinger-Keldysh physics. As we have mentioned before, different bulk fields correspond to different boundary operators. In particular, a bulk gauge field described by a $U(1)$ Maxwell theory is dual to a boundary current. Different aspects of the bulk gauge field have been extensively studied in the literature \cite{Policastro:2001yc, Policastro:2002se, Kovtun:2003wp}, and provided some early results emerging from holography. To describe anomaly in holography, a standard approach is to add a Chern-Simons term. In total, this corresponds to the Maxwell-Chern-Simons theory for a $U(1)$ gauge field. Holographically, this gauge field is dual to an anomalous current on the boundary quantum field theory. We study perturbation around the classical background, and then find an effective action for this perturbation. In the other way around, we derive a two-derivative action whose dynamics is dual to an anomalous current on the dual strongly coupled quantum field theory. After finding the solution of perturbation equation, we use an appropriate transformation to find the solution for the whole holographic Schwinger-Keldysh contour. By constructing the on-shell action, we find the Keldysh functional holographically. \\
The structure of this paper is as follows. In section \ref{setup}, we describe the setup of the system that we study. We also describe some of the important properties of the background there. In section \ref{vecsec}, we analyze the perturbation around the background fields. The perturbation is classified as vector sector, and we show that it is the vector perturbation that gives rise to interesting dynamics and affects the Chern-Simons term in the Lagrangian. We solve the perturbation equation in a derivative expansion, and the steps are spelled out in that section. We also construct a $2$-derivative action variation of which gives rise to the vector perturbation equations of motion. In section \ref{on-shell}, we compute the on-shell action, and renormalize it with appropriate counter-term. Section \ref{SK} is devoted for the Schwinger-Keldysh calculation. First we construct the full solution on the Schwinger-Keldysh radial contour, and then compute the Keldysh functional.
\section{Setup} \label{setup}
In this work, we consider the Maxwell-Chern-Simons theory for a $U(1)$ gauge field in a $5$-dimensional space-time. The action we consider is \footnote{Our notation is the following: beginning small Latin letters (a,b,c...) will denote the bulk indices, Greek letters denote the boundary indices, middle small Latin letters (i,j,m...) will denote the spatial indices. }$^,$\footnote{Convention for the Levi-Civita tensor is $\epsilon^{abcef}=-\frac{\varepsilon^{abcef}}{\sqrt{-g}}$, where $\varepsilon^{rtxyz}=1$. }
\begin{equation}
S=-\int d^5x \sqrt{-g}\left( F_{ab}F^{ab}+\frac{\alpha}{4} \epsilon^{abcde}A_a F_{bc}F_{ef} \right). \label{action}
\end{equation}
Here
\begin{equation}
A=A_a dx^a
\end{equation}
is the $U(1)$ gauge field and
\begin{equation}
F=dA=\frac{1}{2}F_{ab}dx^a\wedge dx^b
\end{equation}
is the corresponding field-strength. The first part of the action \eqref{action} is the usual Maxwell part, whereas the second part is the Chern-Simons term with $\alpha$ being the corresponding coupling. Note that the dimension of $\alpha$ is $-\frac{3}{2}$, and the theory is perturbatively non-renormalizable. \\
We are interested in studying an anomalous current in a quantum field theory at a finite temperature. Holographically, the minimum ingredients we need a black-brane metric, and a non-trivial gauge field. To accomplish this, we take the following background \cite{Banerjee:2008th}
\begin{align}
& ds^2=-r^2 f(r) dv^2+2 dvdr+r^2 \delta_{ij}dx^i dx^i, \label{metans} \\
& A=-\xi(r)dv,
\end{align}
where
\begin{align}
& f(r)=1-\left(1+Q^2 \right) \left( \frac{r_+}{r} \right)^4 +Q^2 \left( \frac{r_+}{r} \right)^{6}, \label{metric}\\ &
\xi(r)=\sqrt{\frac{3}{2}} Q \frac{r_+^{3}}{r^{2}}. \label{bg}
\end{align}
The metric is written in terms of ingoing Eddington-Finkelstein coordinates. The transformation that brings from the radial to Eddington-Finkelstein coordinates is described in Appendix \ref{Coord}. These field configurations satisfy the background equations:
\begin{align}
& \nabla_b F^{ba}-\frac{3\alpha}{16}\epsilon^{abcde}F_{bc}F_{de}=0.
\end{align}
The blackening factor $f(r)$ asymptotes to $1$ when $r\rightarrow \infty$. Near $r\rightarrow\infty$, the metric \eqref{metans} becomes:
\begin{equation}
ds^2\approx -r^2 dr^2+2dv dr+r^2 \delta_{ij}dx^i dx^j
\end{equation}
which is the AdS$_{5}$ metric written in Eddington-Finkelstein coordinates. $r=\infty$ is the conformal boundary of the space-time, and near boundary, the metric asymptotes to the $5$-dimensional Anti-de Sitter space-time. \\
The metric \eqref{metric} describes a black-brane which is charged under the gauge field. By setting $Q=0$, we obtain the uncharged case, which is the pure black-brane metric. The blackening factor has a zero when $r=r_+$ which corresponds to a horizon. Temperature corresponding to this black-brane is given by
\begin{equation}
T=\frac{2-Q^2}{2\pi b}
\end{equation}
where we have defined
\begin{equation}
b=\frac{1}{r_+}.
\end{equation}
Setting $Q=\sqrt{2}$, the Hawking temperature vanishes. This emerges from the fact that, at this particular value of $Q$, there is a double zero of the blackening factor. As a consequence of the presence of a double zero, there will be an AdS$_2$ factor in the near horizon geometry. This corresponds to an extremal case. In general, we will call \eqref{metans} to be Anti-de Sitter-Reissner-Nordstr\"om (AdS-RN) black-brane metric. \\
Equations \eqref{metric}-\eqref{bg} constitute the background for our system. In this work, we do not take gravity to be dynamical. That means we consider the metric to be the background, and any back-reaction on this will be ignored. In other words, we consider the probe limit, and the probe gauge field is described by the action \eqref{action}. Variation of the action gives
\begin{align}
\delta S=& 4\int_{M} d^5x \sqrt{-g} \ \delta A_a\left( \nabla_b F^{ba}-\frac{3\alpha}{16}\epsilon^{abcde}F_{bc}F_{de} \right) \nonumber\\
& -4 \int_{\partial M} d^4x\sqrt{-\gamma}\ n_a \delta A_b \left(F^{ab}+\frac{\alpha}{4} \epsilon^{abcde} A_c F_{de} \right) \label{firstvar}
\end{align}
where $M$ is the bulk manifold and $\partial M$ is its boundary with outward normal vector $n_a$.
The gauge field satisfies the equation of motion
\begin{equation}
\nabla_b F^{ba}-\frac{3\alpha}{16}\epsilon^{abcde}F_{bc}F_{de}=0. \label{gfeom}
\end{equation}
One can easily check that the background \eqref{metric}-\eqref{bg} satisfy Eq. \eqref{gfeom}. In the following section we will study the fluctuation of the gauge field.
\section{Vector perturbation} \label{vecsec}
In this section, we study the perturbation of the gauge field on top of the background given by Eq. \eqref{bg}. The perturbation can be classified as vector or scalar depending on their transformation under $SO(3)$ corresponding to the $3$-dimensional spatial part of the metric \eqref{metric}. Since we are considering a spin-1 field, the perturbation can be either vector or scalar. We start with the vector part, and comment on the scalar part in Appendix \ref{scalar}. \\
In the vector sector, we perturb the gauge field\footnote{$\int_{\omega,k}$ is the shorthand notation for $\int \frac{d\omega}{2\pi} \frac{d^3k}{(2\pi)^3}$}
\begin{equation}
a_m= r \int_{\omega,k} e^{-i \omega v} \left[ \Psi_{E} (r,k,\omega) V_m^{E} (\vec{k}, \vec{x})+\Psi_{O} (r,k,\omega) V_m^{O} (\vec{k}, \vec{x}) \right] . \label{vec}
\end{equation}
where $V_m^{E,O}$ are the vector harmonics satisfying \cite{Gubser:2007nd, Kodama:2003kk}
\begin{equation}
\nabla^2 V_m^{E,O}+k^2 V_m^{E,O}=0, \ \ \partial^m V_m^{E,O}=0. \label{defharmonic}
\end{equation}
Here $\nabla^2=\partial^m \partial_m$ is the $3$-dimensional Laplacian corresponding to the spatial part of the metric. The superscript refer to even (E) and odd (O) based on their behavior under parity transformations $\vec{x}\rightarrow -\vec{x}, \ \vec{k} \rightarrow -\vec{k}$. These vector harmonics have the following property:
\begin{align}
& \partial_m V_n^{E}-\partial_n V_m^{E}=-i k \epsilon_{mnp} V_p^{O}, \label{EO1} \\
& \partial_m V_n^{O}-\partial_n V_m^{O}=i k \epsilon_{mnp} V_p^{E}. \label{EO2}
\end{align}
These equations are consistent in the sense that by using Eq. \eqref{defharmonic}, one equation maps the other. Note the sign difference on the right hand side of Eqs. \eqref{EO1}-\eqref{EO2}. This will manifest itself in the Chern-Simons contribution of the fluctuation equation which we describe below. \\
We use the following normalization condition
\begin{align}
& \int d^3 \vec{x}\ V^E_m \left( \vec{k}_1,\vec{x} \right) V_E^m \left( \vec{k}_2,\vec{x} \right)=\left(2\pi \right)^3 \delta^{(3)}\left( \vec{k}_1+\vec{k}_2\right), \\
& \int d^3 \vec{x}\ V^O_m \left( \vec{k}_1,\vec{x} \right) V_O^m \left( \vec{k}_2,\vec{x} \right)=-\left(2\pi \right)^3 \delta^{(3)}\left( \vec{k}_1+\vec{k}_2\right), \\
& \int d^3 \vec{x}\ V^O_m \left( \vec{k}_1,\vec{x} \right) V_E^m \left( \vec{k}_2,\vec{x} \right)=0.
\end{align}
This is consistent with Eqs. \eqref{EO1}-\eqref{EO2}. Although we did not use any specific representation of the vector harmonics, a particular representation is given by \cite{Gubser:2007nd}
\begin{align}
& V^E_m \left(\vec{k},\vec{x} \right)=\frac{1}{k k_\perp}e^{i\vec{k}\cdot\vec{x}}\left(k_\perp^2\quad -k_xk_y\quad -k_x k_z \right) \\
& V^O_m \left(\vec{k},\vec{x} \right)= \frac{1}{ k_\perp}e^{i\vec{k}\cdot\vec{x}} \left(0\quad -k_z\quad k_y \right)
\end{align}
where
\begin{equation}
\vec{k}=\left(k_x \quad k_y \quad k_z \right), \ \ k_\perp=\sqrt{k_y^2+k_z^2}, \ \ \text{and} \ \ k=\sqrt{k_x^2+k_y^2+k_z^2}.
\end{equation}
Using this representation one can check all the statements made above. \\
Including the perturbation, in the vector sector, the gauge field and the field-strength tensors are:
\begin{align}
& A=-\xi dv+\int_{\omega,k} \left(r\Psi_E V_j^E+r\Psi_O V_j^O \right)e^{-i \omega v}dx^j \\
& F=-\xi' dr\wedge dv+\int_{\omega,k} \left[\left\{ \frac{d}{dr}\left(r\Psi_E \right)V_j^E+\frac{d}{dr}\left(r\Psi_O \right)V_j^O \right\} e^{-i \omega v} dr\wedge dx^j \right. \nonumber \\
&\qquad\qquad\qquad\qquad\qquad -i \omega r \left(\Psi_E V_j^E+\Psi_O V_j^O \right)e^{-i \omega v} dv\wedge dx^j\nonumber \\
& \qquad\qquad\qquad\qquad\qquad\left. -\frac{i kr}{2}\epsilon_{mnp} \left(\Psi_E V^O_p-\Psi_O V^E_p \right)e^{-i \omega v} dx^m\wedge dx^n\right] \label{F}
\end{align}
Using these, we can obtain the equation of motion for the vector fluctuation. They are:
\begin{align}
& \frac{d}{dr}\left[ r^3 f \frac{d}{dr} \left( r \Psi_{E} \right) \right]=\left(k^2+3 i \omega r \right) \Psi_{E}+2 i \omega r^2 \frac{d \Psi_{E}}{dr}+\frac{3 i \alpha k \xi'r}{2} \Psi_{O}, \label{Vec1} \\
& \frac{d}{dr}\left[ r^3 f \frac{d}{dr} \left( r \Psi_{O} \right) \right]=\left(k^2+3 i \omega r \right) \Psi_{O}+2 i \omega r^2 \frac{d \Psi_{O}}{dr}-\frac{3 i \alpha k \xi'r}{2} \Psi_{E}. \label{Vec2}
\end{align}
A quick look at the Eqs. \eqref{Vec1} and \eqref{Vec2} shows that a complex combination of the
perturbations $\Psi_{E,O}$ will satisfy decoupled radial equations. These two equations got coupled because of the presence of the Chern-Simons term. As we mentioned before, the Chern-Simon term mixes the parity and effectively acts as a coupling between the even and odd sector. This is reflected in the equations \eqref{Vec1}-\eqref{Vec2}. Note also the sign difference in front of the Chern-Simons coupling in Eqs. \eqref{Vec1}-\eqref{Vec2}. This is because of Eqs. \eqref{EO1}-\eqref{EO2}, since they have opposite sign in the right hand side.
\subsection{A two-derivative action for the vector perturbation}
In the previous section we had obtained the equation of motion for the vector fluctuations, which obeys a second order differential equation. It is therefore instructive to find a two-derivative action whose variation will yield the fluctuation equation. \\
There are several ways to obtain this action. The straightforward procedure is to start with the action \eqref{action} and expand it upto quadratic order in the fluctuation $\Psi_{E,O}$. The quadratic part is the desired action. Performing this exercise we find:
\begin{align}
& S^{(2)}= -2 \int_{\omega,k}\int dr \left[r^3 f\left\{ \frac{d}{dr}\left( r \Psi_E \right) \frac{d}{dr}\left( r \Psi_E^{(-)} \right)-\frac{d}{dr}\left( r \Psi_O \right) \frac{d}{dr}\left( r \Psi_O^{(-)} \right) \right\} \right. \nonumber \\
& +i \omega r^2 \left\{ \Psi_E^{(-)} \frac{d}{dr}\left( r \Psi_E \right) - \Psi_E \frac{d}{dr}\left( r \Psi_E^{{(-)}}\right) - \Psi_O^{{(-)}} \frac{d}{dr}\left( r \Psi_O \right) + \Psi_O \frac{d}{dr}\left( r \Psi_O^{(-)} \right)\right\}\nonumber \\
& \left. +k^2 r \left(\Psi_E \Psi_E^{(-)}-\Psi_O \Psi_O^{(-)} \right)\right] \nonumber \\
& +ik\alpha \int_{\omega,k}\int dr \left[ \frac{d}{dr} \left\{ r^2 \xi \left(\Psi_E \Psi_O^{(-)}+\Psi_E^{(-)} \Psi_O \right) \right\}-3r^2\xi' \left(\Psi_E \Psi_O^{(-)}+\Psi_E^{(-)} \Psi_O \right) \right] \label{2dev}
\end{align}
where $\Psi^{{(-)}}_{E,O}\left(\omega,k \right)=\Psi_{E,O}\left(-\omega,k \right)|_{\vec{k}\rightarrow -\vec{k}} $.
Note that, the $\alpha$-independent term arises from parity preserving Maxwell part, while the $\alpha$-dependent Chern-Simons term mixes parity. The total derivative part in the last line of Eq. \eqref{2dev} comes from the Chern-Simons part, and being a total derivative term, this does not play a role in the equation of motion. It will, however, contribute to the on-shell action. We will elaborate more on this in section \ref{on-shell}. Starting from the action \eqref{2dev}, a straight-forward exercise reveals that the equation of motion for $\Psi_E$ and $\Psi_O$ are indeed Eqs. \eqref{Vec1}-\eqref{Vec2}.
\subsection{Solution of perturbation equations in the hydrodynamic expansion}
In section \ref{vecsec}, we have found the equations for the vector fluctuation. Since the equations got coupled, this is unlikely that a closed form exact solution can be found. Instead, we work in the gradient expansion. Since the time and $\vec{x}$ dependence is exponential, a gradient expansion is also equivalent to small $\omega$ and $k$ expansion. This is what we refer to the hydrodynamic expansion. \\
In the hydrodynamic expansion, we look for solution in the form
\begin{align}
& \Psi_E=\Psi_E^{(0)}+\left( \omega \Psi_E^{(\omega)}+k \Psi_E^{(k)}\right)+\left(\omega^2 \Psi_E^{(\omega^2)}+k^2 \Psi_E^{(k^2)}+\omega k \Psi_E^{(\omega k)}\right)+\cdots \label{expE} \\
& \Psi_O=\Psi_O^{(0)}+\left( \omega \Psi_O^{(\omega)}+k \Psi_O^{(k)} \right)+\left( \omega^2 \Psi_O^{(\omega^2)}+k^2 \Psi_O^{(k^2)}+\omega k \Psi_O^{(\omega k)}\right)+\cdots \label{expO}
\end{align}
where the superscript denotes the order of the perturbative expansion. For the pure Maxwell Lagrangian, no odd power of $k$ emerged. But due the presence of Chern-Simons term, odd power of $k$ can be non-trivial. This is a novel feature in the Chern-Simons case. \\
We insert the expansions \eqref{expE} and \eqref{expO} into Eqs. \eqref{Vec1}-\eqref{Vec2}. Since the right hand side of Eqs. \eqref{Vec1}-\eqref{Vec2} is higher order in derivative than the left hand side, equating at each order we have equations of the form:
\begin{align}
& \frac{d}{dr}\left[ r^3 f \frac{d}{dr} \left( r \Psi_{E}^{\left( \omega^m k^n\right)
} \right) \right]=J_{E}^{\left( \omega^m k^n \right)
} (r), \label{GenE} \\
& \frac{d}{dr}\left[ r^3 f \frac{d}{dr} \left( r \Psi_{O}^{\left( \omega^m k^n \right)
} \right) \right]=J_{O}^{\left( \omega^m k^n \right)}(r) \label{GenO}
\end{align}
where the right hand side represents the source that is built from the solutions of previous orders. The only exception is the zeroth order for which the equation is sourceless, i.e the right hand side is zero. \\
Let us now take the equations at each order, in which the zeroth order equation is:
\begin{equation}
\frac{d}{dr}\left[ r^3 f \frac{d}{dr} \left( r \Psi_{E,O}^{\left( 0 \right)
} \right) \right]=0.
\end{equation}
This is a second order differential equation, integration of which contains two integration constants. As we want to impose regularity at the horizon $r=r_+$, one integration constant has to be set to zero. Here regularity guarantees the analyticity of the quantity at the horizon. Therefore, the solution to the zeroth order equation is
\begin{equation}
\Psi_{E,O}^{\left( 0 \right) }=\frac{C_{E,O}}{r}.
\end{equation}
Following the usual holographic dictionary, $C_{E,O}$ will be identified as the source on the boundary. \\
For non-zero order, we can write a general expression for the solution of equations Eqs. \eqref{GenE}-\eqref{GenO}. Again, these are second order differential equations, which require two boundary conditions. We demand regularity at the horizon. We also require that $r\Psi_{E,O}$ goes to $C_{E,O}$ at the boundary. Since we have already demanded that $r \Psi_{E,O}^{(0)} $ goes to $C_{E,O}$ at the boundary, this implies that $ r \Psi_{E,O}^{(\omega^m k^n)} $ where both $m,n\neq 0$, goes to zero at the boundary. After imposing these two conditions, some manipulations yield the following general expression for the solution:
\begin{align}
\Psi_{E,O}^{\left( \omega^m k^n \right)}&=-b^2 \frac{1}{br}\int_{br}^{\infty}d\eta \ \frac{1}{\eta^3 f\left(\frac{\eta}{b} \right)} \int_1^\eta d\kappa\ J_{(E,O)}^{\omega^m k^n
}\left( \frac{\kappa}{b} \right) \label{gen} \\
&=-b^2 \frac{1}{br}\int_{br}^{\infty}d\eta \ \frac{1}{\eta^3 f\left(\frac{\eta}{b} \right)} \left\{ F^{\left( \omega^m k^n\right) }(\eta)-F^{\left( \omega^m k^n\right)}(1) \right\} \\
& =b^2 \frac{1}{br}\int_{br}^{\infty}d\eta \ \frac{F^{\left(\omega^m k^n\right)}(1)}{\eta^3 f\left(\frac{\eta}{b} \right)} -b^2 \frac{1}{br}\int_{br}^{\infty}d\eta \ \frac{F^{\left(\omega^m k^n\right) }(\eta)}{\eta^3 f\left(\frac{\eta}{b} \right)}, \label{gensim}
\end{align}
where we have defined
\begin{equation}
F^{\left(\omega^m k^n \right)} (\eta)=\int_1^\eta d\kappa\ J_{(E,O)}^{\omega^m k^n
}\left( \frac{\kappa}{b} \right).
\end{equation}
The first term in Eq. \eqref{gensim} is universal, and can be integrated to find:
\begin{align}
\int_{br}^\infty d\eta \frac{1}{\eta^3 f\left(\frac{\eta}{b} \right)}=&-\frac{i \pi \left(-2 Q^2+2 Z+1\right)}{4 \left(Q^2-2\right) Z}-\frac{\left(1-2 Q^2\right) \tanh ^{-1}\left(\frac{2 b^2 r^2+1}{Z}\right)}{2 \left(Q^2-2\right) Z} \nonumber \\
& -\frac{\log \left(b^4 r^4+b^2 r^2-Q^2\right)-2 \log \left(1-b^2 r^2\right)}{4 \left(Q^2-2\right)}
\end{align}
where we have defined $Z=\sqrt{1+4Q^2}$. Near $r=\infty$, this has an asymptotic expansion:
\begin{equation}
\int_{br}^\infty d\eta \frac{1}{\eta^3 f\left(\frac{\eta}{b} \right)}\underset{r\rightarrow\infty}{=}\frac{1}{2 (b r)^2}+\frac{1+Q^2}{6 (b r)^6}+\mathcal{O}\left(\frac{1}{(br)^7} \right).
\end{equation}
The main difficulty to find an exact solution comes from the second part of Eq. \eqref{gensim}. However, often we need an expansion near the boundary located at $r=\infty$. We circumvent the difficulty by expanding the integrand of the second term in Eq. \eqref{gensim} near $r=\infty$, and then perform the integration. One might worry that the constant term might be different in this way of finding the near boundary expansion. This is fixed since we set the constant term for non-zero order to be zero, which is manifest from the upper limit of Eq. \eqref{gen}. Now we perform the integrations when possible to find the solution, and then find the near boundary expansion. In the case where the integrations are not tractable, we perform the strategy outline above to obtain the near boundary expansion. \\
The details of the calculations are presented in Appendix \ref{details}. Here, we record the main results. For the linear order, we could find the exact solutions. They are:
\begin{align}
& \Psi^{(\omega)}_{E,O}(r)=-i b^2 \frac{C_{E,O} }{br} F_1(br), \label{solomega} \\
& \Psi^{(k)}_{E,O}(r)=\pm \left( \frac{3}{2} \right)^{3/2}i \alpha C_{O,E} Q b^2 \frac{1}{br} F_2 (br), \label{solk}
\end{align}
where we have defined
\begin{align}
& F_1(br)=\int_{br}^\infty d\eta \frac{\eta-1}{\eta^3 f\left(\frac{\eta}{b} \right)}, \\
& F_2 (br)=\int_{br}^{\infty} d\eta \frac{\eta^2 -1}{\eta^5 f\left( \frac{\eta}{b} \right)}.
\end{align}
Note that Eq. \eqref{solk} originates from the Chern-Simons terms. In the absence of Chern-Simons term or for uncharged solution, this term would not be present. \\
The integrations becomes intractable for quadratic terms. However, by the strategy outlined above we have obtained the near boundary $r\rightarrow\infty$ expansion. They are:
\begin{align}
& \Psi_{E,O}^{(\omega^2)}=\frac{b^3}{4} C_{E,O}\left\{\frac{4 G_1-1}{(br)^3}+\frac{2 \ln (br)}{(br)^3} \right\}+\frac{ b^3}{2}C_{E,O} \frac{1}{(br)^4}+\mathcal{O}\left(\frac{1}{r^7} \right), \\
& \Psi_{E,O}^{(k^2)}= -b^3 C_{E,O} \frac{27 \alpha ^2 \ln 2+16 \ln (b r)-16 G_2+8}{32 (br)^3} +\mathcal{O}\left(\frac{1}{r^7} \right) \\
& \Psi_{E,O}^{(\omega k)}=\pm b^3 \left( \frac{3}{2}\right)^{\frac{1}{2}} \alpha C_{O,E} \left\{ \frac{f_3 (Q)}{(br)^3}+\frac{3Q}{4 (br)^4}-\frac{3Q}{8 (br)^6} \right\}+\mathcal{O}\left(\frac{1}{r^7} \right)
\end{align}
where $G_1$ and $G_2$ are the values of integrals $\int d\kappa \sqrt{\kappa}\frac{d}{d\kappa}\left\{ \sqrt{\kappa} F_1(\kappa) \right\} $ and $G_2$ is the value of $\int d\kappa \left\{\frac{1}{\kappa}-\frac{27\alpha^2 Q^2}{4 \kappa^3}F_2 (\kappa) \right\}$ respectively evaluated at $\kappa=1$ which is the horizon. The function $f_3(Q)$ is defined in Eq. \eqref{f3}. This vanishes at $Q=0$. \\
Now we have all the ingredients to write the near boundary expansion of the fluctuation. We find
\begin{align}
\frac{\Psi_{E,O}}{b}=& C_{E,O }\frac{1}{b r}-ib \omega C_{E,O} \left\{ \frac{1}{(br)^2}-\frac{1}{2 (br)^3}+\frac{1+Q^2}{5 (br)^6}+\mathcal{O}\left(\frac{1}{(br)^7} \right) \right\} \nonumber \\
& \pm i b k \left( \frac{3}{2} \right)^{3/2} \alpha C_{O,E} Q \left\{ \frac{1}{(br)^2}-\frac{1}{2 (br)^3}+\frac{1+Q^2}{5 (br)^6} + \mathcal{O}\left(\frac{1}{(br)^7} \right) \right\} \nonumber \\
& +\frac{b^2 \omega^2}{4} C_{E,O} \left\{ \left( \frac{4 G_1-1}{(br)^3}+\frac{2 \ln (br)}{(br)^3} \right)+ \frac{1}{2 (br)^4}+\mathcal{O}\left(\frac{1}{(br)^7} \right) \right\} \nonumber \\
& -b^2 k^2 C_{E,O} \left\{ \frac{27 \alpha ^2 \ln 2+16 \ln (b r)-16 G_2+8}{32 (br)^3} +\mathcal{O}\left(\frac{1}{(br)^7} \right) \right\} \nonumber \\
& \pm b^2 \omega k \left( \frac{3}{2}\right)^{\frac{1}{2}} \alpha C_{O,E} \left\{ \frac{f_3 (Q)}{(br)^3}+\frac{3Q}{4 (br)^4}-\frac{3Q}{8 (br)^6} +\mathcal{O}\left(\frac{1}{(br)^7} \right) \right\}\nonumber\\
& \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad \qquad+\mathcal{O}\left(\omega^3,k^3, \omega k^2, k^2\omega \right).
\end{align}
We will require this expression for various holographic calculations. In the next section, we compute the on-shell action, and then renormalize it.
\section{Quadratic on-shell action}
\label{on-shell}
In this section, we will obtain the on-shell action. Following the AdS/CFT dictionary, this on-shell action corresponds to the generating functional of the dual quantum field theory. We start with the bare action, and then proceed to renormalize by using the holographic renormalization.
\subsection{Bare action}
The bare action can be obtained in a couple of ways. One can start from the first variation of the action recorded in Eq. \eqref{firstvar}, and impose the on-shell condition. The latter kills the first line of Eq. \eqref{firstvar} and we find:
\begin{equation}
S^{(1)}_{\text{on-shell}}=-4 \int_{\partial M} d^4x\sqrt{-\gamma}\ n_a \delta A_b \left(F^{ab}+\frac{\alpha}{4} \epsilon^{abcde} A_c F_{de} \right). \label{firstvaronshell}
\end{equation}
Here $n_a$ is the outward unit normal to the regulated boundary defined via $r=\Lambda$, where $\Lambda\rightarrow \infty$ for the boundary, and $\gamma_{\mu\nu}$ is the induced metric on the boundary. A straight-forward calculation gives
\begin{equation}
n_a=\left(\frac{1}{r\sqrt{f}},0,\vec{0} \right).
\end{equation}
Since the variation of the gauge field does not change the shape of the boundary, the normal vector is also unchanged. A variation of Eq. \eqref{firstvaronshell} gives
\begin{equation}
S^{(2)}_{\text{on-shell}}=-2 \int_{\partial M}d^4x \sqrt{-\gamma} n_a \delta A_b \left\{ \delta F^{ab}+\frac{\alpha}{4} \epsilon^{abcde}\left(\delta A_c \bar{F}_{de}+\bar{A}_c \delta F_{de} \right)\right\}
\end{equation}
where a over-bar denotes the background (unperturbed) value. \\
Alternatively, one can start with the action \eqref{2dev}. There will be some total derivative terms which become the boundary terms. In either way we find a simple result
\begin{align}
S^{(2)}_{\text{on-shell}}=& -2 \int_{\omega,k} r^4 f \left\{ \Psi_E \frac{d}{dr}\left( r\Psi_E^{(-)} \right)-\Psi_O \frac{d}{dr}\left( r\Psi_O^{(-)} \right) \right\} \nonumber \\
& +i \alpha r^2 \xi \int_{\omega,k} k \left( \Psi_E \Psi_O^{(-)}+\Psi_E^{(-)} \Psi_O\right). \label{S2}
\end{align}
The first line comes from the Maxwell part, and the second line is the Chern-Simons contribution. Once again, the Maxwell part is parity preserving whereas the Chern-Simons part
mixes parity. As this is a quadratic action for the perturbations, a suitable diagonalied modes involving the
mixing of the modes $\Psi_{E,O}$ will produce decoupled equation as we had commented earlier. This action is the generating functional of the dual QFT via the AdS/CFT correspondence.
\subsection{Holographic renormalization}
The on-shell action is divergent. To remove the divergences we need appropriate counter-term. As described in \cite{Landsteiner:2011iq}, the required counter-term is:
\begin{equation}
S_{\text{ct}}= u \int d^4x \sqrt{-\gamma}\hat{F}_{(0)\mu\nu}\hat{F}^{(0)\mu\nu} \label{ct}
\end{equation}
where the bulk metric is written in domain wall coordinates as:
\begin{equation}
ds^2=du^2+\gamma_{\mu\nu}dx^\mu dx^\nu.
\end{equation}
In Eq. \eqref{ct}, $\hat{F}_{(0)\mu\nu}$ is the leading term in the near boundary expansion. Now we need a relationship between the coordinates $r$ and $u$ near the boundary. \\
Following Appendix \ref{Coord}, a coordinate transformation
\begin{equation}
du=\frac{dr}{r \sqrt{f(r)}}
\end{equation}
takes the metric from the Eddington-Finkelstein coordinates to the domain wall coordinates. Near the boundary $f(r)\rightarrow 1$, therefore the required coordinate transformation is:
\begin{equation}
u=\ln (mr)+\mathcal{O} \left(\frac{1}{r} \right).
\end{equation}
where $m$ is introduced for dimensional reason. Since the only dimension we have is the temperature, we set the scheme by choosing $m=b$. Therefore, the required counter-term is:
\begin{equation}
S_{\text{ct}}= \ln (br) \int d^4x \sqrt{-\gamma}\hat{F}_{(0)\mu\nu}\hat{F}^{(0)\mu\nu}. \label{ctneeded}
\end{equation}
Since this action is quadratic in the field-strength tensor, the renormalization is required starting from the second order in fluctuation. \\
We can extract $\hat{F}_{\mu\nu}$ from Eq. \eqref{F}. Explicitly,
\begin{align}
\hat{F}=-\int_{\omega,k} \left[\frac{}{} i \omega r \left(\Psi_E V_j^E+\Psi_O V_j^O \right)e^{-i \omega v} dv\wedge dx^j \right. \nonumber \\
\left. +\frac{i kr}{2}\epsilon_{mnp} \left(\Psi_E V^O_p-\Psi_O V^E_p \right)e^{-i \omega v} dx^m\wedge dx^n \right]. \label{Fhat}
\end{align}
From here we can extract:
\begin{align}
& \hat{F}_{vj}=-\int_{\omega,k} \frac{}{} i \omega r \left(\Psi_E V_j^E+\Psi_O V_j^O \right)e^{-i \omega v}, \\
& \hat{F}_{mn}=-\int_{\omega,k}i kr\epsilon_{mnp} \left(\Psi_E V^O_p-\Psi_O V^E_p \right)e^{-i \omega v}.
\end{align}
Using these, a straightforward calculation yields
\begin{equation}
S_{\text{ct}}=-2 r^2\ln (br) \int_{\omega,k} \left(\frac{\omega^2}{\sqrt{f}}-k^2 \sqrt{f}\right) \left(\Psi_E \Psi_E^{(-)}-\Psi_O \Psi_O^{(-)} \right). \label{Sct}
\end{equation}
Note the the counter-term starts from the quadratic order in the hydrodynamic expansion.
\subsection{Renormalized action}
Now we have all the ingredients to compute the renormalized on-shell action. The latter is given by:
\begin{equation}
S^{\text{ren}}_{\text{on-shell}}=S^{(2)}_{\text{on-shell}}+S_{\text{ct}}.
\end{equation}
Using Eqs. \eqref{S2} and \eqref{Sct} we find the following expression for the renormalized on-shell action
\begin{align}
S^{\text{ren}}_{\text{on-shell}}=& \int_{\omega,k} \left[-2 r^4 f \left\{ \Psi_E \frac{d}{dr}\left( r\Psi_E^{(-)} \right)-\Psi_O \frac{d}{dr}\left( r\Psi_O^{(-)} \right) \right\}+i \alpha k r^2 \xi \left( \Psi_E \Psi_O^{(-)}+\Psi_E^{(-)} \Psi_O\right) \right. \nonumber \\
&\left. -2 r^2 \ln (br) \left(\frac{\omega^2}{\sqrt{f}}-k^2 \sqrt{f}\right) \left(\Psi_E \Psi_E^{(-)}-\Psi_O \Psi_O^{(-)} \right)\right]. \label{Srenfin}
\end{align}
This will be the generating functional on the QFT side via the AdS/CFT correspondence. We can observe that both the Maxwell and the counter-term parts of the renormalized on-shell action \eqref{Srenfin} are parity preserving, but the Chern-Simons part mixes parity. \\
We insert the near boundary expansion of $\Psi_{E,O}$ into Eq. \eqref{Srenfin}. Since $\xi\sim \frac{1}{r^2}$ and $\Psi_{E,O} \sim \left\{ \frac{1}{r}+\mathcal{O}\left( \frac{1}{r^2}\right) \right\}$, the Chern-Simons part does not contribute on-shell action directly. It does, however, through the solution. Therefore, all the divergences come from the Maxwell term, which get canceled by the counter-term. After the cancellation we get:
\begin{align}
S_{\text{on-shell}}^{\text{ren}}= \frac{3\sqrt{6} i C_E C_O Q \alpha}{b}k+\frac{C_E^2-C_O^2}{8} \left\{ 32 G_1 \omega^2 +\left(16 G_2-27\alpha^2 \ln 2 \right)k^2 \right\} \nonumber \\
+\mathcal{O}\left(\omega^3, \omega^2 k, \omega k^2, k^3 \right), \label{Srenonshell}
\end{align}
where $G_1$ and $G_2$ are defined in Eqs. \eqref{G1} and \eqref{G2} respectively. In Eq. \eqref{Srenonshell} all the terms proportional to $\alpha$ are coming from the Chern-Simons term. Once again, the parity mixing term begets from the Chern-Simons term. Consequently, the even-odd correlator is non-trivial in this case.
\subsection{The boundary current}
To get more insight about the derivative expansion, in this section, we will compute the boundary current. According to the standard AdS/CFT prescription, the boundary current is given by \cite{Zaanen:2015oix, Ammon:2015wua}:
\begin{equation}
\langle J^\mu \rangle=\sqrt{-\gamma} n_r \left( F^{r\mu}+\frac{\alpha}{4}\varepsilon^{r\mu\nu\rho\sigma}A_\nu F_{\rho\sigma} \right).
\end{equation}
Using the results from previous section, this can be written as
\begin{equation}
\langle J^\mu \rangle=r^3 \left( F^{r\mu}+\frac{\alpha}{4}\varepsilon^{r\mu\nu\rho\sigma}A_\nu F_{\rho\sigma} \right).
\end{equation}
Using different expressions and after some algebra we find:
\begin{align}
& \langle J^v \rangle=r^3 \left[\xi'+\frac{i\alpha}{2r}\int_{\omega_1,\vec{k}_1} \int_{\omega_2,\vec{k}_2} k_2 \left\{ \Psi_{O}^{(1)} \Psi_{E}^{(2)} \vec{V}_{O}^{(1)}\cdot \vec{V}_{O}^{(2)}-\Psi_{E}^{(1)} \Psi_{O}^{(2)} \vec{V}_{E}^{(1)}\cdot \vec{V}_{E}^{(2)} \right. \right. \nonumber \\
&\qquad\qquad\qquad \left. \left. + \Psi_E^{(1)}\Psi_E^{(2)} \vec{V}_E^{(1)}\cdot \vec{V}_O^{(2)}-\Psi_O^{(1)}\Psi_O^{(2)} \vec{V}_O^{(1)}\cdot \vec{V}_E^{(2)} \right\} e^{-i(\omega_1+\omega_2)v} \right] \\
& \langle J^m \rangle=r^3 \left[\int_{\omega,k} \left\{ \left( f \frac{d}{dr} \left( r\Psi_E \right)-\frac{i\omega}{r} \Psi_E\right)V^m_E+ \left( f \frac{d}{dr} \left( r\Psi_O \right)-\frac{i\omega}{r} \Psi_O\right)V^m_O \right\} e^{-i\omega v} \right. \nonumber \\
&\qquad\qquad -\frac{\alpha}{4 r^3} \left\{-2ir\xi \int_{\omega,k} k \left( \Psi_E V^m_O-\Psi_O V^m_E \right) e^{-i\omega v} \right. \nonumber \\
&\qquad\qquad -2ir^2 \int_{\omega_1, k_1} \int_{\omega_2, k_2} \epsilon^{mnp} \omega_1 \left(\Psi_E^{(1)}\Psi_E^{(2)} V_n ^{E(1)} V_p ^{E(2)} +\Psi_O^{(1)}\Psi_O^{(2)} V_n ^{O(1)} V_p ^{O(2)} \right. \nonumber \\
&
\qquad\qquad \left. \left. \left. +\Psi_E^{(1)}\Psi_O^{(2)} V_n ^{E(1)} V_p ^{O(2)}+\Psi_O^{(1)}\Psi_E^{(2)} V_n ^{O(1)} V_p ^{E(2)} \right)e^{-i(\omega_1+\omega_2)v} \right\} \right].
\end{align}
To see the implications of the hydrodynamic expansion, we compute the divergence of the boundary current. After some algebra, we find:
\begin{align}
\langle \partial_\mu J^\mu \rangle&=2 \alpha r^2 \int_{\omega_1,k_1} \int_{\omega_2,k_2} \omega_1 k_2 \left\{ \Psi_{O}^{(1)} \Psi_{E}^{(2)} \vec{V}_{O}^{(1)}\cdot \vec{V}_{O}^{(2)}-\Psi_{E}^{(1)} \Psi_{O}^{(2)} \vec{V}_{E}^{(1)}\cdot \vec{V}_{E}^{(2)} \right. \nonumber \\
& \left. + \Psi_E^{(1)}\Psi_E^{(2)} \vec{V}_E^{(1)}\cdot \vec{V}_O^{(2)}-\Psi_O^{(1)}\Psi_O^{(2)} \vec{V}_O ^{(1)}\cdot \vec{V}_E^{(2)} \right\} e^{-i(\omega_1+\omega_2)v}. \label{ano}
\end{align}
This can be succinctly written
\begin{align}
\langle\partial_\mu J^\mu \rangle=-\frac{\alpha}{4} \epsilon^{\mu\nu\rho\sigma}\hat{F}_{\mu\nu}\hat{F}_{\rho\sigma}=-\alpha \hat{F}\wedge \hat{F}. \label{anoFwF}
\end{align}
That means inserting expression of $\hat{F}$ from Eq. \eqref{Fhat} into Eq. \eqref{anoFwF}, we can retrieve Eq. \eqref{ano}. This is the expression of familiar ABJ anomaly in Quantum Field Theory \cite{Adler:1969gk, Bell:1969ts, Peskin:1995ev}. Note that there was no electric field or magnetic field in the background. It's the fluctuation that induces an effective electric $(\sim \omega)$ and magnetic field $(\sim k)$. Therefore, we can interpret the current $J^\mu$ to be anomalous where the symmetry is broken by the gradient expansion. The Chern-Simons term in the action takes care of this symmetry breaking.
\section{Schwinger-Keldysh contour and the influence functional} \label{SK}
In the previous section we have used the ingoing Eddington-Finkelstein coordinates. To construct the full Schwinger-Keldysh solution we should also patch the the outgoing Hawking solution. The procedure to obtain the outgoing solution from the ingoing solution is presented below \cite{Jana:2020vyx}. \\
The metric we started with is:
\begin{equation}
ds^2=2dvdr-r^2 f(r)dv^2+r^2 dx_i dx^i \ . \label{metricin}
\end{equation}
This was written in ingoing Eddington-Finkelstein coordinates. We define a new coordinate $\chi$ by the following equation:
\begin{equation}
dr=\frac{i\beta}{2} r^2 f(r) d\chi \label{chi}
\end{equation}
where
\begin{equation}
\beta=\frac{2\pi b}{\left[2-Q^2 \right]}
\end{equation}
is the inverse Hawking temperature. The metric \eqref{metricin} can be written as:
\begin{equation}
ds^2=\frac{i\beta}{2} \left(r^2 f \right)^2 dv \left( i\beta d\chi-dv\right)+r^2 dx_i dx^i.
\end{equation}
From here this is manifest that:
\begin{equation}
dv\to i\beta d\chi-dv=dv', \ \omega\to -\omega \ .
\end{equation}
preserves the form of the metric. This coordinate transformation will be referred as the Debye time reversal, and it takes from the ingoing to the outgoing Eddington-Finkelstein coordinates. Under this time reversal the vector perturbations do not change since they only concern the spatial coordinates. Therefore the full solution is given by:
\begin{align}
& \Psi_E^{SK}=\Psi_E^{in}|_{\left\{ C_E\rightarrow C_{\text{in}}^E, C_O\rightarrow C_{\text{in}}^O \right\} }+ e^{-\omega\beta\chi} \Psi_E^{out}|_{\left\{ C_E\rightarrow C_{\text{out}}^E, C_O\rightarrow C_{\text{out}}^O \right\} } , \\
& \Psi_O^{SK}=\Psi_O^{in}|_{\left\{ C_E\rightarrow C_{\text{in}}^E, C_O\rightarrow C_{\text{in}}^O \right\} }+ e^{-\omega\beta\chi} \Psi_O^{out}|_{\left\{ C_E\rightarrow C_{\text{out}}^E, C_O\rightarrow C_{\text{out}}^O \right\} }.
\end{align}
Here $\Psi^{in}_{E,O}$ are the solutions that we have constructed so far in the ingoing Eddington-Finkelstein coordinates. $\Psi^{out}_{E,O}$ is constructed from $\Psi^{in}_{E,O}$ by taking $\omega\rightarrow -\omega$. We can also write the coefficients $C_{in}^{E,O}, C_{out}^{E,O}$ in terms of the Schwinger-Keldysh basis $C_L^{E,O}, C_R^{E,O}$. Holographically this amounts to complexifying the bulk geometry as Fig. \ref{contour} shows.
From Eq. \eqref{chi}, by a straightforward complex integration by using Cauchy integral formula we can write:
\begin{equation}
\chi(\infty+i \epsilon)=0, \quad \chi(\infty-i \epsilon)=1.
\end{equation}
The loci $\chi=0$ and $\chi=1$ will be denoted as the left and right boundaries respectively. We can specify the sources at the left and right boundaries which we denote as the left/right basis. In the left/right basis we have the following equations:
\begin{align}
& C_{in}^{E,O}+C_{out}^{E,O}=C_L^{E,O}, \\
& C_{in}^{E,O}+C_{out}^{E,O}e^{-\omega\beta}=C_R^{E,O}. \label{LR}
\end{align}
Solution of these is given by:
\begin{align}
C_{in}^{E,O}=f_{BE}(\omega) \left(C_R^{E,O} e^{\beta\omega}-C_L^{E,O} \right), \quad C_{out}^{E,O}=f_{BE}(\omega)\left(C_L^{E,O}-C_R^{E,O} \right)e^{\beta\omega},
\end{align}
where
\begin{equation}
f_{BE}(\omega)=\frac{1}{e^{\beta\omega}-1}
\end{equation}
is the Bose-Einstein factor. \\
We will now proceed to calculate the on-shell action \eqref{Srenfin} for the full Schwinger-Keldysh solution. For the complexified geometry, there are two boundaries located at $\chi=0$ and $\chi=1$. To get the full on-shell action we consider the individual on-shell actions, and subtract. We subtract because the unit normals are directed opposite at the boundaries. That means, the Schwinger-Keldysh functional is given by:
\begin{equation}
S_{\text{on-shell}}^{\text{SK}}=S_{\text{on-shell}}^{\text{ren}}|_{\chi=1}-S_{\text{on-shell}}^{\text{ren}}|_{\chi=0}.
\end{equation}
To better understand, we calculate the on-shell actions at the respective boundaries. We find:
\begin{align}
& S_{\text{on-shell}}^{\text{ren}}|_{\chi=0}=\int_{\omega,k}\left[ \frac{3\sqrt{6} i}{b} \left( C^E_{\text{in}}+C^E_{\text{out}} \right)\left( C^O_{\text{in}}+C^O_{\text{out}} \right)\alpha Q k+\frac{1}{8} \left\{ \left( C^E_{\text{in}}+C^E_{\text{out}} \right)^2 \right. \right.\nonumber \\
&\qquad\qquad\qquad\qquad \left. -\left( C^O_{\text{in}}+C^O_{\text{out}} \right)^2 \right\} \left\{ 32 G_1 \omega^2 +\left(16 G_2-27\alpha^2 \ln 2 \right)k^2 \right\} \nonumber \\
& \qquad\qquad\qquad\qquad \left. \frac{}{} +\mathcal{O}\left(\omega^3, \omega^2 k, \omega k^2, k^3 \right)\right] \\
& S_{\text{on-shell}}^{\text{ren}}|_{\chi=1}=\int_{\omega,k}\left[ \frac{3\sqrt{6} i}{b} \left( C^E_{\text{in}}+C^E_{\text{out}} \right)\left( C^O_{\text{in}}+C^O_{\text{out}} \right)\alpha Q k+\frac{1}{8} \left\{ \left( C^E_{\text{in}}+C^E_{\text{out}} \right)^2 \right. \right.\nonumber \\
&\qquad\qquad\qquad\qquad \left. -\left( C^O_{\text{in}}+C^O_{\text{out}} \right)^2 \right\} \left\{ 32 G_1 \omega^2 +\left(16 G_2-27\alpha^2 \ln 2 \right)k^2 \right\} \nonumber \\
& \qquad\qquad\qquad\qquad \left. \frac{}{}+ \frac{4i}{b} \left( C^E_{\text{in}}C^E_{\text{out}}-C^O_{\text{in}}C^O_{\text{out}} \right) \beta\omega^2+\mathcal{O}\left(\omega^3, \omega^2 k, \omega k^2, k^3 \right)\right].
\end{align}
From here, we can obtain
\begin{equation}
S_{\text{on-shell}}^{\text{SK}}=\int_{\omega,k} \left[ \frac{4i}{b} \left( C^E_{\text{in}}C^E_{\text{out}}-C^O_{\text{in}}C^O_{\text{out}} \right) \beta\omega^2+\mathcal{O}\left(\omega^3, \omega^2 k, \omega k^2, k^3 \right)\right].
\end{equation}
We can express this in terms of left/right basis by using Eq. \eqref{LR}. We find
\begin{align}
& S_{\text{on-shell}}^{\text{SK}} =\int_{\omega,k} \left[ -\frac{4i e^{\beta\omega}}{b \left(e^{\beta\omega}-1 \right)^2}\left\{ C^E_L \left( C^E_L-C^E_R \right)-C^O_L \left( C^O_L-C^O_R \right)+\left[ C^O_R \left( C^O_L-C^O_R \right) \right. \right.\right. \nonumber \\
&\qquad\qquad\qquad \left. \left. \left. -C^E_R \left( C^E_L-C^E_R \right)e^{\beta\omega} \right] \beta\omega^2\right\}+\frac{}{}\mathcal{O}\left(\omega^3, \omega^2 k, \omega k^2, k^3 \right) \right] \label{SKLR}
\end{align}
We can also express this in terms of Keldysh (average/difference) basis
\begin{equation}
C^{E,O}_{A}=\frac{1}{2}\left( C^{E,O}_{L}+C^{E,O}_{R} \right), \ \ C^{E,O}_{D}=\left( C^{E,O}_{L}-C^{E,O}_{R} \right).
\end{equation}
In terms of this basis we can write Eq. \eqref{SKLR} as
\begin{align}
& S_{\text{on-shell}}^{\text{SK}} =\int_{\omega,k} \left[ -\frac{2i e^{\beta\omega}}{b \left(e^{\beta\omega}-1 \right)^2} \left\{ C^E_D \left( 2C^E_A+C^E_D \right)-C^O_D \left(2 C^O_A+C^O_D \right)+\left[C^O_D \left( 2 C^O_A-C^O_D\right) \right. \right. \right. \nonumber \\
&\left. \left. \left. -C^E_D \left( 2 C^E_A-C^E_D \right) \right] e^{\beta\omega} \right\} \beta\omega^2 +\frac{}{}\mathcal{O}\left(\omega^3, \omega^2 k, \omega k^2, k^3 \right) \right].
\end{align}
This vanishes when we set $C^E_D=C^O_D=0$ as expected for a Keldysh functional. The important piece of information we can extract is that contribution from the Chern-Simons term gets canceled. Hence, the Schwinger-Keldysh functional is insensitive to the anomaly, at least up to the order that we considered here. Although we have not computed the higher order terms in derivative expansion, we believe the cancellation of anomaly will also persist for the higher order. This statement, however, requires careful scrutiny checks which we keep as a future work.
\section{Conclusion}
In this work we have studied the Maxwell-Chern-Simons theory in a $5$-dimensional space-time: a charged black-brane which asymptotes to a AdS$_5$ near $r=\infty$ which is the boundary. The gauge field only has time component. Via the AdS/CFT correspondence, we are considering a current in the dual field theory side with a non-trivial chemical potential which is the difference between the gauge field value evaluated at the boundary and horizon. \\
We have considered perturbation of the gauge field. This perturbation can be classified as vector and scalar modes depending on their transformation properties with respect to the spatial $\mathbb{R}^3$ metric. In Appendix \ref{scalar}, we have shown that the Chern-Simons term does not contribute for scalar perturbation. Therefore, the non-trivial perturbation comes from the vector sector only. This is further classified as even or odd depending on their transformation under parity. \\
We have found the equations of motion for the vector fluctuations. These equations are different for the even and odd sector. Unlike the Maxwell sector, the presence of the Chern-Simons term mixes this even and odd perturbation. To solve these equations, we have performed a derivative expansion which we call the hydrodynamic expansion. The details of this procedure is presented in Appendix \ref{details}. While in the first order we have found an exact form of the solution, in the second order the integrations become intractable. For this case, we have performed an asymptotic expansion of the integrand, and then we integrate. This procedure enables us to find the asymptotic expansion of the formal solution. \\
We have found a quadratic effective action \eqref{2dev} variation of which gives the vector equations of motion. This come from the Maxwell-Chern-Simons action, and then expanding it to the quadratic order. While the Maxwell part of the action is parity preserving, the Chern-Simons part mixes parity. Later, we show that the expectation value of the divergence of the boundary current is non-zero, and therefore the current is anomalous. We interpret the time derivative as the induced ``electric field", and the spatial derivative to be induced ``magnetic field". Therefore the anomaly is reminiscent of ABJ anomaly of quantum field theory. \\
We have computed the on-shell action which via the AdS/CFT correspondence is the generating functional of the dual current. We have found that parity preserving part comes from Maxwell term. The Chern-Simons term mixes parity, and gives rises to the even-odd correlator. \\
The bulk geometry can be complexified to find the Schwinger-Keldysh functional. We have used a complex radial contour, with the end point boundaries corresponding to the left and right sector of the Schwinger-Keldysh functional. Normal to these boundaries are oppositely oriented, and after the appropriate subtraction of the on-shell actions, we have obtained the Schwinger-Keldysh functional. In the Keldysh basis (average/difference basis), we have found that the Schwinger-Keldysh functional becomes zero when we set the difference source to be zero. This is an expected feature of Keldysh functional. Upon the subtraction of the on-shell actions, we found that the Chern-Simons term gets canceled. This implies that the Schwinger-Keldysh function is evanescent to the anomalous term up to the order we have considered, and we expect that this persists for the higher order terms as well. \\
There are several interesting arenas to explore as future works. In this work, we have not considered the backreaction of the gauge field on the metric. The effects of this gauge field on metric is an interesting problem that we want to study in near future. Depending on the parity of the gauge field, there should be two different dispersion relations. It would be intereseting to study these questions together with the Schwinger-Keldysh analysis. We wish to report the answers to this question in near future. \\
Another study that we wish to pursue is how to address the question of vacuum decay or phase transition in the context of Schwinger-Keldysh physics. Some of the work have been done, and presented in \cite{calzetta_hu_2008, kamenev_2011}, but the effect of strong coupling is not properly understood. Here holography comes into the role, and numerous works have been done under the name Coleman-de Luccia (CdL) tunneling (see \cite{Ghosh:2021lua} for a recent work). It would be interesting to understand the vacuum decay via CdL instanton in the Schwinger-Keldysh formalism. We hope to report this in a future work.
\section*{Acknowledgments}
We would like to thank Karl Landsteiner for clarifying issues with the holographic renormalization to us. We also express our gratitude to R. Loganayagam for illuminating discussions. We also thank Irfan Sadat Ameen for checking some of the calculations at initial stages of this work. \\
JKG dedicates this paper to his beloved mother Mrs. Shitali Ghosh for her unconditional love, and lifelong support.
|
\@startsection{section}{1{\@startsection{section}{1}
\z@{.7\linespacing\@plus\linespacing}{.5\linespacing}
{\normalfont\scshape}}
\makeatother
\begin{document}
\title[Partition-theoretic interpretation for certain truncated series]
{Partition-theoretic interpretation for certain truncated series}
\author{D. S. Gireesh$^1$ and B. Hemanthkumar$^1$}
\address{$^1$ Department of Mathematics, M. S. Ramaiah University of Applied Sciences, Peenya, Bengaluru-560 058, Karnataka, India}
\email{<EMAIL>}\email{<EMAIL>}
\begin{abstract}
In this article, we provide partition-theoretic interpretations for some new truncated pentagonal number theorem and identities of Gauss. Also, we deduce few inequalities for some partition functions.
\end{abstract}
\subjclass[2010]{11P81, 11P83, 11P84, 05A17} \keywords{Partitions; Theta functions; Truncated series}
\maketitle
\@startsection{section}{1{Introduction}\label{Int}
In 2012, Andrews and Merca \cite{AM1} proved truncated pentagonal number theorem
\begin{equation*}\label{APT}
\frac{1}{(q;q)_\infty}\sum_{\tau =0}^{\nu-1}(-1)^\tau q^{\tau (3\tau +1)/2}\left(1-q^{2\tau +1}\right)=1+(-1)^{\nu-1}\sum_{n=1}^{\infty}\dfrac{q^{\binom{\nu }{2}+(\nu +1)n}}{(q;q)_n}\left[\begin{array}{c}n-1\\ \nu -1\end{array}\right],
\end{equation*}
where \[(a;q)_\infty=\prod_{n=0}^{\infty}(1-aq^n),\qquad(a;q)_n=\frac{(a;q)_\infty}{(aq^n;q)_\infty}\]
and
\[\left[\begin{array}{c}M\\ N\end{array}\right]=\left[\begin{array}{c}M\\ N\end{array}\right]_q=\begin{cases}
0,\qquad \qquad\qquad \quad \,\text{if} \,N<0 \, \,\text{or}\, \, N>M,\\
\dfrac{(q;q)_M}{(q;q)_N(q;q)_{M-N}}, \,\text{otherwise}.
\end{cases}\]
As a consequence of the above result they obtained family of inequalities for the partition function $p(n)$ \cite{A}. Namely, for each $\nu \geq1$,
\begin{equation}\label{pe}
(-1)^{\nu -1}\sum_{\tau =0}^{\nu -1}(-1)^\tau \big(p\left(n-\tau (3\tau +1)/2\right)-p\left(n-\tau (3\tau +5)/2-1\right)\big)\geq 0
\end{equation}
with strict inequality if $n\geq \nu (3\nu +1)/2$. They also gave the partition interpretation for the truncated sum in \eqref{pe} as
\begin{equation*}\label{pe1}
(-1)^{\nu -1}\sum_{\tau =0}^{\nu -1}(-1)^\tau \big(p\left(n-\tau (3\tau +1)/2\right)-p\left(n-\tau (3\tau +5)/2-1\right)\big)=M_\nu (n),
\end{equation*}
where $M_\nu (n)$ counts the number of partitions of $n$ in which $\nu $ is the smallest integer that is not a part and there are more parts greater than $\nu $ compared to less than $\nu $.
Inspired by the above work, Guo and Zeng \cite{GZ} considered two idenitities of Gauss
\begin{align}\label{phi}
1+2\sum_{\tau =1}^{\infty}(-1)^\tau q^{\tau ^2}=\dfrac{(q;q)_\infty}{(-q;q)_\infty}=\phi(-q)
\end{align}
and
\begin{align}\label{psi}
\sum_{\tau =0}^{\infty}(-1)^\tau q^{\tau (2\tau +1)}\left(1-q^{2\tau +1}\right)=\dfrac{(q^2;q^2)_\infty}{(-q;q^2)_\infty}=\psi(-q)
\end{align}
to derive the truncated identities
\begin{align}\label{G1}
\dfrac{(-q;q)_\infty}{(q;q)_\infty}\left(1+2\sum_{\tau =1}^{\nu }(-1)^\tau q^{\tau ^2}\right)=1+(-1)^\nu \sum_{n=\nu +1}^{\infty}\dfrac{(-q;q)_\nu (-1;q)_{n-\nu }q^{(\nu +1)n}}{(q;q)_n}\left[\begin{array}{c}n-1\\ \nu \end{array}\right]
\end{align}
and
\begin{align}\label{G2}
\nonumber\dfrac{(-q;q^2)_\infty}{(q^2;q^2)_\infty}&\sum_{\tau =0}^{\nu -1}(-1)^\tau q^{\tau (2\tau +1)}\left(1-q^{2\tau +1}\right)\\& =1+(-1)^{\nu -1}\sum_{n=\nu }^{\infty}\dfrac{(-q;q^2)_\nu (-q;q^2)_{n-\nu }q^{2(\nu +1)n-\nu }}{(q^2;q^2)_n}\left[\begin{array}{c}n-1\\ \nu -1\end{array}\right]_{q^2},
\end{align}
for all $\nu \geq 1$.
Let $\overline{p}(n)$ be the number of overpartitions of $n$ (see Corteel and Lovejoy \cite{CL}) and $\textnormal{pod}(n)$ be the number of partitions of $n$ wherein odd parts are distinct (see Hirschhorn and Sellers \cite{HS}). The reciprocals of \eqref{phi} and \eqref{psi} give the generating functions for $\overline{p}(n)$ and $\textnormal{pod}(n)$, respectively.
Let $p_{2,4}(n)$ be the number of partitions of $n$ with parts $\not\equiv 2\pmod{4}$ and note that the generating function for $p_{2,4}(n)$ is same as that of $\textnormal{pod}(n)$.
From \eqref{G1} and \eqref{G2}, Guo and Zeng \cite{GZ} obtained the following inequalities
\begin{equation}\label{opi}
(-1)^\nu \left(\overline{p}(n)+2\sum_{\tau =1}^{\nu }(-1)^\tau \overline{p}(n-\tau ^2)\right)\geq 0
\end{equation}
with strict inequality if $n\geq(\nu +1)^2$ and
\begin{equation}\label{dopi}
(-1)^{\nu -1}\sum_{\tau =0}^{\nu -1}(-1)^\tau \big(\textnormal{pod}\left(n-\tau (2\tau +1)\right)-\textnormal{pod}\left(n-(\tau +1)(2\tau +1)\right)\big)\geq 0
\end{equation}
with strict inequality if $n\geq \nu (2\nu +1)$.
Andrews and Merca \cite{AM2} revised \eqref{G1} and \eqref{G2} as following:
\begin{equation*}\label{G3}
\dfrac{(-q;q)_\infty}{(q;q)_\infty}\left(1+2\sum_{\tau =1}^{\nu }(-1)^\tau q^{\tau ^2}\right)=1+2(-1)^\nu \dfrac{(-q;q)_\nu }{(q;q)_\nu }\sum_{\tau =0}^{\infty}\dfrac{q^{(\nu +1)(\nu +\tau +1)}(-q^{\nu +\tau +2};q)_\infty}{(1-q^{\nu +\tau +1})(q^{\nu +\tau +2};q)_\infty}
\end{equation*}
and
\begin{equation*}\label{G4}
\dfrac{(-q;q^2)_\infty}{(q^2;q^2)_\infty}\sum_{\tau =0}^{2\nu -1}(-q)^{\tau (\tau +1)/2} =1-(-1)^\nu \dfrac{(-q;q^2)_\nu }{(q^2;q^2)_{\nu -1}}\sum_{\tau =0}^{\infty}\dfrac{q^{\nu (2\tau +2\nu +1)}(-q^{2\tau +2\nu +3};q^2)_\infty}{(q^{2\nu +2\tau +2};q^2)_\infty}.
\end{equation*}
From these two identities they deduced interpretations of the sums in the inequalities \eqref{opi} and \eqref{dopi}:
\begin{equation*}\label{opp}
(-1)^\nu \left(\overline{p}(n)+2\sum_{\tau =1}^{\nu }(-1)^\tau \overline{p}(n-\tau ^2)\right)=\overline{\mu}_\nu (n),
\end{equation*}
where $\overline{\mu}_\nu (n)$ counts the number of overpartitions of $n$ in which the first part greater than $\nu $ appears at least $\nu +1$ times and
\begin{equation*}\label{dopp}
(-1)^{\nu -1}\sum_{\tau =0}^{\nu -1}(-1)^\tau \big(\textnormal{pod}\left(n-\tau (2\tau +1)\right)-\textnormal{pod}\left(n-(\tau +1)(2\tau +1)\right)\big)=MP_\nu (n),
\end{equation*}
where $MP_\nu (n)$ is the number of partitions of $n$ in which the first part greater than $2\nu -1$ is odd and appears exactly $\nu $ times. All other odd parts appear at most once.
To know more about truncated identities one can read Burnette and Kolitsch \cite{BK}, Chan et al. \cite{CH}, He et al. \cite{HJZ}, Kolitsch \cite{K}, Mao \cite{M}, Merca et al. \cite{MWY}, Wang and Yee \cite{WY,WY1}, Yee \cite{Y}.
Furthermore Andrews and Merca \cite{AM2} conjectured that
\begin{conj}
For $\nu $ even or $n$ odd,
\begin{align}\label{co1}
(-1)^{\nu -1}\sum_{i=0}^{\nu -1}(-1)^i\big(p\left(n-i(2i+1)\right)-p\left(n-(i+1)(2i+1)\right)\big)\leq M_\nu (n).
\end{align}
\end{conj}
We establish a partition-theoretic interpretation for the truncated sum in the inequality \eqref{co1} and we show that the sum is bounded above by $p(n-\nu (2\nu +1))$.
In this article, we obtain several partition-theoretic interpretaion of some new truncated sums of pentagonal number theorem and two identities of Gauss. Also, we deduce inequalities for various partition functions.
MacMahon \cite[p.13]{A} introduced what he termed modular partitions. Given positive integers $m$ and $n$, there exist $b\geq 0$ and $1\leq r\leq m$ such that $n=mb+r$. The $m$-modular partitions are a modification of the Young diagram so that $n$ is represented by a row of $b$ boxes with $m$ in each of them and one box with $r$ in it. For convenience, we put $r$ in the first column of the Young diagram. For example, the below figure shows the $3$-modular Young diagram of the partition $16+14+7+5+2$.
\begin{figure*}[!h]
\centering
\ytableausetup{centertableaux,mathmode, boxsize=1cm,mathmode}
\begin{ytableau}
*(gray) 1 & *(gray)3 & *(gray)3 & *(gray)3 & *(gray)3 & 3 \\
*(gray)2 & *(gray)3 & *(gray)3 & *(gray)3 & *(gray)3 \\
1 & 3 & 3 \\
2 & 3 \\
2 \\
\end{ytableau}
\caption{The $3$-modular Young diagram of 16+14+7+5+2}
\end{figure*}
For a Young diagram of partition $\pi$, define the $\nu $-Durfee rectangle, $\nu $ being a nonnegative integer, to be the largest rectangle which fits in the graph whose width minus its height is $\nu $. In Fig. 1, the $3$-Durfee rectangle of the partition is the shaded rectangle of size $2\times 5$.
For a fixed $\nu >0$ and $n\geq 0$, define $M(a,m,\nu ;n)$ to be the number of partitions of $n$ into the parts $\equiv a\pmod{m}$ such that all parts $\leq m\nu +a$ occur as a part at least once and the parts below the $(\nu +2)$-Durfee rectangle in the $m$-modular graph are strictly less than the width of the rectangle. For example, the partition $20+17+11+8+5+2$ is counted by $M(2,3,2;63)$. Whereas, the partition $23+17+17+8+5+5+2$ is not counted by $M(2,3,2;77)$.
\begin{theorem}\label{M3}
For a fixed $\nu \geq 0$ and positive integers $a$, $m$ such that $a<m$, we have
\begin{equation}\label{M13}
\sum_{n=0}^{\infty}M(a,m,\nu ;n)q^n=q^{a+\nu (m\nu +m+2a)/2}\sum_{\tau =0}^{\infty}\frac{q^{\tau (m\nu +m\tau +m+a)}}{(q^m;q^m)_\tau (q^a;q^m)_{\nu +\tau +1}}
\end{equation}
\end{theorem}
For a fixed $\nu >0$ and $n\geq 0$, define $N_\nu (n)$ be the number of partitions of $n$ such that all parts $\leq \nu -1$ occur at least once and the parts below the $\nu $-Durfee rectangle in the Young diagram are strictly less than the width of the rectangle. For example, the partition $6+5+3+3+2+1$ is counted by $N_3(20)$. Whereas, the partition $6+5+5+2+1+1$ is not counted by $N_3(20)$.
\begin{theorem}\label{Nk}
For a fixed $\nu \geq 1$, we have
\begin{equation}\label{nk1}
\sum_{n=0}^{\infty}N_\nu (n)q^n=q^{\nu (\nu -1)/2}\sum_{\tau =0}^{\infty}\frac{q^{\tau (\nu +\tau )}}{(q;q)_\tau (q;q)_{\nu +\tau -1}}.
\end{equation}
\end{theorem}
For a fixed $\nu \geq 0$ and $n\geq 0$, define $p(a,m,\nu ;n)$ to be the number of partitions of $n$ in which parts $\equiv a\pmod{m}$ form a partition counted by $M(a,m,\nu ;n-\beta)$, $\beta$ is the sum of parts $\not\equiv a\pmod{m}$.
\begin{theorem}\label{T2}
For $\nu \geq1$ and $n> \nu (3\nu +5)/2$,
\begin{align}\label{pi}
\nonumber(-1)^\nu &\left(\sum_{\tau =0}^{\nu }(-1)^\tau p\left(n-\tau (3\tau +1)/2\right)-\sum_{\tau =0}^{\nu -1}(-1)^\tau p\left(n-\tau (3\tau +5)/2-1\right)\right)\\&=p(2,3,\nu ;n)+p(1,3,\nu ;n).
\end{align}
\end{theorem}
For example, let $\nu =2$ and $n=17$.
\vspace{0.4cm}
\begin{tabular}{ l|l }
\hline
Partitions counted & Partitions counted \\ by $p(2,3,2;17)$ & by $p(1,3,2;17)$\\
\hline
$8+5+2+2$ & $7+5+4+1$ \\
$8+5+2+1+1$ & $7+4+4+1+1$ \\
&$7+4+3+2+1$\\
&$7+4+3+1+1+1$\\
&$7+4+2+2+1+1$\\
&$7+4+2+1+1+1+1$\\
&$7+4+1+1+1+1+1+1$\\
\end{tabular}
\vspace{0.4cm}
Hence, $p(17)-p(15)+p(10)-p(16)+p(12)=9=p(2,3,2;17)+p(1,3,2;17).$
\begin{theorem}\label{T3}
For $\nu \geq 1$ and $n>\nu (2\nu +3)$,
\begin{align}\label{p1i}
\nonumber(-1)^\nu &\left(\sum_{\tau =0}^{\nu }(-1)^\tau p\left(2n-\tau (2\tau +1)\right)-\sum_{\tau =0}^{\nu -1}(-1)^\tau p\left(2n-(\tau +1)(2\tau +1)\right)\right)\\&=(-1)^\nu p_o(n)+p(3,4,\nu ;2n)+p(1,4,\nu ;2n)
\end{align}
and
\begin{align}\label{p2i}
\nonumber(-1)^\nu &\left(\sum_{\tau =0}^{\nu }(-1)^\tau p\left(2n+1-\tau (2\tau +1)\right)-\sum_{\tau =0}^{\nu -1}(-1)^\tau p\left(2n+1-(\tau +1)(2\tau +1)\right)\right)\\&=p(3,4,\nu ;2n+1)+p(1,4,\nu ;2n+1).
\end{align}
\end{theorem}
\begin{theorem}\label{T4}
For $\nu \geq 1$ and $n> \nu (2\nu +3)$,
\begin{align}\label{dpi}
\nonumber(-1)^\nu &\left(\sum_{\tau =0}^{\nu }(-1)^\tau p_o\left(2n+1-\tau (2\tau +1)\right)-\sum_{\tau =0}^{\nu -1}(-1)^\tau p_o\left(2n+1-(\tau +1)(2\tau +1)\right)\right)\\&= p_o(3,4,\nu ;2n+1)+p_o(1,4,\nu ;2n+1),
\end{align}
where $p_o(a,4,\nu ;n)$ $(a\in\{1,3\})$, counts partitions of $n$ into odd parts in which the parts $\equiv a\pmod{4}$ form a partition counted by $M(a,4,\nu ;n-\alpha)$, $\alpha$ is the sum of parts $\not\equiv a\pmod{4}$.
\end{theorem}
For a fixed $\nu \geq 0$ and $n\geq 0$, $p_{2,4}(a,4,\nu ;n)$ counts partitions of $n$ into the parts $\not\equiv 2\pmod{4}$ in which the parts $\equiv a\pmod{4}$ form a partition counted by $M(a,4,\nu ;n-\alpha)$, $\alpha$ is the sum of parts $\not\equiv a\pmod{4}$.
\begin{theorem}\label{T5}
For $\nu \geq 1$ and $n> \nu (2\nu +3)$,
\begin{align}\label{podi}
\nonumber (-1)^\nu &\left(\sum_{\tau =0}^{\nu }(-1)^\tau p_{2,4}\left(n-\tau (2\tau +1)\right)-\sum_{\tau =0}^{\nu -1}(-1)^\tau p_{2,4}\left(n-(\tau +1)(2\tau +1)\right)\right)\\&=p_{2,4}(3,4,\nu ;n)+p_{2,4}(1,4,\nu ;n).
\end{align}
\end{theorem}
\begin{theorem}\label{T6}
For $\nu \geq 1$ and $n\geq (\nu +1)^2$,
\begin{equation}\label{oppi}
(-1)^\nu \left(\overline{p}(n)+2\sum_{\tau =1}^{\nu }(-1)^\tau \overline{p}(n-\tau ^2)\right)=2 \overline{p}(1,2,\nu ;n),
\end{equation}
where $\overline{p}(1,2,\nu ;n)$ counts overpartitions of $n$ in which the non-overlined odd parts form a partition counted by $M(1,2,\nu ;n-\alpha)$, $\alpha$ is the sum of parts other than non-overlined odd parts.
\end{theorem}
\begin{theorem}\label{T7}
For $\nu \geq 1$ and $n\geq (\nu +1)^2$,
\begin{equation}\label{ppi}
(-1)^\nu \left(pp(2n)+2\sum_{\tau =1}^{\nu }(-1)^\tau pp(2n-\tau ^2)\right)=(-1)^\nu p(n)+2pp_e(2n)
\end{equation}
and
\begin{equation}\label{ppi1}
(-1)^\nu \left(pp(2n+1)+2\sum_{\tau =1}^{\nu }(-1)^\tau pp(2n+1-\tau ^2)\right)=2pp_e(2n+1),
\end{equation}
where $pp_e(n)$ counts bipartitions $(\pi_1,\pi_2)$ such that $\pi_1$ is a unrestricted partition and $\pi_2$ is a partition in which odd parts are counted by $M(1,2,\nu ;n-|\pi_1|-\alpha)$, $|\pi_1|$ is the sum of parts of $\pi_1$ and $\alpha$ is the sum of even parts of $\pi_2$.
\end{theorem}
\begin{theorem}\label{T8}
For each $n\geq 1$,
\begin{equation}\label{eq8}
p(n)=p(2,3,0;n)+p(1,3,0;n)
\end{equation}
and
\begin{equation}\label{eq9}
p_{2,4}(n)=p_{2,4}(3,4,0;n)+p_{2,4}(1,4,0;n).
\end{equation}
\end{theorem}
\begin{theorem}\label{CP}
For $\nu \geq 1$ and $n\geq 0$,
\begin{equation}\label{CP1}
\sum_{\tau =0}^{\infty}(-1)^\tau p\left(n-\dfrac{(\nu +\tau )(\nu +\tau -1)}{2}\right)=N_\nu (n).
\end{equation}
\end{theorem}
For example, let $\nu =2$ and $n=10$. Partitions counted by $N_2(10)$ are $1+1+1+1+1+1+1+1+1+1$, $9+1$, $8+1+1$, $7+2+1$, $7+1+1+1$, $6+2+1+1$, $6+1+1+1+1$, $5+2+2+1$, $5+2+1+1+1$, $5+1+1+1+1+1$, $4+2+2+1+1$, $4+2+1+1+1+1$, $4+1+1+1+1+1+1$, $3+2+2+2+1$, $3+2+2+1+1+1$, $3+2+1+1+1+1+1$, $3+1+1+1+1+1+1+1$, $5+4+1$, $4+4+1+1$.
Hence, $p(9)-p(7)+p(4)-p(0)=19=N_2(10).$
The following inequality is obvious from the last theorem.
\begin{corollary}\label{CI}
For $\nu \geq 1$ and $n\geq 0$,
\begin{equation}
\sum_{\tau =0}^{\infty}(-1)^\tau p\left(n-\dfrac{(\nu +\tau )(\nu +\tau -1)}{2}\right)\geq 0,
\end{equation}
here we assume $p(x)=0$ if $x$ is a negative integer.
\end{corollary}
To prove our theorems, we rely on Gauss hypergeometric series \cite[Eqn. 1.2.14]{GR}
\begin{equation*}\label{GH}
_2\phi_1\left(\genfrac{}{}{0pt}{}{\alpha,\beta}{\gamma};q,z\right)=\sum_{n=0}^{\infty}\frac{(\alpha;q)_n(\beta;q)_n}{(q;q)_n(\gamma;q)_n}z^n
\end{equation*}
and second identity by Heine's transformation of $_2 \phi_1$ series \cite[Eqn. 1.4.5]{GR},
\begin{equation*}\label{HT}
_2\phi_1\left(\genfrac{}{}{0pt}{}{\alpha,\beta}{\gamma};q,z\right)=\dfrac{(\gamma/\beta;q)_\infty(\beta z;q)_\infty}{(\gamma;q)_\infty(z;q)_\infty}\
_2\phi_1\left(\genfrac{}{}{0pt}{}{\alpha \beta z/\gamma,\beta}{\beta z};q,\gamma/\beta\right).
\end{equation*}
Without explicitly mentioning we frequently use these two identities in our proofs.
\@startsection{section}{1{Proofs of main theorems}\label{MR}
In this section, we prove Theorems \ref{M3}--\ref{CP}.
\subsection*{Proof of Theorem \ref{M3}}
For a fixed $\nu \geq 0$, the partitions counted by $M(a,m,\nu ;n)$ with $(\nu +2)$-Durfee rectangle, in a $3$-modular Young diagram, of size $\tau \times (\nu +\tau +2) $ are composed of three pieces. The first piece is Durfee rectangle with enumerator $q^{\tau (m\nu +m\tau +m+a)}$, another is the piece right to the Durfee rectangle which are partitions with $\leq \tau$ parts and the last piece under the Durfee rectangle are the partitions into parts $\leq \nu +\tau+1$. It is clear from the definition of $M(a,m,\nu ;n)$ that the parts below and the parts right to the Durfee rectangle are generated by $q^{a+\nu (m\nu +m+2a)/2}/(q^a;q^m)_{\nu +\tau +1}$ and $1/(q^m;q^m)_\tau $, respectively. Here, $q^{a+\nu (m\nu +m+2a)/2}$ accounts for all parts $\equiv a\pmod{m}$ which are $\leq m\nu +a$.
\subsection*{Proof of Theorem \ref{Nk}}
Proof of Theorem \ref{Nk} is similar to that of \ref{M3}, so we omit the details.
\subsection*{Proof of Theorem \ref{T2}}
From the pentagonal number theorem, we have
\begin{align}\label{eqt1}
\nonumber(q;q)_\infty&=\sum_{\tau =0}^{\infty}(-1)^\tau q^{\tau (3\tau +1)/2}+\sum_{\tau =-1}^{-\infty}(-1)^\tau q^{\tau (3\tau +1)/2}\\&
=\sum_{\tau =0}^{\infty}(-1)^\tau q^{\tau (3\tau +1)/2}-\sum_{\tau =0}^{\infty}(-1)^\tau q^{\tau (3\tau +5)/2+1}.
\end{align}
Thus,
\begin{align}\label{eq1}
\nonumber&\dfrac{1}{(q;q)_\infty}\left(\sum_{\tau =0}^{\nu }(-1)^\tau q^{\tau (3\tau +1)/2}-\sum_{\tau =0}^{\nu -1}(-1)^\tau q^{\tau (3\tau +5)/2+1}\right)\\&
\nonumber=1-\dfrac{1}{(q;q)_\infty}\sum_{\tau =\nu +1}^{\infty}(-1)^\tau q^{\tau (3\tau +1)/2}+\dfrac{1}{(q;q)_\infty}\sum_{\tau =\nu }^{\infty}(-1)^\tau q^{\tau (3\tau +5)/2+1}\\&\nonumber
=1+\dfrac{(-1)^\nu q^{2+\nu (3\nu +7)/2}}{(q;q)_\infty}\sum_{\tau =0}^{\infty}(-1)^\tau q^{\tau (3\tau +7)/2+3\nu \tau }+\dfrac{(-1)^\nu q^{1+\nu (3\nu +5)/2}}{(q;q)_\infty}\sum_{\tau =0}^{\infty}(-1)^\tau q^{\tau (3\tau +5)/2+3\nu \tau }\\&
\nonumber=1+\dfrac{(-1)^\nu q^{2+\nu (3\nu +7)/2}}{(q;q)_\infty}\lim_{\delta\rightarrow 0}\ _2\phi_1\left(\genfrac{}{}{0pt}{}{q^3,\dfrac{q^{3\nu +5}}{\delta}}{0};q^3,\delta\right)\\&\quad
\nonumber+\dfrac{(-1)^\nu q^{1+\nu (3\nu +5)/2}}{(q;q)_\infty}\lim_{\delta\rightarrow 0} \ _ 2\phi_1\left(\genfrac{}{}{0pt}{}{q^3,\dfrac{q^{3\nu +4}}{\delta}}{0};q^3,\delta\right)\\&
\nonumber=1+\dfrac{(-1)^\nu q^{2+\nu (3\nu +7)/2}}{(q;q)_\infty}\lim_{\delta\rightarrow 0}\dfrac{(q^{3\nu +5};q^3)_\infty}{(\delta;q^3)_\infty}\sum_{\tau =0}^{\infty}\dfrac{(-1)^\tau \delta^\tau q^{3\tau (\tau +1)/2}\left(q^{3\nu +5}/\delta;q^3\right)_\tau }{(q^3;q^3)_\tau (q^{3\nu +5};q^3)_\tau }\\&\quad
\nonumber+\dfrac{(-1)^\nu q^{1+\nu (3\nu +5)/2}}{(q;q)_\infty}\lim_{\delta\rightarrow 0}\dfrac{(q^{3\nu +4};q^3)_\infty}{(\delta;q^3)_\infty}\sum_{\tau =0}^{\infty}\dfrac{(-1)^\tau \delta^\tau q^{3\tau (\tau +1)/2}\left(q^{3\nu +4}/\delta;q^3\right)_\tau }{(q^3;q^3)_\tau (q^{3\nu +4};q^3)_\tau }
\\&
\nonumber=1+(-1)^\nu q^{2+\nu (3\nu +7)/2}\dfrac{(q^{3\nu +5};q^3)_\infty}{(q;q)_\infty}\sum_{\tau =0}^{\infty}\dfrac{q^{\tau (3\nu +3\tau +5)}}{(q^3;q^3)_\tau \left(q^{3\nu +5};q^3\right)_\tau }\\&\quad
\nonumber+(-1)^\nu q^{1+\nu (3\nu +5)/2}\dfrac{(q^{3\nu +4};q^3)_\infty}{(q;q)_\infty}\sum_{\tau =0}^{\infty}\dfrac{q^{\tau (3\nu +3\tau +4)}}{(q^3;q^3)_\tau \left(q^{3\nu +4};q^3\right)_\tau }\\&
\nonumber=1+(-1)^\nu q^{2+\nu (3\nu +7)/2}\frac{(q^2;q^3)_\infty}{(q;q)_\infty}\sum_{\tau =0}^{\infty}\frac{q^{\tau (3\nu +3\tau +5)}}{(q^3;q^3)_\tau (q^2;q^3)_{\nu +\tau +1}}\\&\quad
+(-1)^\nu q^{1+\nu (3\nu +5)/2}\frac{(q;q^3)_\infty}{(q;q)_\infty}\sum_{\tau =0}^{\infty}\frac{q^{\tau (3\nu +3\tau +4)}}{(q^3;q^3)_\tau (q;q^3)_{\nu +\tau +1}}.
\end{align}
Using Theorem \ref{M3} and equation \ref{eq1}, we arrive at \eqref{pi}.
\subsection*{Proof of Theorem \ref{T3}}\label{OP}
From the Gauss second identity \eqref{psi}, we have
\begin{align}
\nonumber&\dfrac{1}{\psi(-q)}\left(\sum_{\tau =0}^{\nu }(-1)^\tau q^{\tau (2\tau +1)}-\sum_{\tau =0}^{\nu -1}(-1)^\tau q^{(\tau +1)(2\tau +1)}\right)\\&\nonumber=1-\dfrac{1}{\psi(-q)}\sum_{\tau =\nu +1}^{\infty}(-1)^\tau q^{\tau (2\tau +1)}+\dfrac{1}{\psi(-q)}\sum_{\tau =\nu }^{\infty}(-1)^\tau q^{(\tau +1)(2\tau +1)}
\\&\nonumber
=1+(-1)^\nu \dfrac{q^{3+\nu (2\nu +5)}}{\psi(-q)}\sum_{\tau =0}^{\infty}(-1)^\tau q^{\tau (4\nu +2\tau +5)}+(-1)^\nu \dfrac{q^{1+\nu (2\nu +3)}}{\psi(-q)}\sum_{\tau =0}^{\infty}(-1)^\tau q^{\tau (4\nu +2\tau +3)}\\&\nonumber
=1+(-1)^\nu \dfrac{q^{3+\nu (2\nu +5)}}{\psi(-q)}\lim_{\delta\rightarrow 0}\ _2\phi_1\left(\genfrac{}{}{0pt}{}{q^4,\dfrac{q^{4\nu +7}}{\delta}}{0};q^4,\delta\right)+(-1)^\nu \dfrac{q^{1+\nu (2\nu +3)}}{\psi(-q)}\lim_{\delta\rightarrow 0}\ _2\phi_1\left(\genfrac{}{}{0pt}{}{q^4,\dfrac{q^{4\nu +5}}{\delta}}{0};q^4,\delta\right)
\\&\nonumber
=1+(-1)^\nu \dfrac{q^{3+\nu (2\nu +5)}}{\psi(-q)}\lim_{\delta\rightarrow 0}\dfrac{(q^{4\nu +7};q^4)_\infty}{(\delta;q^4)_\infty}\sum_{\tau =0}^{\infty}\dfrac{(-1)^\tau \delta^\tau q^{2\tau (\tau +1)}(q^{4\nu +7}/\delta;q^4)_\tau }{(q^4;q^4)_\tau (q^{4\nu +7};q)_\tau }\\&\nonumber\quad
+(-1)^\nu \dfrac{q^{1+\nu (2\nu +3)}}{\psi(-q)}\lim_{\delta\rightarrow 0}\dfrac{(q^{4\nu +5};q^4)_\infty}{(\delta;q^4)_\infty}\sum_{\tau =0}^{\infty}\dfrac{(-1)^\tau \delta^\tau q^{2\tau (\tau +1)}(q^{4\nu +5}/\delta;q^4)_\tau }{(q^4;q^4)_\tau (q^{4\nu +5};q)_\tau }\\&\nonumber
=1+(-1)^\nu q^{3+\nu (2\nu +5)}\frac{(q^3;q^4)_\infty}{\psi(-q)}\sum_{\tau =0}^{\infty}\frac{q^{\tau (4\nu +4\tau +7)}}{(q^4;q^4)_\tau (q^3;q^4)_{\nu +\tau +1}}\\&\quad
+(-1)^\nu q^{1+\nu (2\nu +3)}\frac{(q;q^4)_\infty}{\psi(-q)}\sum_{\tau =0}^{\infty}\frac{q^{\tau (4\nu +4\tau +5)}}{(q^4;q^4)_\tau (q;q^4)_{\nu +\tau +1}}.\label{eq2}
\end{align}
Multiplying equation \eqref{eq2} by $(q^4;q^4)_\infty/(q^2;q^2)_\infty$ and using $\psi(-q)=\frac{(q;q)_\infty(q^4;q^4)_\infty}{(q^2;q^2)_\infty}$, we obtain
\begin{align*}
\dfrac{1}{(q;q)_\infty}&\left(\sum_{\tau =0}^{\nu }(-1)^\tau q^{\tau (2\tau +1)}-\sum_{\tau =0}^{\nu -1}(-1)^\tau q^{(\tau +1)(2\tau +1)}\right)\\&\nonumber=\dfrac{(q^4;q^4)_\infty}{(q^2;q^2)_\infty}+
(-1)^\nu q^{3+\nu (2\nu +5)}\frac{(q^3;q^4)_\infty}{(q;q)_\infty}\sum_{\tau =0}^{\infty}\frac{q^{\tau (4\nu +4\tau +7)}}{(q^4;q^4)_\tau (q^3;q^4)_{\nu +\tau +1}}\\&\quad+(-1)^\nu q^{1+\nu (2\nu +3)}\frac{(q;q^4)_\infty}{(q;q)_\infty}\sum_{\tau =0}^{\infty}\frac{q^{\tau (4\nu +4\tau +5)}}{(q^4;q^4)_\tau (q;q^4)_{\nu +\tau +1}},
\end{align*}
which implies that
\begin{align}\label{dop2}
\nonumber(-1)^\nu &\sum_{n=0}^{\infty}p(n)q^n\left(\sum_{\tau =0}^{\nu }(-1)^\tau q^{\tau (2\tau +1)}-\sum_{\tau =0}^{\nu -1}(-1)^\tau q^{(\tau +1)(2\tau +1)}\right)\\&
=(-1)^\nu \sum_{n=0}^{\infty}p_o(n)q^{2n}+\sum_{n=0}^{\infty}p(3,4,\nu ;n)q^n+\sum_{n=0}^{\infty}p(1,4,\nu ;n)q^n.
\end{align}
Equating coefficients of $q^{2n}$ and $q^{2n+1}$ on both sides of \eqref{dop2}, we arrive at \eqref{p1i} and \eqref{p2i}, respectively.
\subsection*{Proof of Theorem \ref{T4}}\label{DOP}
Proof of \eqref{dpi} is similar to that of \eqref{p2i} but instead of multiplying $(q^4;q^4)_\infty/(q^2;q^2)_\infty$, we multiply $(q^4;q^4)_\infty$.
\subsection*{Proof of Theorem \ref{T5}}
Equation \eqref{podi} follows from \eqref{eq2} and Theorem \ref{M3}.
\subsection*{Proof of Theorem \ref{T6}}
From the Gauss identity \eqref{phi}, we have
\begin{align}
\nonumber\dfrac{1}{\phi(-q)}\left(1+2\sum_{\tau =1}^{\nu }(-1)^\tau q^{\tau ^2}\right)&=1-\dfrac{2}{\phi(-q)}\sum_{\tau =\nu +1}^{\infty}(-1)^\tau q^{\tau ^2}
\\&\nonumber=1+\dfrac{2(-1)^\nu q^{(\nu +1)^2}}{\phi(-q)}\sum_{\tau =0}^{\infty}(-1)^\tau q^{\tau ^2+2\tau (\nu +1)}\\&
\nonumber=1+\dfrac{2(-1)^\nu q^{(\nu +1)^2}}{\phi(-q)}\lim_{\delta\rightarrow 0}\ _2\phi_1\left(\genfrac{}{}{0pt}{}{q^2,\dfrac{q^{2\nu +3}}{\delta}}{0};q^2,\delta\right)\\&\nonumber
=1+\dfrac{2(-1)^\nu q^{(\nu +1)^2}}{\phi(-q)}\lim_{\delta\rightarrow 0}\dfrac{(q^{2\nu +3};q^2)_\infty}{(\delta;q)_\infty}\sum_{\tau =0}^{\infty}\dfrac{(-1)^\tau \delta^\tau q^{\tau (\tau +1)}(q^{2\nu +3}/\delta;q^2)_\tau }{(q^2;q^2)_\tau (q^{2\nu +3};q^2)_\tau }\\&\nonumber
=1+\dfrac{2(-1)^\nu q^{(\nu +1)^2}}{\phi(-q)}(q^{2\nu +3};q^2)_\infty\sum_{\tau =0}^{\infty}\dfrac{q^{2\tau ^2+(2\nu +3)\tau }}{(q^2;q^2)_\tau (q^{2\nu +3};q^2)_\tau }\\&\nonumber
=1+\dfrac{2(-1)^\nu q^{(\nu +1)^2}}{(q;q)_\infty(q;q^2)_\infty}(q^{2\nu +3};q^2)_\infty\sum_{\tau =0}^{\infty}\dfrac{q^{\tau (2\nu +2\tau +3)}}{(q^2;q^2)_\tau (q^{2\nu +3};q^2)_\tau }\\&
=1+\dfrac{2(-1)^\nu q^{(\nu +1)^2}}{(q;q)_\infty}\sum_{\tau =0}^{\infty}\dfrac{q^{\tau (2\nu +2\tau +3)}}{(q^2;q^2)_\tau (q;q^2)_{\nu +\tau +1}}.\label{eq6}
\end{align}
Equation \eqref{oppi} follows from \ref{eq6} and Theorem \ref{M3}.
\subsection*{Proof of Theorem \ref{T7}}
Multiplying the equation \eqref{eq6} by $1/(q^2;q^2)_\infty$ and using $\phi(-q)=(q;q)^2_\infty/(q^2;q^2)_\infty$, we obtain
\begin{equation*}\label{op1}
\dfrac{1}{(q;q)^2_\infty}\left(1+2\sum_{\tau =1}^{\nu }(-1)^\tau q^{\tau ^2}\right)=\dfrac{1}{(q^2;q^2)_\infty}+2(-1)^\nu \frac{q^{(\nu +1)^2}}{(q;q)_\infty(q^2;q^2)_\infty}\sum_{\tau =0}^{\infty}\frac{q^{\tau (2\nu +2\tau +3)}}{(q^2;q^2)_\tau (q;q^2)_{\nu +\tau +1}}
\end{equation*}
and the statements follow easily from Theorem \ref{M3} and the above equation.
\subsection*{Proof of Theorem \ref{T8}}
From the equation \eqref{eqt1}, we have
\begin{align}
\nonumber&\dfrac{1}{(q;q)_\infty}=1-\dfrac{1}{(q;q)_\infty}\sum_{\tau =1}^{\infty}(-1)^\tau q^{\tau (3\tau +1)/2}+\dfrac{1}{(q;q)_\infty}\sum_{\tau =0}^{\infty}(-1)^\tau q^{\tau (3\tau +5)/2+1}\\&\nonumber
=1+\dfrac{q^2}{(q;q)_\infty}\sum_{\tau =0}^{\infty}(-1)^\tau q^{\tau (3\tau +7)/2}+\dfrac{q}{(q;q)_\infty}\sum_{\tau =0}^{\infty}(-1)^\tau q^{\tau (3\tau +5)/2}\\&
\nonumber=1+\dfrac{q^2}{(q;q)_\infty}\lim_{\delta\rightarrow 0}\ _2\phi_1\left(\genfrac{}{}{0pt}{}{q^3,\dfrac{q^5}{\delta}}{0};q^3,\delta\right)+\dfrac{q}{(q;q)_\infty}\lim_{\delta\rightarrow 0} \ _ 2\phi_1\left(\genfrac{}{}{0pt}{}{q^3,\dfrac{q^4}{\delta}}{0};q^3,\delta\right)\\&
\nonumber=1+\dfrac{q^2}{(q;q)_\infty}\lim_{\delta\rightarrow 0}\dfrac{(q^5;q^3)_\infty}{(\delta;q^3)_\infty}\sum_{\tau =0}^{\infty}\dfrac{(-1)^\tau \delta^\tau q^{3\tau (\tau +1)/2}\left(q^5/\delta;q^3\right)_\tau }{(q^3;q^3)_\tau (q^5;q^3)_\tau }\\&\quad
\nonumber+\dfrac{q}{(q;q)_\infty}\lim_{\delta\rightarrow 0}\dfrac{(q^4;q^3)_\infty}{(\delta;q^3)_\infty}\sum_{\tau =0}^{\infty}\dfrac{(-1)^\tau \delta^\tau q^{3\tau (\tau +1)/2}\left(q^4/\delta;q^3\right)_\tau }{(q^3;q^3)_\tau (q^4;q^3)_\tau }
\\&
\nonumber=1+q^{2}\dfrac{(q^{5};q^3)_\infty}{(q;q)_\infty}\sum_{\tau =0}^{\infty}\dfrac{q^{\tau (3\tau +5)}}{(q^3;q^3)_\tau \left(q^{5};q^3\right)_\tau }+q\dfrac{(q^{4};q^3)_\infty}{(q;q)_\infty}\sum_{\tau =0}^{\infty}\dfrac{q^{\tau (3\tau +4)}}{(q^3;q^3)_\tau \left(q^{4};q^3\right)_\tau }\\&
=1+q^{2}\frac{(q^2;q^3)_\infty}{(q;q)_\infty}\sum_{\tau =0}^{\infty}\frac{q^{\tau (3\tau +5)}}{(q^3;q^3)_\tau (q^2;q^3)_{\tau +1}}
+q\frac{(q;q^3)_\infty}{(q;q)_\infty}\sum_{\tau =0}^{\infty}\frac{q^{\tau (3\tau +4)}}{(q^3;q^3)_\tau (q;q^3)_{\tau +1}}.\label{eqt8}
\end{align}
Using equation \eqref{eqt8} and Theorem \ref{M3}, we arrive at \eqref{eq8}.
Similarly one can prove \eqref{eq9} by using Gauss second identity \eqref{psi}.
\subsection*{Proof of Theorem \ref{CP}} First, we will prove the following lemma:
\begin{lemma}
For $\nu \geq 1$ and $m\geq 0$, we have
\begin{equation}\label{l1}
\sum_{n=0}^{\infty}N_\nu (n)q^n=\sum_{n=0}^{\infty}p(n)q^n\sum_{\tau =0}^{m}(-1)^\tau q^{(\nu +\tau )(\nu +\tau -1)/2}+(-1)^{m+1}\sum_{n=0}^{\infty}N_{\nu +m+1}(n)q^n.
\end{equation}
\end{lemma}
\begin{proof}
From the Theorem \ref{Nk}, we have
\begin{align}\label{p1}
\nonumber\sum_{n=0}^{\infty}N_\nu (n)q^n&=q^{\nu (\nu -1)/2}\sum_{\tau =0}^{\infty}\frac{q^{\tau (\nu +\tau )}}{(q;q)_\tau (q;q)_{\nu +\tau -1}}\\&\nonumber
=q^{\nu (\nu -1)/2}\sum_{\tau =0}^{\infty}\frac{(1-q^{\nu +\tau })q^{\tau (\nu +\tau )}}{(q;q)_\tau (q;q)_{\nu +\tau }}\\&\nonumber
=q^{\nu (\nu -1)/2}\sum_{\tau =0}^{\infty}\frac{q^{\tau (\nu +\tau )}}{(q;q)_\tau (q;q)_{\nu +\tau }}-q^{\nu (\nu +1)/2}\sum_{\tau =0}^{\infty}\frac{q^{\tau (\nu +\tau +1)}}{(q;q)_\tau (q;q)_{\nu +\tau }}\\&\
=q^{\nu (\nu -1)/2}\sum_{n=0}^{\infty}p(n)q^n-\sum_{n=0}^{\infty}N_{\nu +1}(n)q^n,\quad (\text{from \cite[p. 91]{GH} and \eqref{nk1}})
\end{align}
which is $m=0$ case of \eqref{l1}. Now assume that \eqref{l1} is holds for some $m\geq 0$, then, from \eqref{l1} and \eqref{p1}, we have
\begin{align*}
\sum_{n=0}^{\infty}N_\nu (n)q^n&=\sum_{n=0}^{\infty}p(n)q^n\sum_{\tau =0}^m(-1)^\tau q^{(\nu +\tau )(\nu +\tau -1)/2}\\&\quad+(-1)^{m+1}\left(q^{(\nu +m)(\nu +m+1)/2}\sum_{n=0}^{\infty}p(n)q^n-\sum_{n=0}^{\infty}N_{\nu +m+2 }(n)q^n\right)\\&
=\sum_{n=0}^{\infty}p(n)q^n\sum_{\tau =0}^{m+1}(-1)^\tau q^{(\nu +\tau )(\nu +\tau -1)/2}+(-1)^{m+2}\sum_{n=0}^{\infty}N_{\nu +m+2}(n)q^n,
\end{align*}
which is the $m+1$ case of \eqref{l1}.
\end{proof}
For $n\leq (\nu +m)(\nu +m+1)/2$, equating coefficients of $q^n$ on both sides of \eqref{l1},
\begin{equation}\label{l4}
\sum_{\tau =0}^{m}(-1)^\tau p\left(n-\dfrac{(\nu +\tau )(\nu +\tau -1)}{2}\right)=N_\nu (n).
\end{equation}
Now the statement \eqref{CP1} is obvious from \eqref{l4} for sufficiently large $m$.
\@startsection{section}{1{Inequalities}\label{In}
Let us define $pp(n)$ to be the number of bipartitions of $n$, and $p_o(n)$ be the number of partitions of $n$ into odd parts. The generating function for $pp(n)$ and $p_o(n)$ are given by \[\sum_{n=0}^{\infty}pp(n)=\dfrac{1}{(q;q)^2_\infty}\]
and
\[\sum_{n=0}^{\infty}p_o(n)=\dfrac{1}{(q;q^2)_\infty}=\dfrac{(q^2;q^2)}{(q;q)_\infty},\] respectively.
\begin{corollary}
For $n$ odd or $\nu $ even,
\begin{equation}\label{bpc}
(-1)^\nu \left(pp(n)+2\sum_{\tau =1}^{\nu }(-1)^\tau pp(n-\tau ^2)\right)\geq 0,
\end{equation}
with strict inequality if $n\geq (\nu +1)^2$.
\end{corollary}
\begin{proof}
Using the fact that $\phi(-q)=\dfrac{(q;q)^2_\infty}{(q^2;q^2)_\infty}$ in \eqref{eq6} and after rearranging the terms, we obtain
\begin{align}\label{pp}
\dfrac{(-1)^\nu }{(q;q)^2_\infty}\left(1+2\sum_{\tau =1}^{\nu }(-1)^\tau q^{\tau ^2}\right)=\dfrac{(-1)^\nu }{(q^2;q^2)_\infty}+2B_\nu (q),
\end{align}
where
\[B_\nu (q)=\frac{q^{(\nu +1)^2}}{(q;q)_\infty(q^2;q^2)_\infty}\sum_{\tau =0}^{\infty}\frac{q^{2\tau ^2+(2\nu +3)\tau }}{(q^2;q^2)_\tau (q;q^2)_{\nu +\tau +1}}\]
has nonnegative coefficient. Since $\dfrac{1}{(q^2;q^2)_\infty}$ is an even function of $q$ with positive coefficients; every term in \eqref{pp} is nonnegative if $\nu$ is even, however if $\nu$ is odd every coeeficients of odd powers are nonnegative.
\end{proof}
\begin{corollary}
With strict inequality if $n> \nu (2\nu +3)$. We have
\item[1.] For $n$ odd or $\nu $ even,
\begin{align}\label{p12}
(-1)^{\nu -1}\sum_{\tau =0}^{\nu -1}(-1)^\tau \big(p\left(n-\tau (2\tau +1)\right)-p\left(n-(\tau +1)(2\tau +1)\right)\big)\leq p(n-\nu (2\nu +1)).
\end{align}
\item[2.] For odd $n$,
\begin{align}
(-1)^{\nu -1}\sum_{\tau =0}^{\nu -1}(-1)^\tau \big(p_o\left(n-\tau (2\tau +1)\right)-p_o\left(n-(\tau +1)(2\tau +1)\right)\big)\leq p_o\left(n-\nu (2k+1)\right).
\end{align}
\end{corollary}
\begin{proof}
Proof is similar to that of \eqref{bpc} but instead of using \eqref{eq6}, we use \eqref{eq2}. So we omit the details here.
\end{proof}
|
\section{Theoretical model}\label{model}
We start with the following higher order dimensionless NLH equation describing pulse propagation in the nonparaxial regime \cite{mc2012}
\begin{equation}
\begin{array}{lcl}\label{1.0}
i(\frac{\partial \psi}{\partial \zeta} + \alpha \frac{\partial \psi}{\partial \tau}) + \frac{a_1}{2} \frac{\partial^2 \psi}{\partial \tau^2} + a_2 |\psi|^2 \psi + i[a_3 \frac{\partial^3 \psi}{\partial \tau^3}+ a_4 \frac{\partial (|\psi|^2 \psi)}{\partial \tau} + a_5 \psi \frac{\partial (|\psi|^2)}{\partial \tau}] + a_6 \frac{\partial^2 \psi}{\partial \zeta^2} + a_7 \frac{\partial^4 \psi}{\partial \tau^4} = 0
\end{array}
\end{equation}
where $\psi(\zeta,\tau)$ is the complex pulse envelope, $\zeta$ and $\tau$ are dimensionless space and time variables respectively. The coefficients $a_i, i=1,2,3,7$ are respectively the group velocity dispersion (GVD) , the self phase modulation (SPM), the TOD, and the FOD. $a_2 = 1$, $a_1 = \pm 1$ indicates anomalous and normal GVD respectively. The term proportional to $a_4$ corresponds to
SS \cite{anda}. This causes a temporal pulse distortion, which leads to the formation of an optical shock at the trailing edge of the pulse \cite{oliv} unless balanced by the dispersion. The term proportional to $a_5$ accounts for the delayed Raman response \cite{yui} which is responsible for the pulse undergoing a frequency shift known as SFS \cite{mit}. The term proportional to $a_6 (>0)$ (nonparaxial parameter) refers to the spatial dispersion arising from the nonparaxial effect.\\
The term $i\alpha \frac{\partial \psi}{\partial \tau}$ in (\ref{1.0}) can be transformed away by the Galilean boost
\begin{equation}
t = \tau - \alpha \zeta, ~~~~ x = \zeta
\end{equation}
i.e. the solutions to (\ref{1.0}) describe waves in a frame moving with speed $\frac{1}{\alpha}$ along the $\zeta$ axis. Thus Eq. (\ref{1.0}) is transformed into
\begin{equation}\label{1.1a}
\begin{split}
i\frac{\partial \psi}{\partial x} + \frac{1}{2}(a_1 + 2 a_6 \alpha^2) \frac{\partial^2 \psi}{\partial t^2} - 2 \alpha a_6 \frac{\partial^2 \psi}{\partial x \partial t} + a_2 |\psi|^2 \psi + i[a_3 \frac{\partial^3 \psi}{\partial t^3}+ a_4 \frac{\partial (|\psi|^2 \psi)}{\partial t} + a_5 \psi \frac{\partial (|\psi|^2)}{\partial t}]\\ + a_6 \frac{\partial^2 \psi}{\partial x^2} + a_7 \frac{\partial^4 \psi}{\partial t^4} = 0
\end{split}
\end{equation}
Now we restrict to solutions for which $\alpha a_6 << O(1)$ and $a_6 \alpha^2 << O(1)$. Also the terms $\frac{\partial^2 \psi}{\partial t^2}$ and $\frac{\partial^2 \psi}{\partial x \partial t}$ are assumed to be both $O(1)$. These solutions will satisfy the equation
\begin{equation}\label{1.1b}
\begin{array}{lcl}
i\frac{\partial \psi}{\partial x} + \frac{a_1}{2} \frac{\partial^2 \psi}{\partial t^2} + a_2 |\psi|^2 \psi + i[a_3 \frac{\partial^3 \psi}{\partial t^3}+ a_4 \frac{\partial (|\psi|^2 \psi)}{\partial t} + a_5 \psi \frac{\partial (|\psi|^2)}{\partial t}] + a_6 \frac{\partial^2 \psi}{\partial x^2} + a_7 \frac{\partial^4 \psi}{\partial t^4} = 0
\end{array}
\end{equation}
In what follows we shall take $a_1, a_2 = 1$. It is important to note that the term $a_6 \frac{\partial^2 \psi}{\partial x^2}$ in (\ref{1.1b}) arising from the absence of SVEA, leads to a dispersion relation that supports both forward and backward propagating waves. This is in sharp contrast to NLS \cite{zak1973} and Manakov \cite{mana1974} equations where the backward propagating waves are not allowed \cite{chris06}.
To obtain the dipole solitary waves we use the following traveling wave ansatz
\begin{equation}\label{1.2}
\psi = u(\xi)e^{i\phi(x,t)},~~\phi(x,t)=k x-\omega t+\phi_0,~~\xi = \beta(t-cx),
\end{equation}
where $u$ is the amplitude and $\phi$ is the phase of the solitary wave. These are real functions of $\xi$. The parameter $c, k, \phi_0$ are the inverse velocity, wave number and real constant respectively.
On substituting this ansatz in Eq. (\ref{1.1b}) and separating the real and imaginary parts we obtain
\begin{equation}\label{1.5}
\begin{array}{lcl}
a_7 \beta^4 u''''+ (\frac{1}{2}a_1 + a_6 c^2 + 3 a_3 \omega -6 a_7 \omega^2) \beta^2 u''+(a_2 +
\omega a_4)u^3 -(k + k^2 a_6 + \frac{1}{2} \omega^2a_1 + \omega^3 a_3 - \omega^4 a_7) u = 0\,,
\end{array}
\end{equation}
and further
\begin{equation}\label{1.6}
\begin{array}{lcl}
(a_3 -4\omega a_7) \beta^3 u''' +(3a_4 + 2a_5)\beta u^2 u' -(c+2 k c a_6 + \omega a_1 +3\omega^2 a_3 -4\omega^3 a_7) \beta u' = 0\,
\end{array}
\end{equation}
\section{Dipole/quadrupole solitary waves}\label{sol}
{\bf Solution I. Dipole Soliton}. It is easily checked by direct substitution that Eqs. (\ref{1.5}) and (\ref{1.6}) admit the following dipole solitary wave solution\\
\begin{equation}\label{1.12}
u(\xi) = A sech(\xi) \tanh(\xi)\,,
\end{equation}
provided
\begin{subequations}\label{6}
\begin{align}
2a_5+3a_4 & = 0 \\
a_3-4\omega a_7 & = 0 \\
(1+2ka_6)c & = -\omega(a_1+8\omega^2a_7) \\ \
(a_2+\omega a_4)A^2 & = 120 a_7\b^4
\end{align}
\end{subequations}
and further
\begin{equation}\label{8}
\begin{array}{lcl}
\frac{1}{2}a_1+c^2a_6 =-2(3\omega^2-5\b^2)a_7\\
k+k^2a_6+\frac{1}{2}\omega^2a_1 =(11\b^4-3\omega^4)a_7
\end{array}
\end{equation}
It is relevant to mention that similar constraint relation between the SS and the SFS (Eqn.\ref{6}(a)) has also been obtained in \cite{li,Rel2,chou,triki,Rel5} for obtaining different solitary wave solutions of nonlinear Sch\"odinger equation containing higher order dispersion and non-Kerr nonlinearities and also for obtaining chirped solitary wave solutions of coupled cubic NLH equation in the presence of SS and SFS \cite{naresh}.\\
We now show that the dipole solitary wave
(\ref{1.12}) is indeed a dipole soliton by considering the collision
dynamics as depicted in Fig. 1.
For this the initial condition chosen is
\begin{equation}\label{12}
\begin{array}{rcr
\Psi(0,t) = \psi(0,t) -1.5~sech(0.2t+5)~\tanh(0.2t+5) \displaystyle e^{i(-\omega t+\phi_0)}
\end{array}
\end{equation}
The two dipole solitary waves converge during propagation and at a distance of about $x = 10~ \rm units$ they collide as shown in Fig.1. It is quite evident that the two dipole solitary waves collide elastically and continue their propagation after the collision. The intensity profile of the dipole soliton (\ref{1.12}) with respect to $t$ is shown in Fig.2(a) while the numerical simulation of the intensity profile is shown in Fig.2(c) demonstrating its stable evolution for the chosen values of the parameters. From Fig.2(b) we can see that as the nonparaxial parameter $a_6$ increases, the amplitude $A$ increases but the speed $|c|$ and the pulse width $\frac{1}{\beta}$ both decrease.
\begin{figure}[
\centering
\includegraphics[width=6cm,height=3cm]{interaction_sechtanh.eps}
\caption{Evolution of the intensity profile and collision dynamics of (\ref{1.12}) as a function of propagation distance $x$ for $a_3=\frac{1}{60},~a_4=0.126,~a_5=-0.189,~a_6=0.5,~a_7=0.15,\beta=0.577914,~\omega=0.027778,~\phi_0=0.5,~c=0.023777,~k=-2.169328,~A=1.4195$.
}
\end{figure}
\begin{figure
\centering
\subfloat[]{\includegraphics[width=4.5cm,height=4.0cm]{int_sechtanh.eps}
~~
\subfloat[]{\includegraphics[width=4.5cm,height=4.0cm]{p2_sechtanh.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{evo_sechtanh.eps}}\\%fig2(c)
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{evo_sechtanh10n.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{rob1_sechtanh.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{rob2_sechtanh.eps}
\caption{(a) plot of $|\psi|^2$ versus $t$ at $x=0$ (c), (d) numerical simulations of the intensity profile without and with $10\%$ white noise respectively of the solution (\ref{1.12}) for $a_3=\frac{1}{60},~a_4=0.126,~a_5=-0.189,~a_6=0.5,~a_7=0.15,\beta=0.577914,~\omega=0.027778,~\phi_0=0.5,~c=0.023777,~k=-2.169328,~A=1.4195$ respectively, (b) plots of $|c|$, $\frac{1}{\beta}$, $A$ versus $a_6$ for the solution (\ref{1.12}) for $a_3=-0.18,~a_4=0.126,~a_7=0.15$, (e) plot of the numerical simulation of the intensity profile by adding $0.003$ to the parameters $a_i, i = 1,2, \cdots 7$ and (f) plot of the numerical simulation of the intensity profile by adding the overall factor $0.001$ to the parametric conditions (\ref{6}) \& (\ref{8}) respectively.}
\end{figure}
{\bf Solution II. Periodic wave exhibiting dipole/quadrupole structure within a period}\\ \\
Eqns.(\ref{1.5}) and (\ref{1.6}) also admit the following elliptic solitary wave solution
\begin{equation}\label{1.20}
u(\xi) = A\sqrt{m} sn(\xi, m) dn(\xi, m)\,,
\end{equation}
($m$ being the modulus of the Jacobi elliptic functions $sn, dn$, $0< m <1$ \cite{abra}) provided Eqs.(6a) - (6d) are satisfied and in addition, the following two relations are also satisfied
\begin{equation}\label{1.21}
\begin{array}{lcl}
\frac{1}{2}a_1 + c^2 a_6 = -2[3\omega^2-5(2m-1) \beta^2] a_7\\
k + k^2 a_6 + \frac{1}{2}\omega^2a_1 = [(11+64m -64 m^2) \beta^4 - 3\omega^4] a_7
\end{array}
\end{equation}
It is easily checked that this solution exhibits dipole structure within a period so long $m\leq 0.5$. On the other hand, for values of $m > 0.5$, it exhibits quadrupole structure within a period. Fig.3(a) and 4(a) respectively depict the dipole and the quadrupole intensity profiles of (\ref{1.20}) respectively while Figs. 3(c) and 4(c) depict the numerical simulation of the intensity profiles of the dipole and quadrupole structure respectively. As is evident from the figures, both the dipole and the quadrupole structures exhibit stable evolutions for the chosen parameters. From Figs.3(b) and 4(b), we can see that the speed $|c|$ of both the elliptic waves having dipole and quadrupole structure within a period decreases as $a_6$ increases. Fig.3(b) shows that both the amplitude and the pulse width of the elliptic wave having dipole structure within a period increase for increasing $a_6$ but for elliptic wave having quadrupole structure within a period, while the amplitude increases, the pulse width decreases as $a_6$ increases as shown in Fig. 4(b) . It is worth noting that the solution (\ref{1.12}) is the $m=1$ limit of the elliptic wave solution (\ref{1.20}).\\ \\
\begin{figure}[
\begin{center}
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{int_sndn_d.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{p2_sndn_d.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{evo_sndn_d.eps}}\\%fig3(c)
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{evo_sndn_d10n.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{rob1_sndn_d.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{rob2_sndn_d.eps}
\caption{(a) plot of $|\psi|^2$ versus $t$ at $x=0$ (c), (d) numerical simulation of the intensity profile without and with $10\%$ white noise respectively of the solution (\ref{1.20}) (dipole) for $a_3=-0.165022,~a_4=0.1671553,~a_5=0.250733,~a_6=4.31016,~a_7=-0.006895,~\beta=0.0103625,~\omega=5.982461,~\phi_0=0.5,~c=0.476998,~k=1.301437,~A=0.2141115,~m=0.5$ respectively, (b) plots of $|c|$ , $\frac{1}{\beta}$, $A$ versus $a_6$ for the solution (\ref{1.20}) for $a_3=0.4,~a_4=0.5,~a_7=-0.04,~m=0.4$, (e) plot of numerical simulation of the intensity profile by adding $0.003$ to the parameters $a_i, i = 1,2, \cdots 7$ and (f) plot of numerical simulation of the intensity profile by adding the overall factor $0.001$ to the parametric conditions (\ref{6}) \& (\ref{1.21}) respectively.}
\end{center}
\end{figure}
\begin{figure}[
\begin{center}
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{int_sndn_q.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{p2_sndn_q.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{evo_sndn_q.eps}}\\%fig4(c)
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{evo_sndn_q10n.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{rob1_sndn_q.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{rob2_sndn_q.eps}
\caption{(a) plot of $|\psi|^2$ versus $t$ at $x=0$ (c), (d) numerical simulation of the intensity profile without and with $10\%$ white noise respectively of the solution (\ref{1.20}) (quadrupole) for $a_3=\frac{1}{60},~a_4=0.126,~a_5=-0.189,~a_6=0.95,~a_7=0.15,~\beta=1.291913,~\omega=0.027778,~\phi_0=0.5,~c=-0.004247,~k=2.919107,~A=7.0688,~m=0.6$ respectively, (b) plots of $|c|$, $\frac{1}{\beta}$ , $A$ versus $a_6$ of the solution (\ref{1.20}) for $a_3=0.4,~a_4=0.126,~a_7=0.4,~m=0.75$, (e) plot of numerical simulation of the intensity profile by adding $0.003$ to the parameters $a_i, i = 1,2, \cdots 7$ and (f) plot of numerical simulation of the intensity profiles by adding the overall factor $0.001$ to the parametric conditions (\ref{6}) \& (\ref{1.21}) respectively.}
\end{center}
\end{figure}
{\bf Solution III. Periodic wave exhibiting dipole structure within a period}\\ \\
Eqns.(\ref{1.5}) and (\ref{1.6}) admits yet another elliptic solitary wave
\begin{equation}\label{1.23}
u(\xi) = A m sn(\xi, m) cn(\xi, m)\,,
\end{equation}
($m, 0<m <1$ being the modulus of the Jacobi elliptic functions $sn, cn$, \cite{abra}) provided Eqs.(6a) - (6d) are satisfied and further the following relations are also satisfied
\begin{equation}\label{1.24}
\begin{array}{lcl}
\frac{1}{2}a_1 + c^2 a_6 = -2[3\omega^2-5(2-m) \beta^2] a_7\\
k + k^2 a_6 + \frac{1}{2}\omega^2a_1 = [(11m^2+64m -64) \beta^4 - 3\omega^4] a_7
\end{array}
\end{equation}
It is worth noting that when m=1, this periodic solution becomes the solution (\ref{1.12}). The intensity profile and numerical simulation of the intensity profile of the elliptic wave (\ref{1.23}) are shown in Figs.5(a) and (c) respectively. Stable evolution for the chosen values of the parameters are evident from the Fig. 5(c). From Figs.5(b), it is clear that the speed $|c|$ and the amplitude $A$ of this elliptic wave decrease as $a_6$ increases while the pulse-width $\frac{1}{\beta}$ increases as $a_6$ increases.\\
\begin{figure}[
\begin{center}
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{int_sncn.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{p2_sncn.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{evo_sncn.eps}}\\%fig5(c)
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{evo_sncn10n.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{rob1_sncn.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{rob2_sncn.eps}
\caption{(a) plot of $|\psi|^2$ versus $t$ at $x=0$ (c), (d) numerical simulation of the intensity profile without and with $10\%$ white noise respectively of the solution (\ref{1.23}) for $a_3=\frac{1}{60},~a_4=0.126,~a_5=-0.189,~a_6=0.1,~a_7=0.15,~\beta=0.506763,~\omega=0.027778,~\phi_0=0.5,~c=0.0285981,~k=-9.861067,~A=1.0876,~m=0.7$ respectively, (b) plots of $|c|$, $\frac{1}{\beta}$, $A$ versus $a_6$ of the solution (\ref{1.23}) for $a_3=0.8,~a_4=-0.5,~a_7=0.4,~m=0.7$, (e) plot of numerical simulation of the intensity profile by adding $0.003$ to the parameters $a_i, i = 1,2, \cdots 7$ and (f) plot of numerical simulation of the intensity profile by adding the overall factor $0.001$ to the parametric conditions (\ref{6}) \& (\ref{1.24}) respectively.}
\end{center}
\end{figure}
{\bf Solution IV. Periodic wave with arbitrary amplitude having dipole structure within a period}\\
Remarkably, Eqns.(\ref{1.5}) and (\ref{1.6}) also admit the following trigonometric solitary wave solution with arbitrary amplitude $A$
\begin{equation}\label{1.25}
u(\xi) = A {\rm sin}(\xi)\,,
\end{equation}
provided following relations are also satisfied
\begin{equation}\label{1.26}
\begin{array}{lcl}
3a_4+2a_5=0\\
4(a_3-4\omega a_7)\beta^2+(1+2ka_6)c+(\omega a_1+3\omega^2a_3-4\omega^3a_7)=0\\
\omega a_4+a_2=0\\
a_7(16\beta^4+\omega^4+24\beta^2\omega^2)-a_3\omega(12\beta^2+\omega^2)-2\beta^2(a_1+2a_6c^2)-(k+k^2a_6+\frac{1}{2}\omega^2a_1)=0\\
\end{array}
\end{equation}
The intensity profile and numerical simulation of the intensity profile of the solitary wave (\ref{1.25}) exhibiting stable evolution for the chosen parameters are shown in Figs.6(a) and (c) respectively. Fig. 6(b) shows that the
pulse-width $\frac{1}{\beta}$ increases as $a_6$ increases but $|c|$ decreases with increasing $a_6$.
\begin{figure}[
\begin{center}
\centering
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{int_sin.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{p2_sin.eps}
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{evo_sin.eps}}\\%fig6(c)
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{evo_sin10n.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{rob1_sin.eps}
~~
\subfloat[\label{}]{\includegraphics[width=4.5cm,height=4.0cm]{rob2_sin.eps}
\caption{(a) plot of $|\psi|^2$ versus $t$ at $x=0$ (c), (d) numerical simulation of the intensity profile without and with $10\%$ white noise respectively of the solution (\ref{1.25}) for $a_3=0.1,~a_4=-0.4,~a_5=0.6,~a_6=0.5,~a_7=0.04,~\beta=1.25169,~\omega=2.5,~\phi_0=0.5,~c=0.005,~k=0.0125112,~A=0.1$ respectively, (b) plots of $|c|$, $\frac{1}{\beta}$ versus $a_6$ respectively of the solution (\ref{1.25}) for $a_3=0.1,~a_4=0.5,~a_6=0.04,~a_7=0.8,~k=0.1$, (e) plot of numerical simulation of the intensity profile by adding $0.003$ to the parameters $a_i, i = 1,2, \cdots 7$ and (f) plot of numerical simulation of the intensity profile by adding the overall factor $0.001$ to the parametric conditions (\ref{6}) \& (\ref{1.26}) respectively.}
\end{center}
\end{figure}
\section{Stability analysis}
The physical importance of the solitary wave solutions rests on their stability. In this regard, it is worth noting that Figures 2(c), 3(c), 4(c), 5(c), and 6(c) all refer to an ideal setting, which is not always the case in real-world situations. There are a variety of factors that might induce instability in propagating structures \cite{triki2019}. As a result, it becomes vital to investigate the stability of solutions in the presence of external noise or perturbations. For this, $10\%$ white noise perturbation is added to the initial input,
as $\psi(x=0,t)=u(\beta t)e^{i\phi(x=0,t)}(1+\epsilon)$, where $u(\beta t)$ are exact solutions which are given by (8), (12), (14), (16) and the white noise $\epsilon$ is realized by utilizing a random matrix given by~\cite{zhen},
\begin{equation}\label{111}
\epsilon={\rm 10\%noise}= 0.1\sqrt{2}(rand(numel(t),1)-0.5)(1 + i)
\end{equation}
In MATLAB program, $numel(t)$ represents the number $N$ of discretized values of $t$ (refer Appendix) and $rand(numel(t),1)$ returns a $N\times1$ matrix of uniformly distributed random numbers on the open interval $(0,1)$. The term $(1+i)$ in (\ref{111}) signifies that $\psi(x=0,t)$ is perturbed by complex broadband perturbation.
The results of numerical simulation of the intensity profiles of the dipole soliton (\ref{1.12}), elliptic solitary wave (\ref{1.20}) having the dipole structure within a period for $m=0.5$ and having the quadrupole structure within a period for $m=0.6$ respectively, elliptic solitary wave (\ref{1.23}) having the dipole structure within a period and trigonometric solitary wave (\ref{1.25}) exhibiting the dipole structure within a period, by adding $10\%$ white noise are shown in Figs.2(d), 3(d), 4(d), 5(d), 6(d) respectively. The results reveal that the elliptic solitary wave (\ref{1.20}) having the dipole structure within a period for $m=0.5$ of the Jacobi elliptic function, has very weak instability in its propagation under $10\%$ ((Fig.3(d)) white noise but the quadrupole structure within a period for $m = 0.6$ (Fig.4(d)) shows stable evolution for the values of the parameters considered herein. But the other dipole solitary waves (\ref{1.23}) and (\ref{1.25}) propagate appreciably in a stable manner for the chosen parameters under finite initial white noise perturbations. It is also important to test the robustness of the solutions by violation of the parametric conditions. This is especially essential for experimental implementations where the values of the parameters are always approximate and may also fluctuate \cite{triki}. Here, the violations of the parametric conditions are done in two ways: (i) by adding $0.003$ to all the parameters $a_i, i = 1,2\cdots 7$ (ii) by adding an overall factor $0.001$ to the right side of the parametric conditions 9(a) - (d), (10), (13), (15), (17). The resulting numerical simulations are shown in Fig.2(e)-(f), Fig.3(e)-(f) $\&$ Fig.4(e)-(f), Fig.5(e)-(f), Figs.6(e)-(f) for the solitary waves (\ref{1.12}), (\ref{1.20}) for $m=0.5~\&~m=0.6$, (\ref{1.23}), (\ref{1.25}) respectively. It is evident from the above mentioned figures that the solitary waves are quite robust in both the cases of violations of the parametric conditions. It is to be noted that the nonlinear modes are stable under $10\%$ white noise perturbations and also under violations of the parametric conditions for both positive (Figs. 3, 4, 6 ) and negative (Figs. 1, 2, 5) $k$ values (similar to \cite{kanna}) when the other parameters are kept fixed at the chosen values. The positive and negative $k$ values signify that the direction of the wave vector $\vec{k}$ (in one dimension it is a scalar) for forward propagating wave is opposite to that of the backward propagating wave.
\section {Conclusion}\label{con}
In this article, we have studied the cubic NLH system in the presence of TOD, FOD and also non-Kerr nonlinearity like the SS and the SFS. This system models ultrashort nonparaxial pulse propagation with Kerr and non-Kerr nonlinearities, higher order dispersions, and spatial dispersion which becomes important when SVEA fails. We have demonstrated that this system features dipole soliton and also elliptic waves having the dipole/quadrupole structure within a period. The novelty of the propagation dynamics arising from incorporating the higher order dispersion and non-Kerr nonlinearity lies in the appearance of dipole structure within a period of these elliptic waves. Most interestingly, for one of the elliptic waves, there is a transition from the dipole to the quadrupole structure when the value of the modulus parameter $m$ of the Jacobi elliptic function $m > 0.5$. Also an arbitrary amplitude sinusoidal periodic wave is obtained which exhibits dipole structure within a period. So far as we are aware of, this interesting characteristic feature of the propagation dynamics of nonparaxial periodic waves has not been unveiled till now. These solitary wave solutions exist when the model parameters satisfy certain parametric conditions. These conditions exhibit a subtle balance between dispersion, Kerr, and non-Kerr nonlinearities, which is essential to control the dynamics of these solitary waves. To determine the physical significance of the obtained solutions, the effect of the nonparaxial parameter on the amplitude, pulse-width and speed of solitary waves are investigated. It has been observed that the speed of the solitary waves can be reduced by tuning the nonparaxial parameter. To test the stability of the solitary waves, numerical simulations have been performed by adding $10\%$ white noise to the initial stationary mode. It has been found that the elliptic wave having the dipole structure (depending on the value of the modulus parameter of Jacobi elliptic function) within a period has very weak instability in its propagation pattern under these noises for the chosen values of the parameters while the corresponding quadrupole structure propagate stably. All the other dipole solitary waves has stable propagation under these white noise perturbations. Furthermore, robustness of the solutions are examined by violation of the parametric conditions in two different ways namely, by adding $.003$ to all the parameters $a_i, i = 1,2 \cdots 7$ and also by adding an overall factor $.001$ to the parametric conditions. The solitary waves are found to be quite robust. The applications of dipole and quadrupole solitary wave solutions presented here may be found in nonparaxial optical contexts involving sub picosecond or femtosecond pulses, such as in ultrahigh-bit-rate optical communication systems \cite{krug72}.
The results presented in this article constitute a step forward to the exploration of dipole, multipole, kink, anti-kink etc. solitary waves for inhomogeneous higher order NLH system as is done in \cite{yang} for HNLSE . Also, it would be interesting to construct these types of solitary wave soluions in the presence of quintic, septic Kerr and non-Kerr nonlinearity as obtained in \cite{sarma} and \cite{azz1,triki} respectively for HNLSE . This is important in order to keep up with current developments in high-repetition-rate (beyond ultrashort) fiber-based pulse sources \cite{had}. Some of these issues are currently being investigated, and we hope to report these in the near future.\\
{\bf Data Availability}\\
Data sharing is not applicable to this article as no new data were created or analyzed in this study.\\
{\bf Conflict of Interest}\\
The authors have no conflicts to disclose.
~~\\
{\bf Appendix :}\\
For checking the dynamical evolution of the derived solutions, the finite difference method \cite{tref} is used with
$\delta t=0.1$, $\delta x=10^{-4}$.
and the boundary condition is given by:
$$\psi(x,t)=0,~\text{at the boundary.}$$
A one-dimensional uniform spatial grid is taken consisting of $N$ points labeled by $t_j = -L+(j-1)\delta t$, $j = 1, 2, .. N$, where $\delta t = \frac{2L}{N-1}$ is the lattice spacing and $L$ is the half-width. The left and right boundary points are located at $j = 1$ and $j = N$ respectively. So, the boundary conditions at the edges of the spatial grid are
$$\psi(x,\pm L)=0.$$
The derivatives are discretized as \cite{tref}
\begin{gather*}
\psi_x(x,t)|_{(x_i,t_j)}=\frac{\psi_{i+1,j}-\psi_{i-1,j}}{2\delta x},\\
\psi_{xx}(x,t)|_{(x_i,t_j)}=\frac{\psi_{i+1,j}-2\psi_{i,j}+\psi_{i-1,j}}{\delta x^2},\\
\psi_t(x,t)|_{(x_i,t_j)}=\frac{\psi_{i,j+1}-\psi_{i,j-1}}{2\delta t},\\
\psi_{tt}(x,t)|_{(x_i,t_j)}=\frac{\psi_{i,j+1}-2\psi_{i,j}+\psi_{i,j-1}}{\delta t^2},\\
\psi_{ttt}(x,t)|_{(x_i,t_j)}=\frac{\psi_{i,j+2}-2\psi_{i,j+1}+2\psi_{i,j-1}-\psi_{i,j-2}}{2\delta t^3},\\
\psi_{tttt}(x,t)|_{(x_i,t_j)}=\frac{\psi_{i,j+2}-4\psi_{i,j+1}+6\psi_{i,j}-4\psi_{i,j-1}+\psi_{i,j-2}}{\delta t^4},\\
\end{gather*}
where $\psi(x_i,t_j)=\psi_{i,j}.$
|
\section{Introduction}
The problem of estimating function under monotonicity constraints is generally called isotonic regression or monotone function estimation \citep{barlow1972statistical, robertson1988order, groeneboom2014nonparametric}.
This problem naturally arises in a variety of scientific applications including the dose-response modeling in epidemiological studies \citep{morton2000additive} and estimation of the demand curves in economics \citep{quah2000monotonicity}.
There is a large literature on point estimation for isotonic regression, including the minimization of sums-of-squares errors subject to ordering constraints \citep{brunk1972statistical, robertson1988order}, the use of splines \citep{ramsay1998estimating, wang2008isotonic} and kernels \citep{mammen1991estimating, hall2001nonparametric} with appropriate constraints.
Relaxing the hard constraint on mononicity is known as the nearly-isotonic regression and has also been studied \citep{tibshirani2011nearly,matsuda2021generalized}.
To extend those approaches to allow for uncertainty quantification by posterior analysis, we need to model a prior on the target function. This can be done in multiple ways, including the linear combination of monotone functions \citep{neelon2004bayesian,cai2007bayesian,mckay2011variable,shively2009bayesian}, Gaussian processes combined with monotone projection \citep{lin2014bayesian} and Gaussian processes for derivatives of functions \citep{lenk2017bayesian,kobayashi2021flexible}. Although these methods can realize the posterior distribution of the exactly or approximately monotone function, they have two limitations. First, the models are not flexible enough to express a function with ``jumps,'' or sudden increases/decreases at several points. Second, the posterior computation can be costly and inefficient due to the restriction for monotonicity.
In this paper, we address both problems by considering a prior for the first-order differences of the function of interest at consecutive points. In doing so, we utilize the global-local shrinkage by the half-horseshoe (HH) prior, a heavily-tailed distribution defined on the positive real line, to ensure the exact monotonicity. This distribution is the truncated version of the horseshoe prior \citep{carvalho2010horseshoe} and exhibits the same heavy tail, which enables the adaptive response of the posterior distribution to jumps. In addition, the HH prior can be written as the scale mixture of half-normals (the normal distribution truncated on the positive real line), allowing an efficient posterior computation by a simple Gibbs sampler. The horseshoe and its variants have already been applied to the first-order or higher-order differences in the context of time series analysis \citep{faulkner2018locally,bitto2019achieving,kowal2019dynamic} and shown the successful posterior inference for complex patterns of functional forms. We add the monotone constraint to this series of research.
We clarified the utility of the HH prior for the isotonic regression through the two theoretical results about the Bayes estimator of the target function. First, we verified the posterior tail-robustness under the proposed model. In other words, the more extreme jump we observe, the less shrinkage we apply a posteriori, making the Bayes estimator adaptive to the observed jump. Although this property is well-known for the horseshoe-type prior \citep{carvalho2010horseshoe,hamura2021global}, the evaluation of posterior robustness for isotonic regression is complicated due to the correlation between the function values at two consecutive points. Consequently, the robustness for the isotonic regression by the HH prior is ``weaker'' than those without monotonicity constraint, in terms of the speed of diminishing shrinkage effects.
Second, we proved the Kullbuck-Leibler super-efficiency of the proposed method under sparsity, showing efficiency in estimating functions that are constant on some regions.
The rest of the paper is organized as follows. In Section~\ref{methods}, we provide the model formulation and introduce shrinkage priors for positive valued differences of coefficients. The Gibbs sampler for posterior computation is provided in Section~\ref{methods}.
In Section~\ref{theory}, we show theoretical results on the posterior robustness to jumps and the efficiency in handling sparsity.
Section~\ref{simulation} is devoted to numerical experiments for the investigation of finite sample performance of the proposed methods together with some existing methods.
The application to the detection of the drop of Nile-river water flow is discussed in Section~\ref{data_analysis}. The proofs, algorithms, and additional experimental results are provided in the Appendix.
\section{Methods}
\label{methods}
\subsection{Model formulation}
We assume the $n$ observations $y_1, ..., y_n$ are conditionally independent and modeled by
$$
y_i=f(x_i)+{\varepsilon}_i, \ \ {\varepsilon}_i\sim N(0, \sigma^2), \quad \text{for $i=1, ..., n$},
$$
where $f: \mathbb{R} \to \mathbb{R}$ is monotone. For simplicity, we assume $x_i=i$ and let $\theta_i=f(x_i)$.
Then, the monotonicity constraint translates to $\theta_1 \le \theta_2 \le \cdots \le \theta_n$, and the model can be rewritten as
\begin{equation}
y|\theta , \sigma ^2 \sim N(\theta , \sigma ^2 I_n) ,
\label{model_1}
\end{equation}
where $y = (y_1, ..., y_n)^\top$ and $\theta = (\theta_1, ..., \theta_n)^\top$.
We guarantee the monotonicity of function values by using priors for the first-order differences of function values supported in $(0, \infty)$.
Denote the differences by
$\eta_j = \theta_j - \theta_{j-1}$ for $j=2, ..., n$ and the initial value by $\eta_1 = \theta_1$.
Then we have $\theta = D \eta$ where $\eta = (\eta_1, ..., \eta_n)^\top$ and $D$ is the lower triangular matrix defined below:
\begin{equation}
D = \begin{bmatrix}
1 & 0 & \cdots & 0 & 0 \\
1 & 1 & \cdots & 0 & 0 \\
& & \ddots & & \vdots \\
1 & 1 & \cdots & 1 & 0 \\
1 & 1 & \cdots & 1 & 1
\end{bmatrix}.
\label{D_mat}
\end{equation}
Since $D$ is invertible, a prior for $\eta$ implies
a corresponding prior for $\theta$ via the equality $\theta = D \eta$. The likelihood of $\eta$ is obtained as
\begin{equation*}
y | \eta, \sigma^2\sim N(D\eta ,\sigma ^2 I_n).
\end{equation*}
\subsection{Priors for differences of coefficients}
To model locally adaptive priors for the positive parameters $\eta_2, ..., \eta_n$, we utilize global-local shrinkage priors whose supports are the half real line.
We assume that $\eta _2,\dots ,\eta _n$ are mutually independent and follow the truncated normal distributions on $(0,\infty)$ as
\begin{equation}
\eta_j|\tau_j, \lambda, \sigma^2 \sim N_{+}(0, \sigma^2\lambda\tau_j) \quad \text{and} \quad \tau_j \sim \pi(\tau_j), \quad \text{for $j=2, ..., n$},
\label{prior_form}
\end{equation}
where both $\lambda$ and $(\tau_2, ..., \tau_n)$ are all positive. Here, $\lambda$ is the global parameter that shrinks all $\eta_j$'s toward zero uniformly, while $\tau_j$ is the local parameter that provide a custom shrinkage effect for each individual $j$.
The flexibility of shrinkage effects depends on the choice of priors for the global and local shrinkage parameters, $\lambda$ and $(\tau_2, ..., \tau_n)$.
Following the discussions of \citep{faulkner2018locally}, we consider three types of priors: half-horseshoe, half-Laplace and half normal.
\begin{itemize}
\item \textit{Half-horseshoe prior}.
The horseshoe prior \citep{carvalho2009handling,carvalho2010horseshoe} has the property of very strong shrinkage effect, as represented by the density spikes at near zero. In addition, its density has a Cauchy-like heavy tail, indicating that the shrinkage is not applied to outlying signals. This combination results in excellent performance as a shrinkage prior. We assume a truncated version of horseshoe prior for $\eta_j$.
This is realized by assuming that $\tau_j|\nu_j \sim Ga(1/2, \nu_j)$ and $\nu_j \sim Ga(1/2, 1)$ in (\ref{prior_form}), where $Ga(1/2, \nu_j)$ is the Gamma distribution with shape $1/2$ and rate $\nu_j$ (with mean $1/(2\nu_j)$).
We name this model for $\eta_j$ the half-horseshoe prior.
The density spike at $\eta _j=0$ can also be seen in this half-horseshoe prior, reflecting our belief that $\theta _j \approx \theta _{j+1}$ for most $j$'s. Likewise, the heavy tail is expected to explain a possible large deviation of $\theta_{j+1}$ from $\theta_j$.
\item \textit{Half-Laplace prior}.
The half-Laplace prior, which is merely the exponential prior, is obtained by $\tau_j \sim Ga(1, \nu)$. As the truncated version of the Bayesian Lasso \citep{park2008bayesian}, this model will be compared with the half-horseshoe prior.
\item \textit{Half-normal prior}.
We also consider the half-normal prior, by setting $\tau _j=1$ in (\ref{prior_form}).
The normal distribution is usually not viewed as a shrinkage prior, but we use it for the comparison with the half-horseshoe prior.
\end{itemize}
In the left panel of Figure~\ref{density}, we visually illustrate the marginal distributions of $\eta_j$ under the half-horseshoe, half-Laplace and half-normal priors. It is shown in this figure that the density under the half-horseshoe prior has a spike at the origin and heavier tail than those under the half-Laplace and half-normal priors. In addition,
in the right panel of Figure~\ref{density}, we generate four sets of coefficients $(\theta_1, ..., \theta_{10})$ from the half-horseshoe prior with $\lambda = 1$ and $\sigma^2 = 1$, and illustrate the shapes of the functions.
\begin{figure}[htbp]
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[keepaspectratio, width=75mm]{density.pdf}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[keepaspectratio, width=75mm]{genfunc.pdf}
\end{minipage}
\caption{Shapes of the half-horseshoe, half-Laplace and half-normal densities (left). Examples of shapes of monotone functions generated from the half-horseshoe prior (right).}
\label{density}
\end{figure}
\subsection{Global parameters and posterior computation}
For the observational variance $\sigma^2$, we assign the non-informative prior, $\pi(\sigma ^2)=\sigma^{-2}$. For the global shrinkage parameter, we set $\lambda|\xi \sim Ga(1/2, \xi)$ and $\xi \sim Ga(1/2, 1)$.
We also set the scale-beta prior for $\eta_1$, namely,
$\eta _1 | \tau _1, \sigma ^2 \sim N( 0, \sigma ^2 \tau _1 )$, $\tau _1 | \nu _1 \sim Ga(1,\nu_1)$, and $\nu _1 \sim Ga(1/2,1).$ We use this set of priors for all the models, including the half-horseshoe, half-Laplace and half-normal models for $\eta _{2:n}$.
Although the Bayes estimator of $\theta_i$ is not analytically available, there is an efficient yet simple Markov cain Monte Carlo algorithm for posterior computation.
The full conditional posteriors of all parameters become well-known distributions, so that we can efficiently carry out Gibbs sampling by generating posterior sampling from those distributions. We provide the detailed step-by-step Gibbs samplings in the Appendix.
\subsection{Extension to irregular grids}
Our model formulation is so far restricted to the case where data is observed at equally spaced locations. Here, following \cite{faulkner2018locally}, we generalize the model to allow for data observed at irregular spaced locations.
Let $x_1 < x_2 < \cdots < x_n$ be the locations of observations, and we denote by $w_j = x_j - x_{j-1}$ the distance between adjacent locations.
For the case of irregular grids, we assume the following prior for $\eta_j$:
\begin{equation}
\eta_j|\tau_j, \lambda, \sigma^2 \sim N_{+}(0, \sigma^2\lambda\tau_j w_j) \quad \text{and} \quad \tau_j \sim \pi(\tau_j), \quad \text{for $j=2, ..., n$}.
\label{prior_irr}
\end{equation}
The difference from the regular grid case (\ref{prior_form}) can be seen in the dependence of the variance of $\eta_j$ on $w_j$.
For this dependence, for example, when $w_j$ is small, it is expected to have a stronger shrinkage effect on $\eta_j = \theta_j - \theta_{j-1}$.
The half-horseshoe, half-Laplace and half-normal priors for $\eta_j$ can be applied
in the same way with the regular grids case. The algorithm of the Gibbs sampler in the irregular grid case is trivially obtained and provided in the Appendix.
\section{Theoretical properties}
\label{theory}
\subsection{Posterior robustness for $\eta_j$}
The original horseshoe estimator for sparse signals is robust in the sense that the difference between the estimate and the observation vanishes as the observation becomes extreme \citep{carvalho2010horseshoe}.
This property is known as the tail-robustness.
In this subsection, we show that the posterior mean estimator in our setting is also robust to large positive signals, but in a different sense from the standard setting. This theoretical result can explain the reason that the proposed method provides the posterior of the function that is adaptive to jumps.
Consider the model (\ref{model_1}), where $\sigma^2$ is assumed to be fixed.
We rewrite the model (\ref{model_1}) in terms of the initial value $z_1 = y_1$ and the differences $z_j = y_j - y_{j-1}$ as observations;
$z \sim N(\eta, \Sigma)$, where
$z = D^{-1}y = (y_1, y_2 - y_1, ..., y_n - y_{n-1})^\top, \eta = D^{-1}\theta$ and $\Sigma = \sigma^2 D^{-1}(D^{-1})^\top$, and $D$ is the lower triangular matrix defined in (\ref{D_mat}).
We assign the horseshoe prior for $\eta_1$, and assign the half-horseshoe prior for $\eta_j$ for $j=2, ..., n$, respectively:
\begin{gather*}
\eta_1 | \tau_1 \sim N(0, \tau_1^2), \quad \tau_1 \sim C^+(0, 1), \\
\eta_j | \tau_j \sim N_+(0, \tau_j^2), \quad \tau_j \sim C^+(0, 1), \quad j=2, ..., n,
\end{gather*}
where $C^+(0, 1)$ denotes the half-Cauchy distribution.
For simplicity, we omit the global shrinkage parameter.
We arbitrary choose an index $i_\ast \in \{2, ..., n \}$, where a jump of the function value is observed, and fix the values $z_j \in \mathbb{R}$ for $j \neq i_\ast$. The next theorem shows the behavior of the posterior mean $E[\eta_{i_\ast} | z]$ in the presence of large signals.
\begin{thm}
The posterior mean $E[\eta_{i_\ast} | z]$ of the model described above is weakly tail-robust, in the sense that
\begin{equation*}
\frac{ |E[\eta_{i_\ast} | z] - z_{i_\ast}| }{ z_{i_\ast} } \to 0 \quad \mathrm{as} \ z_{i_\ast} \to \infty.
\end{equation*}
\label{tail_robust}
\end{thm}
The robustness property shown above is weaker than the usual tail-robustness (i.e., $|E[\eta_{i_\ast} | z] - z_{i_\ast}| \to 0$ as $z_{i_\ast} \to \infty$).
As stated in the introduction, the analysis of the posterior means under our setting is more difficult than a standard setting where the location parameters $\eta_j$'s are independent \citep{carvalho2010horseshoe}. The difference from the standard setting, and the source of weakened robustness, is the correlation of observation $z$ (or, equivalently, the correlation of the original location parameter $\theta$). The robustness property of horseshoe estimators under such correlation has not been studied.
It is also noteworthy that the weak tail-robustness is not necessarily achieved when using other priors with thinner density tail than those of the horseshoe prior.
For example, consider the following normal prior model for a single observation: $z_j|\eta _j \sim N(\eta_j, \sigma^2)$ and $\eta_j \sim N(0, 1)$, ignoring both monotonicity constraint and correlation for simplicity. Then we obtain the posterior mean estimator $E[\eta_j | z_j] = (1 + \sigma^2)^{-1}z_j$, which does not have the weak tail-robustness since $| E[\eta_j | z_j] - z_j | / z_j \to \sigma^2 / (1+\sigma^2)$ as $z_j\to\infty$.
This example, together with Thereom~\ref{tail_robust}, shows the necessity of the heavily-tailed prior for $\eta_j$ to achieve the weak tail-robustness.
\subsection{Efficiency in handling sparsity}
The half-horseshoe prior can also handle with sparsity in $\eta_j$ efficiently for its density spike at the origin.
Following \cite{carvalho2010horseshoe}, we formalize this efficiency in terms of the rate of convergence of the Kullback-Leibler divergence between the true sampling density and its posterior mean estimator.
By showing this efficiency, we claim that under the sparse setting, the (half) horseshoe prior can reconstruct the true sampling density faster than any others prior with a bounded density at the origin.
Unlike the weak-tail robustness, the proof completely parallels that of the standard horseshoe prior \citep[Theorem~4]{carvalho2010horseshoe}.
For any parameter $\eta \ge 0$, let denote the probability density function of $x$ under $\eta$ as $f_\eta(x) = N(x | \eta, 1), x \in \mathbb{R}$.
Let $\eta_0$ denote the true value of the parameter, and let $\Delta_{KL}(\eta_0 || \eta) = \Delta_{KL}\{f_{\eta_0} || f_\eta \}$
denote the Kullback-Leibler divergence of $f_\eta$ from $f_{\eta_0}$.
Because $f_\eta$ is assumed to be the normal density, we can express the Kullback-Leibler divergence explicitly as
\[
\Delta_{KL}(\eta_0 || \eta)
=
\frac{(\eta - \eta_0)^2}{2}.
\]
Suppose we have $n$ observations $y_1, ..., y_n$.
Let $p$ be the marginal prior density on $\eta$, and for each $k = 1,\dots,n$, let
$\pi_k(d\eta | y_1, ..., y_k)$ be the posterior density based on $k$ observations $y_1, ..., y_k$, respectively.
Then the posterior mean estimator of the density based on the $k$ observations is defined as
\[
\hat{f}_k(x)
=
\int_0^\infty f_\eta(x) \pi_k(d \eta | y_1, ..., y_k).
\]
As the accuracy of estimation, we use the following Ces\'{a}ro-average risk:
\[
R_n
=
\frac{1}{n} \sum_{k=1}^n \Delta_{KL}\{f_{\eta_0} || \hat{f}_k \}.
\]
The following theorem shows that the half-horseshoe prior leads to a super-efficinet rate of convergence when $\eta_0 = 0$.
\begin{thm}
Suppose the true sampling model is $N(\eta_0, 1)$, and we assume the half-horseshoe prior for $\eta_0$.
Then, the rates of convergence of Ces\'{a}ro-average risk $R_n$ under $\eta_0 = 0$ and $\eta_0 > 0$ are given as follows
\[
R_n = \begin{cases}
O[n^{-1}\{\log n - \log \log n \}], \quad \eta_0 = 0, \\
O\{n^{-1}\log n \}, \quad \eta_0 > 0.
\end{cases}
\]
\label{kl_risk}
\end{thm}
As stated in \cite{carvalho2010horseshoe}, the optimal convergence rate of Ces\'{a}ro-average risk is $O\{n^{-1}\log n \}$ when using a priors with bounded density. Therefore,
Theorem \ref{kl_risk} implies that under the sparse setting $\eta_0 = 0$, the density estimator based on the half-horseshoe prior converges to the true density faster than those based on other bounded priors on $[0, \infty)$.
Theorem \ref{kl_risk} also implies that for $\eta_0 > 0$, the density estimator based on the half-horseshoe prior should converge at least as fast as those based on other bounded priors.
\section{Simulation study}
\label{simulation}
We here investigate the finite sample performance of the proposed method together with other methods. We generated $n=100$ observations from $y_i \sim N(\theta_i, 0.25) $ where $\theta_i$ is the value of a monotone function $f(x)$ at $x=i$. In generating the true monotone function $f$, we adopted the following five scenarios:
\begin{align*}
&\text{(I)}\,\ f(x) = 2, \\
&\text {(II)}\,\ f(x) =
\begin{cases}
0 & (0 \le x \le 25)\\
2.5 & (25 < x \le 80) \\
3 & (80 < x \le 100)
\end{cases}, \\
&\text{(III)}\,\ f(x) = 0.04x, \\
&\text{(IV)}\,\ f(x) =
\begin{cases}
0.02x & (0 \le x \le 20)\\
0.02x + 1 & (20 < x \le 50) \\
0.02x + 1.5 & (50 < x \le 80) \\
0.02x + 1.75 & (80 < x \le 100)
\end{cases}, \\
&\text{(V)}\,\ f(x) =
\frac{1}{4.4}\exp\{ 0.05x - 2\} + 1.
\end{align*}
These true functions are constant, piecewise constant, linear, piecewise linear and exponential shape, respectively. In scenarios (I) and (III), the true function has less frequent increments in comparison to the other scenarios. Note that in scenarios (II) and (IV), there are a few jumps in the function values.
We estimated $\theta_i$'s using the following six methods:
\begin{itemize}
\item HH: The proposed method with the half-horseshoe prior for $\eta_j$.
\item HL: The method with the half-Laplace prior for $\eta_j$.
\item HN: The method with the half-normal prior for $\eta_j$.
\item TF: The Bayesian trend-filtering with the horseshoe prior proposed by \cite{faulkner2018locally}.
\item GP: The Gaussian process regression method (e.g., \cite{williams2006gaussian}).
\item GPP: The Gaussian process projection method proposed by \cite{lin2014bayesian}.
\end{itemize}
To implement the posterior analysis in the GP and GPP methods, we employed Python package "GPflow".
On the covariance kernel of Gaussian process, we assigned the same prior used in the simulation in \cite{lin2014bayesian}.
All the methods require computations by Markov chain Monte Carlo methods. For each dataset, we generated 2,500 posterior samples after discarding 500 samples as a burn-in period.
We computed the posterior means of $\theta_i$'s as their point estimates in all methods. Note that the estimated function is guaranteed to be monotonic in the HH, HL, HN and GPP methods, while it is not necessarily monotonic in the TF and GP methods, they are not.
The performance of these point estimators were evaluated by the root mean squared error (RMSE) defined as $\sqrt{n^{-1} \sum_{i=1}^n (\hat{\theta}_i - \theta_i)^2}$. We also computed $95\%$ credible intervals of $\theta_i$'s based on all the methods, and evaluated their performance using the coverage probability (CP) and average length (AL). We repeated this process for $1,000$ times.
In Table \ref{table_regular}, we presented the averaged values of the RMSEs, CPs and ALs in all the methods. We can see that under scenarios (I), (III) and (V), the GPP method performs well in terms of RMSE, CP and AL, while the proposed HH method is comparable in RMSE. For scenarios (II) and (IV), where the true function has a few jumps, the HH method performs the best in RMSE, as well as achieves the high coverage rates by the interval estimates of reasonable lengths. Seeing this result, we confirms that the
proposed half-horseshoe prior can flexibly represent functions with abrupt changes, as expected.
In addition, we reported an example of posterior fits under three methods HH, TF and GPP in Figure~\ref{example_fits_1}.
We can see that the HH method is most adaptive to the abrupt increases in scenarios (II) and (IV).
The same graphical illustration for the other three methods (the HL, HN and GP methods) is reported in the Appendix.
\begin{table}[!htbp]
\begin{center}
\begin{tabular}{cclcccccc} \toprule
Scenario & & HH & HL & HN & TF & GP & GPP \\ \midrule
& RMSE & 0.034 & 0.044 &0.057 & 0.094 & 0.030 & 0.021 \\
(I) & CP &91.9 &90.3 &83.4 &99.7 &98.1 &99.0 \\
& AL &0.121 &0.127 &0.151 &0.469 &0.138 &0.118 \\ \midrule
& RMSE &0.087 &0.312 &0.517 &0.115 &0.222 &0.200 \\
(II) & CP &91.9 &90.0 &83.9 &98.6 &91.3 &90.4 \\
& AL &0.220 &0.345 &0.411 &0.501 &0.535 &0.315 \\ \midrule
& RMSE &0.079 &0.055 &0.057 &0.090 &0.05 &0.05 \\
(III) & CP &98.8 &99.4 &99.5 &99.3 &94.0 &94.4 \\
& AL &0.357 &0.284 &0.257 &0.467 &0.193 &0.190 \\ \midrule
& RMSE &0.081 &0.139 &0.166 &0.127 &0.156 &0.141 \\
(IV) & CP &97.1 &77.1 &66.7 &96.6 &81.0 &82.2 \\
& AL &0.322 &0.291 &0.275 &0.493 &0.357 &0.321 \\ \midrule
& RMSE &0.110 &0.159 &0.231 &0.094 &0.075 &0.072 \\
(V) & CP &89.8 &65.8 &43.0 &99.5 &94.0 &94.2 \\
& AL &0.352 &0.308 &0.311 &0.460 &0.279 &0.256 \\ \bottomrule
\end{tabular}
\end{center}
\caption{Averaged values of root mean squared errors (RMSEs) of point estimators, and coverage probability (CP) and average length(AL) of $95\%$ credible intervals under five scenarios with $n=100$.}
\label{table_regular}
\end{table}
\begin{figure}[htbp]
\centering
\includegraphics[keepaspectratio, width=5.65in]{example1.pdf}
\caption{Example fits by the HH, TF and GPP methods under the five scenarios. Plots show true functions (dashed red lines), posterior means(solid black lines), and associated $95 \%$ Bayesian credible interval (gray bands) for each $\theta_i$. Values between observed locations are interpolated for plotting.}
\label{example_fits_1}
\end{figure}
Next, we present a simulation result when data is observed at irregular spaced locations. We randomly picked 25 elements from $\{1, 2, ..., 100\}$ and sorted them as $x_1 < x_2 < \cdots < x_{25}$. We generated 25 observations from $y_k \sim N(f(x_k), 0.25), k=1, ..., 25$ under the five scenarios for the true monotone function. We estimated the 100 function values $f(i)$, $i=1, ..., 100$ from the 25 observations $y_k$'s, including those without the corresponding observations, by using the HH, GP and GPP methods. To interpolate $f(x)$ for $x\in \{ x_1,\dots ,x_{25}\}$, we first estimated $f(x_k)$'s by the method described in Section 2.4, then computed the function values between the observed locations by the linear interpolation.
In Table \ref{table:irregular}, we reported the averaged values of RMSEs, CPs and ALs in this setting.
We can observe a similar result to the case of regular grids. For scenarios (II) and (IV), the HH method performs quite well in terms of RMSE and has short credible intervals with good coverage probabilities.
\begin{table}[!htbp]
\centering
\begin{tabular}{cclccc} \toprule
Scenario & & HH & GP & GPP \\ \midrule
& RMSE &0.084 &0.056 &0.042 \\
(I) & CP &88.0 &100.0 &99.3 \\
& AL &0.278 &0.300 &0.278 \\ \midrule
& RMSE &0.262 &0.385 &0.364 \\
(II) & CP &81.6 &91.5 &89.3 \\
& AL &0.420 &1.147 &0.762 \\ \midrule
& RMSE &0.139 &0.097 &0.095 \\
(III) & CP &96.0 &98.4 &98.5 \\
& AL &0.566 &0.445 &0.446 \\ \midrule
& RMSE &0.178 &0.243 &0.224 \\
(IV) & CP &92.2 &78.4 &78.5 \\
& AL &0.556 &0.605 &0.566 \\ \midrule
& RMSE &0.194 &0.210 &0.204 \\
(V) &CP &92.1 &96.0 &96.4 \\
& AL &0.592 &0.700 &0.615 \\ \bottomrule
\end{tabular}
\caption{The averaged values of the RMSEs, CPs and ALs in the case of irregular grids.}
\label{table:irregular}
\end{table}
\section{Data analysis}
\label{data_analysis}
In this section, we estimate the trend of the
yearly volume of Nile river at Aswan by applying the proposed method.
Our data consists of measurements of the annual flow of the Nile river from $1871$ to $1970$, obtained from \cite{durbin2012time} and publicly available in $\mathsf{R}$-pakcage \textit{datasets} \citep{datasets}.
This data has a decreasing trend as a whole and has an apparent change point near $1898$ when a dam was built.
In this study, we are interested in whether or not the estimated trends can detect the change point.
To this end, we apply the proposed isotonic regression method with the HH prior to this data, as well as the HL, HN and GPP methods.
In each Gibbs sampler, we generate $5,000$ posterior samples after discarding $1,000$ posterior samples as burn-in.
Figure \ref{nile_plots} shows the posterior means and 95\% credible intervals of the function values.
It is observed in this figure that the proposed method with the half-horseshoe prior successfully detects the abrupt change in $1898$. In contrast, this change is less emphasized, or not at all detected by the other methods. This is another example where the sufficient flexibility of the proposed method is utilized to allow large jumps.
\begin{figure}[htbp]
\centering
\includegraphics[keepaspectratio, width=5.65in]{nile.pdf}
\caption{Estimates for the Nile river data by different methods. Solid lines are posterior means and dashed lines are $95\%$ credible intervals. Dashed vertical lines indicate the changepoint (1898).}
\label{nile_plots}
\end{figure}
\section{Concluding remarks}
In this research, we proposed the half-horseshoe prior for the robust posterior inference of monotone functions. It should be noted that the half-horseshoe prior is a continuous distribution, where we place no probability mass on $\{ \theta _j = \theta _{j+1} \}$, or $\{\eta_j = 0\}$. The extension to the point-mass mixture, known as the spike-and-slab prior, is of great interest. It is well known, however, that the ability of handling of sparsity under the spike-and-slab prior comes at the increased computational cost for the full posterior inference. This problem is expected to be inherited to the problem of estimating monotone functions. The trade-off of the model flexibility and computational cost is left for the future research.
The application of the half-horseshoe prior to the higher-order differences is trivial.
Although the first-order difference is extensively studied in this research to focus on the monotonicity, the second-order difference of function values is also of great interest to realize convex constraints on the target function.
Investigating the methodological and theoretical property of such methods for convex function estimation would be an important topic of future research.
In addition, the proposed method can be extended easily to situations having multiple covariates.
For example, when $p$ covariates $(x_{i1},\ldots,x_{ip})$ are obtained, we may consider an additive model, $y_i=\sum_{k=1}^p f_k(x_{ik})+{\varepsilon}_i$.
If any functions of $f_1,\dots ,f_p$ are known to be monotone, then we can model the monotonicity by assigning the proposed prior independently for those functions.
The posterior computation for such an additive model is straightforwardly implemented by Bayesian backfitting \citep{hastie2000bayesian}, combined with the conditional conjugacy of the proposed half-horseshor priors.
\clearpage
|
\section{Introduction}
By an \emph{exponential Diophantine equation}
\cite{STbook} we often mean
an equation of the form
\begin{equation}\label{eqn:exp}
f(X)=Y^n
\end{equation}
where $f$ is a polynomial with integer or rational coefficients,
and the unknowns $X$, $Y$ are taken to be either integral or rational.
The exponent $n$ is also an unknown, usually taken to run through
the integers $n \ge 2$. The epithet \lq\lq exponential\rq\rq\ simply
refers to the fact that the exponent is a variable (if the exponent
is fixed the equation is called superelliptic).
A famous theorem of Schinzel and Tijdeman \cite{SchinzelTijdeman}
concerning such equations asserts that if $f \in \mathbb{Z}[X]$
as at least two distinct roots, then there are finitely
many solutions $(X,Y,n)$ to \eqref{eqn:exp} with $X$, $Y$, $n \in \mathbb{Z}$, $n \ge 3$
and $\lvert Y \rvert \ne 1$.
One of the earliest examples of an exponential Diophantine equation
is
\begin{equation}\label{eqn:Lebesgue}
X^2+1=Y^n, \qquad X,~Y,~n \in \mathbb{Z}, \quad n \ge 2;
\end{equation}
In 1850 Victor Lebesgue \cite{Lebesgue} proved
that the only solutions are $(X,Y)=(0,1)$ if $n$ is odd,
and $(X,Y)=(0,\pm 1)$ if $n$ is even.
Lebesgue's argument is elementary, and some version
of this argument, with the help of the Primitive Divisor
Theorem of Bilu, Hanrot and Voutier \cite{BHV} is capable
of resolving many equations of the form
$X^2+C=Y^n$ (e.g. \cite{Cohn}). In contrast, the equation $X^2+7=Y^n$
(which generalizes the famous Ramanujan--Nagell equation)
apparently cannot be tackled by elementary arguments,
and was solved by
Bugeaud, Mignotte and Siksek \cite{BMS} using a combination
of tools from Diophantine approximation and Galois
representations of elliptic curves.
We can think of the Catalan equation
\[
X^m-Y^n=1, \qquad X,~Y,~m,~n \in \mathbb{Z}, \quad m,~n \ge 2
\]
as an infinite family of exponential Diophantine
equations of the form \eqref{eqn:exp}
by rewriting it as $X^m+1=Y^n$,
giving an exponential Diophantine equation
for each value of $m$.
The famous Catalan conjecture, proved
by Mih\u{a}ilescu \cite{Mihailescu} in 2004,
states the only solution with $XY \ne 0$ is $3^2-2^3=1$.
We can view equation \eqref{eqn:exp} as the question
of which perfect powers belong to $f(\mathbb{Z})$
(if the unknowns $X$, $Y$ belong to $\mathbb{Z}$)
or to $f(\mathbb{Q})$
(if the unknowns $X$, $Y$ belong to $\mathbb{Q}$).
It is therefore natural to ask which
whether there is a restriction
on the set of perfect powers
that can belong to $f(\mathbb{Z})$
or $f(\mathbb{Q})$.
Indeed, at the recent
\lq\lq Rational Points\rq\rq\ conference (Schney, April 2022),
Siksek posed the following two questions.
\begin{question}\label{question:1}
Let
\[
\mathscr{P}_\mathbb{Z} \; = \; \{ a^n \; : \; a \in \mathbb{Z}, \quad n \ge 2\}
\]
be the set of perfect powers in $\mathbb{Z}$. Let $S$ be a finite
subset of $\mathscr{P}_\mathbb{Z}$. Is there a polynomial $f_S \in \mathbb{Z}[X]$
such that $f_S(\mathbb{Z}) \cap \mathscr{P}_\mathbb{Z}=S$?
\end{question}
\begin{question}\label{question:2}
Let
\begin{equation}\label{eqn:cQ}
\mathscr{P}_\mathbb{Q} \; = \; \{ \alpha^n \; : \; \alpha \in \mathbb{Q}, \quad n \ge 2\}
\end{equation}
be the set of perfect powers in $\mathbb{Q}$. Let $S$ be a finite
subset of $\mathscr{P}_\mathbb{Q}$. Is there a polynomial $f_S \in \mathbb{Q}[X]$
such that $f_S(\mathbb{Q}) \cap \mathscr{P}_\mathbb{Q}=S$?
\end{question}
Question~\ref{question:1} was answered
affirmatively by Gajovi\'{c} \cite[Theorem 3.1]{Gajovic},
and we briefly recall (and slightly simplify) his elegant argument which
yields an
explicit polynomial $f_S$ in terms of $S$.
Let $S=\{b_1,b_2,\dotsc,b_r\} \subseteq \mathscr{P}_\mathbb{Z}$.
Let
\[
g(X)\; =\; \prod_{i=1}^r (X-b_i)^2+1,
\qquad
h(X) \; = \; (X-1) \cdot g(X) +1,
\qquad
f_S(X) \; =\; g(X) \cdot h(X).
\]
It is clear that $f_S(b_i)=b_i$, so $S \subseteq f_S(\mathbb{Z}) \cap \mathscr{P}_\mathbb{Z}$.
To prove the reverse inclusion let $x \in \mathbb{Z}$
such that $f_S(x) =y^n$ for some $y \in \mathbb{Z}$ and $n \ge 2$;
it is enough to prove that $x=b_i$ for some $i$.
Write $c=\prod_{i=1}^r(x-b_i)$.
We consider two cases:
\begin{itemize}
\item $y=0$. Then $h(x)=0$, so $x-1=-1/(c^2+1)$.
Since $x \in \mathbb{Z}$, we have $c=0$, and so $x=b_i$ for some $i$.
\item $y \ne 0$. Note that $g(x)>0$ and $h(x)$
are coprime integers whose product
is $y^n$. Thus
$g(x)=z^n$ for some integer $z$,
and so $c^2+1=z^n$.
Whence $c=0$, by the aforementioned theorem of Lebesgue concerning
\eqref{eqn:Lebesgue}. Hence $x=b_i$ for some $i$,
completing the proof.
\end{itemize}
We point out that for the last step Gajovi\'{c} invokes
Mih\u{a}ilescu's theorem \cite{Mihailescu}
(the Catalan conjecture), although it is enough
to invoke the special (and elementary) case due to
Lebesgue.
\bigskip
The purpose this paper is to
give an affirmative answer to Question~\ref{question:2}.
Concisely, our theorem can be stated as follows.
\begin{thm}\label{thm:main}
Let $S$ be a finite subset of $\mathscr{P}_\mathbb{Q}$.
Then there is a polynomial $f_S \in \mathbb{Z}[X]$
such that $f_S(\mathbb{Q}) \cap \mathscr{P}_\mathbb{Q}=S$.
\end{thm}
We give a recipe for $f_S$ in terms of a certain effectively
computable (though currently inexplicit) constant.
Instead of Lebesgue's theorem, or Mih\u{a}ilescu's theorem,
our proof that $f_S(\mathbb{Q}) \cap \mathscr{P}_\mathbb{Q}=S$ crucially relies on
the following theorem due to Ellenberg,
to Bennett, Ellenberg and Ng,
and to Bruin.
\begin{thm}[Ellenberg et al.]\label{thm:Ellenberg}
Let $n \ge 4$. Then the equation $a^2+b^4=c^n$
has no solutions in coprime non-zero integers.
\end{thm}
The theorem is due to Bruin \cite{Bruin} for $n=6$,
who treats this case using an elliptic Chabauty argument.
It is due to Ellenberg \cite{Ellenberg24p} for prime $n \ge 211$,
and to Bennett, Ellenberg and Ng \cite{BEN}
for all other $n \ge 4$. These two papers make
use of deep results in theory of Galois representations
of $\mathbb{Q}$-curves as well as a careful study of
critical values of Hecke $L$-functions of modular forms.
We shall also need a famous theorem
on perfect powers
in non-degenerate binary recurrence sequences,
due to
Peth\H{o} \cite{Petho} and independently to
Shorey
and Stewart \cite{ShSt83}.
We shall not need the theorem in its full generality,
so we only state a special case.
\begin{thm}[Peth\H{o}, Shorey and Stewart]\label{thm:ST}
Let $a$, $b$, $\alpha$, $\beta$ be non-zero integers
with $\alpha \ne \pm \beta$. Let
\[
u_t \; = \; a \alpha^t + b \beta^t, \qquad t \in \mathbb{Z}, \qquad t \ge 0.
\]
Then there is an effectively
computable constant $C(a,b,\alpha,\beta)$
such that
$u_t \notin \mathscr{P}_\mathbb{Z}$
for $t > C(a,b,\alpha,\beta)$.
\end{thm}
The theorem is proved using
the theory of lower bounds for linear forms
in logarithms of algebraic numbers.
\section{Preliminary results}
We shall need the following consequence of Theorem~\ref{thm:ST}.
\begin{lem}\label{lem:prelim}
Let $\gamma \in \mathbb{Q}$, $\gamma \ne 0$. Then there is an effectively
computable constant $D(\gamma)$
such for $t > D(\gamma)$,
\[
\gamma - 2^t \notin \mathscr{P}_\mathbb{Q}.
\]
\end{lem}
\begin{proof}
Write $\gamma=u/v$ where $u$, $v \in \mathbb{Z}$, $u \ne 0$, $v \ge 1$
and $\gcd(u,v)=1$. We let
\[
D(\gamma) \; =\; \max\{0,\log_2{\lvert \gamma \rvert},
\;
C(u,-v,1,2)
\}
\]
where $C(a,b,\alpha,\beta)$ is as in Theorem~\ref{thm:ST}.
Note that since $C(a,b,\alpha,\beta)$ is
effectively computable, so is $D(\gamma)$.
Let $t > D(\gamma)$.
We suppose $\gamma -2^t \in \mathscr{P}_\mathbb{Q}$ and derive
a contradiction.
Since $t>\log_2{|\gamma|}$, we have $\gamma-2^t \ne 0$.
Observe that $\gamma-2^t=(u-2^t v)/v$
where the numerator $u-2^t v$
and the denominator $v$ are coprime. As $\gamma -2^t \in \mathscr{P}_\mathbb{Q} \setminus \{0\}$
we conclude that $u-2^t v \in \mathscr{P}_\mathbb{Z}$.
We apply Theorem~\ref{thm:ST}
with $a=u$, $b=-v$, $\alpha=1$, $\beta=2$.
Since $t>C(u,-v,1,2)$, the theorem tells us
that $u-2^t v \notin \mathscr{P}_\mathbb{Z}$ giving a contradiction.
\end{proof}
We also need the following two corollaries to
Theorem~\ref{thm:Ellenberg}.
\begin{cor}\label{cor:Ellenberg1}
Let $n \ge 2$. Then the only solutions
to the equation $A^4+B^4=2C^n$ with $A$, $B$, $C \in \mathbb{Z}$
and $\gcd(A,B)=1$ satisfy $A=\pm 1$, $B=\pm 1$.
\end{cor}
\begin{proof}
Note that $A$, $B$ are both odd. Write
\[
U \; =\; AB, \qquad V \; =\; \frac{A^4-B^4}{2}.
\]
Then $U$, $V$ are coprime integers and satisfy
\[
U^4+V^2 \; = \; \left(\frac{A^4+B^4}{2} \right)^2 \; = \; C^{2n}.
\]
By Theorem~\ref{thm:Ellenberg} we have $UVC=0$.
However, $U$ is odd so $U \ne 0$. Thus $C \ne 0$.
Hence $V=0$, so $A^4=B^4$. Since $A$, $B$
are coprime, $A=\pm 1$, $B=\pm 1$ as required.
\end{proof}
\begin{cor}\label{cor:Ellenberg2}
Let $n \ge 2$. Then the only solutions
to the equation $A^4+B^4=C^n$ with $A$, $B$, $C \in \mathbb{Z}$
and $\gcd(A,B)=1$ satisfy $A=0$, $B=\pm 1$, or $B=0$, $A=\pm 1$.
\end{cor}
\begin{proof}
For $n=2$ this is a famous result of Fermat, proved by infinite descent.
For $n=3$ it is in fact a result of Lucas
\cite[Section 5]{BCDY}.
Suppose $n \ge 4$. Then we can rewrite the
equation as $(A^2)^2+B^4=C^n$
and conclude from Theorem~\ref{thm:Ellenberg}.
\end{proof}
\section{Proof of Theorem~\ref{thm:main}}\label{sec:proof}
Let $S$ be a finite subset of $\mathscr{P}_\mathbb{Q}$.
We would like to give a polynomial $f_S \in \mathbb{Z}[X]$
such that $f_S(\mathbb{Q}) \cap \mathscr{P}_\mathbb{Q}=S$.
For $S=\emptyset$
we may take any constant polynomial whose value is not a perfect power. Thus we may suppose $S \ne \emptyset$.
We write
\[
S\; = \; \{ \beta_1,\dotsc,\beta_r \}
\]
where $r \ge 1$, and $\beta_i \in \mathscr{P}_\mathbb{Q}$.
As the $\beta_i$ are
rational numbers, we can write
\[
\beta_i \; = \; \frac{a_i}{c_i}, \qquad
a_i,~c_i \in \mathbb{Z}, \qquad c_i \ge 1, \qquad \gcd(a_i,c_i)=1.
\]
Let $p_1,p_2,\dotsc,p_t$ be the distinct prime
divisors of $\prod_{i=1}^nc_i$, and let
\begin{equation}\label{eqn:kdef}
k\; = \; \lcm(4,p_1 -1, p_2-1, \dotsc,p_t-1).
\end{equation}
Let
\begin{equation}\label{eqn:F}
F(X) \; =\; \prod_{i=1}^r (c_i X-a_i)^2 \; -1.
\end{equation}
Since $c_i \ge 1$ we note that $F$ has degree $2r$,
and hence at most $2r$ roots. We are only interested
in the non-zero rational roots of $F$, and we let these
be $\delta_1,\delta_2,\dotsc,\delta_m$.
We let $s$ be an integer satisfying the following:
\begin{gather}
\text{$s \, =\, 2^{\kappa}-1$ for some $\kappa \ge 1$}, \label{eqn:sdef1}\\
s \; \ge \; D(4\delta_j), \qquad j=1,\dotsc,m.
\label{eqn:sdef3}
\end{gather}
Here $D(\cdot)$ is as in Lemma~\ref{lem:prelim}.
It is clear that such an $s$ is effectively computable.
Let
\begin{gather}
g(X) \; =\; \prod_{i=1}^r (c_i X-a_i)^k \; + \; 1,
\qquad
h(X) \; = \; (X-2^s)g(X) \, + \, 2^s, \label{eqn:ghdef} \\
f_S(X) \; = \; g(X) \cdot h(X) \; \in \mathbb{Z}[X].
\label{eqn:fSdef}
\end{gather}
To prove Theorem~\ref{thm:main}
we need to show that $f_S(\mathbb{Q}) \cap \mathscr{P}_\mathbb{Q}=S$.
Observe that $g(\beta_i)=1$, $h(\beta_i)=\beta_i$
and so $f_S(\beta_i)=\beta_i$.
Thus $\beta_i \in f_S(\mathbb{Q}) \cap \mathscr{P}_\mathbb{Q}$, and therefore
$S \subseteq f_S(\mathbb{Q}) \cap \mathscr{P}_\mathbb{Q}$.
We would like to prove the reverse inclusion.
Thus let $x \in \mathbb{Q}$ and suppose that $f_S(x) \in \mathscr{P}_\mathbb{Q}$.
To prove Theorem~\ref{thm:main} it will
be enough to show that $x=\beta_i$ for some $1 \le i \le r$.
Write
\begin{equation}\label{eqn:uvdef}
x=\frac{u}{v}, \qquad u,~v \in \mathbb{Z}, \quad v \ge 1, \quad
\gcd(u,v)=1.
\end{equation}
Let
\begin{equation}\label{eqn:ABdef}
A\; =\; \prod_{i=1}^r (c_i u - a_i v), \qquad
B\; = \; \frac{A}{\gcd(A,v^r)}, \qquad w \; = \; \frac{v^r}{\gcd(A,v^r)},
\end{equation}
and note that
\begin{equation}\label{eqn:gcdBw}
\gcd(B,w)=1, \qquad w \ge 1.
\end{equation}
Recall that we would like to show that $x=\beta_i$
(or equivalently $u/v=a_i/c_i$) for some $1 \le i \le r$.
This is equivalent to $A=0$, which is equivalent to
$B=0$. Thus our proof will be complete on showing that $B=0$.
We note that
\begin{equation}\label{eqn:gh}
g(x) \; = \; \frac{A^k + v^{kr}}{v^{kr}} \; = \;
\frac{B^k+w^k}{w^k},
\qquad
h(x) \; = \; \frac{(u-2^s v) \cdot (B^k+w^k) + 2^s v w^k}{v w^k}.
\end{equation}
\begin{lem}\label{lem:2pow}
$\gcd(B^k+w^k,v)$ is a power of $2$.
\end{lem}
\begin{proof}
Let $p$ be an odd prime dividing both $B^k+w^k$ and $v$.
From \eqref{eqn:ABdef} we see that $(B^k+w^k)$
divides $A^k+v^{kr}$. Thus $p \mid (A^k+v^{kr})$
and so $p \mid A$.
From the product definition of $A$ in \eqref{eqn:ABdef},
and since $\gcd(u,v)=1$ from \eqref{eqn:uvdef}, we see that $p \mid c_i$
for some $1 \le i \le r$. It follows from the definition of $k$ in \eqref{eqn:kdef}
that $(p-1) \mid k$. Since $\gcd(B,w)=1$ \eqref{eqn:gcdBw}, we have that $p\nmid w$ and $p\nmid B$, so $B,w\in \mathbb{Z}_p^*$. Then we have $B^k+w^k\equiv 0\mod p$, so $(Bw^{-1})^k\equiv -1\mod p$. This contradicts Fermat's little theorem.
\end{proof}
\begin{lem}\label{lem:not2}
$B^k+w^k \ne 2$.
\end{lem}
\begin{proof}
Recall that $k$ is an even integer by \eqref{eqn:kdef},
and that $w \ge 1$ from \eqref{eqn:gcdBw}.
Suppose $B^k+w^k=2$.
Then $\lvert B \rvert=w=1$.
In particular, from \eqref{eqn:gh} we have $g(x)=2$.
Thus
\begin{equation}\label{eqn:fS0}
f_S(x)\; = \; g(x) \cdot h(x) \; = \; 4x-2^{s+1}
\end{equation}
from \eqref{eqn:ghdef} and \eqref{eqn:fSdef}.
First we show that $x \ne 0$.
Recall our assumption that $f_S(x) \in \mathscr{P}_\mathbb{Q}$,
and so $f_S(x)=y^n$ for some rational $y$
and some integer $n \ge 2$.
If $x=0$, then $y^n=-2^{s+1}=-2^{2^\kappa}$
by \eqref{eqn:sdef1}. Thus $n \mid 2^{\kappa}$
and so $n$ is even. Hence $y^n >0$
giving a contradiction. We conclude that $x \ne 0$.
Recall, from \eqref{eqn:uvdef}, that $v \ge 1$
and $x=u/v$. From \eqref{eqn:ABdef}, as $\lvert B \rvert=w=1$,
we see that
\[
\lvert A \rvert \; = \; \gcd(A,v^r) \; = \; v^r.
\]
Thus $A = \pm v^r$.
Dividing both sides of $A = \pm v^r$ by $v^r$ gives
\[
\prod_{i=1}^r (c_i x- a_i) \; = \; \frac{A}{v^r} \; =\; \pm 1,
\]
from the product expression for $A$ in \eqref{eqn:ABdef}.
Thus $x$ is a non-zero root of the polynomial $F(X)$
given in \eqref{eqn:F}. We have previously
labelled the roots of $F$ by $\delta_1,\delta_2,
\dotsc,\delta_m$. Thus $x=\delta_j$ for some $1 \le j \le m$.
By \eqref{eqn:sdef3} we have
$s+1 > D(4 \delta_j)$.
Hence, by Lemma~\ref{lem:prelim}
and \eqref{eqn:fS0} we have
$f_S(x) =4x-2^{s+1}=4 \delta_j-2^{s+1} \notin \mathscr{P}_\mathbb{Q}$,
giving a contradiction.
\end{proof}
\begin{lem}\label{lem:zero}
If $f_S(x)=0$ then $B=0$.
\end{lem}
\begin{proof}
Suppose $f_S(x)=0$ and recall that $f_S(x)=g(x)h(x)$.
Note that $k$ is even by \eqref{eqn:kdef},
and also $w \ge 1$ by \eqref{eqn:gcdBw}.
Thus $B^k+w^k$ is a positive integer
and hence, from \eqref{eqn:gh}
we have $g(x) \ne 0$.
Hence $h(x)=0$,
whence from \eqref{eqn:gh},
\begin{equation}\label{eqn:-vw}
(u-2^s v) \cdot (B^k+w^k) \; = \; -2^s v w^k.
\end{equation}
We claim that $B^k+w^k$ is a power of $2$. To prove
this claim let $p \mid (B^k+w^k)$ be an odd prime.
As $\gcd(B,w)=1$ by \eqref{eqn:gcdBw},
we have $p \nmid w$.
From Lemma~\ref{lem:2pow} we note that $p \nmid v$.
However, from \eqref{eqn:-vw}, we have $p \mid 2^s v w^k$
giving a contradiction.
Thus $B^k+w^k$ is indeed a power of $2$.
Since $k$ is even and $\gcd(B,w)=1$ we see
that $4 \nmid (B^k+w^k)$ and so
$B^k+w^k=1$ or $2$.
However, $B^k+w^k \ne 2$
from Lemma~\ref{lem:not2}.
We conclude that $B^k+w^k=1$.
As $w \ge 1$, we obtain $w=1$ and $B=0$
as required.
\end{proof}
Recall that we have supposed that $f_S(x) \in \mathscr{P}_\mathbb{Q}$
and to complete the proof of Theorem~\ref{thm:main},
it is enough to show that $B=0$.
Lemma~\ref{lem:zero} establishes this if $f_S(x)=0$.
Thus we may suppose $f_S(x) \ne 0$.
Hence $f_S(x)=y^n$ where $y$ is a non-zero rational and $n \ge 2$.
We claim that $B^k+w^k=z^n$ or $2 z^n$ for some odd positive
integer $z$. To prove this let $p$ be an odd prime dividing
$B^k+w^k$. Then, as before $p \nmid w$ since $\gcd(B,w)=1$,
and $p\nmid v$ from Lemma~\ref{lem:2pow}.
From the expression for $h(x)$ in \eqref{eqn:gh}
we see that $p$ divides neither the numerator nor the denominator of $h(x)$,
and hence $\ord_p(h(x))=0$.
However $g(x)h(x)=f_S(x)=y^n$, so
\[
\ord_p(B^k+w^k)
\; = \; \ord_p\left(\frac{B^k+w^k}{w^k} \right)\; =\; \ord_p(g(x)) \; = \; \ord_p(y^n) \; \equiv\; 0 \pmod{n}.
\]
As this is true for every odd prime dividing $B^k+w^k$
we have $B^k+w^k=2^e z^n$ for some $e \ge 0$ and some odd integer $z$.
Now as before $4 \nmid (B^k+w^k)$, since $k$ is
even and $\gcd(B,w)=1$. Thus $B^k+w^k=z^n$
or $B^k+w^k=2 z^n$ where $z$ is odd.
Suppose first that $B^k+w^k=2z^n$.
Recall that $4 \mid k$ by \ref{eqn:kdef}.
By Corollary~\ref{cor:Ellenberg1} we have $B=\pm 1$, $w=1$,
contradicting Lemma~\ref{lem:not2}.
Thus $B^k+w^k=z^n$. We apply Corollary~\ref{cor:Ellenberg2}
to conclude that $B=0$ or $w=0$. However, $w \ge 1$,
so $B=0$ completing the proof.
|
\section{Introduction}
Let $k$ be an algebraically closed field of characteristic zero. By a variety we mean an integral separated scheme of finite type. We let $\OO_X$ be its structure sheaf and $K_X$ be its the field of rational functions so that $K_X=\fract \OO_X(U)$ for any affine open set $U\subset X$.
We also let $\GM$ and $\GA$ be the multiplicative group and the additive group over $k$, respectively. It is well known that regular additive group actions on affine varieties are in one-to-one correspondence with certain derivations called locally nilpotent \cite{Fre06,Dai03}. Indeed, letting $X$ be an affine variety a locally nilpotent derivation on $X$ is a $k$-derivation $\partial\colon \OO_X(X)\to \OO_X(X)$ such that for every $f\in \OO_X(X)$ there exists $i\in \ZZ_{\geq 0}$ with $\partial^i(f)=0$. All derivations in this paper are $k$-derivations so we will call them simply derivations. Given a locally nilpotent derivations $\partial$ on $X$ we obtain a regular $\GA$-actions $\varphi\colon \GA\times X\to X$ on $X$ via the exponential map
$$\varphi^*=\exp(z\partial)\colon \OO_X(X)\to\OO_X(X)[z] \quad\mbox{given by}\quad f\mapsto \sum_{i\geq0}\frac{z^i\partial^i(f)}{i!}\,.$$
On the other hand, given a regular $\GA$-action $\varphi\colon \GA\times X\to X$ on $X$ we obtain a locally nilpotent derivations $\partial$ on $X$ via
$$\partial\colon \OO_X(X)\to \OO_X(X) \quad\mbox{given by}\quad f\mapsto \operatorname{ev}_{0}\circ \frac{d}{dz}\circ \varphi^*(f)\,,$$
where $\operatorname{ev}_{0}\colon\OO_X(X)[z]\to \OO_X(X)$ is the evaluation morphism in $z=0$.
In \cite{DuLi16} Dubouloz and the second author, introduced a class of rationally integrable derivations that generalized locally nilpotent derivations to the rational setting. Indeed, a derivation $\partial\colon K_X\to K_X$ is called rationally integrable if the exponential map
$$\exp(z\partial)\colon K_X\to K_X[|z|] \quad\mbox{given by}\quad f\mapsto \sum_{i\geq0}\frac{z^i\partial^i(f)}{i!}\,,$$
factors through the ring $K_X(z)\cap K_X[|z|]$. Their main theorem provides a one-to-one correspondence between rationally integrable derivations $\partial\colon K_X\to K_X$ and rational $\GA$-actions $\varphi\colon \GA\times X\dashrightarrow X$ on $X$. The correspondence is given similarly to above via
$$\varphi^*=\exp(z\partial) \quad \mbox{and}\quad \partial= \operatorname{ev}_{0}\circ \frac{d}{dz}\circ \varphi^*(f)\,,$$
after recalling that $K_X(z)\cap K_X[|z|]=\{r(z)\in K_X(z)\mid \operatorname{ord}_0(r)\geq 0\}$ so that $\operatorname{ev}_{0}$ is well defined.
In this paper we expand and generalize the results in \cite{DuLi16} to allow a classification of rational $\GM$-action. It is well known that regular $\GM$-action on an affine variety $X$ are in one-to-one correspondence with semisimple derivations $\partial\colon \OO_X(X)\to \OO_X(X)$ having integer eigenvalues. Recall that such a derivation is semisimple if there exists a basis $\{a_i\}_{i\in I}$ of $A$ as vector space such that $\partial(a_i)=\lambda_i a_i$ with $\lambda_i\in k$.
In Definition~\ref{def:rat-semisimple}, we introduce rational semisimple derivations. A derivation $\partial\colon K_X\to K_X$ on an algebraic variety $X$ is rational semisimple if there exist a generating set $\{a_i\}_{i\in I}$ of $K_X$ (as field) such that $\partial(a_i)=\lambda_i a_i$ with $\lambda_i\in \mathbb{Z}$. Every semisimple derivation whose eigenvalues are integers numbers is rational semisimple. Our main result in this paper is Theorem~\ref{main-th} establishing a one-to-one correspondence between rational $\GM$-actions $\varphi\colon \GM\times X\dashrightarrow X$ on $X$ and rational semisimple derivations $\partial\colon K_X\to K_X$ on $X$. The correspondence is as follows: we prove in Corollary~\ref{Cor5.2} that the image of $\varphi^*$ is contained in $K_X(t)\cap K_X[|t-1|]=\{r(t)\in K_X(t)\mid \operatorname{ord}_1(r)\geq 0\}$ and so we obtain $\partial$ from $\varphi$ via
$$\partial\colon K_X\to K_X \quad\mbox{given by}\quad f\mapsto \operatorname{ev}_{1}\circ \frac{d}{dt}\circ \varphi^*(f)\,.$$
As for the other direction, letting $\sigma$ be the isomorphism of formal power series rings given by the logarithmic power series
$$\sigma\colon K_X[|z|]\to K_X[|t-1|],\quad \mbox{given by} \quad z\mapsto \sum_{i\geq 1}(-1)^{i+1}\frac{(t-1)^i}{i}\,.$$
we recover the rational action $\varphi$ from $\partial$ via $\varphi^*=\sigma\circ\exp(z\partial)$.
As a consequence of our main result, we prove in Corollary~\ref{rational-slice} the existence of a rational slice for every rationally semisimple derivation, i.e., an element $s\in K_X$ such that $\partial(s)=s$. Moreover, we prove in Proposition~\ref{cor desc} that $K_X=K_X^{\GM}(s)$ and that under this isomorphism the derivation $\partial$ is given by $\partial=s\frac{d}{ds}$. Here, $K_X^{\GM}$ is the field of invariant of the $\GM$-action.
Finally, we provide in Proposition~\ref{prop:Regular-actions} a characterization of regular actions of the multiplicative group on the class of varieties that are proper over the spectrum of its ring of global regular functions. Such characterization agrees with the one recalled above in the particular case of affine varieties.
\subsection*{Acknowledgements}
Part of this work was done during a stay of both authors at IMPAN in Warsaw. We would like to thank IMPAN and the organizers of the Simons semester ``Varieties: Arithmetic and Transformations'' for the hospitality.
\section{Rational $\GM$-action}
Let $\mu\colon\GM\times \GM\to \GM$ the morphism given by the group law $(t_1,t_2)\mapsto t_1t_2$, and let $\operatorname{e}_{\GM}\colon\,\operatorname{Spec}(k)\to \GM$ be the neutral element map. A rational action of the multiplicative group is a rational map $\varphi\colon\GM\times X \dashrightarrow X , (x,t)\mapsto t\cdot x$ such that the following diagrams are commutative:
\begin{eqnarray}
\xymatrix@=4.5em{ \GM\times \GM\times X \ar@{-->}[r]^{\mathrm{id}_{\GM}\times \varphi} \ar[d]_{\mu\times \mathrm{id}_X} & \GM\times X \ar@{-->}[d]^{\varphi} & \mathrm{Spec}(k)\times X \ar[r]^{e_{\GM}\times \mathrm{id}_X} \ar[dr]_{\mathrm{pr}_2} & \GM\times X\ar@{-->}[d]^{\varphi} \\ \GM\times X \ar@{-->}[r]^{\varphi} & X & & X.}
\label{eq:GeomDiagrams2}
\end{eqnarray}
We let $\dom(\varphi)$ be largest open subset of $\GM\times X$ where $\varphi$ is well defined. If $(g,x)\in \dom(\varphi)$, we denote $\varphi(g,x)$ simply by $g\cdot x$. The next lemma show that $\dom(\varphi)\cap (\{g\} \times X)$ is a non-empty open subset of $ \{g\} \times \GM$. In particular for $g=1$, $\dom(\varphi)\cap (\{1\} \times X)$ is a open subset not empty. This will allow us to exhibit a criterion for the existence of rational action in terms of the function field of $X$.
\begin{lemma}\label{lem-demazure}
Let $X$ be an algebraic variety endowed with a rational $\GM$-action. For every fixed $g\in \GM$, we define $V_{g}=\dom(\varphi) \cap(\{g\} \times X)$.
$V_{g}$ is a non-empty open subset of $\{g\} \times X \simeq X$ and the morphism
\begin{eqnarray*}
\varphi_{g}\colon V_{g} & \dashrightarrow & X\\
x & \mapsto & \varphi(g,x)
\end{eqnarray*}
is dominant.
\end{lemma}
\begin{proof}
Following Demazure in \cite{Dem70} we consider the morphism
$u_{g}\colon \GM\to \GM$
, $h\mapsto gh^{-1}$ whose inverse is $(u_{g})^{-1}\colon \GM\to \GM$
, $h\mapsto h^{-1}g$. We define $\beta_{g}=\varphi\circ(u_{g}\times \operatorname{id}_{X})\circ(\operatorname{pr}_1,\varphi)$,where $\operatorname{pr}_1$ is the projection in the first coordinate and $\operatorname{id}_{X}$ is the identity in $X$. As
\begin{center}
$\beta_{g}\colon\GM\times X \stackrel{ (\operatorname{pr}_{1},\varphi) }{\displaystyle\dashrightarrow} \GM\times X \stackrel{u_{g}\times\operatorname{Id}_{X} }{\displaystyle\dashrightarrow} \GM\times X \stackrel{\varphi }{\displaystyle\dashrightarrow} X$
\end{center}
We have $\beta_{g}\colon \GM\times X \dashrightarrow X$, $(h,x)\mapsto \varphi(g,x)$ is dominant for all choice of $g$.
Note that this map only depend of $x$.
Then $\dom(\beta_{g})=\GM \times U$, with $U \subseteq X$ an open subset. We have $\beta_{g}=\varphi_{g}\circ\operatorname{pr}_{2}$, where $\operatorname{pr}_2$ is the projection of $\GM\times X$ into the second coordinate. Hence, $U=\operatorname{pr}_{2}(V_{g})$. Since $\varphi_{g}$ is a rational map and $V_{g}=\{g\}\times \dom(\varphi_{g})$, we obtain $V_{g}\not=\emptyset$. Moreover $\beta_{g}=\varphi_{g}\circ\operatorname{pr}_{2}$, which means $\varphi_{g}$ is dominant because $\operatorname{Im}(\beta_{g})\subseteq \operatorname{Im}(\varphi_{g})$.
\end{proof}
We can apply the case $g=1$ to obtain the following characterization.
\begin{corollary}\label{Cor5.2}
Let $\varphi\colon \GM\times X \dashrightarrow X$ be a rational action then $\operatorname{Im}(\varphi^{\ast})\subseteq \mathcal{O}_{\nu}$, where $\mathcal{O}_{\nu}=\{ r(t) \in K(t) \,\,| \, \,
\operatorname{ord}_{1}(r)\geq 0 \}$ is the discrete valuation ring of
$K(t)$ and $\ord_1$ is the order of vanishing in $t=1$.
\end{corollary}
\begin{proof}
Let $V'\subseteq \GM\times X$ and $X'\subseteq X$ be affine open sets such that $\varphi|_{V'}\colon V'\to X'$ is regular. We let $\OO_{\GM\times X}(V')= A\subseteq K(t)$ and $\OO_X(X')= B\subseteq K$. By Lemma~\ref{lem-demazure}, we may and will assume $V'\cap V_1\neq \emptyset$ and so it is an dense open set of $\{1\}\times X$. Now, the composition $V'\cap V_1\to V'\to X'$ is dominant by Lemma~\ref{lem-demazure} and induces algebra homomorphisms
$$B\stackrel{\varphi^*}{\longrightarrow} A\longrightarrow A/A(t-1)\,,$$
and this composition is injective. This in turn shows that $\ord_1(\varphi^*(b))=0$ for every $b\in B$ different from $0$. Since $K=\fract B$ we have $\ord_1(\varphi^*(f))=0$ for every $f\in K$, which proves the corollary.
\end{proof}
Given a rational action $\varphi\colon \GM\times X \dashrightarrow X$, for every fixed $g\in \GM$ we obtain a birational automorphism $$\varphi_{g}\colon X \dashrightarrow X, \qquad x\mapsto \varphi(g,x)$$
since $\varphi(1,x)=\operatorname{Id}_X$ and for every $g,g'\in \GM$ we have $\varphi_{g} \circ \varphi_{g'}(x)=\varphi_{gg'}(x)$. Moreover, the map $\GM\to \operatorname{Bir(X)}$ sending $g$ to $\varphi_g$ is a group homomorphism. Finally, a rational action $\varphi\colon \GM\times X \dashrightarrow X$ such that $\dom(\varphi)=\GM\times X$ is a regular action.
\subsection{Criterion for existence of $\GM$-rational actions}
A $\GM$-rational action $\varphi\colon\GM\times X \dashrightarrow X$ in a variety $X$,is equivalent to the co-action homomorphism $\varphi^{\ast}_t\colon K_X\to k(\GM\times X)=K_X(t)$ where the map $\varphi^{\ast}_t$ factors through the subalgebra $\mathcal{O}_{\nu}=\{ r(t) \in K_X(t) \,\,| \, \, \operatorname{ord}_{1}(r)\geq 0 \}$, with $\mathfrak{m}_{\nu}=\{ r(t) \in K_X(t) \,\,| \, \, \operatorname{ord}_{1}(r)> 0 \}$. The co-action morphism is characterized by the commutativity of the follow diagrams:
\begin{eqnarray} \label{comorphism}
\xymatrix@=4.5em{ K_X \ar@{->}[r]^{\varphi^{\ast}_{t_1}} \ar[d]_{\varphi^{\ast}_{t_2}} & K_X (t_1) \ar@{->}[d]^{t_1\mapsto t_1t_2} & & K_X\simeq \mathcal{O}_{\nu}/ \mathfrak{m}_{\nu} \ar[r]^{\varphi^{\ast}_t} \ar[dr]_{\operatorname{Id}_{K_X}} & \mathcal{O}_{\nu}\ar@{->}[d]^{\operatorname{ev}_1} \\ K_X(t_2) \ar@{->}[r]^{\widetilde{\varphi}} & K_X(t_1,t_2) & & & K_X\simeq \mathcal{O}_{\nu}/ \mathfrak{m}_{\nu}}
\label{eq:GeomDiagrams3}
\end{eqnarray}
The following proposition is classical.
\begin{proposition}\label{prop}
Le $X$ be a variety. $X$ admits a nontrivial rational $\GM$-action if only if it is birationally isomorphic to $Y\times\mathbb{P}^1$ for some $k$-variety $Y$.
\end{proposition}
\begin{proof}
See \cite{Ros56,Mat63,Pop16}.
\end{proof}
Mimicking the case of $\GA$-actions on affine geometry, we we have the following definition:
\begin{definition} \label{def:rational-slice}
Let $X$ be a $\GM$-variety, an element $s\in K_X$ such that $\varphi^{\ast}_t(s)=ts$ is called rational slice.
\end{definition}
For each faithful rational action, a rational slice always exists. Indeed, since the action is faithful, there exists two semi-invariant $a,b\in K_X$ whose weights $n,m$ are relatively prime, i.e., $\varphi^{\ast}_t(a)=t^{n}a$ and $\varphi^{\ast}_t(b)=t^{m}b$. By Bezout theorem we have that there exist $c,d\in \mathbb{Z}$ such that $nd+mc=1$. Hence $s=a^db^c$ satisfies $\varphi^{\ast}_t(s)=st$ and so is a rational slice.
A derivation on $k$-algebra $A$, is a linear map $\partial\colon A\to A$ such that satisfy the Leibniz rules, $\partial(ab)=a\partial(b)+b\partial(a)$. We define the kernel of a derivation $\partial$ as its kernel as a linear map, i.e., $\operatorname{ker}(\partial):=\{a\in A| \partial(a)=0\}$. The set of derivations over $A$ is denoted by $\operatorname{Der}(A)$.
We say $\partial$ is a semisimple derivation on $A$ if there exists a basis $\{a_i\}_{i\in I}$ of $A$ as $k$-vector space of eigenvalues such that $\partial(a_i)\in ka_i$, we will focus in the subset of semisimple derivations whose eigenvalues are integers numbers. For more details over semisimple derivations see \cite{Now01,Now94}. Is known that the regular $\GM$-actions are in correspondence with the set of semisimple derivations whose eigenvalues are integers numbers.
Analogously with the definition of semisimple derivations over a $k$-algebra, we will give a definition of derivation over $K_X$, we will call rational semisimple.
\begin{definition} \label{def:rat-semisimple}
Let be $K_X$ the field of rational function associated to algebraic variety $X$, $\partial\in \operatorname{Der}(K_X)$. We say $\partial$ is rational semisimple if there exist a generating set $\{a_i\}_{i\in I}$ of $K_X$ (as field) such that $\partial(a_i)=\lambda_i a_i$ with $\lambda_i\in \mathbb{Z}$.
\end{definition}
Let $\partial\colon K_X\to K_X$ a $k$-derivation. Denoting the $i$-th iteration of $\partial$ by $\partial^{i}$ and $\partial^0$ is the identity map, we define the exponential map
$$\exp(z\partial)\colon K_X\to K_X[|z|],\quad \mbox{given by} \quad f\mapsto \sum_{i\geq 0}\frac{z^i\partial^{i}(f)}{i!}\,.$$
Furthermore, since $\partial$ is a $k$-derivation, the following proposition shows that $\exp(z\partial)$ is a $k$-rings homomorphism.
\begin{proposition}\label{prop2.7}
Let be $\partial$ a $k$-derivation on $K_X$, then exponential map $\exp(z\partial)\colon K_X\to K_X[|z|]$ is a $k$-ring homomorphism.
\end{proposition}
\begin{proof}
The exponential map $\exp(z\partial)$ is $k$-linear since $\partial$ is $k$-linear. As for the product structure, we have
\begin{align*}
\exp(z\partial)(fg)&=\sum_{i\geq 0}
\frac{\partial^i(fg)z^{i}}{i!}\\
&=\sum_{i \geq 0}
\frac{1}{i!}\displaystyle\left(\sum^{i}_{j = 0}{n \choose j}\partial^{j}(f)\partial^{i-j}(g)\right)z^{i}\\
&=\sum_{i \geq 0}
\displaystyle\left(\sum^{i}_{j = 0}\frac{\partial^{j}(f)}{j!}\frac{\partial^{i-j}(g)}{(i-j)!}\right)z^{i}\\
&=\sum_{i\geq 0}
\frac{\partial^i(f)z^{i}}{i!}\sum_{l\geq 0}
\frac{\partial^l(g)z^{l}}{l!}\\
\exp(z\partial)(fg)&=\exp(z\partial)(f)\exp(z\partial)(g)
\end{align*}
\end{proof}
Furthermore, there is an isomorphism of formal power series rings
\begin{align*}\sigma\colon K_X[|z|]\to K_X[|t-1|],\quad \mbox{given by} \quad z\mapsto \sum_{i\geq 1}(-1)^{i+1}\frac{(t-1)^i}{i}\,.
\end{align*}
The inverse of this isomorphism is given by
\begin{align*}
\sigma^{-1}\colon K_X[|t-1|]\to K_X[|z|],\quad \mbox{given by} \quad t-1\mapsto \sum_{i\geq 1}\frac{z^i}{i!}\,.
\end{align*}
This corresponds to the logarithmic series and its inverse is the exponential series. In the sequel we will show that the image of $\sigma\circ\exp(z\partial)$ is contained in $K_X(t-1)=K_X(t)$ and that $\sigma\circ\exp(z\partial)$ is the comorphism of a rational $\GM$-action if and only if $\partial$ is a rational semisimple derivation. With this in view, for every we denote $\phi^*_{\partial}=\sigma\circ\exp(z \partial)$ so that its comorphism corresponds to a map $\phi_\partial\colon \GM\times X\dashrightarrow X$.
\begin{lemma} \label{contained-action}
Let be $\partial\colon K_X\to K_X$ a rational derivation. Assume that the image of $\phi^*_\partial$ is contained in $K_X(t)$, then $\phi_{\partial}$ is an action of the multiplicative group.
\end{lemma}
\begin{proof}
From Proposition~\ref{prop2.7} we have $\phi^*_{\partial}$ is a ring homomorphism. Let
$$t_1=\sum_{i\geq 0}\frac{z_1^i}{i!}\quad\mbox{and}\quad t_2=\sum_{i\geq 0}\frac{z_2^i}{i!}$$
We have the identity
$$t_1t_2=\displaystyle\sum_{i\geq 0}\frac{1}{i!}(z_1+z_2)^i$$
By Lemma~\ref{lem-demazure}, for every $t\in \GM$ we can specialize $\phi_\partial^*$ to obtain a field automorphism $\phi_\partial^*(t)\colon K_X\to K_X$. We extend $\partial$ to a derivation $\partial\colon K_X(z_1,z_2)\to K_X(z_1,z_2)$ by setting $\partial(z_1)=\partial(z_2)=0$. Since now the derivations $z_1\partial$ and $z_2\partial$ commute, by \cite[Proposition~2.4.2]{Now94} we have
\begin{align*}
\phi^*_{\partial}(t_1t_2)(f)&=\sigma\circ\exp((z_1+z_2)\partial)(f)\\
&=\sigma\circ\big(\exp(z_1\partial)\circ \exp(z_2\partial)\big)(f)\\
&=\phi^*_{\partial}(t_1)\circ \phi^*_{\partial}(t_2)(f)
\end{align*}
Furthermore, by definition of $\phi^*_\partial\colon K_X\to K_X[|t-1|]$ the composition
\begin{align*}
\operatorname{ev}_1\circ\phi^*_\partial&=\operatorname{ev}_1\circ\sigma\circ\exp(z\partial)\\
&=\operatorname{ev}_0\circ\exp(z\partial)\\
&=\operatorname{ev}_0\circ\sum_{i\geq 0}\frac{z^i\partial^i}{i!}\\
\operatorname{ev}_1\circ\phi^*_\partial &=\operatorname{Id}_{K_X}.
\end{align*}
Hence, by \eqref{comorphism} we have that $\phi_\partial^*$ is the comorphism of a multiplicative group action.
\end{proof}
\begin{proposition}\label{prop2.8}
The following are equivalent.
\begin{enumerate}[$(i)$]
\item $\partial\colon K_X\to K_X$ is rational semisimple.
\item The image of $\phi^*_{\partial}$ is contained in $K_X(t)$.
\end{enumerate}
\end{proposition}
\begin{proof}
Assume first that $\partial$ is rational semisimple. Hence, there exists a set of field generators $\{a_i\}_{i\in I}$ of $K_X$ such that
$\partial(a_j)=\lambda_ja_j$ with $\lambda_j\in \mathbb{Z}$. For these generators we have $$\phi^*_{\partial}(a_j)=\sigma\circ\exp(z\partial)(a_j)=\sigma\left(\sum_{i\geq 0}\frac{z^i\partial^i(a_j)}{i!}\right)=\sigma\left(\sum_{i\geq 0}\frac{z^i\lambda^i_ja_j}{i!}\right)=\sigma\left(\sum_{i\geq 0}\frac{(z\lambda_j)^i}{i!}a_j\right)=t^{\lambda_j} a_j\,.$$
Let now $f,g\in k[a_i,i\in I]$ with $g\neq 0$. Now,
\begin{align*}
\phi^*_{\partial}\left(\frac{f}{g}\right) &=
\sigma\circ\exp(z\partial)\left(\frac{f}{g}\right)\\
&=\sigma\circ\exp(z\partial)\left(\frac{f(a_i,i\in I)}{g(a_i,i\in I)}\right) \\
&=\frac{\sigma\circ\exp(z\partial)(f(a_i,i\in I))}{\sigma\circ\exp(z\partial)(g(a_i,i\in I))} \\
&=\frac{f(t^{\lambda_i} a_i,i\in I)}{g(t^{\lambda_i}a_i,i\in I)}\in K_X(t)
\end{align*}
This proves $(i)\to (ii)$. Inspired by Koshevoii \cite{Kos67}, to prove the converse assertion, we let $f\in K_X$ and we let
$$\phi^*_{\partial}(f)=t^\ell\cdot \frac{\sum_i a_it^i}{\sum_i b_it^i},$$
with $\ell \in \ZZ$ and $a_0,b_0\neq 0$. We also may and will assume that the representation is irreducible meaning that $\sum_i a_it^i$ and $\sum_i b_it^i$ are relatively prime. Furthermore, such a representations of $\phi^*_{\partial}(f)$ is unique if we further assume that $b_0=1$.
By Lemma~\ref{contained-action}, we have $\phi^*_{\partial}(s)\circ\phi^*_{\partial}(t)(f)=\phi^*_{\partial}(st)(f)$. This yields
$$t^\ell\dfrac{\sum_{i} \phi^*_{\partial}(s)(a_i)t^i}{\sum_{i} \phi^*_{\partial}(s)(b_i)t^i}=t^\ell \dfrac{\sum_{i} a_i(st)^i}{\sum_{i} b_i(st)^i}=t^\ell \dfrac{\sum_{i} (a_is^i)t^i}{\sum_{i} (b_is^{i})t^i}.$$
This yields
$\phi^*_{\partial}(a_j)=a_jt^j$ and $\phi^*_{\partial}(b_j)=b_jt^j$. In particular, this implies that
$$\phi^*_{\partial}(a_j)=\sigma\circ\exp(z\partial)(a_j)=\sigma\left(\sum_{i\geq 0}\frac{z^i\partial^i(a_j)}{i!}\right)\,.$$
while
$$t^{j}a_j=\sigma\left(\sum_{i\geq 0}\frac{(zj)^i}{i!}\right)\cdot a_j=\sigma\left(\sum_{i\geq 0}\frac{(zj)^i}{i!}a_j\right)=\sigma\left(\sum_{i\geq 0}\frac{z^ij^ia_j}{i!}\right)$$
We conclude that
$$\sum_{i\geq 0}\frac{z^i\partial^i(a_j)}{i!}=\sum_{i\geq 0}\frac{z^ij^ia_j}{i!}$$
In particular, taking equality in the term with $i=1$ we obtain that $\partial(a_j)=ja_j$. Finally, since $f=\dfrac{\sum_{i} a_j}{\sum_{i} b_i}$ we obtain that the set
$$\left\{a_i,b_i\in K_X\mid t^\ell\cdot \phi_\partial^*(f)=\frac{\sum_i a_it^i}{\sum_i b_it^i} \mbox{ for some } f\in K_X\right\}$$
generates $K_X$ and so $\partial$ is a rational semisimple derivation.
\end{proof}
Given a rational $\GM$-action $\varphi$ on $X$ with comorphism $\varphi^*\colon K_X\to K_X(t)$, we define the map
$$D_\varphi\colon K_X\to K_X \quad\mbox{given by}\quad f\mapsto \operatorname{ev}_{1}\circ \frac{d}{dt}\circ \varphi^*(f)\,,$$
generalizing the usual definition of the infinitesimal generator of a 1-parameter group action. The following lemmas will be required in our proof of our main result.
\begin{lemma} \label{derivation-from-action}
Let $\varphi$ be a rational $\GM$-action on $X$ with comorphism $\varphi^*\colon K_X\to K_X(t)$. Then the follwing hold:
\begin{enumerate} [$(i)$]
\item The map $D_\varphi$ is a derivation.
\item We have an equality $D_\varphi=\displaystyle \operatorname{ev}_1\circ\frac{d}{dt}\circ\varphi^*=\operatorname{ev}_0\circ\frac{d}{dz}\circ\sigma^{-1}\circ\varphi^*$.
\end{enumerate}
\end{lemma}
\begin{proof}
By Corollary \ref{Cor5.2}, we know that the image of $\varphi^*$ is contained in $K_X[|t-1|]$. Given $f$ and $g$ in $K_X$, we let $\varphi^*(f)=\sum_{i\geq 0}a_i(t-1)^i$ and $\varphi^*(g)=\sum_{i\geq 0}b_i(t-1)^i$. We have $a_0=f$ and $b_0=g$ since $\varphi^*$ is the comorphism of a $\GM$-action. Moreover, we have
\begin{align} \label{eq:D-is-a1}
D_\varphi(f)=\operatorname{ev}_1\circ\frac{d}{dt}\circ\varphi^*(f)=\operatorname{ev}_1\left(\sum_{i\geq 1}a_i\cdot i(t-1)^{i-1}\right)=a_1\,,
\end{align}
so that the map $D_\varphi(f)$ corresponds to the first order term of $\varphi^*(f)$. To prove $(i)$, remark that the composition $D_\varphi=\operatorname{ev}_1\circ \frac{d}{dt}\circ\varphi^*$ is $k$-linear and maps $k$ to $0$. Furthermore, the term of first order term of $fg$ is $a_0b_1+a_1b_0$. This yields the Leibniz rule since $$D_\varphi(fg)=a_0b_1+a_1b_0=fD_\varphi(g)+D_\varphi(f)g\,.$$
Assertion $(ii)$ follows by the following straightforward computation:
\begin{align*}
\operatorname{ev}_0\circ\frac{d}{dz}\circ\sigma^{-1}\circ\varphi^*(f)&=\operatorname{ev}_0\circ \frac{d}{dz}\circ \sigma^{-1}\left(\sum_{i\geq 0}a_i\cdot (t-1)^i\right)\\
&=\operatorname{ev}_0\circ \frac{d}{dz}\left(\sum_{i\geq 0}a_i \left(\sum_{j\geq 1}\frac{z^j}{j!}\right)^i\right)\\
&=\operatorname{ev}_0\left[\sum_{i\geq 1}a_i\cdot i \left(\sum_{j\geq 1}\frac{z^j}{j!}\right)^{i-1}\cdot\left(\sum_{j\geq 0}\frac{z^j}{j!}\right)\right]=a_1=D_\varphi(f)\,.
\end{align*}
\end{proof}
We will now prove that the map $\psi^*:=\sigma^{-1}\circ\varphi^*\colon K_X\to K_X[|z|]$ is the germ of a rational $\GA$-action in $X$. Indeed, letting $s$ and $w$, in the following diagram put as subscript the transendental element over $K_X$ in the target ring.
\begin{align*}
& K_X\stackrel{\varphi_t^*}{\longrightarrow} K_X[|t-1|]\stackrel{\sigma^{-1}_z}{\longrightarrow} K_X[|z|] \\
& K_X\stackrel{\varphi_s^*}{\longrightarrow} K_X[|s-1|]\stackrel{\sigma^{-1}_w}{\longrightarrow} K_X[|w|]
\end{align*}
We also let $\psi_z^*:=\sigma_z^{-1}\circ\varphi_t^*$ and
$\psi_w^*:=\sigma_w^{-1}\circ\varphi_s^*$. With these definitions we now proof the following lemma.
\begin{lemma} \label{germ-Ga}
With the above notation, we have $\operatorname{ev}_0\circ\psi^*_z=\operatorname{Id}_{K_X}$ and $\psi_z^*\circ \psi_w^*=\psi_{z+w}^*$.
\end{lemma}
\begin{proof}
Letting $f\in K_X$ we assume
$$\varphi^*(f)=\sum_{i\geq 0}a_i(t-1)^i \quad \mbox{so that} \quad \psi^*_z(f)=\sum_{i\geq 0}a_i\left(\sum_{j\geq 1}\frac{z^j}{j!}\right)^i\,.$$ Since $\varphi^*$ is the comorphism of a $\GM$-action we have $a_0=f$ and so $\operatorname{ev}_0\circ\,\psi^*_z(f)=a_0=f$. To prove the second assertion, remark that
\begin{align*}
\psi^*_z\circ\psi^*_w(f)&=\sigma^{-1}_z\circ\varphi^*_t\circ \sigma^{-1}_w\circ\varphi^*_s\\
&=\sigma^{-1}_z\circ\sigma^{-1}_w\circ\varphi^*_t\circ \varphi^*_s(f)\\
&=\sigma^{-1}_z\circ\sigma^{-1}_w\circ \varphi^*_{st}(f)\\
&=\sigma^{-1}_z\circ\sigma^{-1}_w\left(\sum_{i\geq 0}a_i(st-1)^i\right)\\
&=\sigma^{-1}_z\circ\sigma^{-1}_w\left(\sum_{i\geq 0}a_i\bigg[\big(s-1+1\big)\big(t-1+1\big)-1\bigg]^i\right)
\end{align*}
Applying now $\sigma^{-1}_z\circ\sigma^{-1}_w$ amounts to replace $t-1$ by $\sum_{i\geq 1}\frac{z^i}{i!}$ and $s-1$ by $\sum_{i\geq 1}\frac{w^i}{i!}$. Hence, we obtain
\begin{align*}
\psi^*_z\circ\psi^*_w(f)&=\sum_{i\geq 0}a_i\left[\left(\sum_{j\geq 1}\frac{w^j}{j!}+1\right)\left(\sum_{j\geq 1}\frac{z^j}{j!}+1\right)-1\right]^i\\
&=\sum_{i\geq 0}a_i\left[\left(\sum_{j\geq 0}\frac{w^j}{j!}\right)\left(\sum_{j\geq 0}\frac{z^j}{j!}\right)-1\right]^i\\
\end{align*}
Now, by the usual properties of the exponential sum, we obtain
\begin{align*}
\psi^*_z\circ\psi^*_w(f)&=\sum_{i\geq 0}a_i\left[\left(\sum_{j\geq 0}\frac{(w+z)^j}{j!}\right)-1\right]^i\\
&=\sum_{i\geq 0}a_i\left[\left(\sum_{j\geq 1}\frac{(w+z)^j}{j!}\right)\right]^i\\
\psi^*_z\circ\psi^*_w(f) &=\psi^*_{w+z}(f)
\end{align*}
\end{proof}
\begin{lemma} \label{derivaciones-z-t}
Let $\varphi$ be a rational $\GM$-action on $X$ with comorphism $\varphi^*\colon K_X\to K_X(t)$. Then the iterations $D^i_\varphi$ satisfy $\displaystyle D^i_\varphi=\operatorname{ev}_{0}\circ \frac{d^i}{dz^i}\circ \sigma^{-1}\circ\varphi^*$
\end{lemma}
\begin{proof}
We can now use the argument as in \cite[Proposition~4.10]{Dai03} applied to $\psi^*=\sigma^{-1}\circ\varphi^*$. For the convenience of the reader we copy the argument here. Letting $\psi^*_z(f)=\sum_{i\geq 0}a_iz^i$, we have
\begin{align*}
\sum_{i\geq 0}\psi^*_{z}(a_i)w^i=\psi^*_z\circ\psi^*_w(f)=\psi^*_{w+z}(f)
&=\sum_{\ell\geq 0}a_\ell(z+w)^\ell \\
&=\sum_{\ell\geq 0}a_\ell\sum_{i+j=\ell\geq 0}\binom{\ell}{i}z^jw^i\\
&=\sum_{i\geq 0}\left( \sum_{j\geq 0}a_{i+j}\binom{i+j}{i}z^j\right)w^i
\end{align*}
Hence we obtain
$$\psi^*(a_i)=\sum_{j\geq 0}a_{i+j}\binom{i+j}{i}z^j\mbox{ and in particular } \psi^*(a_1)=\sum_{j\geq 0}a_{j+1}(j+1)z^j=\sum_{j\geq 1}a_{j}jz^{j-1}$$
Since $D_{\varphi}(f)=a_1$ by \eqref{eq:D-is-a1}, we have $\varphi^*\circ D_\varphi=\frac{d}{dz}\circ\varphi^*$. Indeed,
$$\psi^*(D_{\varphi}(f))=\psi^*(a_1)=\sum_{j\geq 1}a_{j}jz^{j-1}=\frac{d}{dz}\left(\sum_{j\geq 0}a_jz^j\right)=\frac{d}{dz}(\psi^*(f))$$
Hence, we have $\psi^*\circ D^i_{\varphi}=\frac{d^i}{dz^i}\circ\psi^*$ for all $i\geq 0$. Composing on the left with $\operatorname{ev}_0$ we obtain
$$D^i_\varphi=\operatorname{ev}_0\circ \frac{d^i}{dz^i}\circ\sigma^{-1}\circ\varphi^*\,,$$
since $\operatorname{ev}_0\circ\,\psi^*=\operatorname{Id}_{K_X}$ by Lemma~\ref{germ-Ga}.
\end{proof}
In the following proposition, we show that the derivation $D_\varphi$ is rational semisimple.
\begin{proposition}\label{prop-comp-1}
Let $\varphi\colon \GM\times X\dasharrow X$ be a $\GM$-rational action on $X$. Then the following hold
\begin{itemize}
\item[$(i)$] $D_\varphi$ is a rational semisimple derivation.
\item[$(ii)$] The composition $\phi^*_{D_\varphi}=\sigma\circ\exp(zD_\varphi)$ equals $\varphi^*$.
\end{itemize}
\end{proposition}
\begin{proof}
The assertion $(i)$ follows directly from $(ii)$ and Proposition~\ref{prop2.8} since in this case the image of $\phi^*_{D_\varphi}$ equals the image of $\varphi^*$ which is contained in $K_X(t)$.
To prove $(ii)$, let $f\in K_X$. By Corollary~\ref{Cor5.2}, we have that $\varphi^*(f)\in K_X[|t-1|]$ and so $\sigma^{-1}\circ \varphi^*(f)\in K_X[|z|]$. Let $\sigma^{-1}\circ\varphi^*(f)=\sum_{j\geq 0}a_jz^j$. By Lemma~\ref{derivaciones-z-t} we have that
\begin{align*}
\phi^*_{D_\varphi}(f)&=\sigma\circ\sum_{i\geq 0}\frac{z^iD^i_{\varphi}(f)}{i!} \\
&=\sigma\circ\sum_{i\geq 0}\frac{z^i}{i!}\cdot \operatorname{ev}_0\circ\frac{d^i}{dz^i}\circ\sigma^{-1}\circ\varphi^*(f) \\
&=\sigma\circ\sum_{i\geq 0}\frac{z^i}{i!}\cdot \operatorname{ev}_0\circ\frac{d^i}{dz^i}\left(\sum_{j\geq 0}a_jz^j\right) \\
&=\sigma\circ\sum_{i\geq 0}\frac{z^i}{i!}\cdot \operatorname{ev}_0\circ\left(\sum_{j\geq i}a_j\frac{j!}{(j-i)!}z^{j-i}\right) \\
&=\sigma\circ\sum_{i\geq 0}\frac{z^i}{i!}\cdot a_i \cdot i! \\
&=\sigma\circ\sum_{i\geq 0}a_i z^i=\sigma\circ\sigma^{-1}\circ\varphi^*(f)=\varphi^*(f)
\end{align*}
\end{proof}
\begin{proposition}\label{prop-comp-2}
Let $\partial\colon K_X\to K_X$ be a rational semisimple derivation. Then the composition $D_{\phi_\partial}=\operatorname{ev}_1\circ\frac{d}{dt}\circ\sigma\circ\exp(z\partial)$ equals $ \partial$.
\end{proposition}
\begin{proof}
Let $\partial$ be a rational semisimple derivation. Then by Lemma~\ref{derivation-from-action}~(ii), we have
\begin{align*}
D_{\phi_\partial}=\operatorname{ev}_1\circ\frac{d}{dt}\circ\sigma\circ\exp(z\partial)=\operatorname{ev}_0\circ\frac{d}{dz}\circ\sigma^{-1}\circ\sigma\circ\exp(z\partial)=\operatorname{ev}_0\circ\frac{d}{dz}\circ\exp(z\partial)
\end{align*}
Letting now $f\in K_X$ we obtain
\begin{align*}
D_{\phi_\partial}(f)&=\operatorname{ev}_0\circ\frac{d}{dz}\circ\exp(z\partial)(f)
=\operatorname{ev}_0\circ\frac{d}{dz}\circ \sum_{i\geq 0}\frac{z^i\partial^i(f)}{i!}
=\operatorname{ev}_0\circ \sum_{i\geq 1}\frac{iz^{i-1}\partial^i(f)}{i!}=\partial(f)\,.
\end{align*}
This proves the proposition.
\end{proof}
The following is our main theorem in this paper establishing a one-to-one correspondence between rational $\GM$-actions on $X$ and rational semisimple derivations on $K_X$.
\begin{theorem} \label{main-th}
Let $X$ be an algebraic variety. There exists a one-to-one correspondence between the rational $\GM$-actions over $X$ and rational semisimple derivations on $K_X$ given by
\begin{align*}
\Big\{\mbox{Rational semisimple derivations on $K_X$}\Big\} \quad &\longleftrightarrow \quad\Big\{\mbox{Rational $\GM$-actions on $X$}\Big\} \\
\partial \quad & \longrightarrow \quad \phi_\partial \\
D_\varphi \quad &\longleftarrow \quad \varphi
\end{align*}
\end{theorem}
\begin{proof}
Let $X$ be an algebraic variety. If $\partial\colon K_X\to K_X$ is a rational semisimple derivation, then $\phi^*_\partial$ is a rational $\GM$-action by Lemma~\ref{contained-action} and Proposition~\ref{prop2.8}. On the other hand, if $\varphi$ is a rational $\GM$-action on $X$, then, by Proposition~\ref{prop-comp-1}~$(i)$, we have that $D_\varphi$ is rational semisimple. The fact that these maps are mutually inverse to each other is proven in Proposition~\ref{prop-comp-1}~(ii) and Proposition~\ref{prop-comp-2}.
\end{proof}
\section{Examples and applications}
In this section we provide several examples and applications of our main theorem. To perform the computations in the sequel, we need the following technical lemma proving that conjugation of a rational semisimple derivation by an automorphism $\varphi$ amounts to conjugation of the corresponding $\GM$-action by the same automorphism $\varphi$.
\begin{lemma}\label{lemma-ejemplos}
Letting $\partial\colon K_X\to K_X$ be a rational semisimple derivation and $\varphi^*\colon K_X\to K_X$ be a $k$-automorphim, we have
$$\phi^*_{\varphi^*\circ \partial\circ(\varphi^*)^{-1}}=\varphi^*\circ\phi^*_{\partial}\circ(\varphi^*)^{-1}\,$$
\end{lemma}
\begin{proof}
Since $(\varphi^*\circ \partial\circ (\varphi^*)^{-1})^i=\varphi^*\circ \partial^i\circ(\varphi^*)^{-1}$ and $\sigma$ commutes with $\varphi^*$, we have
\begin{align*}
\phi_{\varphi^*\circ \partial\circ (\varphi^*)^{-1}}&=\sigma\circ\sum_{i\geq 0}\frac{z^i(\varphi^*\circ \partial\circ (\varphi^*)^{-1})^i}{i!}
=\sigma\circ \varphi^*\circ\left(\sum_{i\geq 0}\frac{z^i\partial^i}{i!}\right)\circ(\varphi^*)^{-1}
=\varphi^*\circ \phi^*_{\partial}\circ (\varphi^*)^{-1}
\end{align*}
\end{proof}
\begin{example}
Letting $X=\AF^2=\spec k[x,y]$ we let $E$ be the Euler derivation given by
$$E=ax\frac{\partial}{\partial x}+by\frac{\partial}{\partial y} \quad \mbox{with}\quad a,b\in \mathbb{Z}\,.$$
This derivation is a regular semisimple derivation corresponding to the linear $\GM$-action on $X$ given by
$$\GM\times X\to X \quad\mbox{where}\quad (t,(x,y))\mapsto (t^ax,t^by)\,.$$
If we conjugate $E$ with the birational map
$$\varphi\colon X\dashrightarrow X \quad\mbox{given by}\quad (x,y)\mapsto ((x-1)(y-1)+1,y)\,,$$
whose inverse is
$$\varphi^{-1}\colon X\dashrightarrow X \quad\mbox{given by}\quad (x,y)\mapsto \left(\frac{x-1}{y-1}+1,y\right)\,.$$
We define the rational semisimple derivation $\partial=\varphi^*\circ E\circ(\varphi^*)^{-1}$. A straightforward computation shows that
$$\partial=\left( \frac{a((x-1)(y-1)+1)-b(x-1)y}{y-1}\right)\frac{\partial}{\partial x} + by \frac{\partial}{\partial y}$$
By Lemma~\ref{lemma-ejemplos}, we have $\phi^*_{\partial}=\varphi^*\circ \phi^*_{E}\circ (\varphi^*)^{-1}$. More explicitly we obtain:
$$\phi_{\partial}\colon \GM\times X\dashrightarrow X \quad\mbox{given by}\quad (x,y)\mapsto\left(\frac{t^a((x-1)(y-1)+1)-1}{t^by-1}+1,t^by\right)\,.$$
We can recover the rational derivation $\partial$ by computing $\operatorname{ev}_1\circ\frac{d}{dt}\circ\phi^*_{\partial}$. Indeed, a tedious computation shows that
\begin{align*}
\frac{d}{dt}(\phi^*_{\partial}(x))&=\frac{\left(at^{a-1}[(x-1)(y-1)+1]\right)(t^by-1)-bt^{b-1}y[t^a[(x-1)(y-1)+1]-1]}{(t^by-1)^2} \\
\frac{d}{dt}(\phi^*_{\partial}(y))&=\frac{d}{dt}(t^by)=bt^{b-1}y\,.
\end{align*}
So that
\begin{align*}
\operatorname{ev}_1\circ \frac{d}{dt}\circ\phi^*_{\partial}(x)&=\frac{a((x-1)(y-1)+1)-b(x-1)y}{y-1} \\
\operatorname{ev}_1\circ\frac{d}{dt}\circ\phi^*_{\partial}(y)&=by\,,
\end{align*}
recovering the initial derivation $\partial$.
\end{example}
Recall that a rational slice of a rational $\GM$-action $\varphi$ is a function $s\in K_X$ such that $\varphi^*(s)=ts$. We can also characterize slices in terms of the corresponding rational semisimple derivation as we show in the following lemma.
\begin{corollary} \label{rational-slice}
Letting $X$ be an algebraic variety with fields of the rational function $K_X$, we let $\partial\colon K_X\to K_X$ be a rational semisimple derivation. Then $s\in K_X$ is a rational slice of $\phi^*_\partial(t)$ if and only if $\partial(s)=s$.
\end{corollary}
\begin{proof}
Assume first that $\partial(s)=s$. Then $\partial^i(s)=s$ for every $i\geq 0$ and so we have
\begin{align*}
\phi^*_{\partial}(s)&=\sigma\circ\exp{z\partial}(s)=s\cdot\sigma\left(\sum_{i\geq 0}\frac{z^i}{i!}\right)=ts\,.
\end{align*}
Now if $\phi^*_{\partial}(s)=ts$ then $\partial(s)=\operatorname{ev}_1\circ\frac{d}{dt}\circ\phi^*_{\partial}(s)=s$
\end{proof}
Furthermore, a slice provide a ruling of the field $K_X$ over the field of invariants $K^{\GM}_X$ of the $\GM$-action.
\begin{proposition}\label{cor desc}
If $s$ is a rational slice for an faithful rational $\GM$-action $\varphi\colon \GM\times X\dashrightarrow X$, then $s$ is transcental over $K^{\GM}_X$, $K_X=K^{\GM}_X(s)=(\ker\partial)(s)$ and $\partial=s\frac{d}{d s}$ on $K^{\GM}_X(s)$.
\end{proposition}
\begin{proof}
We define the set $T=\{a\in K^*_X \mid \varphi^*(a)=t^{i}a \mbox{ with } i \in \mathbb{Z}\}$. In the proof Proposition~\ref{prop2.8} we proved that $T$ generates $K_X$. Moreover, $T$ is a group under multiplication. We have $\varphi^*(a)=t^0a$ for all $a\in K^{\GM}_X$. Therefore $T_0:=K^{\GM}_X\setminus\{0\}$ is a subgroup of $T$. Let now $T/T_0$, given $a,b\in T$ such that $\varphi^*_t(a)=t^ia$ and $\varphi^*_t (b)=t^ia$, for some $i\in\ZZ$, we have $\varphi^*(ab^{-1})=ab^{-1}$ implies $ab^{-1}\in T_0$. Hence, $a$ and $b$ differ by a element of $T_0$.
We define the group homomorphism $T\to \ZZ$ given by $a\mapsto i$ where $\varphi^*(a)=t^ia$. This homomorphism is surjective by the existence of a rational slice, see Definition~\ref{def:rational-slice} and below. Moreover, its kernel is $T_0$. We conclude that $T/T_0\simeq\ZZ$ with a rational slice $s$ as generator. Since $T_0$ generated the field $K^{\GM}_X$ and $T$ generates the field $K_X$ we obtain that $K_X=K^{\GM}_X(s)$.
Assume now that $s$ is algebraic over $K^{\GM}_X$, i.e., assume that there exists a non trivial polynomial $P\in K^{\GM}_X[x]$ such that $P(s)=0$, then $\varphi^*(P(s))=P(ts)=0$. This is a contradiction since $t$ is is transcendental over $K_X$ and so the same holds over the subfield $K^{\GM}_X$. We conclude that therefore $s$ is transcendental over $K^{\GM}_X$. Finally, the structure of $\partial$ given as $\partial=s\frac{d}{d s}$ on $K^{\GM}_X(s)$ follows since $\partial(K^{\GM}_X)=0$ and $\partial(s)=s$.
\end{proof}
As in the rational case treated above, given a regular $\GM$-action $\varphi$ on $X$, we define a derivation of the structure sheaf $D_\varphi\colon \OO_X\to \OO_X$ given over every affine open set $U\subseteq X$ by
$$D_\varphi\colon \OO_X(U)\to \OO_X(U) \quad\mbox{given by}\quad f\mapsto \operatorname{ev}_{1}\circ \frac{d}{dt}\circ \varphi^*(f)\,.$$
Any derivation $\partial\colon \OO_X\to\OO_X$ induces a derivation $K_X\to K_X$ simply by extending $\partial\colon \OO_X(U)\to\OO_X(U)$ to the field of fractions $\fract(\OO_X(U))=K_X$ for any affine open set via the Leibniz rule. We denote this derivation also by the same symbol $\partial\colon K_X\to K_X$.
In the next proposition, we characterize the derivations of the structure sheaf $X$ that come from a regular $\GM$-action in the case where $X$ is semi-affine. Recall that a variety $X$ is called semi-affine if the canonical
morphism $X\to \spec \OO_{X}(X)$
is proper. In this case $\mathcal{O}_{X}(X)$ is finitely generated and so $\spec\mathcal{O}_{X}(X)$ is an affine variety \cite[corollary 3.6]{GoLa73}. For
instance, complete or affine $k$-varieties are semi-affine, blow-ups of semi-affine varieties are also semi-affine.
\begin{proposition}\label{prop:Regular-actions}
Regular $\GM$-actions on a semi-affine variety $X$ are in one-to-one correspondence with rational
semisimple derivations $\partial\colon\mathcal{O}_{X}\to\mathcal{O}_{X}$
such that the derivation on global sections $\partial_X\colon\OO_{X}(X)\to\OO_{X}(X)$
on the ring of global regular functions is semisimple with integers eigenvalues.
\end{proposition}
\begin{proof}
By Rosenlicht theorem \cite{Ros56}, for any regular $\GM$-action on $X$ there exists of a nonempty $\GM$-invariant affine open subset $U$. Hence, $\partial_U\colon \OO_X(U)\to \OO_X(U)$ is semisimple with integer eigenvalues and since $\OO_{X}(X)\subset\OO_{X}(U)$
it follows that $\partial_X$ is a semisimple
derivation of $\OO_{X}(X)$ with integer eigenvalues.
Conversely, let $\partial\colon\mathcal{O}_{X}\to\mathcal{O}_{X}$
be a derivation such that $\partial_{0}=\partial_X\colon\OO_{X}(X)\to \OO_{X}(X)$
is semisimple with integer eigenvalues. Then $\partial_{0}$ induces a possibly trivial
regular $\GM$-action $\varphi_{0}\colon\GM\times X_{0}\to X_{0}$
on $X_{0}=\spec \OO_X(X)$ for which the
canonical morphism $p\colon X\to X_{0}$ is $\GM$-equivariant.
In particular, for every point $x\in X$, letting $\xi=\varphi\mid_{\GM\times\left\{ x\right\} }\colon\GM\dashrightarrow X$,
$t\mapsto\varphi(t,x)$ and $\xi_{0}=\varphi_{0}\mid_{\GM\times p(x)}\colon\GM\to X_{0}$,\,$t\mapsto\varphi_{0}(t,p(x))$
, we have a commutative diagram \[\xymatrix {\GM \ar@{-->}[r]^{\xi} \ar[dr]_{\xi_0} & X \ar[d]^{p} \\ & X_0.}\]
Since $p$ is proper, we deduce from the valuative criterion for properness
applied to the local ring of every closed point $t\in\GM$
that $\varphi$ is defined at every point $\left(x,t\right)\in\GM\times X$
whence is a regular $\GM$-action on $X$.
\end{proof}
\bibliographystyle{alpha} |
\section{Introduction}
\label{sec:intro}
Point clouds are sparse 3D signals composed by geometry and attributes information (which may include color, reflectance, normal vectors, etc.), and have been in the spotlight of researchers in recent years for its usability in applications such as AR/VR (augmented and virtual reality) \cite{Schwarz2019REALTIMEDA,mpeg:m49235}, telecommunications \cite{mpeg:m38136}, autonomous vehicle \cite{mpeg:m43647} and world heritage \cite{mpeg:m37240}.
Because of the need for compression for either storage or transmission, the Motion Picture Experts Group (MPEG) has been directing efforts for the compression of point clouds in two fronts, i.e., geometry-basde point cloud compression (G-PCC) and video-based point cloud compression (V-PCC).
Concerning the geometric information, each point may be expressed as a list $V$ of unordered ternary information, such that the $n$-th point is $\textbf{v}_n = (x_n, y_n, z_n)$. As for its coding, while V-PCC uses video codecs for encoding its projection in a plane, G-PCC uses the \emph{octree} structure \cite{article:danilo:2020}.
Geometry coding under G-PCC may be either lossless or lossy. To encode geometry in a lossy way, G-PCC prunes the octree in order to eliminate nodes after a certain pre-determined level.
In this process, G-PCC uses a coordinate transformation at the encoder, which down-scales the original point cloud geometry $V$ such that, for the $n$-th point of $V_d$
\begin{equation}
\textbf{v}_{d_n}=\texttt{round}\left(\frac{\textbf{v}_n-T}{s}\right),
\label{eq:downsampling}
\end{equation}
where, $\texttt{round}(\cdot)$ is the function that rounds the components of a vector to the nearest integer,
\begin{equation}
T=(\min x_n, \min y_n, \min z_n),
\end{equation}
and $s>1$ is the scale factor, given by the inverse of the \texttt{positionQuantizationScale} parameter \cite{mpeg:codec_gpcc}.
This scaling reduces the number of points to be encoded due to rounding and to duplicate point removal, making $V_d$ a coarser geometry when compared to $V$. The larger the scale factor $s$, the lower the encoding rate, and consequently, the lower the number of output points, and the coarser the geometry.
At the decoder side, the scaled geometry is expanded and shifted back to its original position using the same values of $s$ and $T$.
Some previous techniques have been proposed in order to improve coding efficiency for these lossy cases, such as the use of slicing interpolation \cite{mpeg:m42519} inside G-PCC or even completely new approaches, such as using dyadic decomposition \cite{inproc:Eduardo}.
On the other hand, the use of a lookup table (LUT) based on neighborhood inheritance has been used for context generation on geometry coding of dynamic voxelized point clouds \cite{article:diogo:2020} as a replacement to the octree.
Most recently, Borges \emph{et al.~} \cite{article:tomas:2022} proposed the use of LUTs relating the downsampled neighbourhood of a given voxel with its children occupancy for super-resolution (SR)
of voxelized point clouds downsampled at arbitrary fractional scales.
Assuming that self-similarities are somewhat maintained at different scales, and taking into account the irregulaties of fractional downsampling, they estimate children occupancy based on the neighborhood of downsampled versions of a single input point cloud.
Results have shown that the fractional SR provides better results than other solutions, such as using nearest-neighbours interpolation (NNI) and NNI followed by laplacian smooting \cite{inproc:LS}.
Figure \ref{fig:method_example} shows an example of such results.
As a continuation of that work, we apply the fractional SR technique as a post-processing tool for point clouds decoded with G-PCC.
\begin{figure*}[t]
\centering
\begin{minipage}[b]{0.33\linewidth}
\centering
\centerline{\includegraphics[width=\textwidth]{figs/rb_org}}
\centerline{(a)}\medskip
\end{minipage}
\begin{minipage}[b]{0.33\linewidth}
\centering
\centerline{\includegraphics[width=\textwidth]{figs/rb_dec}}
\centerline{(b)}\medskip
\end{minipage}
\begin{minipage}[b]{0.33\linewidth}
\centering
\centerline{\includegraphics[width=\textwidth]{figs/rb_sr}}
\centerline{(c)}\medskip
\end{minipage}
\caption{Example of point clouds in the post-processing method: (a) $V$; (b) $V_{d}$ with NNI; (c) $V_{sr}$.}
\label{fig:method_example}
\end{figure*}
\section{Post-Processing Method}
\label{sec:format}
Since the fractional SR technique of Borges \emph{et al.~} \cite{article:tomas:2022} can be employed for arbitrary scale factors, it can be used to improve quality of point clouds downsampled using Eq. \eqref{eq:downsampling}, provided that $s$ is known.
Thus, it can be used to post-process point clouds geometric information, after being encoded and decoded using the G-PCC codec in lossy-geometry configuration.
Recognizing that the lossy encoding of geometry in G-PCC is but a downsampling, one can improve the downsampled point cloud geometry $V_d$ using the fractional scale factor $1<s\leq 2$ from the encoder, to generate $V_{sr}$.
Figure \ref{fig:method} shows the general idea of the proposed post-processing method. For some coding conditions, the value of $s$ may be greater than 2. In this situations, the fractional SR may be applied successively.
\begin{figure}[htb]
\centering
\includegraphics[scale=1]{figs/method.pdf}
\caption{The fractional SR is applied to the decoded point cloud using the same scale factor as used in the downsampling of the G-PCC codec.}
\label{fig:method}
\end{figure}
The SR method relies on the use of the adjacent neighbourhood of a given voxel in order to predict its children nodes. For some sparser point clouds, the immediate neighbours may not be available for some (or maybe all) voxels. In order to surpass this problem, we down-scale the input point cloud by a factor $s'$ prior to the encoding. Then, a subsequent up-scaling of the cloud $V_{sr}$ by the same factor is necessary, after the ``frac SR'' step, as depicted in Fig. \ref{fig:dus}.
\begin{figure}[htb]
\centering
\includegraphics[scale=1]{figs/dus.pdf}
\caption{For sparse clouds, a down-scaling is applied previously to the encoding a a up-scaling is applied after the SR process.}
\label{fig:dus}
\end{figure}
\section{Experimental Results}
\label{sec:pagestyle}
\begin{figure}[htb]
\centering
\begin{minipage}[t]{\linewidth}
\centering
\centerline{\includegraphics[width=7.5cm]{figs/dancer_vox11_00000001_plot_d1}}
\centerline{(a)}
\medskip
\end{minipage}
\begin{minipage}[t]{\linewidth}
\centering
\centerline{\includegraphics[width=7.5cm]{figs/longdress_vox10_1300_plot_d1}}
\centerline{(b)}
\medskip
\end{minipage}
\caption{Comparative results under D1 PSNR metric for solid point clouds: (a) \textit{dancer}, (b) \textit{longdress}.}
\label{fig:rd_curves1}
\end{figure}
\begin{figure}[htb]
\begin{minipage}[b]{\linewidth}
\centering
\centerline{\includegraphics[width=7.5cm]{figs/house_without_roof_00057_vox12_plot_d1}}
\centerline{(a)}\medskip
\end{minipage}
\begin{minipage}[b]{\linewidth}
\centering
\centerline{\includegraphics[width=7.5cm]{figs/staue_klimt_vox12_plot_d1}}
\centerline{(b)}\medskip
\end{minipage}
\caption{Comparative results under D1 PSNR metric for non-solid point clouds: (a) \textit{house\_without\_roof}, (b) \textit{staue\_klimt}.}
\label{fig:rd_curves2}
\end{figure}
In order to test the post-processing method, we propose two distinct experiments, according to the sparsity of the input point clouds, for solid (voxelized point clouds with continuous surface) and non-solid clouds, according to the categories used in MPEG \cite{mpeg:m55485}.
In the first one, we apply the fractional SR directly to the decoded versions of solid point clouds (as in Fig. \ref{fig:method}). The scaling factor is defined by the MPEG G-PCC's Common Test Conditions (CTC) \cite{mpeg:ctc_gpcc} according to the point cloud's geometry for each rate point to be tested, as shown in Table \ref{tab:rCTC}.
The greater the rate id (as in ``R6''), the greater the bit rate, the lower the value of $s$.
\begin{table}[htb]
\small
\centering
\begin{tabular}{||c c c c c c c||}
\hline
Geometry precision & R6 & R5 & R4 & R3 & R2 & R1 \\ [0.5ex]
\hline\hline
10 & 15/16 & 7/8 & 3/4 & 1/2 & 1/4 & 1/8 \\
11 & 7/8 & 3/4 & 1/2 & 1/4 & 1/8 & 1/16\\
\hline
\end{tabular}
\caption{The \texttt{positionQuantizationScale} parameter used in G-PCC's CTC for different geometry precision to generate each rate point.}
\label{tab:rCTC}
\end{table}
In the second experiment, for testing non-solid point clouds, we apply a downscaling prior to encoding. The scaling factor $s'$ is defined as the highest power of 2 that makes $V_d$ keep approximately the same number of points as $V$, i.e., we enforce a downscaling only to densify the input geometry, avoiding over-decimating its points. The values are shown in Table \ref{tab:rCTC2}, for the two tested clouds. According to the G-PCC CTC's ``lossy-geometry'' section, the largest possible value used for this scaling would be 2048 (defined for point clouds with 20 bits in geometry precision). This means we only need 4 bits to transmit $s'$ as side information. Since the bitrate is calculated in bits per input points, i.e., the number of points in the original point cloud (which vary from hundreds of thousands to millions), these 4 bits are negligible.
\begin{table}[htb]
\small
\centering
\begin{tabular}{||c c||}
\hline
Point cloud & $s'$ \\ [0.5ex]
\hline\hline
house\_without\_roof\_00057\_vox12 & 2\\
statue\_klimt\_vox12 & 4 \\
\hline
\end{tabular}
\caption{Down-scaling factor for non-solid clouds.}
\label{tab:rCTC2}
\end{table}
For both experiments, we compare our results to those of G-PCC as well as to machine learning (ML) based end-to-end compression techniques \cite{quach2020improved,wang2020multiscale,article:AIPCC:ADL}, with test results provided by Zaghetto \cite{mpeg:zaghetto:AI1}. It is important to note that we only bring the results for the sequences used in the ML tests, for a direct comparison. Some of the clouds in the G-PCC's CTC where used in the training of some of these models, and therefore could not be used for testing.
Figures \ref{fig:rd_curves1} and \ref{fig:rd_curves2} show the rate-distortion curves for some of the obtained results, for solid and non-solid clouds, respectively. The rate is measured in bits per input point (the number of points in the original input point cloud), while the distortion is measured with the point-to-point PSNR (D1 PSNR) metric \cite{Girardeau2005}.
We note that our results bring great improvement over G-PCC, specially in the higher rates region. We also note that they are comparable to those of ML-based techniques, except for PCGCv2, which outperforms all others. However, we state the fact that our method doesn't require any type of training, while all the ML techniques require the training of models for each rate point shown in the curves.
We also notice a drop in the lower rates region of the curves.
At this region, the scale factor $s$ is greater than 2, requiring to use the fractional SR more than once. This is not ideal because every time we apply it, we imply the neighbourhood for the next application, which may propagate errors.
\begin{table*}[t]
\small
\centering
\begin{tabular}{||l c c c c c||}
\hline
Cloud name & ADL-PCC & PCC-GEO-CNN-v2 & PCGCv2 & frac SR & frac SR (high-rates) \\ [0.5ex]
\hline\hline
dancer\_vox11\_00000001 \cite{Owlii2017} & -77.04 & -87.81 & \textit{-93.63} & -59.79 & \textbf{-91.77} \\
longdress\_vox10\_1300 \cite{8iDataset} & -75.67 & -77.06 & \textit{-89.09} & -62.39 & \textbf{-80.30} \\
loot\_vox10\_1200 \cite{8iDataset} & -75.86 & -74.88 & \textit{-90.38} & -64.40 & \textbf{-81.17} \\
queen\_0200 & -75.93 & -79.46 & \textit{-92.85} & -70.23 & \textbf{-85.29} \\
redandblack\_vox10\_1550 \cite{8iDataset} & -74.11 & \textbf{-75.88} & \textit{-88.62} & -62.52 & -73.89 \\
soldier\_vox10\_0690 \cite{8iDataset} & -76.99 & -75.30 & \textit{-89.39} & -65.63 & \textbf{-80.66} \\
\hline
\end{tabular}
\caption{BR-Rate comparison to G-PCC for solid clouds, in rate \%.}
\label{tab:solid}
\end{table*}
\begin{table*}[t]
\small
\centering
\begin{tabular}{||l c c c c||}
\hline
Cloud name & ADL-PCC & PCC-GEO-CNN-v2 & PCGCv2 & frac SR+DUS\\ [0.5ex]
\hline\hline
house\_without\_roof\_00057\_vox12 & \textbf{-41.14} & 26.91 & \textit{-87.77} & -30.84 \\
statue\_klimt\_vox12 & \textit{-41.16} & 207.18 & \textbf{-39.50} & -37.80 \\
\hline
\end{tabular}
\caption{BR-Rate comparison to G-PCC for non-solid clouds, in rate \%.}
\label{tab:sparse}
\end{table*}
In Tables \ref{tab:solid} and \ref{tab:sparse} we bring the BD-rate comparison of all results to that of G-PCC for solid and non-solid clouds, respectively. In Table \ref{tab:solid}, we also show a column comparing our results considering only the four higher rate points (``frac SR (high-rates)''), i.e. the four rightmost points from the graphs shown in Figure \ref{fig:rd_curves1}, where the rate ranges are closer to most of those from the ML techniques, bringing therefore a more adequate for comparison.
Table \ref{tab:sparse} shows ``frac SR+DUS'' (down-up-scaling) to highlight the difference between the two experiments.
We also show an example of a viewpoint from \textit{longdress} for $s=2$ in Fig. \ref{fig:longdress_prints}.
Figure \ref{fig:longdress_prints}(a) shows the original uncompressed cloud.
In Figs. \ref{fig:longdress_prints}(b) and \ref{fig:longdress_prints}(c), we see the decoded point cloud under G-PCC ($V_d$) and the latter with the post-processing ($V_{sr}$), respectively.
Important to note that the colors shown in \ref{fig:longdress_prints}(c) are obtained with a recoloring tool, only for a better observation.
Analysing these two figures, one can clearly see the increase in the number of points and the filling of holes.
This is further emphasized in Figs. \ref{fig:longdress_prints}(d) and (e), where the Euclidean distance is shown for each voxel comparing the distorted and the original clouds, first between $V^{orig}$ and $V_d$, then between $V^{orig}$ and $V_{sr}$.
The colorbar scale ranges from 0 to 3.0.
\section{Conclusions}
\label{sec:typestyle}
We have presented a method for post-processing point clouds which were encoded and decoded using MPEG's G-PCC codec by applying a previously proposed fractional SR technique. We compared this method with the originally decoded clouds as well as with ML-based end-to-end coding techniques. The results have shown that our method brings great improvement in quality (D1 PSNR metric) over G-PCC, specially for higher rate values (small scaling factor) in solid point clouds. In all tested cases, the method is compatible to those based on ML, except for one, which outperforms all others. However, our method does not need any sort of previous training, which is required for the ML techniques.
\begin{figure}[!h]
\begin{minipage}[b]{0.48\linewidth}
\centering
\centerline{\includegraphics[width=0.95\linewidth]{figs/img_org_crop}}
\centerline{(a)}\medskip
\end{minipage}
\begin{minipage}[b]{0.48\linewidth}
\centering
\centerline{\includegraphics[width=0.95\linewidth]{figs/img_r3_crop}}
\centerline{(b)}\medskip
\end{minipage}
\begin{minipage}[b]{0.48\linewidth}
\centering
\centerline{\includegraphics[width=0.95\linewidth]{figs/img_r3_sr_crop}}
\centerline{(c)}\medskip
\end{minipage}
\begin{minipage}[b]{0.48\linewidth}
\centering
\centerline{\includegraphics[width=0.95\linewidth]{figs/img_r3_org_crop}}
\centerline{(d)}\medskip
\end{minipage}
\begin{minipage}[b]{0.48\linewidth}
\centering
\centerline{\includegraphics[width=0.95\linewidth]{figs/img_r3_sr_org_crop}}
\centerline{(e)}\medskip
\end{minipage}
\caption{Example of comparative results for \textit{longdress} compressed with $s=2$: (a) uncompressed ($V^{orig}$); (b) compressed with G-PCC ($V_{d}$); (c) compressed with G-PCC+SR ($V_{sr}$); (d) $V_{d}$ \textit{vs} $V^{orig}$; (e) $V_{sr}$ \textit{vs} $V^{orig}$.}
\label{fig:longdress_prints}
\end{figure}
\bibliographystyle{./_config/IEEEbib}
|
\section*{Abstract}
A model needs to make verifiable predictions to have any scientific value. In opinion dynamics, the study of how individuals exchange opinions with one another, there are many theoretical models which attempt to model opinion exchange, one of which is the Martins model, which differs from other models by using a parameter that is easier to control for in an experiment. In this paper, we have designed an experiment to verify the Martins model and contribute to the experimental design in opinion dynamic with our novel method.
\section*{Introduction}
The field of opinion dynamics has a wide variety of theoretically derived models that potentially describe human interactions and the resulting change in opinions. Despite the appeal of these models, there is a dearth of empirical evidence to support their utility \cite{Mcloserlinktoreality2009Sobkowicz}. For a model to be scientifically verifiable, the model needs to make testable predictions about the outcome of an experiment. While modern examples of research \cite{EDmultidimopinion2021Friedkin,EDrecursivecommunication2020Pescetelli} demonstrate an effective method to investigate opinion dynamics models, most theoretical opinion dynamics models don't offer predictions on behaviours, which makes it challenging if not impossible to create controlled experiments which can verify these models \cite{COsocialinfluencefrontiers2017Flache}. Consider the bounded confidence model \cite{BCboundedconfidence2002Hegselmann,BCmeetdiscussseg2002WeisbuchDeffuant}, which includes the parameter $\epsilon$ limiting agent interactions. Certain values of $\epsilon$ can create polarisation. But because $\epsilon$ is an abstract (and highly subjective) measure in opinion space, it is difficult to create an experimental condition to control $\epsilon$. In general, the level of abstraction in the opinion dynamics models' parameterisations limits the design and implementation of experiments for testing model validity. Further, opinion dynamics models created from data are also difficult to verify because, as stated in \cite{COsocialinfluencefrontiers2017Flache}, models fitted to the data of experiments rarely make testable predictions about future data. We break this trend by designing and executing an experiment testing the claims made by the Martins model \cite{BSbayesianupdating2009Martins}.
The Martins model \cite{BSbayesianupdating2009Martins} represents opinions as probability density functions such that a person has an opinion $x \in \mathbb{R}$ and an associated uncertainty $\sigma \in \mathbb{R}^+$. Their opinion and uncertainty represent a Gaussian density function with mean $x$ and standard deviation $\sigma$. When two agents interact in the Martins model, they share their opinions (and in the extended model \cite{BSmistrust2021Adams} their uncertainties), and the two agents then update their opinion and uncertainty via Bayesian updating. A key parameter in the model is $p\in [0,1]$, which is the propensity for agents to believe that other agents have useful information and use that information to update their opinions. When $p=1$ consensus is always reached, whereas values of $p<1$ polarisation emerges. Compared to the bounded confidence model's $\epsilon$, the parameter $p$ is much more interpretable and controllable (in an experimental setting) than the parameter $\epsilon$.
We present, in this article, a comprehensive literature review of previous empirical studies in opinion dynamics. Then we outline a design for an experiment which can test whether the Martins model can predict opinion shifts of individuals. We executed such an experiment and, in this article, present the results of the experiment. In the results, we found two distinct phenomena occurring in the experiment: when two individuals are close in opinion, the Martins model made a reasonable prediction of the opinion shift; when two individuals are far in opinion, the observed opinion shift followed a what would be expected from discrete opinion choice model. We concluded by discussing our novel results and identifying the limitations of our experiment.
\section*{Previous experiments}
There is limited evidence of direct use in opinion dynamics of experimental data to either verify hypotheses based on model predictions or construct empirical models. This scarcity of evidence is partly due to the difficulty of designing an experiment that accurately replicates real-world interactions while controlling the experimental conditions and has resulted in empirical data collection in opinion dynamics evolving independently from the theory.
\subsection*{Experimental Data Collection}
Many empirical investigations into opinion dynamics draw inspiration from psychology studies that investigated opinion change \cite{EDunderminewiscrowd2011Lorenz,EDpeoplestraforopinchange2009Soll,EDextremememberspolarization2009Swol,EDhumandecisionprocesses2004Yaniv}. All of these studies served as guidelines for the experimental design of the later opinion dynamics studies. For example, the study \cite{EDextremememberspolarization2009Swol} aimed to test two hypotheses: ``Extreme members will contribute more to the group discussion than less extreme members. (1a) They will use more words than less extreme members, and (1b) they will take more turns than the less extreme member," and (2) ``There should be greater group polarisation in a group containing an extreme member than in groups not containing an extreme member." The authors of \cite{EDextremememberspolarization2009Swol} tested these by dividing 129 participants into 43 groups of three. Participants in each group were asked their opinion and knowledge on the legalisation of marijuana before the experiment. The participants read material related to the legalisation of marijuana and then discussed the issue within their group until the group reached a compromise. After the discussion, participants reevaluated their opinion. This experimental design formed the bases for the approach to collecting data in the opinion dynamics literature.
\subsection*{Building Empirical Models}
While Opinion Dynamic's inception began in the 1950's \cite{COsocialpower1956French}, one of the first significant studies focused on developing a model of opinion change using experimental data was published in 2013 \cite{EDcollectivedynamics2013Moussaid}.
The experimental design in \cite{EDcollectivedynamics2013Moussaid} draws direct inspiration from the previous psychology literature, but the study generated a model of human behaviour from the collected data rather than prove any specific hypothesis. Participants were asked general knowledge questions with a real number answer, e.g. ``How long is the Mississippi river?" and rated their confidence in their answer on a 1 to 6 scale, with lower meaning less confident. Participants only saw one other participant's answer and confidence at a time.
The authors of \cite{EDcollectivedynamics2013Moussaid} used the experimental data to create an influence map of the experimental subjects' behaviours. An influence map is a surface in relative opinion and relative uncertainty space, which describes the opinion change of an individual according to their relative opinion and relative uncertainty with a hypothetical interaction partner. The authors used the influence map to create a decision tree model. Depending on where an interaction fell on the influence map, the model specifies three ways an agent could update their opinion after an interaction with another agent: \textit{rejecting} where there is no change in opinion; \textit{compromising} where the opinion shifted `halfway' towards the other opinion; and \textit{adopting} where the opinion changed to be the other opinion. The resulting model is related to the bounded confidence model \cite{BCboundedconfidence2002Hegselmann,BCmeetdiscussseg2002WeisbuchDeffuant} such that the regions of rejecting, compromising and adopting could be used to determine an $\epsilon$, but the influence map of \cite{EDcollectivedynamics2013Moussaid} implies a more nuanced picture which the bounded confidence model cannot address.
More modern models like the Martins \cite{BSbayesianupdating2009Martins} and relative agreement models \cite{Mrelativeagree2002Deffuant} produce similar behaviour seen in the influence map generated by \cite{EDcollectivedynamics2013Moussaid}. But it is difficult to precisely confirm whether the models like Martins or relative agreement can accurately predict the behaviour observed in \cite{EDcollectivedynamics2013Moussaid}. Specifically, the Martins and relative agreement models rate confidence/uncertainty as continuous values, which conflicts with the discrete 1-6 scale \cite{EDcollectivedynamics2013Moussaid} used to measure confidence, therefore making the empirical data incompatible with the theoretical models. The goal of \cite{EDcollectivedynamics2013Moussaid} was to use the data to generate a model, not to verify an existing model.
\subsection*{Other Experimental Designs}
Modern studies have improved the experimental design of \cite{EDcollectivedynamics2013Moussaid}. For example, the work \cite{EDrecursivecommunication2020Pescetelli} provided a novel contribution where participants interacted in pairs through digital displays and exchanged their opinions, but participants could see their interaction partner update their opinion in real-time. The new method proved a controlled, yet realistic, environment to test ideas about opinion exchange and revealed new behaviour in which participants became more confident upon observing that their interaction partner changed their opinion. This empirical evidence provides clues to produce theoretical models which can predict opinion exchange more effectively.
More significant is the work of \cite{EDmultidimopinion2021Friedkin}. Specifically, the study \cite{EDmultidimopinion2021Friedkin} investigates how groups assessed threatening objects and developed a model similar to the French model \cite{COsocialpower1956French} which establishes seven testable predictions ranging from conditions on how individuals modify their threat assessments to predicting the process in which society might reach consensus. Agents in the model of the study, like in the original French model \cite{COsocialpower1956French}, weighted their neighbours such that when the model progressed, the agents would adopt the weighted average opinion of their neighbours according to the weighting the agent assigned each neighbour. For example, consider a three agent simulation with Agents 1, 2 and 3 holding opinions $x_1,x_2,x_3$ respectively, and Agent 1 weighting every agent in the simulation according to this vector $[0.2,\ 0.3,\ 0.5]$. When the model updates Agent 1's opinion will be $0.2x_1+0.3x_2+0.5x_3$. The study measured these weightings by first giving 100 chips to each participant after the group discussion. Next, participants distributed their chips according to how much they were convinced by other group members that an object was ``threatening". Participants were allowed to keep chips if they were not convinced by the group. The chip distribution provided by each participant directly measured the weightings necessary for the modified French model. The study concludes by evaluating the model's ability to predict an individual's threat assessments. The work of \cite{EDmultidimopinion2021Friedkin} demonstrates an effective method to evaluate the opinion dynamics model, which this paper hopes to emulate.
The side of opinion dynamics concerned with measuring the most influential individual has already produced empirical studies that seek to verify theoretical models' predictions. The work of \cite{SNoldindegree2012Corazzini} offshoots from the French model \cite{COsocialpower1956French} by imposing that the weights of the French be related to the in-degree agents, i.e. how well listened an agent is. This weighting scheme relies on a parameter $\rho$ such that when $\rho = 0$ in-degree does not affect opinion dissemination when $\rho = 1$ neighbours are weighted proportionally to an agent's personal in degree and when $\rho \rightarrow \infty$ agents only listen to the neighbour with the most in-degree. In the study, \cite{SNindegree2015Battiston} the authors developed an experiment that isolated the effect of in-degree. Specifically, the authors developed a social network for participants that controlled for in-degree. The result of the experiment was the rejection of the null hypotheses, i.e. $\rho = 0$, suggesting that in-degree has a role in opinion dissemination. With the experiment of this study, we seek to accomplish a similar goal on the inter-personal level and ascertain whether mistrust influences interpersonal communication as the Martins model describes.
\section*{Materials and Methods}
In comparison with previous work, we designed this study's experiment to be more abstract. This abstraction allows for a more direct comparison between the model variables, i.e. agent uncertainty and opinion, and the data collected from the experiment. In addition, the abstraction minimises the impact of cultural bias. Consider the general knowledge question used in previous experiments. The questions limited the pool of participants to those somewhat knowledgeable of the topic, e.g. \ a question like ``How long is the Mississippi river?" limits participants to those from the US. We naturally avoided this problem with our experiment. We took advantage of this flexibility to make the study a snowball sample study; participants are encouraged to invite others to participate, to increase the sample size for the study.
\subsection*{Ethics Statement}
The following experimental design and experiment was approved by the Queensland University of Technology (QUT) Human Research Ethics Committee (UHREC) as Negligible-Low Risk. Reference number: 2000000739.
\subsection*{Recruitment of Participants}
As stated previously, our recruitment strategy for this experiment was a snowball sampling strategy. We advertised the experiment on the social media websites Facebook and YouTube and through the mailing list and Slack workgroups of the QUT mathematics school. When a participant finished the experiment, we recorded the IP address associated with the device they used for the experiment as part of the data. We recorded IP addresses to determine the number of unique participants in the experiment. We recorded no other personal data on the participant. A total of $257$ unique participants participated in the experiment, assuming that participants are unique to each IP address.
\subsection*{The Experiment}
The experiment entailed playing a game on the internet hosted on QUT servers. The goal of the game was to find a hidden dot inside a black box on screen. Participants were given information to find the dot in the context of a social interaction. A participants would first see a blue circle which was explained as information they was always reliable. The hope being that a participant would internalise the blue circle as their opinion. Next a participant would see a red circle which was explained as not being reliable all the time. The idea being that participants would interpret the red circle as rumours which are subject to being false. Lastly a participant would draw a new circle in response to the red and blue circles. We directly controlled the reliable of the red circle while informing the participants which allowed us to control for $p$ in the Martins model. See \nameref{S1_File} for the source code of the website and in total we had 3760 games played.
\subsubsection*{Participant Instructions}
Before any participant played a game, they first saw instructions for the game. The instructions described the game similarly to how it is described in this paper except without the probability terminology, e.g.\ confidence intervals, to avoid confusion for the participants. The instructions developed a backstory to the game to encourage participants to role-play so that the participants responded realistically. The instructions described an eccentric Flemish trillionaire, Monsieur Dotte, as hosting the game, and they wanted the world to indulge in his passion for puzzles and social deduction. So, M. Dotte offered a `cash prize' for those who do well at his game of finding the dot. This framing allowed us to communicate specific information to participants, e.g.\ the reliability of the red circle at different traffic light signals and the 80\% chance a reliable circle had in containing the dot, while keeping the scenario plausible in the minds of the participants. It was made clear that there was no monetary reward for playing the game and the `cash' was just their score after finishing the game and held no fiscal value. See \nameref{S1_Fig} for the instructions we gave to the participants on the game website.
\subsubsection*{The Game}
Initially, a participant would see an empty black box. The participant would then click on the box resulting in the blue circle appearing, e.g.\ they could see \nameref{S2_Fig}. The blue circle represented an 80\% confidence interval which was explained to the participants as an 80\% to contain the dot. The blue circle was 100\% reliable and always gave information on the dot's location. The dot could be outside the blue circle, but because the blue circle was 80\% confidence interval, the dot would appear close to the circle.
When participants clicked again, the red circle would appear, e.g.\ they could see \nameref{S3_Fig}. Like the blue circle, the red circle purports to be an 80\% confidence interval of the dots' position, but the red circle has a probability of being unreliable, i.e. drawn at random and independent of the dot's actual location. To communicate the unreliability of the red circle a traffic light above the box would light up such that: when the traffic light was red, the red circle had a reliability of 20\%; when the light was yellow, the red circle had a reliability of 50\%; when the light was green, the red circle had a reliability of 80\%. The confidence interval of the red circle remained constant when the red circle was randomly determined to be reliable. Otherwise, the red circle would be drawn randomly inside the box. These reliability probabilities were communicated to the participants in the instruction and were our attempt to exactly quantify $p$ to allow for more definitive model predictions.
Finally, the game directed the participant to consider the position and reliability of the circles and draw a new circle (hereafter, the `user circle') that they believed contained the dot. After the participant finished drawing the user circle they were scored based on their accuracy (whether the dot was in their circle) and their precision (how small their circle was) relative to the blue circle and were encouraged to play again. Then we recorded the final game states, including the size and position of all three circles, the reliability of the red circle, a unique session id and the participant's IP address. \nameref{S4_Fig} shows an example of a finished game.
\subsubsection*{Collected Data}
The Martins model \cite{BSbayesianupdating2009Martins} and its extension \cite{BSmistrust2021Adams} predicts an individual's shift in opinion based on the parameter $p$, the opinions $x$ and the uncertainties $\sigma$ of both individuals involved in an interaction. As part of the experiment we recorded the specific values for $p$, $x_i$, $x_j$, $\sigma_i$ and $\sigma_j$ of every simulated interaction. i.e.\ every game that a participant played and Table \ref{table1} describes how we organised that data.
\begin{table}[!ht]
\centering
\caption{
{\bf Data collected from the experiment}}
\begin{tabular}{|l+l|}
\hline
{\bf Variable name } & \textbf{Description} \\\thickhline
$p$ & The probability of the red circle giving useful information \\\hline
$\Main{x}$ & The x-coordinate of center of the blue circle \\\hline
$\Main{y}$ & The y-coordinate of center of the blue circle \\\hline
$\Main{r}$ & The radius of the blue circle \\\hline
$\Other{x}$ & The x-coordinate of center of the red circle \\\hline
$\Other{y}$ & The y-coordinate of center of the red circle \\\hline
$\Other{r}$ & The radius of the red circle \\\hline
$\User{x}$ & The x-coordinate of center of the user circle \\\hline
$\User{y}$ & The y-coordinate of center of the user circle \\\hline
$\User{r}$ & The radius of the user circle \\\hline
\end{tabular}
\begin{flushleft} The blue circle is the first circle seen by the participant and represents the knowledge already acquired. The red circle is the second circle seen by the participant and represents the opinion and conjecture of another actor. The user circle is the circle drawn by the participant.
\end{flushleft}
\label{table1}
\end{table}
Except for $p$, the data in Table \ref{table1} are in units of pixels, whereas the Martins model deals in the abstract (unitless) opinion space. For clarity and similitude, we scaled all the relevant data removing the unit of pixels. We calculated the scaling factor by finding the radius of the circle of area equivalent to an HD monitor display (1920 by 1080). We then divided the radius by the number of standard deviations to produce an 80\% confidence interval, resulting in 634 pixels per standard deviation. We used this factor to scale the data and remove the unit pixels.
\subsubsection*{Scoring}
We encouraged participants to play multiple games by giving the participant a score per attempt. Scoring a participant's game follows these steps:
\begin{enumerate}
\item Calculate an accuracy $\User{A}\in \mathbb{R}$ and precision $\User{P}\in \mathbb{R}$ rating for the participant based on the circle they drew.
\item Produce an overall rating for the participant, $\User{R}$, as a weighted sum of $\User{A}$ and $\User{P}$ with weights $w_A$ and $w_P$, respectively, e.g.\ in the experiment $w_A = 0.1$ and $w_P=70$.
\item Repeat steps 1 and 2 for the blue circle, producing a rating for the blue circle $\Main{R}$.
\item Find the relative rating $R$ for the participant's circle
$$R = \User{R} - \Main{R} + R_0,$$
where $R_0$ is the rating given for guessing equally well as the blue circle, e.g.\ for the experiment $R_0 = 2\times 10^{-3}$. Giving a participant a rating relative to the blue circle encourages participants to guess better than the guaranteed information they start with.
\item Calculate a score $S \in [0,S_\mathrm{max}]$ as a sigmoid function of $R$, e.g.\ the experiment used
$$\frac{S_\mathrm{max}}{1+e^{-R/2}}.$$
The constant $S_\mathrm{max}$ is the maximum score achievable when playing the game, e.g.\ in the experiment $S_\mathrm{max}$ is one hundred thousand.
\end{enumerate}
\begin{remark}
To calculate accuracy and precision we used the following
\begin{align}
&A(e,r) = \frac{1}{1/L_\mathrm{max} + e} + U_\mathrm{bonus} \frac{r - r_\mathrm{min}}{r_\mathrm{min}},
\label{eqn:accuracy} \\
&P(r) = P_\mathrm{factor}\frac{1}{\pi \left(r - r_\mathrm{min}\right)^2} - \frac{\pi \left(r - r_\mathrm{min}\right)^2}{B_\mathbf{area} - \mathrm{min}\left[B_\mathbf{area},\pi \left(r - r_\mathrm{min}\right)^2\right]},
\label{eqn:precision}
\end{align}
where $e$ is the error of the circle which is the distance between the centre of the circle and the dot, $r$ is the radius of the circle, $r_\mathrm{min}$ is the radius of the smallest possible circle that can be drawn, $L_\mathrm{max}$ is the maximum accuracy score achievable when getting $e=0$, $U_\mathrm{bonus}$ controls how much the circle radius factors into accuracy, $B_\mathbf{area}$ is the area of the box containing the dot and $P_\mathrm{factor}$ controls for how precise a circle of a given area is. For the experiment $L_\mathrm{max} = 100$, $U_\mathrm{bonus} = 0.01$ and $P_\mathrm{factor} = 1$.
We rated accuracy and precision this way for two main reasons. First was so that accuracy and precision would be completely unrelated to the Martins model because if it were, participants would be encouraged to guess more like the Martins model, thus biasing the data. The second was for the ratings to produce a ``fair'' score by relating it to tangible concepts, e.g.\ a circle the size of the box would be considered very imprecise and thus would give no score, i.e.\
$$\User{P}\rightarrow -\infty \implies \User{R}\rightarrow -\infty \implies S = 0.$$
A score that a participant considers fair would encourage them to continue playing, at the very least not dissuade them.
\end{remark}
\subsection*{Model and Data Predictions}
The Martins model predicts that a user's opinion will fall on the line segment connecting the centres of the blue and red circles. The predicted distance from the centre of the blue circle to the user's opinion is
\begin{align}
h_{\mathrm{expected}}
&= p^* \frac{d_x}{1 + R^2_\sigma}, \label{ChangeInX}
\end{align}
where $d_x$ is the distance between the centres of both the blue and red circle, $R_\sigma$ is the ratio of both the red and blue circles' confidences which in this case means the ratio of both circles' radiuses and $p^*$ is a variable in the model dependent on $d_x$ relative to $R_\sigma$ (see \nameref{S1_Appendix} for more details). Similarly, the difference between the variances of the user circle and the blue circle is predicted to be
\begin{align}
k_{\mathrm{expected}}
&= p^*\left(\frac{1}{1 + R^2_\sigma}\right) \left( (1-p^*)\frac{(d_x)^2}{1 + R^2_\sigma} - \sigma_i^2\right). \label{ChangeInSig}
\end{align}
We can multiply this quantity by $\pi$ and $1.29^2$ to get the expected change in circle area, where $1.29$ is the number of standard deviations away from the mean required to construct an 80\% confidence interval. These theoretical values can be directly compared to the observed data and assessed for the goodness of fit.
We calculate the observed shift towards the red circle $h$ and change in circle area $k$ as
\begin{align}
h_{\mathrm{observed}}
&= \frac{(\Main{x} - \User{x})(\Main{x} - \Other{x}) + (\Main{y} - \User{y})(\Main{y} - \Other{y})}{d_x}, \\
k_{\mathrm{observed}}
&= \left(\Main{r}^2-\User{r}^2\right)\pi.
\end{align}
To compare the Martins model with the data, we calculate the following: $d_x$, the distance between $x_i$ and $x_j$; $\sigma_i$, the uncertainty of initial belief; $R_\sigma$, the ratio of $i$ and $j$'s confidences; and the Martins model quantity $p^*$. The value $d_x$ is the distance between the centers of both the blue and the red circles and is thus
$$ d_x = \sqrt{(\Main{x} - \Other{x})^2 + (\Main{y} - \Other{y})^2}. $$
The circles are presented to the participants as 80\% confidence intervals, therefore
$$ \sigma_i = \mathrm{\Main{r}}/1.29$$
where $1.29$ is the number of standard deviations from the mean required to get 80\% confidence. The ratio of $i$ and $j$'s confidences is
$$ R_\sigma = \frac{\mathrm{\Main{r}}}{\Other{r}}. $$
The quantity $p^*$ is a function of $d_x$, $\sigma_i$ and $R_\sigma$ and is
\begin{align}
p^* = \frac{p\phi\left(d_x,\sigma_i\sqrt{1 + R^2_\sigma}\right)}{p\phi\left(d_x,\sigma_i\sqrt{1 + R^2_\sigma}\right) + (1 - p)} \label{pStarMod}
\end{align}
where
\begin{align}
\phi\left(d_x,\sigma_i\sqrt{1 + R^2_\sigma}\right) = \left(1/\left(\sigma_i\sqrt{2\pi(1 + R^2_\sigma)}\right)\right)e^{-(d_x)^2/2\sigma^2_i(1 + R^2_\sigma)}. \label{phi}
\end{align}
Intuitively these predictions mean, even when $p$ is low, we expect to see participants shift towards the red circle more and reduce the size of their circle (relative to the blue) more when the red and blue circle are close to each other. Likewise, even when $p$ is high, we expect to see participants remain close to the blue while keeping the same radius as the blue circle when the red and blue circles are very far apart. This is due to the influence of $p^*$, because $p^*$ controls the degree an agent incorporates a new opinion and $p^*$ depends on $d_x$, i.e.\ the difference in opinion, and $\sigma_i^2 + \sigma_j^2$, i.e.\ the total variance of both agents' opinion. The reliability $p$ only effects the speed at which agents in the model effectively trust other agents and hence doesn't produce significantly different behaviour for values of $p$ between $0.2$ and $0.8$ (except for extreme case like when $p\rightarrow 1$ \cite{BSmistrust2021Adams}). Given the Martins model prediction on the circle participants draw in the game we supply the hypotheses
\begin{enumerate}
\item There is no correlation between the observed and expected shifts away from the blue circle.
\item There is no correlation between the observed and expected change in circle area from the blue circle.
\end{enumerate}
\section*{Results}
For each game, we calculated the user's predicted shift from the blue to the red circle and the change in their circle area compared to that of the blue circle using the Martins model. We compared the predicted results with the observations in Fig \ref{fig1} (for the raw dataset see \nameref{S2_File}). Upon initial observation, we see in the data a few outliers where participants drew large circles in random places relative to the blue circle. We surmise that participants were likely trying to `break' the game in the experiment by drawing the largest circle possible. More interestingly, we can see two patterns emerge from the data. First is the linear relationship we expect to see between the observation and what the Martins model predicted. The second is a tendency to shift and draw a larger circle when the model predicted no change. Likely, two phenomena are simultaneously occurring in this experiment, one that the Martins model can explain and the other the model cannot explain. We suspected that the two phenomena could be divided based on parameters determined in the game, and we developed this filter to separate the data
\begin{align}
d_x < 0.8\left(\Main{r} + \Other{r}\right). \label{eqn:filter}
\end{align}
Eqn \ref{eqn:filter} separates the data based on whether the two circles shown to the participant were overlapping by $20\%$ of their radiuses.
\begin{figure}[!h]
\includegraphics[scale=0.2]{Fig1.jpg}
\caption{{\bf Scatter plot of expected v.s. observed shift towards the red circle and the change in circle area relative to the blue circle.} \\
(A) Shift towards the red circle. (B) Change in circle area. }
\label{fig1}
\end{figure}
When we apply Eqn \ref{eqn:filter} to the data, we find these results:
\begin{enumerate}
\item When the red and blue circles overlap and there is medium to high reliability, we can reject both hypotheses: (1) There is no correlation between the observed and expected shifts away from the blue circle; and (2) There is no correlation between the observed and expected change in circle area from the blue circle.
\item When the red and blue circles do not overlap, participants are making a choice to either stay with the blue circle, adopt the red circle or compromise 44\% with the red circle when trust is high.
\item Participants tend to shift away from the red circle when trust is low, and the red and blue circles overlap.
\end{enumerate}
\subsection*{Data Explained by the Martins model}
Fig \ref{fig2} is the result of applying Eqn \ref{eqn:filter} to the data. In Fig \ref{fig2}A we can see that the filtered observations broadly match expectation with the linear model producing an $R^2 = 0.14$ (see Table \ref{table2}). In Fig \ref{fig2}B the model performs noticeably worse with an $R^2 = 0.0029$ (see Table \ref{table2}). The $p$-values for the slopes from the linear models are statistically significant compared to a Type I Error Rate of $\alpha=0.05$ for the Medium and High trust scenarios this indicates that in these cases there is sufficient evidence to reject our null hypotheses and conclude that the observed results do concur with the extended Martins model predictions. In the Low trust scenarios, the $p$-values are not significant, indicating that in these scenarios, the extended Martins model results are not good predictors of the observed behaviour. In the Low trust scenarios individuals seem to move \textit{away} from the red circle, which is counter to the assumptions of the extended Martins model. We elaborate more this negative shifting in its own section. Investigating Fig \ref{fig2}A reveals the effects of a confounding variable bounding observed values to a minimum value. We suspect this confounding variable to be the minimum size participants can draw their circle, which creates an artificial limit on circle size reduction.
\begin{figure}[!h]
\includegraphics[scale=0.2]{Fig2.jpg}
\caption{{\bf Expected v.s.\ observed filtered based on games where the blue and red circle overlapped.}
The solid red line is the line of best with intercept set to zero. \\
(A) Shift towards the red circle relative to the blue circle. (B) Change in circle area relative to the blue circle. }
\label{fig2}
\end{figure}
\begin{table}[!ht]
\begin{adjustwidth}{-2.25in}{0in}
\centering
\caption{
{\bf The slope and $R^2$ values for Figs \ref{fig2} and \ref{fig3}.}}
\begin{tabular}{|c+l|l|l+l+l|l|l+l|}
\hline
& \multicolumn{4}{c+}{\bf Shift towards Red} & \multicolumn{4}{c|}{\bf Change in circle area} \\
& Low & Mid & High & Total & Low & Mid & High & Total \\ \thickhline
slope & $0.17$ & $0.94$ & $1.14$ & $1.06$ & $-0.02$ & $0.23$ & $1.08$ & $0.27$ \\ \hline
$p$-value & $0.58$ & $\expnumber{1.06}{-43}$ & $\expnumber{1.07}{-114}$ & $\expnumber{1.07}{-141}$ & $0.91$ & $0.03$ & $\expnumber{6.14}{-39}$ & $\expnumber{1.16}{-5}$ \\\hline
$R^2$ & $0.01$ & $0.09$ & $0.15$ & $0.14$ & $0.01$ & $0.01$ & $0.05$ & $0.003$ \\ \hline
\end{tabular}
\begin{flushleft} The low, mid and high headings in the table refer to trust at low, medium and high values respectively, specifically they refer to parameter values $p=0.2$, $0.5$ and $0.8$. Total is taking the data as a whole.
\end{flushleft}
\label{table2}
\end{adjustwidth}
\end{table}
We partitioned the data further by the reliability of the red circle, and Fig \ref{fig3} shows such a partition. In general, the model predicts high trustworthiness interactions more accurately, and most of the outliers in the data lie within the low trustworthiness games. In particular Fig \ref{fig3}B, when $p=0.2$ and $p=0.5$, contain most of the unusually large data points compared with $p=0.8$. We can explain this outlier behaviour as participants attempt unorthodox strategies to get the highest score since the red circle isn't a reliable source of information in those cases.
\begin{figure}[!h]
\includegraphics[scale=0.2]{Fig3.jpg}
\caption{{\bf Expected v.s. Observed filtered based on games where the blue and red circle overlapped separated into different levels of trustworthiness.}
The solid red line is the line of best with intercept set to zero. \\
(A) Shift towards the red circle relative to the blue circle. (B) Change in circle area relative to the blue circle. }
\label{fig3}
\end{figure}
\subsection*{Data Unexplained by the Martins Model}
After investigating the data of the overlapping circles, we shifted focus to the data of the non-overlapping circles. Since the red and blue circles weren't overlapping in this case, participants would see two distinct circles. Therefore, we hypothesised that participants were choosing either to stick with the blue circle (what they know to be true), to `take a leap of faith' and adopt the red circle as their new opinion, or to comprise with the red circle and draw their circle in between the red and blue circles. We tested this hunch by scaling the observed shift towards the red circle by the distance between the blue and red circles $d_x$ resulting in Fig \ref{fig4}. From Fig \ref{fig4} we can see that the majority of the unpredicted shifts (73\%) were between $0$ and $d_x$ with most of these close to $0$, which is consistent with the proposed explanation that the participants made a discrete choice between three options.
\begin{figure}[!h]
\includegraphics[scale=0.2]{Fig4.jpg}
\caption{{\bf Histogram of the shift towards the red circle relative to the total distance between the other and blue circle when the blue and red circles where not significantly overlapping.}
}
\label{fig4}
\end{figure}
Similar to the data explained by the Martins model, we can separate the unexplained data based on trustworthiness which results in Fig \ref{fig5}. The tendency we expected to see, i.e.\ sticking with the known by drawing over the blue circle, is confined to the low trust scenarios of Figs \ref{fig5}, i.e.\ $p=0.2$ and $p=0.5$, but there is a tendency to comprise and a smaller chance to fully trust the red circle which contributes to a rightwards skew, particularly in the medium trust case of $p=0.5$. The tendency to adopt or comprise with the red circle is unsurprisingly common in the high trust case of $p=0.8$, and we can see distinguished peaks when $p=0.8$, suggesting that participants are making a discrete choice concerning whether to fully, partially or not trust the red circle. Furthermore, we note the case when $p=0.8$ relates closely to data collected by \cite{EDcollectivedynamics2013Moussaid} when participants decided to comprise. In the study, \cite{EDcollectivedynamics2013Moussaid} participants chose to adopt on average 40\% of their interaction partner's opinion into their own, which is congruent with the average opinion shift in Table \ref{table3} of $0.44$.
\begin{figure}[!h]
\includegraphics[scale=0.2]{Fig5.jpg}
\caption{{\bf Histogram of the shift towards the red circle relative to the total distance between the other and blue circle when the blue and red circles where not significantly overlapping separated into different levels of trustworthiness.} }
\label{fig5}
\end{figure}
\begin{table}[!ht]
\centering
\caption{
{\bf The summary statistics for Figs \ref{fig4} and \ref{fig5} in pixels.}}
\begin{tabular}{|l+l|l|l+l|}
\hline
& Low & Mid & High & Total \\ \thickhline
Mean & $0.13$ & $0.22$ & $0.44$ & $0.23$ \\ \hline
Median & $0.02$ & $0.07$ & $0.45$ & $0.06$ \\ \hline
Min & $-2.05$ & $-0.91$ & $-0.47$ & $-2.05$ \\ \hline
Max & $4.21$ & $1.39$ & $2.63$ & $4.21$ \\ \hline
1st Quantile & $-0.01$ & $0.00$ & $0.09$ & $0.00$ \\ \hline
3rd Quantile & $0.15$ & $0.39$ & $0.72$ & $0.42$ \\ \hline
\end{tabular}
\begin{flushleft} The low, mid and high headings in the table refer to trust at low, medium and high values respectively, specifically they refer to parameter values $p=0.2$, $0.5$ and $0.8$. Total is taking the data as a whole.
\end{flushleft}
\label{table3}
\end{table}
\subsection*{Negative Opinion shifting in the results}
In the data, there has been a noticeable number of participants shifting away from the red circle, i.e. a negative opinion shift, which the Martins model does not predict. We have tabulated the number of negative opinion shifts in Table \ref{table4} and most negative shifts occur within five pixels of 0. A notable exception is when trust is low, i.e.\ $p=0.2$, and the red and blue circles overlapped (explained data), but the majority of negative shifts that occurred were still below $50$ pixels ($10\%$ the height of the play area and $14\%$ the width) with only $10\%$ of shifts being further than $50$ pixels. Over both the explained and unexplained data sets, low trust games resulted in more negative shifts, whereas high trust games, i.e.\ $p=0.8$, resulted in less negative shifts and both the explained and unexplained data produced similar proportional amounts of negative shifts. We can conclude that most of the negative opinion shifts occurring are from participants attempting to draw their circle on the blue circle, i.e.\ shifting by $0$. In low trust scenarios, however, the negative shift could be more intentional, particularly for the explained data.
\begin{table}[!ht]
\begin{adjustwidth}{-2.25in}{0in}
\centering
\caption{
{\bf Number of games that feature negative opinion shifting from the participants.}}
\begin{tabular}{|c+l|l|l+l+l|l|l+l+l|}
\hline
\multirow{2}{*}{\makecell{\bf Observed Shifts \\ \bf (pixels)}} & \multicolumn{4}{c+}{\bf Explained (games)} & \multicolumn{4}{c+}{\bf Unexplained (games)} & \multirow{2}{*}{\makecell{\bf Total \\ \bf (games)}}\\
& Low & Mid & High & Total & Low & Mid & High & Total & \\ \thickhline
$h_{\mathrm{observed}}\geq 0$ & $163$ & $454$ & $895$ & $1512$ & $574$ & $447$ & $352$ & $1373$ & $2885$ \\ \hline
$-5\leqh_{\mathrm{observed}}<0$ & $36$ & $58$ & $69$ & $163$ & $141$ & $68$ & $26$ & $235$ & $398$ \\ \hline
$-50\leqh_{\mathrm{observed}}<-5$ & $69$ & $82$ & $89$ & $240$ & $95$ & $58$ & $16$ & $169$ & $409$ \\ \hline
$h_{\mathrm{observed}}<-50$ & $31$ & $9$ & $11$ & $51$ & $14$ & $3$ & $0$ & $17$ & $68$ \\ \hline
\end{tabular}
\begin{flushleft} The low, mid and high headings in the table refer to trust at low, medium and high values respectively, specifically they refer to parameter values $p=0.2$, $0.5$ and $0.8$. Total is the total games in a particular category, either explained, unexplained or across the whole experiment.
\end{flushleft}
\label{table4}
\end{adjustwidth}
\end{table}
\subsection*{Breakdown of Individual Participant Involvement}
To ascertain the influence of individual participants in the experimental data, we have developed Fig \ref{fig6}, which shows the number of games played versus the number of players. We uniquely identified participants through their IP addresses and used that information to count how many games a participant played. The median games played was $10$ games whereas the mean games played was $14.63$ games this suggests there is a skew in the histogram Fig \ref{fig6}. Furthermore, the top $10\%$ of participants (in terms of games played) are responsible for only $35\%$ of all games played in the experiment and $80\%$ of participants played $23$ or fewer games. The high number of games played by a small number of players presents a potential issue because their ``learning" could bias the results if we assume that games are independent trials for the purposes of analysis. That is, if players' scores improve as they played additional games, the independence assumption would be invalid, tainting the results of our analyses. To investigate whether participants were learning to play the game, we compiled Table \ref{table5} showing the mean scores for the $n$th game played. We can see that participants' score doesn't increase as they play more games suggesting that participants are not ``learning" to play the game better or at least not learning to improve their scores.
\begin{figure}[!h]
\includegraphics[scale=0.2]{Fig6.jpg}
\caption{{\bf Histogram of the participant frequency on the number of games played.} \\
}
\label{fig6}
\end{figure}
\begin{table}[!ht]
\centering
\caption{
{\bf The mean score of participants by attempts.}}
\begin{tabular}{|l+l|}
\hline
{\bf Attempt} & \textbf{Mean Score} \\\thickhline
1st & 3230 \\\hline
2nd & 1983 \\\hline
3rd & 4896 \\\hline
4th & 2533 \\\hline
5th & 3563 \\\hline
6th & 5072 \\\hline
7th & 3195 \\\hline
8th & 2028 \\\hline
9th & 4358 \\\hline
10th & 3690 \\\hline
11th-15th & 4148 \\\hline
16th-20th & 4039\\\hline
21st-25th & 4212 \\\hline
26th-30th & 5124 \\\hline
31st-40th & 3136 \\\hline
41st-50th & 3007 \\\hline
51st-88th & 3909 \\\hline
\end{tabular}
\label{table5}
\end{table}
\section*{Discussion}
The data we have collected has produced surprising and interesting results. We have identified two different types of behaviour in the experiment. The first type of behaviour is congruent with the predictions made by the Martins model, while the second fell outside the scope of the Martins model, and we were able to distinguish between the two behaviours by developing Eqn \ref{eqn:filter}, which divides the data based on the red and blue circle overlap. When investigating the second dataset, we developed Figs \ref{fig4} and \ref{fig5}, and we concluded that participants are treating the problem of finding the dot as a discrete choice, i.e. it must be in either the red circle or the blue. Adopting the red circle opinion is contrary to the Martins model, which considers distant opinions as ``untrustworthy'' even with a $p$ close to 1. Thus, according to the Martins model, a participant should always ignore the red circle. There are multiple causes for participants to switch to a discrete choice mindset. It is either the result of cognitive bias to simplify the problem or devised from the instructions we gave to participants. The instructions explained that M. Dotte is the one who reveals the blue circle. Despite the instructions explaining that M. Dotte is always reliable, participants might still doubt M. Dotte and not fully internalise the blue circle as their opinion. Although when we compare the unexplained results to the results in the literature, we see startling agreement. We observed in Fig \ref{fig5} that when reliability is low, participants tended to keep their opinions close to the blue circle, shifting towards the red circle at an average of $0.2d_x$. But with increased reliability, participants began to ``compromise'' with the red circle by drawing their circle at $0.4d_x$ from the blue to the red circle. This $0.4$ magnitude shift agrees closely with the results in \cite{EDcollectivedynamics2013Moussaid}, when participants decided to compromise.
The Martins model appears ineffective in predicting the change in the circle area. We note in Fig \ref{fig2}B that a confounding variable is bounding the observed change in the circle area, thus forcing a minimum value. We posit that the confounding variable is the minimum circle size (a five-pixel radius) relative to the size of the play area, i.e.\ the box. We can see a linear tread exists in Fig \ref{fig2}B and is cut by the minimum circle size boundary. The outliers are more extreme in Fig \ref{fig2}B than for \ref{fig2}A, and we can surmise that the outliers for Fig \ref{fig2}B were participants' attempts to `break' the game. Essentially the participants were testing if drawing a large circle would net a substantial number of points. When the data is separated based on reliability, it is clear that the Martins model predicts high-reliability scenarios for circle area change more accurately. In that case, the prediction for circle area change is small enough so that participants can draw circles of those sizes, thus above the minimum circle size boundary.
There is much debate over whether opinions can be ``negatively'' influenced, i.e. when individuals' opinion difference is large, the distance between the individuals' opinions increases after an interaction (i.e. negative opinion shift). Negative opinion influence is not to be confused with a negative opinion shift, which is an unconditional shift away from an interaction partner's opinion, i.e. negative opinion shifts may occur without negative influence. Some theoretical opinion dynamics models \cite{NIdynamicflows2007Mason, NIemergencenorms2006Kitts,NIculturecompetition2003Mark} rely on negative opinion influence to create polarisation, and others like \cite{DAtrust2013Martins} predict negative opinion shift resulting from negative opinion influence (but not necessarily resulting polarisation) in a discrete opinion context. In contrast, empirical experiments that attempted to measure negative opinion influence have so far failed. For example, \cite{EDdislikingnoshift2016Takacs}, although finding evidence of negative opinion shifts, found no evidence of negative opinion influence.
The data we collected conforms with the results in \cite{EDdislikingnoshift2016Takacs}, we observed in our data negative opinion shifts, but it was localised to when the red and blue circles were overlapping, not when the circles were distant. Most negative shifts resulted from participants attempting to draw onto the blue circle and were within 5 pixels of the blue circle. Only when reliability was low and the red and blue circles overlapped did it appear like participants intentionally shifted away from the red circle. The Martins model only predicts positive opinion shifts and thus does not explain the negative opinion shifting occurring at low reliability. The negative opinion shifting is likely the phenomenon which causes the Martins model to be a poor predictor of low-reliability interactions. We theorise that participants, when presented with a low-reliability red circle close to the blue circle, believe that the red circle reduces the chances that the dot is in the blue circle and thus moves away from the red, which the Martins model does not consider. Although, we see similar behaviour in the model developed in \cite{DAtrust2013Martins}.
\section*{Conclusion}
In this paper, we aimed to verify whether the Martins model is an accurate model of opinion exchange. We can conclude from the data that the Martins model is only accurate in specific circumstances. Specifically, we can reject the two null hypotheses when the red and blue circles overlap for medium to high reliability. Furthermore, we identified two phenomena occurring in this experiment; along with the phenomena explained by the Martins model, we observed participants making discrete choices. The discrete choice behaviour exclusively occurred when the red and blue circles didn't significantly overlap. We conjectured that the discrete choice occurred due to the human need to simplify the problem or participants not completely trusting the blue circle as their own opinion. Either way, this highlights the multifaceted nature of opinion exchange and illustrates the context-sensitivity of human behaviour. For a model of opinion exchange to sufficiently capture the complexities of interactions, the model would need to navigate the context of an interaction. Essentially the model needs to switch between discrete and continuous opinions when appropriate creating a complete synthesis of a discrete and continuous opinion model.
The Martins model predicted the opinion shifts of participants with reasonable precision when only considering the data explained by the Martins model. The $R^2$ for the linear tread lines are low because of the presence of outliers and a significant number of interactions that resulted in negative opinion shift, but from Fig \ref{fig2} it is clear that there exists a linear trend. We can conclude that the Martins model predicts the general behaviour of the participants when there is a significant overlap between the red and blue circles. This conclusion is weak, and a more robust experiment with more participants is needed to determine whether the Martins model predicts human behaviour. Due to the simplicity of our experiment design, it should be easy to recreate this experiment at scale.
\section*{Supporting information}
\paragraph*{S1 Fig.}
\label{S1_Fig}
{\bf The instruction given to every participant.}
\includegraphics[scale=0.5]{S1Fig.jpg}
\paragraph*{S2 Fig.}
\label{S2_Fig}
{\bf An example of the game a participant would of played when a participant is exposed to the blue circle.}
\includegraphics[scale=0.5]{S2Fig.jpg}
\paragraph*{S3 Fig.}
\label{S3_Fig}
{\bf An example of the game a participant would of played when a participant is exposed to the red circle.}
\includegraphics[scale=0.5]{S3Fig.jpg}
\paragraph*{S4 Fig.}
\label{S4_Fig}
{\bf An example of a completed the game for a participant.}
\includegraphics[scale=0.5]{S4Fig.jpg}
\paragraph*{S1 File.}
\label{S1_File}
{\bf The experiments website source code.}
\paragraph*{S2 File.}
\label{S2_File}
{\bf Raw data collected from the experiment.}
\paragraph*{S1 Appendix.}
\label{S1_Appendix}
{\bf Equations of the Martins model.}
\begin{align*}
&x_i(t+1) = p^* x_i(t) +p^* \frac{x_i(t)/\sigma_i(t) + x_j(t)/\sigma_j(t)}{1/\sigma_i(t) + 1/\sigma_j(t)}, \\
&\sigma_i^2(t+1) = \left(1 - \frac{\sigma_i^2(t)}{\sigma_j^2(t)+\sigma_i^2(t)}\right)\sigma_i^2(t) + p^* (1-p^*) \left(\frac{x_i(t) - x_j(t)}{1+\sigma_j^2(t)/\sigma_i^2(t)}\right)^2,
\end{align*}
where
$$
p^* = \frac{p \phi\left(x_i(t) - x_j(t),\sqrt{\sigma_i^2(t)+\sigma_j^2(t)}\right)}{p \phi\left(x_i(t) - x_j(t),\sqrt{\sigma_i^2(t)+\sigma_j^2(t)}\right) + 1-p},
$$
and
$$
\phi(\mu,\sigma) = \frac{1}{\sigma\sqrt{2\pi}} e^{\frac{-\mu^2}{2\sigma^2}}
$$
\section*{Acknowledgments}
We would like to acknowledge the contribution from the QUT eResearch team: Ryan Bennett, Mitchell Haring, Yvette Wyborn, Craig Windell and Adam Smith, for their contribution in setting up, running and testing the website involved in this study.
|
\section{Introduction}
To leverage the wisdom of multiple experts in predictions, various fields employ the approach of coordinating multiple teams who independently submit their projections. Such multi-model prediction efforts are common in infectious disease modeling~\cite{borchering2021modeling,FluSight,sharma_dev_mangla_wadhwa_mohanty_kakkar_2021} and climate modeling~\cite{tebaldi2007use,tegegne_melesse_2020,najafi_robertson_massah_2021}. Often, multiple projections are performed under various ``scenarios'' produced by the presence and absence of a decision in the future, e.g., no vaccinations (scenario A) vs vaccinations (scenario B) available in the future. For example, the US/CDC COVID-19 Scenario Modeling Hub coordinates the task of long-term public health impacts under different scenarios \cite{covid_scenario_hub}. The participating models \cite{sikjalpha,bucky,CovidSP,mobs} in this effort predict COVID-19 cases, hospitalization, and deaths for a scenario in weeks ahead as a random variable and output 23 quantiles for each prediction.
The goal of generating scenario projections based on a future decision is to assess the impact of that decision. It is crucial to identify confidence intervals on the impact. However, this is difficult mainly because the two sets of scenario projections $A$ and $B$ are independently generated -- the joint distribution of any given outcome $X$(e.g., number of deaths on a certain date) under the two scenarios is not known to enable computing $P(X|B - X|A)$. Also, the distribution is available as a set of quantiles rather than a continuous cumulative distribution function (CDF). Asking the models to generate a joint distribution of outcomes adds to the challenges. First, the modeling teams already spend a significant amount of time on modeling and projections. Additional work of computing the joint distribution, every time a new decision is to be evaluated, may create barriers to joining the multi-model effort and releasing timely projections. Further, submitting joint distributions $P(X|A, X|B)$ will quadratically increase the space complexity. Currently, the submissions in US Scenario Modeling Hub with only marginal distributions $P(X|A)$ for four scenarios for 50 states and 52 week-long projections and three targets (cases, deaths, and hospitalizations) result in a file size over 100MB.
Our goal is to identify non-trivial bounds on the difference of outcomes under two scenarios without asking for any additional work from the participating modeling teams. We assume that we have quantiles for targets for at least two scenarios of which the difference is to be computed. We make some realistic assumptions supported by observed data to derive an arbitrary $\alpha$-confidence interval for the difference of two scenarios.
We start with a strong assumption (Assumption~\ref{Assumption1}) on the models that result in what we define as \textit{zero-violation models}. Under this assumption, we devise a method to find arbitrary $\alpha$-confidence. We then relax assumption 1 and define \textit{$\epsilon-$violation as model} as a \textit{model} that partially satisfies Assumption~\ref{Assumption1} and claim any arbitrary $\alpha$-confidence interval on scenario difference can be still obtained if the \textit{$\epsilon$-violation model} follows Assumption~\ref{Assumption2}. We demonstrate indications that the assumption is reasonable through experiments on models submitted to the US Scenario Modeling Hub.
\section{Methodology}
In this section, we describe a methodology to bound the difference in scenarios for any arbitrary confidence interval $\alpha$ under some assumptions for the models. It should be noted that methods proposed in this section apply to different scenario modeling problems where Assumption \ref{Assumption1} holds. In this paper, we focus on the context of COVID-19 multi-model predictions.
\subsection{Problem Setting}
Define a \textit{scenario} $S_{t_{\text{app}}} \in \mathcal{S}$ to be an environment in which some events occurred at time $t_{\text{app}}$. For instance, a scenario could be recommending vaccination for children at 2 weeks ahead of the time of prediction, so $t_{\text{app}} = 2 + t_0$, where $t_0$ is the time when the prediction is made. Notice that it is possible that $t_0 \neq t_{\text{app}}$, \textit{i.e} events that distinguish scenarios need not immediately occur at the time of prediction. Within each scenario, each unique set of latent variables $P_0, P_1, ..., P_n$ and the time being predicted $t$ is associated with a separate result (COVID-19 case, hospitalization, deaths). For instance, if we are predicting COVID-19 cases, latent variables $P_0, P_1, ..., P_n$ could be temperature, human mobility, percentage of masked population, \textit{etc}.
A \textit{model} in the multi-model effort takes in a future time $t \geq t_0$ and a \textit{scenario} $S_{t_\text{app}}$ as inputs, and outputs a stochastic prediction of the COVID-19 cases, hospitalizations, and deaths associated with the scenario. We can interpret the \textit{model} that predicts the (COVID cases, deaths, or hospitalizations) for 2 scenarios as a random vector constructed by repeatedly uniformly sampling form the at most countably infinite universe $U(t) = \{(x_i,y_j)\}_N$, where $U$ is dependent the time $t$ that we are predicting for. Each $x_i,y_j \in \mathbb{R}$ is one possible prediction for the two scenarios being compared and is affected by the latent variables $P_1,P_2,...,P_n$ as indicated in figure \ref{fig:tree}. For \textit{models} capable of predicting more than 2 scenarios, interpret them as uniformly sampling the random vectors of length $|\mathcal{S}|$, where $|\mathcal{S}|$ is the number of possible scenarios. Then, for each vector, truncate it by selecting only elements $x_i$ and $y_i$ to form the 2-vector $(x_i,y_j)$, corresponding to the two scenarios of interest. This interpretation of \textit{models} is equivalent to a Monte-Carlo simulation with the underlying latent variables. Figure \ref{fig:tree} provides a detailed illustration of the interpretation of a \textit{model}.
Now, label $x_i$ and $y_j$ such that $x_0 \leq x_1, ..., \leq x_N$ and $y_0 \leq y_1, ..., \leq y_N$, and let $X$, $Y$ be the respective random variables obtained by sampling $x_i$ and $y_j$ separately and independently from $U = \{(x_i,y_j)\}_N$. In our multi-model effort, the \textit{model} only provides information for $X$ and $Y$ in the form of 23 distinctive quantiles, denote them as $\mathbf{Q}^{X} = \{Q^{X}_{q_1}, Q^{X}_{q_2}, ..., Q^{X}_{q_{23}}\}$ and $\mathbf{Q}^{Y} = \{Q^{Y}_{q_1}, Q^{Y}_{q_2}, ..., Q^{Y}_{q_{23}}\}$ where $\{q_1, q_2, ..., q_{23}\} = \mathbf{q}$. Also, let $F_X(k)$ and $F_Y(k)$ denote the cumulative distribution function of random variable $X$ and $Y$ evaluate at $k$, respectively.
\begin{definition}
The tuple $(x_i,y_j)$ is a matching if and only if $(x_i,y_j) \in U(t)$ for some $t$. Denote this matching by $x_i \match y_j$.
\end{definition}
At the right of the figure \ref{fig:tree}, the elements $x_i$ and $y_j$ with the same color represent a \textit{matching} pair $(x_i,y_j)$. Observe that $x_i$ and $y_j$ having the same set of latent variables $(P_1,P_2,...,P_n,t)$ is a necessary but not a sufficient condition for $x_i \match y_j$.
From this interpretation, our problem of bounding difference of scenarios can be formulated as the following:
\begin{tcolorbox}
\textit{Given $\alpha \in (0,1)$ and $t \in \mathbb{N}$, give an $\alpha$-confidence interval for $Z$, the random variable obtained from uniformly sampling $x_i - y_j$ where $x_i \match y_j \in U(t)$}
\end{tcolorbox}
To do this, we need to extract the \textit{matching} information from the 23 quantiles given. We thereby propose the following assumption on the scenario:
\begin{assumption}\label{Assumption1}
(Monotonic Impact of Latent Variables): Changing any subset of latent variables will impact the cases, hospitalizations, and deaths of all scenarios in the same direction.
\end{assumption}
This assumption also follows the real-life observations of epidemics. Usually, the latent factors include more transmissible variants, lack of medical resources, social mobility, population awareness, etc, combining these factors should give us monotonic impact on the potency of the epidemic. It should be noted that, however, this assumption is strong, and we will discuss the limitation of this assumption and relax it in the part where \textit{violation models} are discussed.
\subsection{Zero-Violation Models}
For any \textit{models} predicting scenarios that satisfy assumption \ref{Assumption1}, we would expect it to demonstrate a similar behavior in its predictions. That is, we can reasonably expect for all chosen pair $(x_i,y_j) \in U(t)$ from the \textit{model}, changing the latent variables $P_0, ..., P_n$ will only impact $x_i$ and $y_j$ in the same direction. In other words, any combination of interventions that could have reduced $x_i$ could not increase $y_i$, prompting the following definition:
\begin{definition}
A zero-violation \textit{model} is a model that satisfies assumption \ref{Assumption1}.
\end{definition}
Due to the stochasticity of some models, the leaf nodes in figure \ref{fig:tree} are not necessarily deterministic but are rather random vectors $(X_i,Y_i)$. Taking account of this type of model and reconciling for scenarios where assumption \ref{Assumption1} does not hold, \textit{violation models} is defined and discussed in later sections. For now, keep in mind that \textit{zero-violation models} have deterministic leaf nodes in figure \ref{fig:tree}.
\begin{lemma}\label{lem:1}
(Well-Orderedness of \textit{zero-violation models}): For $x_i \match y_j$ produced by a \textit{zero-violation model}, we must have $i=j$. That is, the rank of \textit{matching} $x_i$ and $y_j$ must be equal in their respective ordered list.
\end{lemma}
\begin{proof}
The proof can be found in Appendix \ref{lem:1proof}
\end{proof}
Lemma \ref{lem:1} implies that $F_X(X \leq x_i) = F_Y(Y \leq y_i) = \frac{i}{N}$ for any \textit{zero-violation model}. Therefore, to sample $Z = X - Y$, it is sufficient to choose $i$ uniformly at random from $\{1,2,...N\}$ and calculate $x_i - y_i$. In our problem, however, only 23 $x_i$ and $y_i$ are given with their corresponding ranks (provided by the 23 quantiles). To adapt, we propose algorithm \ref{alg:IIMB} to find the set of possible upper and lower bound on $Z$, from which any arbitrary $\alpha$-confidence interval on $Z$ can be obtained.
\begin{algorithm}
\begin{flushleft}
\textbf{Input}: Quantile labels $\mathbf{q}$, quantiles $\mathbf{Q^{X_t}}$ and $\mathbf{Q^{Y_t}}$ \\
\textbf{Output}: Upper and lower bound on $Z = x_i - y_j$, $Z^U$ and $Z^L$, at time $t$
\end{flushleft}
\caption{Iterative Zero-Violation Model Bound}\label{alg:IIMB}
\begin{algorithmic}[1]
\For{$j = \{1,2,...,10^5\}$}
\State{Sample $\frac{i}{N}$ from Uniform$(0,1)$}
\State{$q_l \gets \max_{q \in \mathbf{q}}\{q \leq \frac{i}{N}\}$}
\State{$q_u \gets \min_{q \in \mathbf{q}}\{q \geq \frac{i}{N}\}$}
\State{Append $\mathbf{Q^X}_{q_u} - \mathbf{Q^Y}_{q_l}$ to $Z^U$}
\State{Append $\mathbf{Q^X}_{q_l} - \mathbf{Q^Y}_{q_u}$ to $Z^L$}
\EndFor
\end{algorithmic}
\end{algorithm}
In Algorithm\ref{alg:IIMB}, since $Q^X_{q_l} \leq x_i \leq Q^X_{q_u}$ and $Q^Y_{q_l} \leq y_i \leq Q^Y_{q_u}$, it is easy to see that
\begin{align}
Q^X_{q_l} - Q^Y_{q_u} \leq x_i &- y_i \leq Q^X_{q_u} - Q^Y_{q_l}
\intertext{that is,}
Z^L \leq &Z \leq Z^U
\end{align}
Since $\text{I\kern-0.10em P}(l\leq Z\leq u) = \text{I\kern-0.10em P}(Z \leq u) - \text{I\kern-0.10em P}(Z \leq l)$, and $\text{I\kern-0.10em P}(Z \leq u) \geq \text{I\kern-0.10em P}(Z^U \leq u)$, $\text{I\kern-0.10em P}(Z\leq l) \leq \text{I\kern-0.10em P}(Z^L \leq l)$, we arrive at the identity $\text{I\kern-0.10em P}(l \leq Z \leq u) \geq P(Z^{U} \leq u) - \text{I\kern-0.10em P}(Z^L \leq l)$, which can be applied to find the confidence interval. Namely, an $\alpha$-confidence interval is obtained by finding the appropriate $u$ and $l$ such that $\text{I\kern-0.10em P}(l \leq Z \leq u) \leq \text{I\kern-0.10em P}(Z^{U} \leq u) - \text{I\kern-0.10em P}(Z^L \leq l) = \alpha$.
In addition, as the number of available quantiles increases uniformly on $[0,1]$, $q_l$ converges to $q_u$ ($q_l \to q_u$). Thus, $Q^X_{q_l} \to Q^X_{q_u}$, $Q^Y_{q_l} \to Q^Y_{q_u}$, and $Z^U - Z^L = Q^X_{q_u} - Q^X_{q_l} + Q^Y_{q_l} - Q^Y_{q_u} \to 0$; that is, $Z^L \to Z^U = Z$ eventually when large number of quantiles become available.
\begin{figure*}
\begin{center}
\includegraphics[width=0.8\linewidth]{Model_Representation.pdf}
\end{center}
\caption{Illustration for the interpretation of the \textit{zero-violation model} at time $t$. The root node represents the time $t$ that the \textit{model} is generating predictions for, while every other non-leaf nodes are latent variables $P_i$ affecting predictions. As it might not be feasible for the \textit{model} to capture the precise dynamics of the environment(scenario), $P_i$ in the \textit{model} can be different from the $P_i$ of the scenario. The leaf nodes are vectors in $\mathbb{R}^{|\mathcal{S}|}$, where $|\mathcal{S}|$ is the total number of scenarios the model predicts (in this figure, $|\mathcal{S}| = 2$). The leaf nodes are affected by the variables from all of its ancestor nodes. Each unique path in this tree generates a distinct vector, and the collection of which forms the universe at time $t$. The two lists $X$ and $Y$ depicted at the right of $U$ are the ranked independent observations of $x_i$ and $y_j$. }
\label{fig:tree}
\end{figure*}
\subsection{Violation Models}
A \textit{violation model} is a \textit{model} that only partially satisfies assumption \ref{Assumption1}. From Lemma \ref{lem:1}, a \textit{zero-violation model} is expected to have $x_i \match y_j, \sforall i = j$. Equivalently, for \textit{zero-violation models} with scenarios whose $t_{\text{app}} > t_0$, $x_i = y_i$ is expected because $x_i \match y_i$ and the events that differentiates the scenarios has not occurred at time $t \leq t_{\text{app}}$, which entails $Q^X_{q_i} = Q^Y_{q_i} \sforall q_i \in \mathbf{q}$. However, this is not observed in practice for two reasons:
\begin{enumerate}
\item Some models have stochasticity, and the leaf nodes in figure \ref{fig:tree} becomes random variables $(X_i,Y_j)$ instead of the deterministic vector $(x_i,y_j)$. From repeated sampling $x_i$ and $y_j$ from $X_i$ and $Y_j$, we can not guarantee $x_i = y_i$ for $t \leq t_{\text{app}}.$
\item Assumption 1 is violated in the models. This can happen in many ways. For instance, vaccines can be more potent under some specific circumstances.
\end{enumerate}
In reality, a weaker version of assumption \ref{Assumption1} is more likely to hold: cases, hospitalizations, and deaths are monotonic with changing one latent variable, instead of changing a set of latent variables. That is, each vaccination, human mobility, weather, etc. has a monotonic impact, but the monotonic impact is not guaranteed when the changes are combined. Empirically, however, we later demonstrate that the extent to which the assumption is violated is small.
For \textit{models} that violates assumption \ref{Assumption1} or have stochasticity (random vectors as leaf nodes in figure \ref{fig:tree}), we define them as \textit{violation models}. To quantitatively examine this type of \textit{models}, define \textit{$\epsilon$-violation} as a measure on the degree of violation for \textit{violation models}.
\subsubsection{$\epsilon$-Violation}
A mismatch occurs when $x_i \match y_j$ with $i \neq j$. There are two kinds of mismatch that can happen: either $i > j$ or $i < j$. In other words, $x_i$ is matched to an element that is above its rank in $Y$'s ordered list or vice versa. The right hand side of Figure \ref{fig:tree} illustrates this: depicted in orange, $\epsilon_u$ is the maximum upper mismatch of $x_i \match y_i$, while $\epsilon_l$ is the maximum lower mismatch depicted in green. To bound $x_i - y_j$ for $x_i \match y_j$ of a violation model, we therefore need to take the maximum of $x_i - x_{i + \text{mismatch}}$ and $x_i - x_{i - \text{mismatch}}$. The following definition formally defines mismatch.
\begin{definition}\label{def:ep}
The $\epsilon$-violation of a \textit{violation} model is the largest difference in quantiles for $x_i \match y_j$ in their respective ordered lists, \textit{i.e}
\begin{align}
\epsilon_u &= \max_{x_i \match y_j,t}\{F_Y(x_i) - F_X(y_j)\}\\
\epsilon_l &= -\min_{x_i \match y_j,t}\{F_Y(x_i) - F_X(y_j)\}\
\end{align}
where $t$ in max and min represents all predictions over time.
\end{definition}
Intuitively, the larger the $\epsilon_l$ and $\epsilon_u$, the higher the mismatch, and thus more uncertainty the model's predictions. Also, since $x_i \match y_j \Rightarrow i = j$ for \textit{zero-violation models}, we have $F_X(x_i) = F_X(x_j) = F_Y(y_j)$ for all $x_i \match y_j$ for \textit{zero-violation models}, prompting the following corollary:
\begin{corollary}
A model has $\epsilon_u = \epsilon_l = 0$ if and only if it is a \textit{zero-violation} model.
\end{corollary}
To bound $\{x_i - y_j:x_i\match y_j\}$ for \textit{violation models} with algorithm \ref{alg:IIMB}, we need to account for the cases where $y_j$ exceeds $Q^Y_{q_u}$ or falls below $Q^Y_{q_l}$, as shown in the right hand side of figure \ref{fig:tree}. Similar adjustments need to be made for $x_i$ as well, prompting the need for wider bound. To do this, instead of sampling $q_l$ and $q_u$ such that $q_l \leq \frac{i}{N} \leq q_u$, sample $q_l$ and $q_u$ to be the tightest quantiles such that $q_l \leq \max\{\frac{i}{N} - \epsilon\ , \min\{\mathbf{q}\}\}$ and $q_u \geq \min\{\frac{i}{N} + \epsilon, \max\{\mathbf{q}\}\}$, where taking the $\max$ and $\min$ with $\mathbf{q}$ enforces $q_u$ and $q_l$ to be within $[\min\{\mathbf{q}\},\max\{\mathbf{q}\}]$, the range of available quantiles. Sampling this way guarantees the $y_j$ which $x_i$ is supposed to be \textit{matched} with lies within $[Q^Y_{q_l}$, $Q^Y_{q_u}]$, and the $x_k$ that $y_j$ is supposed to be \textit{matched} with is in the interval $[Q^X_{q_l},Q^X_{q_u}]$. Then, similar to the \textit{zero-violation models}, we attain the confidence interval by finding the appropriate $u$ and $l$ such that $\text{I\kern-0.10em P}(l \leq Z \leq u) \geq \text{I\kern-0.10em P}(Z^{U} \leq u) - \text{I\kern-0.10em P}(Z^L \leq l)$. Algorithm \ref{alg:ISMB} illustrates the sampling of $Z^U$ and $Z^L$ in detail.
\begin{algorithm}
\begin{flushleft}
\textbf{Input}: Violation parameters $\epsilon_u$ and $\epsilon_l$, quantile labels $\mathbf{q}$, quantiles $\mathbf{Q^X}$ and $\mathbf{Q^Y}$ \\
\textbf{Output}: Upper and lower bounds $Z^U$ and $Z^L$ at time $t$
\end{flushleft}
\caption{Iterative Violation Model Bound}\label{alg:ISMB}
\begin{algorithmic}[1]
\For{$j = \{1,2,...,10^5\}$}
\State{Sample $\frac{i}{N}$ from Uniform$(0,1)$}
\State{$q_l \gets \max_{q \in \mathbf{q}}\{q \leq \max\{\frac{i}{N} - \epsilon_l,\min\{\mathbf{q}\}\}\}$}
\State{$q_u \gets \min_{q \in \mathbf{q}}\{q \geq \min\{\frac{i}{N} + \epsilon_u,\max\{\mathbf{q}\}\}\}$}
\State{Append $\mathbf{Q^X}_{q_u} - \mathbf{Q^Y}_{q_l}$ to $Z^U_t$}
\State{Append $\mathbf{Q^X}_{q_l} - \mathbf{Q^Y}_{q_u}$ to $Z^L_t$}
\EndFor
\end{algorithmic}
\end{algorithm}
In algorithm \ref{alg:ISMB}, since the CDF is a monotonically increasing function, and $\epsilon_u$ and $\epsilon_l$ is strictly positive,
\begin{equation}
x_i - F^{-1}_Y(\frac{i}{N} + \epsilon_u) \leq x_i - y_j \leq x_i - F^{-1}_Y(\frac{i}{N} - \epsilon_l)
\end{equation}
by the definition of $\epsilon_u$ and $\epsilon_l$ for $\forall x_i \match y_j$. Then, since $Q^Y_{q_u}$ and $Q^Y_{q_l}$ are sampled such that $Q^Y_{q_u} \geq F^{-1}_Y(\frac{i}{N} + \epsilon_u)$ and $Q^Y_{q_l} \leq F^{-1}_Y(\frac{i}{N} - \epsilon_l)$, we have
\begin{align}
x_i - Q^Y_{q_u} &\leq x_i - F^{-1}_{Y}(\frac{i}{N} + \epsilon_u) \\
&\leq x_i - y_j \\
&\leq x_i - F^{-1}_Y(\frac{i}{N} - \epsilon_l) \\
&\leq x_i - Q^Y_{q_l}
\intertext{since $Q^X_{q_l} \leq x_i \leq Q^X_{q_u}$,}
Q^X_{q_l} - Q^Y_{q_u} \leq x_i &- y_j \leq Q^X_{q_u} - Q^Y_{q_u}
\intertext{that is,}
Z^L \leq &Z \leq Z^U
\end{align}
The task remains to extract $\epsilon$ from the quantiles. Since only 23 quantiles and their respective \textit{matchings} are known, it is not possible to obtain information on $F_X(x_i) - F_Y(y_j)$ for the full distribution. Therefore, $\epsilon$ can be only estimated with the \textit{matchings} $x_i \match y_j$ observable from the quantiles. As the \textit{matchings} are only known for scenarios with $t_{\text{app}} \geq t_0$ ($Q_i^X$ should be equal to $Q_i^Y$ since scenario has not taken effect. If not, all misalignment are due to stochasticity), only those scenarios are valid for the approximation of $\epsilon$. Now, we propose the concrete method to estimate such $\epsilon$ from scenarios with $t_{\text{app}} \geq t_0$.
\subsubsection{Estimating $\epsilon$ for Violation Models}
Due to the above-mentioned reasons, additional assumptions on the behavior of the violation measure need to be proposed in order to bound the difference in scenarios with $t > t_{\text{app}}$.
\begin{assumption}\label{Assumption2}
A \textit{violation model} is said to be \textit{well-behaved} if its outputs satisfies both
\begin{align}
\argmax_{t}\{F_Y(y_j) - F_X(x_i)\} \leq t_{\text{app}}\\
\intertext{and}
\argmin_{t}\{F_Y(y_j) - F_X(x_i)\} \leq t_{\text{app}}
\end{align}
for $\forall x_i \match y_j$ \hfill (Non-increasing $\epsilon$ after $t_{\text{app}}$)
\end{assumption}
As discussed in the problem setting, both $F_Y(y_i)$ and $F_X(x_i)$ are sampled from $U(t)$ and is dependent on $t$. Here, with a slight abuse of notation, we take $\argmax$ and $\argmin$ on $t$ to restrict the behavior of the violation across time. Intuitively, assumption \ref{Assumption2} restricts the time at which maximum violation occurs to be before $t_{\text{app}}$, so that the maximum violation is observable. In effect, this assumption enforces the upper-bound estimated on violation for $t \leq t_{\text{app}}$ is still an upper bound for $t > t_{\text{app}}$.
\begin{definition}
Without loss of generality, let $X$ be the scenario such that $Q^X_{q_i} \geq Q^Y_{q_i}$, depending on $i$. Define the estimated upper and lower violation $\tilde{\epsilon_l}$ and $\tilde{\epsilon_u}$ for a \textit{violation model} as the following:
\begin{equation}
\tilde{\epsilon}_l = \max_{i}\{q_i - q_\alpha\}
\text{, where }
\alpha = \max_{k\leq i-1}{\{Q^X_{q_k} \ge Q^Y_{q_{i-1}}\}}
\end{equation}
and
\begin{equation}
\tilde{\epsilon}_u = -\min_{i}\{q_i - q_\beta\}
\text{, where }
\beta = \min_{k\geq i+1}{\{Q^Y_{q_k} \ge Q^X_{q_{i+1}}\}}
\end{equation}
for all $Q^X$ and $Q^Y$ corresponding to $t < t_{\text{app}}$
\end{definition}
\begin{lemma}\label{lem:2}
For the estimated upper and lower violation $\epsilon_u$ and $\epsilon_l$ and the approximated violation $\tilde{\epsilon}_u$ and $\tilde{\epsilon}_l$, we have
\begin{equation}
\epsilon_u \leq \tilde{\epsilon}_u \text{ and } \epsilon_l \leq \tilde{\epsilon}_l
\end{equation}
\end{lemma}
\begin{proof}
The proof can be found in Appendix \ref{lem:2proof}
\end{proof}
The implementation detail of estimating $\tilde{\epsilon}_u$ and $\tilde{\epsilon}_l$ is given in algorithm \ref{alg:estimatingeuel}. The estimated $\epsilon$ obtained can be then plugged into algorithm \ref{alg:IIMB} to evaluate the $\alpha$-confidence interval.
\begin{algorithm}
\caption{Estimating $\epsilon_l$ and $\epsilon_u$}\label{alg:estimatingeuel}
\begin{flushleft}
\textbf{Input:} Quantile labels $\mathbf{q}$, quantiles $\mathbf{Q^{X_t}}$ and $\mathbf{Q^{Y_t}}$, $t_{\text{app}}$\\
\textbf{Output}: $\tilde{\epsilon}_u$ and $\tilde{\epsilon}_l$
\end{flushleft}
\begin{algorithmic}[1]
\State {$\mathbf{\epsilon}_u \gets \text{[ ]}$}
\State {$\mathbf{\epsilon}_l \gets \text{[ ]}$}
\For{$t = 1,2,...,t_{app}$}
\For{$i = 1,2,...,|\mathbf{q}|$}
\If{$\mathbf{Q^{X_t}}[i] \geq \mathbf{Q^{Y_t}}[i]$}
\State{$\mathbf{Q^{U_t}} \gets \mathbf{Q^{X_t}}$}
\State{$\mathbf{Q^{L_t}} \gets \mathbf{Q^{Y_t}}$}
\Else
\State{$\mathbf{Q^{U_t}} \gets \mathbf{Q^{Y_t}}$}
\State{$\mathbf{Q^{L_t}} \gets \mathbf{Q^{X_t}}$}
\EndIf
\State{$\alpha \gets \max_{k\leq i}\{\mathbf{Q^{U_t}}[k] \geq \mathbf{Q^{L_t}}[i-1]\}$}
\State{$\beta \gets \min_{k \geq i}\{\mathbf{Q^{L_t}}[k] \geq \mathbf{Q^{U_t}}[i+1]\}$}
\State{Append $\mathbf{q}[i] - \mathbf{q}[\alpha]$ to $\mathbf{\epsilon}_u$}
\State{Append $\mathbf{q}[\beta] - \mathbf{q}[i]$ to $\mathbf{\epsilon}_l$}
\EndFor
\EndFor
\State{$\tilde{\epsilon}_l \gets \max\{\mathbf{\epsilon}_l\}$}
\State{$\tilde{\epsilon}_u \gets \max\{\mathbf{\epsilon}_u\}$}
\end{algorithmic}
\end{algorithm}
\subsubsection{Approximating $\tilde{\epsilon}$}
As shown in \ref{lem:2}, both $\tilde{\epsilon}_u$ and $\tilde{\epsilon}_l$ are guaranteed to be over-estimations for $\epsilon_u$ and $\epsilon_l$.
In the scenarios where the $x_i$ and $y_j$ are not clustered around certain values, their cumulative distribution functions should be relatively "smooth". As the estimated $\epsilon$ is always maximized in the worst-case scenario (all the values are clustered so that the CDF looks stair-like), it is reasonable to approximate the estimation of $\epsilon$ to obtain a smaller uncertainty bound. Naturally, the CDF of $X$ and $Y$ can be interpolated to obtain more quantiles. The monotonicity of Cubic Hermite Interpolating Polynomial (PCHIP) makes it a suitable candidate for interpolating CDF. As stated earlier, the smaller the $\epsilon$, the closer $Z^{L}$ is to $Z^{U}$ in algorithm \ref{alg:IIMB}, resulting in a tighter confidence interval.
\begin{definition}
Depending on the choice of $i$, let $X$ be the scenario associated with a higher quantile $(Q^X(i) > Q^Y(i))$. let $\textsc{PCHIP}_X(x_i):X \to q \in [0,1]$ be the interpolated CDF of scenario X (i.e $F_X$). Define the approximated $\tilde{\epsilon}_l$ and $\tilde{\epsilon}_u$ as
\begin{align}
\varepsilon_u &= \max_{q}\{\text{PCHIP}_Y(x) - \text{PCHIP}_X(x)\}\\
\varepsilon_l &= -\min_{q}\{\text{PCHIP}_Y(x) - \text{PCHIP}_X(x)\}
\end{align}
for $\forall x \in [\min\{\mathbf{Q^Y},\mathbf{Q^X}\},\max\{\mathbf{Q^X},\mathbf{Q^Y}\}]$.
\end{definition}
As more quantiles become available, the PCHIP interpolation in \ref{alg:approximateep} becomes a more accurate approximation of the inverse CDF. Since PCHIP is always a refined approximation of the quantiles and combining lemma \ref{lem:2}, the relations $\varepsilon_u \leq \tilde{\epsilon}_u \leq \epsilon_u$ and $\varepsilon_l \leq \tilde{\epsilon}_l \leq \epsilon_l$ hold.
\begin{algorithm}
\caption{Approximating $\tilde{\epsilon}$ with PCHIP}\label{alg:approximateep}
\begin{flushleft}
\textbf{Input}: Quantile labels $\mathbf{q}$, quantiles $\mathbf{Q^{X_t}}$ and $\mathbf{Q^{Y_t}}$; \\
\textbf{Output}: Approximated $\varepsilon_l$ and $\varepsilon_u$
\end{flushleft}
\begin{algorithmic}[1]
\State {$\mathbf{\varepsilon_u} \gets \text{[ ]}$}
\State {$\mathbf{\varepsilon_l} \gets \text{[ ]}$}
\For {$t = 0, 0.001, 0.002, ..., 0.999, 1.000\}$}
\State{$\tilde{F}_X \gets \text{PCHIP}(\mathbf{Q^{X_t}},\mathbf{q})$}
\State{$\tilde{F}_Y \gets \text{PCHIP}(\mathbf{Q^{Y_t}},\mathbf{q})$}
\For{$i = 1,2,...,|\mathbf{q}|$}
\If{$\tilde{F}_X(i)$ > $\tilde{F}_X(i)$}
\State{$\tilde{F}_Y,\tilde{F}_X$ = $\tilde{F}_X,\tilde{F}_Y$}
\EndIf
\State{$\mathbf{\tilde{\varepsilon}}.\text{append}(\tilde{F}_Y(i) - \tilde{F}_X(i))$}
\EndFor
\EndFor
\State{$\varepsilon_l \gets \text{max}\{\mathbf{\tilde{\varepsilon}}\}$}
\State{$\varepsilon_u \gets -\text{min}\{\mathbf{\tilde{\varepsilon}}\}$}
\end{algorithmic}
\end{algorithm}
\subsection{Approximating $Z$}
The approach of using interpolated quantiles to approximate $\tilde{\epsilon}_l$ and $\tilde{\epsilon}_u$ can be used to approximate the bounds for $Z = x_i - y_j$ for $\forall x_i \match y_j$ as well. Again, in scenarios where the CDF for $X$ and $Y$ are ``smooth'', we can use the PCHIP interpolator to obtain a modified version of algorithm \ref{alg:ISMB}. For \textit{zero-violation models}, recall $Z^L \to Z^U = Z$ as the number of quantiles approaches infinity uniformly over the range $[0,1]$. Thus using PCHIP to approximate $Z^U_t$ and $Z^L_t$ would result in $Z^L_t = Z^U_t$, as reflected in the bottom right subplot of \ref{fig:Z_different_e}. The implementation details are given in algorithm \ref{alg:AIMB}. To approximate a \textit{zero-violation model}, all that's needed is to apply algorithm \ref{alg:AIMB} with $\epsilon_u = \epsilon_l = 0$.
\begin{algorithm}
\begin{flushleft}
\textbf{Input}: Quantile labels $\mathbf{q}$, quantiles $\mathbf{Q^{X}}$ and $\mathbf{Q^{Y}}$\\
\textbf{Output}: Discrete Random Variables $Z^U$ and $Z^L$
\end{flushleft}
\caption{Approximated Iterated Model Bound}\label{alg:AIMB}
\begin{algorithmic}[1]
\State{$\mathbf{Z_t^U} \gets [\text{ } ];\text{ } \mathbf{Z^L_t} \gets [\text{ }]$}
\State{$\tilde{Q}^{X} \gets \text{PCHIP}^{-1}(\mathbf{q},\mathbf{Q^{X}})$}
\State{$\tilde{Q}^{Y} \gets \text{PCHIP}^{-1}(\mathbf{q},\mathbf{Q^{Y}})$}
\For{$k = 1,2,...,10^5$}
\State{Sample $\frac{i}{N}$ from Uniform$(0,1)$}
\State{$q_l \gets \frac{i}{N} - \epsilon_l ;\text{ } q_u \gets \frac{i}{N} + \epsilon_u$}
\State{Append $\tilde{Q}^{X}(q_u) - \tilde{Q}^{Y}(q_l)$ to $\mathbf{Z^U_t}$}
\State{Append $\tilde{Q}^{X}(q_l) - \tilde{Q}^{Y}(q_u)$ to $\mathbf{Z^L_t}$}
\EndFor
\Return {$\mathbf{Z^U_t}, \mathbf{Z^U_t}$}
\end{algorithmic}
\end{algorithm}
\section{Experiments}
We evaluate our proposed methods on the multi-model prediction results for round 9 and round 11, where round 9 has $t_{\text{app}} > t_0$ and round has $t_{\text{app}} = t_0$; that is, the divergence of scenarios in round 9 is delayed while round 11 is immediate, and each increment in the timestamp $t$ represents a week. We first use round 9 to demonstrate the estimation and approximation of $\epsilon$. As it is only possible to estimate $\epsilon$ for a scenario with $t > t_{\text{app}}$. Then, we devise a reasonable guess of $\epsilon$ and evaluate the difference of scenarios with that $\epsilon$.
\subsection{Estimation and Approximation of $\epsilon$}
\begin{tabular}{c|c|c|c|c}
Model & $\tilde{\epsilon}_l$ & $\tilde{\epsilon}_u$ & $\varepsilon_l$ & $\varepsilon_u$ \\
\hline
USC SIkJalpha \cite{sikjalpha}& \textbf{0.05} & \textbf{0.1} & 0.013 & 0.013 \\
Ensemble & \textbf{0.05} & \textbf{0.1} & 0.038 & 0.037 \\
Ensemble LOP & \textbf{0.05} & \textbf{0.1} & 0.025 & 0.025 \\
Ensemble LOP Untrimmed & \textbf{0.05} & \textbf{0.1} & 0.02 & 0.025 \\
JHUAPL Bucky \cite{bucky}& \textbf{0.05} & \textbf{0.05} & 0.00 & 0.00 \\
MOBS NEU-GLEAM COVID \cite{mobs}& \textbf{0.05} & \textbf{0.1} & 0.05 & 0.05 \\
\end{tabular}
\makeatletter\def\@captype{table}\makeatother
\caption{Summary of $\epsilon$ for round 9 scenario `B' and `A' cumulative cases. For each row, the largest of $\tilde{\epsilon}$ and $\varepsilon$ are marked in bold}\label{table:ep}
In round 9 $t_{\text{app}} = 4$. In other words, with prediction starting at $t_0 = 0$, there is a total of 4 weeks' data available before $t_{\text{app}}$ for obtaining $\epsilon$. We demonstrate the estimation and approximation of both $\epsilon_u$ and $\epsilon_l$ for a list of models participating in the multi-model COVID-19 prediction effort. In figure \ref{fig:epsilon_models}, we choose scenarios `A' and `B' as the scenarios of interest; namely, scenario `A' refers to adopting childhood vaccination and no new COVID-19 variant, while scenario `B' refers to no childhood vaccination and no new COVID-19 variant, all of which are for cases cumulative data. We observe from the left subplot that the estimated epsilon violation $\tilde{\epsilon}_l$ and $\tilde{\epsilon}_u$ is small and clusters for all the models being examined. For the approximations $\varepsilon_l$ and $\epsilon_u$, a linear relationship is observed between the two values for any given \textit{model}. The numerical results are given in table \ref{table:ep}. Notice that for all of the \textit{models} examined, $\varepsilon \leq \tilde{\epsilon}$ as expected.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{rslt8.pdf}
\caption{$\tilde{\epsilon}$ and $\varepsilon$ plots for various models for round 9 scenarios `A' and `B. For the top subplot, $\tilde{\epsilon}_l$ and $\tilde{\epsilon}_u$ are clustered at two coordinates on the graph. This is understandable as $\tilde{\epsilon}_u$ and $\tilde{\epsilon}_l$ are always greater or equal to the smaller difference in quantiles; that is, for models with violations smaller than the smallest difference in quantiles, the estimated violation gets rounded up. For approximated $\varepsilon$, the bottom subplot, clustering is not observed as $\varepsilon$ are calculated from interpolated quantiles where the smallest difference in the quantiles is close to 0.}
\label{fig:epsilon_models}
\end{figure}
\subsubsection{Distribution of $\tilde{\epsilon}$ and $\varepsilon$}
Since we obtain $\tilde{\epsilon}$ and $\varepsilon$ via maximization on $t \leq t_{\text{app}}$, it is worthwhile to examine the distribution of both $\tilde{\epsilon}$ and $\varepsilon$ versus $t$ to make sure the maximums obtained are not outliers. Again, we use the same round and environment settings as before. From figure \ref{fig:e_dist}, the value of $\tilde{\epsilon}_l$ stays constant in the $4$ weeks that are examined, and although slight variations in $\tilde{\epsilon}_u$ is present, no significant outlier exist to have a significant impact on the overall $\tilde{\epsilon}_u$.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{rslt6.pdf}
\caption{Distribution of $\tilde{\epsilon}$ over time for round 9 cases cumulative scenarios `A' and `B'. As observed, since both $\tilde{\epsilon}$ and $\varepsilon$ are maximized over time, the estimates for $\epsilon$ are reasonable as no significant variation is observed over time.}
\label{fig:e_dist}
\end{figure}
For the distribution of $\varepsilon$ in figure \ref{fig:vare_dist}, the clustering effect is significantly reduced as compared to the non-interpolated estimate of $\epsilon$ in figure \ref{fig:e_dist}. From observation, although $\varepsilon$ has a higher variance across weeks for any particular model, the overall distribution remains fairly uniform and no significant outlier exists for all of the models examined.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{rslt5.pdf}
\caption{Distribution of $\varepsilon$ over time for round 9 cases cumulative scenarios `A' and `B'. Comparing to the $\tilde{\epsilon}$ in figure \ref{fig:e_dist}, $\varepsilon$ are not lower-bounded by the smallest difference in quantiles. Therefore, as expected, more variation is observed across time.} \label{fig:vare_dist}
\end{figure}
\subsection{Confidence Intervals of Z}
After obtaining an estimated $\epsilon$ and examining its distribution in detail, we now attempt to use algorithm \ref{alg:ISMB} and algorithm \ref{alg:AIMB} to bound the difference of cumulative cases for round 11 scenario `B' $-$ `A'. Specifically, scenario `B' refers to ``Optimistic severity and high transmissibility increase", while `A' refers to ``Optimistic severity and low transmissibility increase", with both scenarios starting at 2021-12-21. Unlike round 9, round 11 has $t_0 = t_{\text{app}}$, and we therefore cannot estimate nor approximate $\epsilon$ with the algorithm proposed. Yet, we should be able to make reasonable speculation of $\epsilon$ from the previous plots of $\epsilon$ distributions. Instead of taking a conclusive guess, we demonstrate the effect of different $\alpha$ on the resulting $\alpha$-confidence interval. In particular, we use $\alpha = 0.8$. The result from both exact (algorithm \ref{alg:ISMB}) $Z$ and estimated (algorithm \ref{alg:AIMB}) $\alpha$-confidence interval for $Z$ are plotted side by side in figure \ref{fig:Z_different_e}.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{rslt3.pdf}
\caption{0.8-confidence interval for scenario B - A cases cumulative round 11 for the model USC-SikJ$\alpha$. The left column of subplots results from algorithm \ref{alg:ISMB} and the right column are the approximated bounds via algorithm \ref{alg:AIMB}. This plot demonstrates the effects of varying levels of violation have on the resulting bound.}
\label{fig:Z_different_e}
\end{figure}
Intuitively, the effect of simultaneously increasing $\epsilon_u$ and $\epsilon_l$ has a two-sided effect on the $\alpha$-confidence interval obtained for some fixed alpha. When $\epsilon_l = \epsilon_u = 0$ for the approximated method (i.e interpolating the quantiles), it has been previously shown that $Z^U$ converges to $Z^L$; in other words, the medians and mean of both random variables collide. This can be seen from the bottom left subplot of figure \ref{fig:Z_different_e}.
From the plots, we are provided with a way to interpret the results generated by the \textit{model} ``USC SIKJalpha": With at least 0.8 probability, the difference in the projections of highly transmissible and low transmissible variants of COVID-19 is going to fall within the blue region ascribed in figure \ref{fig:Z_different_e}.
\section{Conclusion}
The goal of generating scenario projections based on a future decision is to assess the impact of that decision. In practice, however, as the joint distribution of decisions is unknown, obtaining tight bounds from probabilistic projections is often infeasible without significant changes to the model. Without doing additional work on the model, the methods proposed can find arbitrary $\alpha$-confidence intervals for scenario differences, under some assumptions. The second half of the paper is dedicated to relaxing the assumptions by incorporating a quantitative measure $\epsilon$ on the degree of violation of the assumption. Finally, a method is proposed to reliably approximate the difference in the probabilistic projections of scenarios.
\section*{Acknowledgement}
This work was supported by the Centers for Disease Control and Prevention and the National Science Foundation under the awards no. 2135784 and 2223933. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author and do not necessarily reflect the views of the National Science Foundation or the Center for Disease Control and Prevention.
\bibliographystyle{plainnat}
|
\section{Introduction}
Simultaneous Localisation and Mapping (SLAM) research aims to concurrently estimate both the scene geometry of the unknown environment as well as the robot pose within it from the data of its on-board sensors only. It has rapidly progressed from sparse SLAM~\cite{Davison:etal:PAMI2007, Klein:Murray:ISMAR2007} into dense SLAM~\cite{Newcombe:etal:ISMAR2011, Vespa:etal:RAL2018}, and recently into semantic object-level SLAM~\cite{McCormac:etal:ICRA2017, McCormac:etal:3DV2018}. This fast-evolving research has enabled many robotic applications. Despite this, most SLAM research still assumes a static scene, where points in the 3D world maintain constant spatial positions in a global coordinate. Any information violating this assumption, such as moving objects in the environment, would be treated as outliers and are intentionally ignored in tracking and mapping steps.
This setup, however, can only handle a small amount of dynamic elements, excluding itself from many real-world applications as environments, particularly where humans are involved, are continually changing. A robust SLAM system capable of handling highly dynamic environments, therefore, is desirable. Most current dynamic SLAM research can be categorised into three main directions. One maps the whole changing world in a non-rigid deformable representation to deal with the changing topology of deformable/moving objects~\cite{Newcombe:etal:CVPR2015}. The second aims at improving the robustness and accuracy of camera tracking by ignoring all possibly moving objects and building a single static background model~\cite{Jaimez:etal:ICRA2017, Scona:etal:ICRA2018, Bescos:etal:RAL2018}. The third models dynamic environments by creating object-centric maps for each possibly moving rigid object in the scene while fusing corresponding information into these object-level maps~\cite{Runz:Agapito:ISMAR2018, Xu:etal:ICRA2019}. Object-level tracking and mapping can be conducted for each object and camera motion against the static part of the map. This paper aligns with the last direction as we believe that, similar to human perception, an instance-awareness of the surrounding environment can help intelligent robots perceive the scene changes and enables meaningful interactions with the surrounding environment.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]
{figures/real/24-09-2021-12-47-59/teaser-crop.pdf}
\caption{Given RGB-D images, our system builds object-level dense dynamic maps that can robustly track camera pose and object poses while completing the missing sensor information using object priors. Compared to the classic TSDF maps, our object maps fill in unobserved parts and their latent codes can be optimised jointly with object poses. Interfered regions by humans can be detected and intentionally removed in the system. The background pointclouds are projected for pure visualisation purpose.}
\label{fig:teaser_completion}
\end{figure}
By far most existing object-level dynamic SLAM systems mentioned above adopt the classic map representation that have been exploited in the static SLAM systems, such as pointclouds~\cite{Bescos:etal:RAL2021}, surfels~\cite{Runz:Agapito:ISMAR2018} or volumetric maps~\cite{Xu:etal:ICRA2019}. This leads to partial or incomplete object maps as only the observable information can be fused into the object models. Information in unseen parts can not be filled unless an object or the sensor is moved actively. Contrary to reconstructing objects from scratch, some works recently explored learning-based category-level object shape priors and build object-level maps based on learned shape priors~\cite{Sucar:etal:3DV2020, Wang:etal:3DV2021}. The object geometry and pose are usually optimised via differentiable rendering. However, most of these systems are only applicable in static scenes. Besides, despite being able to generate complete object geometry, object shape priors cannot capture complex geometry details as the bottleneck of its latent representation can only interpolate shapes inside the training datasets~\cite{Park:etal:CVPR2019}. When combined with dense image alignment, such as photometric or ICP residuals, this inconsistency between the measurement and the object prior model inevitably leads to inaccurate object motion trajectory estimates.
This work stands in the middle between reconstructing object geometry from scratch and mapping using object shape priors. We reconstruct the observable part by continuously fusing depth measurements into a volumetric canonical space and predict the complete geometry by conditioning it on the fused volume. The resulting object geometry can preserve the details that have been observed in the past and simultaneously complete the missing geometry information. We also verified that this completed object geometry can further improve the accuracy of object tracking. The main contributions in this paper~can be summarised as follows:
\begin{enumerate}
\item we present, to the best of our knowledge, the first RGB-D object-level dynamic mapping system that can complete unseen parts of objects using a shape prior encoded in neural networks while still reconstructing observed parts accurately;
\item a joint optimisation of object pose and shape geometry based on geometric residuals and differentiable rendering;
\item extensive experiments of object tracking and reconstruction components on synthetic and real-world data to evaluate the benefits of object geometry completion for object-level SLAM.
\end{enumerate}
\section{Related Works}
\paragraph{Object-level dynamic SLAM}
Although object-level dynamic SLAM research can be dated back to \cite{Wang:etal:ICRA2003}, visual dense object-level dynamic SLAM has only been explored recently. From RGB-D sensor inputs, Co-Fusion~\cite{Runz::Agapito::ICRA2017} segments objects by either ICP motion segmentation or semantic segmentation and then tracks objects separately based on ElasticFusion~\cite{Whelan:etal:IJRR2016}. MaskFusion~\cite{Runz:Agapito:ISMAR2018} segments objects using a combination of instance segmentation from Mask-RCNN and geometric edges, and tracks objects using the same approach as Co-Fusion. Both Co-Fusion and MaskFusion use surfels to represent map models, which is memory efficient but cannot directly provide free space information in the map, and neither surface connectivity. DynSLAM-II~\cite{Bescos:etal:RAL2021} extends from ORB-SLAM II~\cite{Mur-Artal:etal:TRO2017} and formulates object tracking using sparse feature descriptor matching. Object maps are represented using clusters of pointclouds, which can bring object poses and geometries into the pose graph optimisation but also lack space connection awareness.
Instead, MID-Fusion~\cite{Xu:etal:ICRA2019} uses memory-efficient octree-based volumetric signed distance field (SDF) representation for objects and re-parametrises tracking residuals in object coordinates and weights. EM-Fusion~\cite{Strecke:Stuckler:CVPR2019} similarly uses volumetric SDF object maps but formulates object tracking as direct alignment of input frames with the SDF representations. Their following work \cite{Strecke:Stuckler:CVPR2020} infers the missing object geometry by penalising the hull and intersection constraints. However, it did not explore shape prior information and requires heavy computation to optimise SDF field explicitly. Instead, we fuse the depth measurement into object-level SDF maps and predict completed object geometries by incorporating a shape prior in continuous occupancy fields.
\paragraph{SLAM with shape prior maps}
Instead of estimating both object geometry and poses from scratch, some approaches use a shape prior to represent objects. Since the coordinates of object shape priors and the run-time measurement are not necessarily aligned, a relative rigid transformation needs to be estimated. This is analogous to the localisation-only problem in SLAM. SLAM++~\cite{Salas-Moreno:etal:CVPR2013} is one of the pioneering object-level RGB-D SLAM systems. It scanned objects in advance and then maps the detected instances at run-time by jointly optimising a pose graph of camera and object poses. Relying on pre-scanned objects, however, limits its ability to scale to unknown object models. Rather than using instance-level shape priors, several following works exploited category-level shape priors as there is limited variance in certain object categories. The category-level shape prior can be learnt in various representations, such as PCA models as in DirectShape~\cite{Wang:etal:ICRA2020}, occupancy grids as in Deep-SLAM++ \cite{Hu:etal:ARXIV2019}, variational autoencoders as in NodeSLAM~\cite{Sucar:etal:3DV2020}, or autodecoders as in DSP-SLAM~\cite{Wang:etal:3DV2021}. However, most of these works only target static environments, as multi-view consistency of static world points is required to localise the shape prior models. \cite{Li:etal:RAL2021} relax this restriction using a Bayesian filter to associate object detections on different frames and fuse the prior model by simply averaging the latent codes from each frame. However, object shape deviations cannot always be captured by the shape prior interpolation. The object tracking accuracy would be affected by the discrepancy between the prior shape model and the online measurement.
We address this challenge by conditioning the completion network on the integrated 3D reconstruction model.
\paragraph{Object-level tracking}
To track moving objects in RGB-D sequences, several pioneering object-level works adopt the frame-to-model tracking methods from RGB-D SLAM systems \cite{Runz::Agapito::ICRA2017} and parametrise them for object tracking~\cite{Xu:etal:ICRA2019, Bescos:etal:RAL2021}. The classic photometric residual, however, is difficult to deal with as object lighting changes; several approaches explore using learning-based robust features to formulate object tracking in direct \cite{Xu:etal:RAL2021} and in-direct ways \cite{Wen:Bekris:IROS2021}. Parallel to learning category-level shape priors, Wang et al.~\cite{Wang:etal:CVPR2019N} proposed to learn category-level pixel-wise correspondence from RGB-D images to the canonical space. The shape is implicitly defined from this correspondence and the frame-to-canonical transformation can be estimated from this noisy correspondence. Rempe et al.~\cite{Rempe:etal:NIPS2020} further proposed to generate more stable correspondences by accumulating temporal information from RGB-D sequences. Recently, Muller et al.~\cite{Muller:etal:CVPR2021} proposed to track moving objects and predict their complete geometry using such canonical correspondence representation. In this work, the object pose is initially predicted using canonical correspondence regression. However, we found it does not necessarily yield alignment to the canonical space and we further optimise the pose tracking using geometric residual and differential rendering.
\begin{figure*}[tbp]
\centering
\includegraphics[width=0.7\linewidth]{figures/pipeline/mapping-crop.pdf}
\caption{The overview of our object geometry representation.}
\label{fig:mapping_overview}
\end{figure*}
\begin{figure}[tbp]
\centering
\includegraphics[width=\linewidth]{figures/pipeline/pipeline-crop.pdf}
\caption{The pipeline of the proposed method}
\label{fig:pipeline}
\end{figure}
\section{Method}
\subsection{Notations and Preliminaries}
In this work, we will use the following notation: a coordinate frame is denoted as $\cframe{A}$. The homogeneous transformation from $\cframe{B}$ to $\cframe{A}$ is denoted as $\T{A}{B}$, which is composed of a rotation matrix $\Crot{A}{B}$ and a translation vector $\trans{A}{A}{B}$.
Every detected object is represented in its individual object coordinate frame $\cframe{O_n}$, with $n \in \{ 0\ldots, N\}$, where $N$ is the total number of objects (excluding background) and 0 denotes background. We assume a canonical static volumetric model is stored in each object coordinate frame, forming the basis of our multi-instance SLAM system. To leverage the category-level shape prior, we need to align the canonical space with the one defined in training, otherwise the completion performance will be deteriorated since it cannot fully take advantage of the shape variances of the objects in the same category. The relative transformation between the current world coordinate and the corresponding object canonical space is defined as a joint state composed of a rigid transformation $\T{W}{O_n}$ and the object scale $s_{O_n}$. We define the object pose using this joint state. $\T{W}{O_n}$ needs to be continuously updated for a moving object but the object scale should be consistent across different frames.
\subsection{System Overview}
\cref{fig:pipeline} shows the pipeline of our proposed system.
Each input RGB-D image is processed by Mask R-CNN to perform instance segmentation.
The camera pose is tracked against background regions, excluding the human mask area and moving objects, similar to what has been proposed in MID-Fusion~\cite{Xu:etal:ICRA2019}.
The object geometry is composed of two nodes with a shared object pose: prior node and posterior node, as shown in \cref{fig:mapping_overview}. The prior node represents its category-level shape prior using a latent code $\mbf z_0 \in \mathbb{R}^{64}$. It can be used to express the continuous SDF field $s$ on any query 3D location in object canonical coordinate $\mbf v \in \mathbb R^3$ using a DeepSDF shape prior network $F_0$~\cite{Park:etal:CVPR2019} as
\begin{equation}
\label{eq:deepsdf}
s = F_0(\mbf v, \mbf z_0).
\end{equation}
The prior node is used to initialise the object pose and re-localise the object model when object tracking is lost as its shape is not affected by measurements. The posterior node encodes a fused partial TSDF volume and its associated TSDF weight volume into TSDF feature volume $\theta_t$ and TSDF confidence volume $\theta_c$ separately using 3D-UNet~\cite{Cciccek:etal:MICCAI2016}. Then a complete occupancy field $o$ can be predicted on any given 3D position $\mbf v \in \mathbb R^3$ using a shape completion network $F_1$:
\begin{equation}
\label{eq:conv-onet}
o = F_1(\mbf v, \theta_t[\mbf v], \theta_c[\mbf v], \mbf z_1).
\end{equation}
where $\theta_t[\mbf v] \in \mathbb{R}^{32} $ and $\theta_c[\mbf v] \in \mathbb{R}^{1}$ denote the feature vectors tri-linearly interpolated at the point $\mbf v$ inside the volumes $\theta_t$ and $\theta_c$, respectively.
We additionally condition it on a latent code $\mbf z_1 \in \mathbb{R}^{32}$ so that the hidden space can be optimised to generate novel shapes.
The shape completion network $F_1$ shares a similar architecture to the CONet~\cite{Peng:etal:ECCV2020}, but also takes extra inputs of TSDF confidence weight and an instance-level latent code.
We perform an efficient Axis Aligned Bounding Box (AABB) ray intersection test~\cite{Majercik:etal:JCGT2018} to find all the visible existing object models in the current viewpoint and render object masks for each visible models. An Intersection of Union (IoU) between the detections on the current frame and the rendered model masks is computed to build data associations between the current frame detections and existing object models. Then we track each object model and complete its geometry by performing a joint optimisation of pose and geometry (\cref{subsec: joint_opt}). Using estimated poses of the camera and objects, new depth measurements are fused into an object model and a complete shape geometry can be predicted by conditioning on the fused model.
New objects are created for unmatched detections by initialising their initial object pose using object prior models.
\subsection{Joint Optimisation of Object Pose and Geometry}
\label{subsec: joint_opt}
Instead of defining an arbitrary canonical space for object coordinates~\cite{Xu:etal:ICRA2019}, we need to estimate the 7DoF relative transformation, which is composed of $\T{W}{O_n}$ and $s_{O_n}$, to align the initialised object coordinate to the training canonical space for each object detection in order to take advantage of the learned prior information.
\paragraph{Initialisation}
\label{method:init}
Given an RGB-D frame $(I_L, D_L)$ and detected object mask $M_n$, we predict their positions in the canonical space $\mbf v$ and associated confidences $w$ from back-projected pointcloud using a modified normalised object correspondence network $F_n$ from \cite{Rempe:etal:NIPS2020}:
\begin{equation}
\generalThree{C_L}{v}{}(\pixel{L}) = \pi^{-1}(\pixel{L}, D_L[\pixel{L}]), \forall \pixel{L} \in M_n,
\end{equation}
\begin{equation}
\label{eq: nocs_regression}
F_n\left( \generalThree{C_L}{v}{} \right) \rightarrow \mbf v, w
\end{equation}
Then we solve the 7-DoF relative transformation, scale $s_{O_n}$, rotation $\Crot{C_L}{O_n}$, and translation $\trans{C_L}{C_L}{O_n}$ from the regressed correspondences using the Umeyama algorithm~\cite{Ummenhofer:PAMI1991} with SVD decomposition:
\begin{equation}
\label{eq: init loss}
\operatornamewithlimits{argmin}_{s_{O_n}, \T{C_L}{O_n}} \sum_{\pixel{L} \in M_n} w \left( \mbf v - \frac{1}{s_{O_n}}\T{C_L}{O_n}^{-1} \generalThree{C_L}{v}{} (\pixel{L}) \right).
\end{equation}
For the latent codes $\mbf z_0$ and $\mbf z_1$, we use both zero code initialisations.
We only run this pose initialisation step for new unmatched object detections. The initial pose solved from SVD decomposition, however, is necessary to be close to the ground-truth canonical pose, due to the unseen shapes or viewpoints, affecting the performance of shape completion.
\paragraph{Coarse Estimation}
\label{method:coarse}
To refine object canonical poses from the initial pose prediction, we jointly optimise it with the prior latent code $\mbf z_0$ to minimise the 3D SDF loss $E_\mathrm{SDF}$ and 2D rendering loss $E_\mathrm{render}$:
\begin{equation}
\label{eq: coarse_loss}
E_\mathrm{coarse} = \lambda_\mathrm{s} E_\mathrm{SDF}+ \lambda_\mathrm{r0} E_\mathrm{render} + \lambda_\mathrm{z0} ||\mbf z_0||.
\end{equation}
The 3D SDF loss is defined to encourage the back-project depth points to align with the object surface, where the zero SDF value is defined
\begin{equation}
\label{eq: sdf_loss}
E_\mathrm{SDF} = \sum_{\pixel{L} \in M_n} F_0 \left( \frac{1}{s_{O_n}} \T{C_L}{O_n}^{-1} \generalThree{C_L}{v}{} (\pixel{L}), \mbf z_0 \right).
\end{equation}
We cannot compute SDF residuals for empty space since ground-truth SDF values are not available at test time. Instead, for the non-surface areas, we use differentiable rendering to encourage the rendered depth $D_L$ to be close to the measured depth $\tilde D_L$. We compute the rendering loss for the visible 3D space inside the object 3D bounding box:
\begin{equation}
\label{eq: render_loss}
E_\mathrm{render} = \sum_{\pixel{L} \in B_n} D_L[\pixel{L}] - \tilde D_L[\pixel{L}],
\end{equation}
where
\begin{equation}
D_L[\pixel{L}] = \sum_{i=1}^N w_i d_i,
\end{equation}
and $w_i$ is the ray-termination probability \cite{Sucar:etal:3DV2020} of sample $i$ at depth $d_i$ along the ray from the pixel $\pixel{L}$,
\begin{equation}
w_i = o_i \prod_{j=1}^{i-1} (1-o_j),
\end{equation}
and $B_n$ is the 2D bounding box rendering from the estimated object 3D bounding box on this frame.
A continuous occupancy field can be extracted from the continuous prior SDF field as proposed in \cite{Wang:etal:3DV2021}:
\begin{equation}
\label{eq: sdf_to_occ}
o_i = -\frac{1}{2\sigma} F_0 \left( \frac{1}{s_{O_n}} \T{C_L}{O_n}^{-1} \left( \pi^{-1}(\pixel{L}, d_i) \right), \mbf z_0 \right),
\end{equation}
where $\sigma$ is the truncation distance to control the transition.
By freezing the network weight in $F_n$, the cost function in \cref{eq: coarse_loss} can be iteratively solved using Gauss-Newton optimisation with analytical Jacobians. Since the prior shape is not necessarily aligned with the actual observation, it is unnecessary to sample every pixel ray. Instead, we run this optimisation on sparse ray samples, which can speed up the optimisation without losing much accuracy.
\paragraph{Dense Refinement}
\label{method:dense}
After the coarse estimation, we have a rough alignment of the object model with the canonical space. However, the optimised object prior $\mbf z_0$ cannot necessarily capture the details of depth measurements. To further align the object model and to complete the hidden part, we jointly optimise the posterior latent code $\mbf z_1$ with the object pose by minimizing a 3D occupancy loss $E_\mathrm{occ}$ on both occupied and empty space (excluding the unknown 3D space) and a similar 2D rendering loss $E_\mathrm{render}$:
\begin{equation}
\label{eq: fine loss}
E_\mathrm{refine} = \lambda_\mathrm{o} E_\mathrm{occ} + \lambda_\mathrm{r1} E_\mathrm{render} + \lambda_\mathrm{z1} ||\mbf z_1||.
\end{equation}
The occupied space is defined on the back-projected points and the empty space is uniformly sampled in the background space as well as the foreground space before the depth measurement. The occupancy loss is defined using the binary cross-entropy loss between the predicted occupancy value $o_{\mbf v}$ from the shape completion network using \cref{eq:conv-onet} and the ground-truth occupancy values $o_{\mbf v}^*$ (0.5 for the occupied space and 0 for the empty space) for sampled points $\mbf v$ inside the occupied and empty space:
\begin{equation}
\label{eq: occupied}
E_\mathrm{occ} = -\sum_{{\mbf v}} [o_{\mbf v} \log(o_{\mbf v}^*) + (1-o_{\mbf v}) \log(1-o_{\mbf v}^*)].
\end{equation}
Similar to the coarse estimation, we can also evaluate the 2D rendering loss using \cref{eq: render_loss}. The difference is here we use the decoded continuous occupancy value for the sampled $d_i$ using \cref{eq:conv-onet}, instead of converting it from the SDF field in \cref{eq: sdf_to_occ}. The refined object pose $\T{C_L}{O_n}$ is used to integrate the current depth frame into the corresponding TSDF volume and weight volume \cite{Newcombe:etal:ISMAR2011}.
\subsection{Training Setup}
The learnable network parameters in this work includes three part, canonical correspondence network $F_n$, shape prior network $F_0$, shape posterior network $F_1$.
We train the canonical correspondence network using the partial pointcloud generated from the synthetic shapenet dataset~\cite{Shapenet:ARXIV2015}. During training, we augmented the input pointcloud with random object pose and solve the 7DoF object poses using \cref{eq: init loss}. To help network prediction robust to outliers, we also added random depth outliers in the pointcloud generation to learn the correspondence confidence $w$ in a self-supervised way. The solved pose is compared to the augmented ground-truth pose and the whole network is trained end-to-end since the estimation is differentiable.
We use the pre-trained off-shelf network weights for the category-level shape prior network $F_0$ \cite{Park:etal:CVPR2019}, which was also trained in the shapenet dataset~\cite{Shapenet:ARXIV2015}. To train the posterior shape completion network, we rendered depth maps for each object in the shapenet dataset~\cite{Shapenet:ARXIV2015} and trained the shape completion network using the partial depth observation. We use the occupancy loss defined in \cref{eq: occupied} to encourage the completed shape to be similar to the ground-truth one. Similar to the latent code training in DeepSDF~\cite{Park:etal:CVPR2019}, different object shapes have their own latent codes, which are trained together with the network. We make different partial observations of the same object shape share the same latent code.
\section{Experiments}
\subsection{Quantitative Reconstruction Evaluations}
\subsubsection{Experimental Setup}
We validate the reconstruction quality of our method on object-level surface reconstruction tasks. We conduct a comparison on the chair category of the ShapeNet~\cite{Shapenet:ARXIV2015} dataset. The split of train/val/test sets follows the same setting in \cite{Peng:etal:ECCV2020}. We randomly select 50 models from the test set to conduct quantitative evaluations. We generate input depth images by rendering images using uniformly sampled virtual camera viewpoints surrounding the CAD model. The hyperparameters used in inference optimization are chosen as $\sigma = 0.05$, $\lambda_\mathrm{s} = 100$, $\lambda_\mathrm{r0} = 2.5$, $\lambda_\mathrm{z0} = 5$, $\lambda_\mathrm{o} = 100$, $\lambda_\mathrm{r1} = 1$, and $\lambda_\mathrm{z1} = 1$.
\subsubsection{Baseline Methods}
To evaluate the object mapping, we compare with the following baseline methods:
\begin{itemize}
\item TSDF-fusion: We fuse the depth measurements into a TSDF volume grid as in \cite{Newcombe:etal:ISMAR2011}.
\item DeepSDF mapping: We use the pre-trained decoder weight in \cite{Park:etal:CVPR2019}. As the shape completion code is not provided, we optimise the SDF loss on the input pointcloud as well as the empty space constraint proposed in IGR~\cite{Gropp:etal:ICML2020}.
\item CONet: We use the weights trained from partial pointcloud input in \cite{Peng:etal:ECCV2020} and pass the accumulated pointcloud in the canonical space to generate the continuous occupancy field where the meshes are extracted.
\end{itemize}
\subsubsection{Metrics}
To quantitatively evaluate the quality and completeness of the shape reconstruction, we use the following metrics:
\begin{itemize}
\item IoU:
We sample 100k points uniformly in the bounding box and evaluate on both the reconstructed and the ground-truth meshes whether each point is inside or outside. The final value is the fraction of intersection over union. Higher is better.
\item Chamfer Distance:
we sample 100k points on the surface of both the ground-truth and the reconstructed mesh. We compute the closest points from the reconstructed to the ground-truth mesh using kD-tree and vice-versa. We then compute the average of the L1 distances to the closest points in each direction. Lower is better.
\item (In-)completeness: As the completeness of the object map is essential in this work, we also report completeness, which is the one-way chamfer distance from the ground-truth meshes to the reconstructed ones. This is to measure the closest distance from each ground-truth mesh points to the reconstruction. Lower is better.
\end{itemize}
\subsubsection{Results and Discussions}
We quantitatively evaluate how the view number of depth measurement would impact the reconstruction results of different methods. \cref{fig:eval_reconstruction} reports the result. It can be seen that our proposed method consistently show better reconstruction results from single view depth completion to multiple views. When the view number is limited, classic TSDF-Fusion shows worse result as it can only reconstruct the visible parts. CONet completes some missing information, but still struggles as it heavily depends on the input pointcloud. DeepSDF does not condition on the input and the latent code optimisation can fit the few depth measurement and shows better completion and reconstruction results. Our proposed method uses both the input information and shape prior information, yielding best performance. When more depth measurements are received, TSDF-Fusion and CONet start to fill in the missing information while DeepSDF struggles to leverage more measurement information. Our result also improves since we can also take advantage of the measurement information. \cref{fig:eval_map_vis} shows an examples of reconstruction results by each method in the view number case of 1, 5, and 10.
\begin{figure}[htb]
\centering
\subfloat[][IoU]
{\includegraphics[width=0.48\linewidth]{figures/eval_mapping/eval_reconstruction_iou.png}
}
\hfill
\subfloat[][Chamfer distance (L1)]
{\includegraphics[width=0.48\linewidth]{figures/eval_mapping/eval_reconstruction_chamfer-L1.png}
} \\
\subfloat[][(In-)Completeness (lower is better)]
{\includegraphics[width=0.48\linewidth]{figures/eval_mapping/eval_reconstruction_completeness.png}
} \\
\caption{Quantitative comparison of reconstruction quality and completion of our proposed methods v.s.\ classic TSDF-Fusion, learning-based DeepSDF and CONet. Our proposed method consistently show better reconstruction results from single view depth completion to multiple views.}
\label{fig:eval_reconstruction}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=0.9\linewidth]{figures/eval_mapping/eval_map-crop.pdf}%
\caption{Qualitative Results on reconstructions. Our method is superior to all other methods in completing missing information and reconstructing fine details.}
\label{fig:eval_map_vis}
\end{figure}
\subsection{Quantitative Tracking Evaluations}
\subsubsection{Experimental Setup}
To quantitatively evaluate the object-level tracking and mapping performance, we randomly select 10 object models from the test split of the chair category in the ShapeNet~\cite{Shapenet:ARXIV2015} and render 200 frames using Blender. To ensure diversity of object motion, texture, and illuminations, we randomise four point light sources, camera viewpoint, and object trajectories. We then subsample the sequences using sampling intervals {1, 2, 4} in order to create small, medium and large motion subsets.
\subsubsection{Baseline Methods}
To evaluate the object tracking, we compare with the following baseline methods:
\begin{itemize}
\item RGB-D VO: a non-learning-based visual odometry method proposed in \cite{Steinbrucker:etal:ICCVW2011}, which minimises the photometric loss between two frames. We re-parametrised it for object tracking.
\item Point-to-Plane ICP: a non-learning geometric registration method \cite{Rusinkiewicz:Levoy:3DMIN2001}
\item Color ICP: a non-learning registration method using both color and geometric information \cite{Park:etal:ICCV2017}
\item Prior: a state-of-the-art object pose estimation using DeepSDF shape prior model. It is originally proposed in \cite{Wang:etal:3DV2021} for static object pose estimation and we re-parametrised it for estimating moving objects. This is equivalent to an ablation study of using the prior model only without conditioned completion refinement from Section III-C c).
\item NOCS: a state-of-the-art learning-based canonical correspondence regression method. We adopted the network architecture proposed in \cite{Rempe:etal:NIPS2020}. This is equivalent to an ablation study of using only the initial prediction from Section III-C a).
\end{itemize}
\subsubsection{Metrics}
To quantitatively evaluate the accuracy of the object tracking, we use the following metrics:
\begin{itemize}
\item ATE: Absolute. Trajectory Error defined in \cite{Sturm:etal:IROS2012}, in the unit of of meters
\item RPE\_t: relative pose error (RPE) metrics in translation defined in \cite{Sturm:etal:IROS2012}, in the unit of of metres
\item RPE\_R: relative pose error (RPE) metrics in rotation defined in \cite{Sturm:etal:IROS2012}, in the unit of of degrees
\item R\_err: mean orientation error on each frame individually, in the unit of degrees
\item t\_err: mean translation error on each frame individually, in the unit of metres
\end{itemize}
The above metrics all indicate better tracking performance when the values are lower. To analyse the trajectory, we align the first frame of the estimated object pose to the ground-truth canonical space.
\subsubsection{Results and Discussions}
\cref{table: quan-track} reports the experimental results. It shows that our approach consistently outperforms both the learning-based and non-learning-based methods in small and large motion situations. For non-learning approaches, RGB-D VO~\cite{Steinbrucker:etal:ICCVW2011}, Point-to-Plane ICP~\cite{Rusinkiewicz:Levoy:3DMIN2001}, and Color ICP~\cite{Park:etal:ICCV2017} only leverage the depth and intensity information from two-view measurements, without taking into account any object shape prior information. The single view canonical correspondence prediction from NOCS~\cite{Rempe:etal:NIPS2020} only considers shape prior information and does not take advantage of the multiview constraint. Our proposed method instead combines both multi-view constraint and shape prior information into object pose estimation for better tracking accuracy. Similar to ours, the shape prior method~\cite{Wang:etal:3DV2021} adopts category-level shape prior from DeepSDF~\cite{Park:etal:CVPR2019} and uses differential rendering to estimate object poses. However, latent code optimisation cannot necessarily capture the geometry deviation between training space and test shapes and thus affects the accuracy of pose estimation.
\begin{table}[tb]
\centering
\setlength{\tabcolsep}{2.5pt}
\subfloat[][Keyframe gap-1]{
\label{table: track_kf1}
\begin{tabular}{lcccccc}
\toprule
method [unit] & ATE [m] & RPE\_t [m] & RPE\_R [$^{\circ}$] & R\_err [$^{\circ}$] & t\_err [m]\\
\midrule
\textbf{Ours} & \textbf{0.030} & \textbf{0.027} & \textbf{3.845} & \textbf{3.931} & \textbf{0.040} \\
Prior & 0.044 & 0.047 & 8.200 & 6.269 & 0.068 \\
RGBD & 0.254 & 0.106 & 18.47 & 32.25 & 0.314 \\
Point2Plane & 0.047 & 0.035 & 4.672 & 5.970 & 0.064 \\
Color ICP & 0.254 & 0.114 & 29.12 & 56.18 & 0.320 \\
NOCS & 0.074 & 0.059 & 23.46 & 37.87 & 0.085 \\
\bottomrule
\end{tabular}
}
\hfill
\subfloat[][Keyframe gap-2]{
\label{table: track_kf2}
\begin{tabular}{lcccccc}
\toprule
method [unit] & ATE [m] & RPE\_t [m] & RPE\_R [$^{\circ}$] & R\_err [$^{\circ}$] & t\_err [m]\\
\midrule
\textbf{Ours} & \textbf{0.033}& \textbf{0.032} & \textbf{5.243} & \textbf{4.224} & \textbf{0.043} \\
Prior & 0.046 & 0.052 & 11.91 & 8.063 & 0.063 \\
RGBD & 1.068 & 0.403 & 30.89 & 50.95 & 1.309 \\
Point2Plane & 0.070 & 0.056 & 8.570 & 9.384 & 0.087 \\
Color ICP & 0.536 & 0.351 & 36.69 & 60.56 & 0.568 \\
NOCS & 0.074 & 0.074 & 21.65 & 37.78 & 0.084 \\
\bottomrule
\end{tabular}
}
\hfill
\subfloat[][Keyframe gap-4]{
\label{table: track_kf4}
\begin{tabular}{lcccccc}
\toprule
method [unit] & ATE [m] & RPE\_t [m] & RPE\_R [$^{\circ}$] & R\_err [$^{\circ}$] & t\_err [m]\\
\midrule
\textbf{Ours} & \textbf{0.034} & \textbf{0.038} & \textbf{6.767} & \textbf{4.834} & \textbf{0.044} \\
Prior & 0.043 & 0.050 & 17.20 & 9.885 & 0.061 \\
RGBD & 1.942 & 0.866 & 43.34 & 68.86 & 2.177 \\
Point2Plane & 0.807 & 0.442 & 18.22 & 20.16 & 0.892 \\
Color ICP & 2.786 & 1.885 & 42.73 & 77.89 & 2.802 \\
NOCS & 0.073 & 0.085 & 26.95 & 35.41 & 0.083 \\
\bottomrule
\end{tabular}
}
\caption{Quantitative evaluation of object tracking method on the synthetic moving objects dataset.}
\label{table: quan-track}
\end{table}
\subsection{Timing analysis}
We implemented our system in PyTorch. The average inference time for a pair of RGB-D image in the resolution of 320 $\times$ 240 is 1.337s on a RTX 3090 platform. A more-detailed breakdown of computation time for each component is shown in \cref{table: time-system-component}. A further breakdown of computation time on tracking components is shown in \cref{table: time-tracking-component}.
We would like to highlight that our current implementation is prototyped in Python. We believe a real-time system can be achieved by exploiting C++ and further GPU parallelisation.
\begin{table}[tb]
\centering
\subfloat[][System components]{
\label{table: time-system-component}
\begin{tabular}{c|c|c|c}
\hline
Components & Tracking & Integration & Completion (visualization) \\ \hline
Time (s) & 1.284 & 0.003 & 0.474 \\ \hline
\end{tabular}
}
\hfill
\subfloat[][Object tracking components]{
\label{table: time-tracking-component}
\begin{tabular}{c|c|c|c}
\hline
Components & Initialization & Coarse est. & Dense refinement \\ \hline
Time (ms) & 0.643 & 0.150 & 1.129 \\ \hline
\end{tabular}
}
\caption{Run-time analysis (s)}
\vspace{-2em}
\end{table}
\subsection{Qualitative Evaluations}
\begin{figure}[tb]
\centering
\subfloat[][Completion of a red chair]
{\includegraphics[width=0.9\linewidth]
{figures/real/24-09-2021-12-51-15/sitting/sitting-crop.pdf}
}\\
\subfloat[][Completion of a blue chair]
{\includegraphics[width=0.9\linewidth]{figures/real/24-09-2021-12-14-38/completion_blue_chair-crop.pdf}
} \\
\caption{Qualitative comparison of classic TSDF volume representation (gray), DeepSDF shape prior representation (blue), and our conditioned completion representation (green): our representation can reconstruct the observed part more correctly than a shape prior and completes the unseen part where TSDF fusion fails.}
\label{fig:shape_completion}
\vspace{-1em}
\end{figure}
\begin{figure}[tb]
\centering
\includegraphics[width=0.9\linewidth]{figures/real/24-09-2021-12-51-15/moving/moving_chairs-crop.pdf}%
\caption{Segment, track, reconstruct and complete a moving chair. Background pointclouds are just for visualization.}
\label{fig:moving_chairs}
\end{figure}
We further demonstrate our proposed method in various real-world scenarios. \cref{fig:shape_completion} shows the results in two different scenes. For each input image, we provide object reconstructions from the currently estimated camera viewpoint to visualise the observed part and from the top-down viewpoint to visualise the hidden part. As a qualitative comparison, we also show the reconstructions using classic TSDF fusion \cite{Newcombe:etal:ISMAR2011} and the learned category-level DeepSDF object prior \cite{Park:etal:CVPR2019}, it can be seen that TSDF-Fusion can only reconstruct the visible parts, leaving many empty holes in the object models. DeepSDF, on the other hand, has watertight reconstructions, but does not match the measurement necessarily, especially for the objects that deviate from the training space. On the contrary, our system can maintain highly detailed reconstructions and generate watertight meshes by filling in the missing parts using category-level shape priors thanks to the conditioned completion. \cref{fig:moving_chairs} also shows a scene where our system can reconstruct the visible parts and complete the hidden information of a moving object. The object pose and object geometry for the moving object are optimised jointly.
\section{Conclusions}
We present a novel approach for object-level tracking and mapping system in dynamic scenes by incorporating learned category-level shape priors. It enables to reasonably complete the object geometry of unseen parts based on the prior knowledge, and provide more robust and accurate tracking accuracy, even under large frame-by-frame motion and in dynamic environments with moving human involved. Experimental results in various scenarios demonstrate the effectiveness of our method. We hope our method can create a deeper understanding of exploring object prior information in object-level SLAM and benefit robots interacting with their surrounding environments. Continue from here, we would like to extend our system to a full graph-based object SLAM system.
\section*{ACKNOWLEDGMENT}
We thank Xingxing Zuo for fruitful discussions. This research is supported by Imperial College London, Technical University of Munich, and the EPSRC grant ORCA Stream B - Towards Resident Robots. Binbin Xu holds a China Scholarship Council-Imperial Scholarship.
|
\section{Introduction}\label{section-introduction}
Smooth representations of ${\mathrm{GL}}_n(F)$, for a nonarchimedean local field $F$ such as ${\mathbb Q}_p$ or $\mathbb F_p((t))$, is of great interest in number theory, for instance due to the local Langlands correspondence. However, most such representations are infinite-dimensional, so finding a finite measure of their size is important. One such measure is the Gelfand-Kirillov dimension:
\begin{defn}
Let $\pi$ be a (complex) admissible representation of ${\mathrm{GL}}_n(F)$. The \emph{Gelfand-Kirillov dimension} (henceforth, GK-dimension) of $\pi$ is a real number $r\ge0$, denoted $\dim_{GK}(\pi)$, such that
\[
\dim(\pi^{K_N})\sim q^{rN},
\]
where $K_N\colonequals 1+M_n({\mathfrak p}_F^N)$ is a compact open subgroup of ${\mathrm{GL}}_n(F)$ and $q$ is the order of the residue field ${\mathscr O}_F/{\mathfrak p}_F$. Here, for functions $f(N)$ and $g(N)$, denote $f(N)\sim g(N)$ if there exists real numbers $0<a<b$ such that for sufficiently large integers $N>0$, the inequality $af(N)<g(N)<bf(N)$ is satisfied.
\end{defn}
In fact, Harish-Chandra and Howe's local character formula says this function $\dim(\pi^{K_N})$ is a polynomial in $q^{N-1}$ for $N$ large (see Corollary~\ref{fixed-poly}), which we call the \emph{growth polynomial} of $\pi$. Now, $\dim_{GK}(\pi)$ is exactly the degree of the growth polynomial (and, in particular, is an integer). A more ambitious goal would be to calculate the entire growth polynomial. The coefficients of the growth polynomial contain important information about the representation, as they behave well under instances of Langlands functoriality (see Section~\ref{section-functoriality}.)
Irreducible representations of ${\mathrm{GL}}_n(F)$ were classified by Bernstein and Zelevinsky \cite{BZ2} in terms of multisegments---combinatorial data---of supercuspidal representations (those corresponding to irreducible representations under the Langlands correspondence). Roughly, arbitrary representations are obtained by inducing from parabolic subgroups, and supercuspidal representations are those not obtained from such a procedure.
For supercuspidal representations $\rho$ of ${\mathrm{GL}}_n(F)$, it is a result of Howe \cite{howe} that $\dim_{GK}(\rho)=\frac{n(n-1)}2$, and in fact, the leading term of the growth polynomial is known (see Proposition~\ref{cusp-asymp}):
\[\dim(\rho^{K_N})=(1+o(1))[n!]_qq^{\frac{n(n-1)}2(N-1)},\]
where $[n!]_q\colonequals [1]_q\cdots[n]_q=\frac{q-1}{q-1}\cdots\frac{q^n-1}{q-1}$ is the $q$-factorial. This was extended to generic representations by Rodier \cite{rodier2}.
We further extend these results, and calculate the leading term of the growth polynomial for \emph{arbitrary} admissible representations of ${\mathrm{GL}}_n(F)$, based on the combinatorial datum of the Bernstein-Zelevinsky classification \cite{BZ2}:
\begin{thm}\label{intro-thm}
Let $\pi$ be an arbitrary admissible irreducible representation of ${\mathrm{GL}}_n(F)$. Then, there exists a multisegment $a=\{\Delta_1,\dots,\Delta_m\}$ such that $\rho=\langle a\rangle$, where $\Delta_i=[\rho_i,\nu^{r_i-1}\rho_i]$ for supercuspidal representations $\rho_i$ of ${\mathrm{GL}}_{n_i}(F)$. Then,
\[
\dim(\pi^{K_N})=(1+o(1))\frac{[n!]_q}{[r_1!]_{q^{n_1}}\cdots[r_m!]_{q^{n_m}}}(q^{N-1})^{\frac12(n^2-n_1r_1^2-\cdots-n_mr_m^2)},
\]
so in particular,
\[
\dim_{GK}(\pi)=\frac12(n^2-n_1r_1^2-\cdots-n_mr_m^2).
\]
\end{thm}
In fact, we are able to express the growth polynomial of arbitrary admissible representations of ${\mathrm{GL}}_n(F)$ \emph{exactly} in terms of growth polynomials for supercuspidal representations. As a consequence, we deduce that $\dim(\pi^{K_N})$ is an \emph{integer-coefficient} polynomial in $q^{N-1}$ (see Corollary~\ref{int-coeff}), generalizing a result of Howe \cite{howe}.
The difficulty is that although every irreducible representation of $G={\mathrm{GL}}_n(F)$ appears as a subquotient of some induced representation $\nInd_P^G(\rho)$ with $\rho$ a supercuspidal representation of the parabolic subgroup $P\subset G$, such representations may have \emph{multiple} irreducible subquotients, and we can a priori only detect the maximal GK-dimension (see Lemma~\ref{exact-seq}). We identify exactly the irreducible subquotient of $\nInd_P^G(\rho)$ with maximal GK-dimension (see Lemma~\ref{main-lemma}) and for each irreducible representation $\pi$ find a pair $(P,\rho)$ such that $\pi$ has maximal GK-dimension among the irreducible subquotients of $\nInd_P^G(\rho)$. Here, $\rho$ may no longer be supercuspidal, but are products of representation of the form $\langle\Delta\rangle$ for some segment $\Delta$, whose growth polynomial is calculated in Subsection~\ref{subsection-delta}.
The paper is structured as follows: In section~\ref{section-terms} we gather the notation used throughout the paper. In section~\ref{section-arbitrary} we prove the main result, the calculation of the leading term of the growth polynomial for arbitrary irreducible representations of ${\mathrm{GL}}_n(F)$. In section~\ref{section-more-examples} we calculate some examples, corresponding to those multisegments with particularly simple combinatorial structure. In section~\ref{section-local-char} we look at Howe's local character formula, and reproduce a proof that indeed $\dim(\pi^{K_N})$ is a polynomial in $q^{N-1}$. Finally, in section~\ref{section-functoriality} we explore how growth polynomials behave under some instances of Langlands functoriality, such as Jacquet-Langlands, base change, and automorphic induction.
We also briefly consider the growth polynomial for representations of ${\mathrm{SL}}_n(F)$ in appendix~\ref{section-sl-n}, which we deduce from the theory for ${\mathrm{GL}}_n(F)$. In appendix~\ref{appendix-cusp} we provide a more direct proof of $\dim_{GK}(\rho)=\frac{n(n-1)}2$ for $\rho$ supercuspidal, based on Bushnell and Kutzko's classification \cite{Bush-Kutz}. Finally, in appendix~\ref{appendix-level-0} we provide explicit calculations for growth polynomials of level zero representations of ${\mathrm{GL}}_2$ and ${\mathrm{GL}}_3$, based on the techniques developed in appendix~\ref{appendix-cusp}.
\section{Terminology and Conventions}\label{section-terms}
Throughout the paper, $F$ is a nonarchimedean local field of residual characteristic $p$. That is, $F$ is a finite extension of ${\mathbb Q}_p$ or $\mathbb F_p((T))$. Let ${\mathscr O}_F$ denote the ring of integers of $F$, and let ${\mathfrak p}_F\subset{\mathscr O}_F$ denote the unique maximal ideal. Moreover, let $q_F$ denote the size of the residue field $k_F\colonequals{\mathscr O}_F/{\mathfrak p}_F$. Subscripts are dropped whenever $F$ is clear from context.
Note that $F$ is assumed to have characteristic zero in Sections~\ref{section-local-char}~and~\ref{section-functoriality}.
Fix a nontrivial character $\psi\colon F\to{\mathbb C}^\times$ of level zero, i.e., $\psi$ is trivial on ${\mathfrak p}_F$ but not on ${\mathcal O}_F$.
By a \emph{representation} of ${\mathrm{GL}}_n(F)$, we always mean a complex, admissible, smooth representations. That is, we will only consider representations $(\pi,V)$ such that $\pi\colon{\mathrm{GL}}_n(F)\to{\mathrm{GL}}_{\mathbb C}(V)$ is continuous with respect to the discrete topology on ${\mathrm{GL}}_{\mathbb C}(V)$, and such that for any compact open subgroup $K\subset{\mathrm{GL}}_n(F)$ the space of fixed vectors
\[
V^K\colonequals \{v\in V:\pi(k)v=v\text{ for any }k\in K\}
\]
is finite-dimensional. The notation $\pi$ is used for arbitrary admissible smooth representations of ${\mathrm{GL}}_n(F)$, while $\rho$ is reserved for supercuspidal representations. We will denote induction by $\Ind$, compact induction by $\cInd$, and normalized induction by $\nInd$ (i.e., $\nInd_H^G(\pi)\colonequals \Ind_H^G(\delta_G^{1/2}\delta_H^{-1/2}\otimes\pi)$).
Let $\mathcal R$ be the Grothendieck ring of representations of ${\mathrm{GL}}$, i.e., the free abelian group generated by representations of ${\mathrm{GL}}_n$, modulo $\phi-\pi-\sigma$ for any exact sequence
\[
0\to\pi\to\phi\to\sigma\to0
\]
of representations of ${\mathrm{GL}}_n$. It becomes a graded commutative ring by the product defined in Section~\ref{section-arbitrary} (see also \cite[Thm~1.9]{BZ2}).
We define a series of compact open subgroups of ${\mathrm{GL}}_n(F)$, by $K_0\colonequals {\mathrm{GL}}_n({\mathscr O}_F)$, the maximal compact subgroup, and for $N>0$, by
\[
K_N\colonequals 1+M_n({\mathfrak p}_F^N)=\ker(K_0\to{\mathrm{GL}}_n({\mathscr O}_F/{\mathfrak p}_F^N)).
\]
The subgroups $\{K_N\}_{N\ge0}$ form a fundamental system of neighborhoods of $1\in{\mathrm{GL}}_n(F)$.
\begin{defn}
Let $\pi$ be a representation of ${\mathrm{GL}}_n(F)$. Then, the \emph{growth polynomial} of $\pi$ is the polynomial $G_\pi(X)\in{\mathbb C}[X]$ such that for $N\gg0$,
\[
\dim(\pi^{K_N})=G_\pi(q^{N-1}).
\]
Such a polynomial exists by Corollary~\ref{fixed-poly} and Lemma~\ref{exact-seq}. Denote by $\dim_{GK}(\pi)$ the degree of $G_\pi$.
\end{defn}
Finally, let $\mathrm{rec}_F\colon \mathbf{Irr}({\mathrm{GL}}_n(F))\to \mathcal G_n(F)$ be the local Langlands correspondence, where $\mathbf{Irr}({\mathrm{GL}}_n(F))$ is the set of irreducible representations of ${\mathrm{GL}}_n(F)$ and $\mathcal G_n(F)$ is the set of semisimple Deligne representations of the Weil group $\mathcal W_F$ (a close relative to the absolute Galois group $\Gal(\overline F/F)$ of $F$).
\section{Reduction to supercuspidal representations}\label{section-arbitrary}
Notation and terminology in this section mostly follows \cite{BZ2}, so $\nu\colon F^\times\to{\mathbb C}^\times$ is the character $\nu(g)\colonequals|\det(g)|$, and a \emph{segment} is a set of supercuspidal representations of the form $[\rho,\nu^{r-1}\rho]\colonequals\{\rho,\nu\rho,\dots,\nu^{r-1}\rho\}$ for some supercuspidal representation $\rho$ and integer $r>0$. By \cite[Thm~6.1]{BZ2} irreducible representations of ${\mathrm{GL}}_n$ are classified by \emph{multisegments}---multisets of segments.
A key construction in Bernstein and Zelevinsky's classification is inducing from parabolic subgroups $P\subset G$. Conjugacy classes of $P$ correspond to partitions of $n$. Indeed, a partition $n=n_1+\cdots+n_r$, denoted $\lambda$, corresponds to
\[
P_\lambda\colonequals \begin{pmatrix}
{\mathrm{GL}}_{n_1}&*&&*\\
&{\mathrm{GL}}_{n_2}&*\\
&&\ddots&*\\
0&&&{\mathrm{GL}}_{n_r}
\end{pmatrix}\subset G.
\]
Let $\pi_1,\dots,\pi_r$ be admissible representations of ${\mathrm{GL}}_{n_1},\dots,{\mathrm{GL}}_{n_r}$, respectively. Following the notation in \cite{BZ2}, we denote
\[
\pi_1\times\cdots\times\pi_r\colonequals \nInd_{P_\lambda}^G(\pi_1\boxtimes\cdots\boxtimes\pi_r),
\]
where $\pi_1\boxtimes\cdots\boxtimes\pi_r$ is inflated via the surjection $P_\lambda\to{\mathrm{GL}}_{n_1}\times\cdots\times{\mathrm{GL}}_{n_r}$, which quotients out the unipotent radical $N_\lambda\subset P_\lambda$.
The following lemma on the size of the double cosets $P_\lambda\backslash G/K_N$ is useful:
\begin{lemma}\label{parabolic-size}
Let $\lambda$ be the partition $n=n_1+\cdots+n_r$. Then,
\[
|P_\lambda\backslash G/K_N|=\begin{bmatrix}n\\ n_1,\cdots,n_r\end{bmatrix}_q(q^{N-1})^{(n^2-n_1^2-\cdots-n_r^2)/2}
\]
for $N>0$, where $\begin{bmatrix}n\\n_1,\dots,n_r\end{bmatrix}_q\colonequals \frac{[n!]_q}{[n_1!]_q\cdots[n_r!]_q}$, and $[n!]_q$ is the $q$-analogue of $n!$, defined in Section~\ref{section-introduction}.
\end{lemma}
\begin{proof}
Since $G=P_\lambda K_0$, we have
\[
P_\lambda\backslash G/K_N\cong P_\lambda({\mathscr O})\backslash K_0/K_N\cong P_\lambda({\mathscr O}/{\mathfrak p}^N)\backslash{\mathrm{GL}}_n({\mathscr O}/{\mathfrak p}^N),
\]
the partial flag variety over ${\mathscr O}/{\mathfrak p}^N$ corresponding to the partition $\lambda$. Now, the subgroup
\[
\begin{pmatrix}
I_{n_1}&0&\cdots&0\\
{\mathfrak p}/{\mathfrak p}^N&I_{n_2}&&\vdots\\
\vdots&&\ddots&0\\
{\mathfrak p}/{\mathfrak p}^N&\cdots&{\mathfrak p}/{\mathfrak p}^N&I_{n_r}
\end{pmatrix}\subset {\mathrm{GL}}_n({\mathscr O}/{\mathfrak p}^N
\]
acts simply on $P_\lambda({\mathscr O}/{\mathfrak p}^N)\backslash{\mathrm{GL}}_n({\mathscr O}/{\mathfrak p}^N)$ from the right, with quotient $P_\lambda(k)\backslash{\mathrm{GL}}_n(k)$, so
\begin{align*}
|P_\lambda\backslash G/K_N|&=|P_\lambda(k)\backslash{\mathrm{GL}}_n(k)|\cdot (q^{N-1})^{(n^2-n_1^2-\cdots-n_r^2)/2}\\
&=\begin{bmatrix}n\\ n_1,\cdots,n_r\end{bmatrix}_q(q^{N-1})^{(n^2-n_1^2-\cdots-n_r^2)/2}.
\end{align*}
Here, the last equality is because the quotient $P_\lambda(k)\backslash{\mathrm{GL}}_n(k)$ is the partial flag variety over $k$, and the projective space ${\mathbb P}_k^{n-1}$ has size $[n]_q=\frac{q^n-1}{q-1}$.
\end{proof}
\begin{rmk}
Here, $\dim(N_\lambda)=\frac12(n^2-n_1^2-\cdots-n_r^2)$, where $N_\lambda\subset P_\lambda$ is the nilpotent radical.
\end{rmk}
The following Mackey-type formula is also useful, and will be used repeatedly throughout the paper:
\begin{lemma}\label{fixed}
Let $G$ be a locally profinite group, $H\subset G$ a closed subgroup, and $K\subset G$ a compact open subgroup. Let $S\subset G$ be a set of representatives for $H\backslash G/K$. Then, for a smooth representation $\lambda$ of $H$,
\begin{align*}
(\cInd_H^G\lambda)^K&\xrightarrow{\sim}\bigoplus_{g\in S}\lambda^{H\cap gKg^{-1}}\\
f&\mapsto (f(g))_{g\in S}.
\end{align*}
\end{lemma}
\begin{proof}
We will first show well-definedness. Let $f\in(\cInd_H^G\lambda)^K$. For each $g\in S$ and $h\in H\cap gKg^{-1}$, we have:
\[
f(g)=f(h\cdot g\cdot g^{-1}h^{-1}g)=\lambda(h)f(g),
\]
so $f(g)\in\lambda^{H\cap gKg^{-1}}$. Moreover, there is a subset $T\subset S$ such that $\supp f=\sqcup_{g\in T}HgK$, where each $HgK$ is open (it is a union of $K$-cosets). Thus, \[H\backslash\supp f=\bigsqcup_{g\in T}H\backslash HgK,\] so since $H\backslash\supp f$ is compact, $T\subset S$ must be finite. That is, the image lies in the direct sum.
We will construct an inverse. Let $\mathbf a=(a_g)_{g\in S}\in\bigoplus_{g\in S}\lambda^{H\cap gKg^{-1}}$ be arbitrary. For any $u\in G=HSK$, there is a $h\in H$, $g\in S$, and $k\in K$ such that $u=hgk$. Let $f_{\mathbf a}(u)\colonequals \lambda(h)a_g$. We claim $f_{\mathbf a}\colon G\to\lambda$ is well-defined. Indeed, if $u=hgk=h'gk'$, then
\[
\lambda(h)a_g=\lambda(h')\lambda((h')^{-1}h)a_g=\lambda(h')a_g,
\]
since $(h')^{-1}h=gk'k^{-1}g^{-1}\in H\cap gKg^{-1}$. Now, $f_{\mathbf a}\in(\Ind_H^G\lambda)^K$ is clear. Finally, $f_{\mathbf a}$ is compactly supported modulo $H$, since
\[
H\backslash \supp f=\bigcup_{g\in T}H\backslash HgK
\]
where $T=\{g\in S:a_g\ne 0\}$ is finite, and each $H\backslash HgK$ is the continuous image of the compact set $K$.
\end{proof}
Now, the growth polynomials (defined in Section~\ref{section-terms}) of induced representations of the form $\pi_1\times\cdots\times\pi_r$ are readily calculated:
\begin{lemma}\label{induced-dim}
Let $n=n_1+\cdots+n_r$ be a partition and let $\pi_1,\dots,\pi_r$ be admissible representations of ${\mathrm{GL}}_{n_1},\dots,{\mathrm{GL}}_{n_r}$, respectively. Then,
\[
G_{\pi_1\times\cdots\times\pi_r}(X)=\begin{bmatrix}n\\ n_1,\cdots,n_r\end{bmatrix}_qX^{(n^2-n_1^2-\cdots-n_r^2)/2}\prod_{i=1}^rG_{\pi_i}(X),
\]
so in particular,
\[
\dim_{GK}(\pi_1\times\cdots\times\pi_r)=\frac12(n^2-n_1^2-\cdots-n_r^2)+\sum_{i=1}^r\dim_{GK}(\pi_i).
\]
\end{lemma}
\begin{proof
Let $\pi=\pi_1\times\cdots\times\pi_r$. The normalization may be ignored since the character $\delta_G^{1/2}\delta_P^{-1/2}$ is trivial on a sufficiently small neighborhood of $P$. By Lemma~\ref{fixed} we have
\[
\rho^{K_N}=\bigoplus_{g\in P\backslash G/K_N}(\rho_1\boxtimes\cdots\boxtimes\rho_r)^{P(F)\cap gK_Ng^{-1}}.
\]
Now, by Iwasawa decomposition $G=BK_0=PK_0$, so the representatives $g\in P\backslash G/K_N$ can be chosen so that $g\in K_0$. Then, $gK_Ng^{-1}=K_N$, so
\begin{align*}
\pi^{K_N}&=\big((\pi_1\boxtimes\cdots\boxtimes\pi_r)^{P(F)\cap K_N}\big)^{\oplus|P\backslash G/K_N|}\\
&=\big((\pi_1\boxtimes\cdots\boxtimes\pi_r)^{K_N\times\cdots\times K_N}\big)^{\oplus|P\backslash G/K_N|}\\
&=\big(\pi_1^{K_N}\otimes\cdots\otimes\pi_r^{K_N}\big)^{\oplus|P\backslash G/K_N|}.
\end{align*}
Now the result follows from Lemma~\ref{parabolic-size}.
\end{proof}
Moreover, growth polynomials behave well with short exact sequences:
\begin{lemma}\label{exact-seq}
Let
\[
0\to\pi\to\phi\to\sigma\to0
\]
be an exact sequence of admissible representations of $G$. Then,
\[
G_\phi(X)=G_\pi(X)+G_\sigma(X),
\]
so in particular, $\dim_{GK}(\phi)=\max\{\dim_{GK}(\pi),\dim_{GK}(\sigma)\}$.
\end{lemma}
\begin{proof}
Follows immediately from \cite[Section~2.3]{Bush-Henn}~Corollary~1, that $(-)^{K_N}$ is an exact functor.
\end{proof}
\begin{rmk}
A nice formalization of Lemma~\ref{induced-dim} and Lemma~\ref{exact-seq}, together with Corollary~\ref{fixed-poly} is the following: the map from the Grothendieck ring of representations of ${\mathrm{GL}}$
\begin{align*}
\mathscr I\colon\mathcal R&\to{\mathbb C}[X,X^{-1}]\\
\pi\in\mathrm{Rep}({\mathrm{GL}}_n)&\mapsto [n!]_q^{-1}X^{-\frac{n^2-n}2}G_\pi(X)
\end{align*}
is a well-defined ring homomorphism. That is, $\mathscr I$ normalizes $\dim(\pi^{K_N})$ by the size of the full flag variety over ${\mathscr O}/{\mathfrak p}^N$.
\end{rmk}
\subsection{The growth polynomial of $\langle\Delta\rangle$}\label{subsection-delta} We calculate the growth polynomial for representations of the form $\langle\Delta\rangle$ for some segment $\Delta$, which is of particular importance since they generate the Grothendieck ring $\mathcal R$ \cite[Cor~7.5]{BZ2}. Thus, growth polynomial for arbitrary representations can be deduced from that of $\langle\Delta\rangle$, via Lemma~\ref{induced-dim} and Lemma~\ref{exact-seq}, as carried out in the next subsection.
We first review the theory of Bernstein-Zelevinsky derivatives, as defined in \cite[Section~3.5]{BZ1}:
\begin{defn}
Let $M_n\subset{\mathrm{GL}}_n$ be the \emph{mirabolic subgroup}, the group of matrices whose last row is $(0,\cdots,0,1)$. For an admissible representation $(\tau,U)$ of $M_n$ and $0<i\le n$, define the \emph{$i$-th derivative} $\tau^{(i)}$ of $\tau$, which is a representation of ${\mathrm{GL}}_{n-i}$, as
\[
\tau^{(k)}\colonequals U/(\theta_i(h)v-\pi(h)v:v\in U,h\in N_i),
\]
where $N_i$ consists of elements of the form
\begin{comment}\[
h=\left(\begin{array}{@{}c|c@{}}
\begin{matrix}1&\\
&\ddots\\
&1\end{matrix}
& \begin{matrix}
* & * &*\\
* & *&*\\
v_{n-i}&*&*
\end{matrix} \\
\hline
\begin{matrix}
{} & {} & {}\\
{} & {} & {}\\
{} & {} & {}
\end{matrix} &
\begin{matrix}
1 & v_{n-i+1}& *\\
&\ddots&v_{n-1}\\
&& 1
\end{matrix}
\end{array}\right)
\]
\end{comment}
\[
h=\left(\begin{array}{@{}c|c@{}}
\begin{matrix}\begin{matrix}
1 \\
&\ddots &\\
&& 1
\end{matrix}\\ \hline
\begin{matrix}
&\\
&\\
&
\end{matrix}\end{matrix} &
\begin{matrix}
* & * &*\\
\vdots & \vdots&\vdots\\
a_{n-i}&*&*\\
1 & a_{n-i+1}& *\\
&\ddots&a_{n-1}\\
&& 1
\end{matrix}
\end{array}\right)
\]
and $\theta_i(h)\colonequals\psi(a_{n-i+1}+\cdots+a_{n-1})$, where $\psi$ is the fixed character of $F$ as defined in Section~\ref{section-terms}. Here, the top-left block has size $(n-i)\times(n-i)$, and the bottom-right block has size $i\times i$. Moreover, for representations of ${\mathrm{GL}}_n$ define the $i$-th derivative as the $i$-th derivative of the restriction to $M_n\subset{\mathrm{GL}}_n$.
\end{defn}
\begin{defn}
Let $\pi$ be a representation of ${\mathrm{GL}}_n(F)$. Then, let the \emph{total derivative} be the formal sum in the Grothendieck ring
\[
\mathscr{D}(\pi)\colonequals\pi+\pi^{(1)}+\cdots+\pi^{(n)}\in\mathcal R.
\]
Extending $\mathscr D$ to $\mathcal R$ by linearity defines a ring homomorphism $\mathcal R\to\mathcal R$, by \cite[Lem~4.5]{BZ1}.
\end{defn}
Now, the following calculations of total derivatives are of particular interest to us:
\begin{lemma}[Bernstein-Zelevinsky~{\cite[Cor~4.14~(a)]{BZ1}}]\label{der-calc1}
Let $\pi$ be a representation of ${\mathrm{GL}}_m$, and let $\varphi$ be a representation of the parabolic subgroup $P_{(n-1)+1}\subset{\mathrm{GL}}_n$. Then, the derivatives of
\[
\sigma\colonequals\nInd_{P_{m+(n-1)+1}}^{P_{(m+n-1)+1}}(\pi\boxtimes\varphi)
\]
are given by, for $i>0$,
\[
\sigma^{(i)}=\pi\times\varphi^{(i)}.
\]
\end{lemma}
\begin{lemma}[Zelevinsky~{\cite[Thm~3.5]{BZ2}}]\label{der-calc2}
Let $\Delta=[\rho,\nu^{r-1}\rho]$ be a segment, where $\rho$ is a supercuspidal representation. Then,
\[
\mathscr{D}(\langle\Delta\rangle)=\langle\Delta\rangle+\langle\Delta^-\rangle,
\]
where $\Delta^-\colonequals[\rho,\nu^{r-2}\rho]$.
\end{lemma}
Now, we calculate the growth polynomial for representations of the form $\langle\Delta\rangle$ for some segment $\Delta$, heavily utilizing the fact that they remain irreducible when restricted to $M_n$ \cite[Rmk~3.6]{BZ2}:
\begin{prop}\label{delta-dim}
Let $\rho$ be a supercuspidal representation of ${\mathrm{GL}}_{n_1}(F)$, and let $\Delta=[\rho,\nu^{r-1}\rho]$, so $\langle\Delta\rangle$ is a representation of ${\mathrm{GL}}_n(F)$, with $n=n_1r$. Then,
\begin{align*}
G_{\langle\Delta\rangle}(X)&=[r!]_{q^{n_1}}^{-1}\begin{bmatrix}n\\n_1,\dots,n_1\end{bmatrix}_qX^{\frac{n_1(n_1-1)r(r-1)}2}G_\rho(X)^r\\
&=\frac{[n!]_q}{[r!]_{q^{n_1}}}X^{\frac{n(n-r)}2}+(\text{lower order terms}),
\end{align*}
so in particular,
\[
\dim_{GK}\langle\Delta\rangle=\frac{n(n-r)}2.
\]
\end{prop}
\begin{proof}
Throughout the proof, we abbreviate the parabolic subgroup $P_\lambda$ as the partition $\lambda$.
There is a $M_n$-homomorphism
\begin{align*}
i\colon \langle\Delta\rangle&\hookrightarrow\langle\Delta^-\rangle\times\nu^{r-1}\rho=\nInd_{(n-n_1)+n_1}^{n}(\langle\Delta^-\rangle\boxtimes\nu^{r-1}\rho)\\&\xrightarrow{res}\sigma\colonequals\nInd_{(n-n_1)+(n_1-1)+1}^{(n-1)+1}(\langle\Delta^-\rangle\boxtimes\nu^{r-1}\rho|_{(n_1-1)+1}),
\end{align*}
where the inclusion $\langle\Delta\rangle\hookrightarrow\langle\Delta^-\rangle\times\nu^{r-1}\rho$ is the $G$-homomorphism defined by the characterization of $\langle\Delta\rangle$ as the unique irreducible submodule of $\rho\times\cdots\times\nu^{r-1}\rho$ \cite[Section~3.1]{BZ2}, and $res$ is the surjection given by restricting functions $f\colon G\to\langle\Delta^-\rangle\boxtimes\nu^{r-1}\rho$ to $P_{(n-1)+1}$. We calculate their derivatives using Lemma~\ref{der-calc1} and Lemma~\ref{der-calc2}:
\begin{align*}
\mathscr{D}(\langle\Delta\rangle)&=\langle\Delta^-\rangle+\langle\Delta\rangle\\
\mathscr{D}(\langle\Delta^-\rangle\times\nu^{r-1}\rho)&=\langle\Delta^{--}\rangle+\langle\Delta^-\rangle+\langle\Delta^{--}\rangle\times\nu^{r-1}\rho+\langle\Delta^-\rangle\times\nu^{r-1}\rho\\
\mathscr{D}(\sigma)&=\langle\Delta^-\rangle+\sigma.
\end{align*}
Here $i$ is nonzero, since both $\langle\Delta\rangle$ and $\sigma$ have $n_1$-th derivative $\langle\Delta^-\rangle$, but the $n_1$-th derivative $\langle\Delta^-\rangle+\langle\Delta^{--}\rangle\times\nu^{r-1}\rho$ of $\langle\Delta^-\rangle\times\nu^{r-1}\rho$ only contains the irreducible representation $\langle\Delta^-\rangle$ with multiplicity one. Now, since $\langle\Delta\rangle$ is an irreducible $M_n$-representation by \cite[Cor~7.9]{BZ2}, the map $i$ must be injective.
Moreover, if $\tau$ is the cokernel of $i$, then its total derivative is
\begin{align*}
\mathscr D(\tau)&=\mathscr{D}(\sigma)-\mathscr{D}(\langle\Delta\rangle)\\
&=\sigma-\langle\Delta\rangle\\
&=\tau.
\end{align*}
That is, $\tau$ has no derivatives of degree $>0$, so by \cite[Section~3.5]{BZ2} we have $\tau=0$. Thus, $i$ is an isomorphism of $M_n$-representations.
The overall strategy is as follows: Use the isomorphism $i$ to calculate the dimension of $\langle\Delta\rangle^{K_N\cap M_n}\supseteq \langle\Delta\rangle^{K_N}$. In fact, we can show \[\langle\Delta\rangle^{K_N\cap M_n}=\langle\Delta\rangle^{K_N\cap{\mathrm{GL}}_{n-1}}=\langle\Delta\rangle^{K_N\cap {^t}M_n},\]
so
\[
\langle\Delta\rangle^{K_N}=\langle\Delta\rangle^{K_N\cap M_n}\cap \langle\Delta\rangle^{K_N\cap {^t}M_n}=\langle\Delta\rangle^{K_N\cap M_n},
\]
which allows us to calculate $\dim(\langle\Delta\rangle^{K_N})$.
First, for $N_1\ge N$ let
\[
J_{N_1}\colonequals\left(\begin{array}{@{}c|c@{}}
\begin{matrix}
\\
K_N\cap{\mathrm{GL}}_{n-1}\\
{}
\end{matrix} &
\begin{matrix}
{\mathfrak p}^{N_1}\\
\vdots\\
{\mathfrak p}^{N_1}
\end{matrix}\\\hline
0\quad \cdots\quad 0&1
\end{array}\right)\subset M_n
\]
be compact open subgroups which approximate the subgroup $K_N\cap{\mathrm{GL}}_{n-1}$ of $M_n$, with $J_N=K_N\cap M_n$. Here,
\begin{equation}\label{symmetric-fixed}
\langle\Delta\rangle^{K_N\cap{\mathrm{GL}}_{n-1}}=\bigcup_{N_1\ge N}\langle\Delta\rangle^{J_{N_1}},
\end{equation}
where $\supseteq$ is obvious and $\subseteq$ is from smoothness of the representation $\langle\Delta\rangle|_{M_n}$. This allows us to use Lemma~\ref{fixed} on the groups $J_{N_1}$.
For each $N_1\ge N$ we have
\begin{equation}\label{double-coset-calc}
P_{(n-n_1)+(n_1-1)+1}\backslash P_{(n-1)+1}/J_{N_1}\cong P_{(n-n_1)+(n_1-1)}\backslash {\mathrm{GL}}_{n-1}/K_N\cap{\mathrm{GL}}_{n-1}
\end{equation}
has size $\begin{bmatrix}n-1\\n_1-1\end{bmatrix}_qq^{n(n_1-1)(N-1)}$ by Lemma~\ref{parabolic-size}, which notably does not depend on the choice of $N_1$. Since $i$ is a $M_n$-isomorphism,
\begin{align}\label{integer-coeff-eq}\dim(\langle\Delta\rangle^{J_{N_1}})&=\dim(\sigma^{J_{N_1}})\nonumber\\
&=|P_{(n-n_1)+(n_1-1)+1}\backslash P_{(n-1)+1}/J_{N_1}|\dim\big(\langle\Delta^-\rangle^{K_N}\otimes(\nu^{r-1}\rho)^{K_N}\big)\\
&=\begin{bmatrix}n-1\\n_1-1\end{bmatrix}_qq^{n(n_1-1)(N-1)}\dim(\langle\Delta^-\rangle^{K_N})\dim(\rho^{K_N}),\nonumber\end{align}
where the second equality is by a similar argument as in Lemma~\ref{induced-dim}. By the independence on the choice of $N_1$, the chain of inclusions
\[
\langle\Delta\rangle^{J_N}\subseteq \langle\Delta\rangle^{J_{N+1}}\subseteq\langle\Delta\rangle^{J_{N+2}}\subseteq\cdots
\]
must all be equalities, so in particular, \eqref{symmetric-fixed} tells us that $\langle\Delta\rangle^{K_N\cap M_N}=\langle\Delta\rangle^{K_N\cap {\mathrm{GL}}_{n-1}}$.
On the other hand, by \cite[Thm~7.3]{BZ} we have $\langle\Delta\rangle\circ\iota\cong\langle\Delta\rangle^\vee\cong\langle[\nu^{1-r}\rho^\vee,\rho^\vee]\rangle$, where $\iota(g)\colonequals {^t}g^{-1}$ is the involution on ${\mathrm{GL}}_n(F)$ corresponding to the nontrivial automorphism of the Dynkin diagram $A_{n-1}$. Thus, for any subgroup $H\subset M_n$ we have
\[
\langle\Delta\rangle^{{^t}H}\cong(\langle\Delta\rangle^\vee)^H.
\]
Symmetrically, all the same arguments, with $\rho$ replaced by $\nu^{1-r}\rho^\vee$, tells us that $\langle\Delta\rangle^{K_N\cap {^t}M_N}=\langle\Delta\rangle^{K_N\cap {\mathrm{GL}}_{n-1}}$.
Since the groups $K_N\cap M_n$ and $K_N\cap {^t}M_n$ generate $K_N$,
\[
\langle\Delta\rangle^{K_N}=\langle\Delta\rangle^{K_N\cap M_n}\cap\langle\Delta\rangle^{K_N\cap {^t}M_n}=\langle\Delta\rangle^{K_N\cap M_n}.
\]
But this dimension was calculated in equation~\ref{integer-coeff-eq}, thus giving us a recursive formula for $\langle\Delta\rangle^{K_N}$.
By downward induction on $r$, and observing that
\[
\begin{bmatrix}n-1\\n_1-1\end{bmatrix}_q=\frac{[n_1]_q}{[n]_q}\cdot\begin{bmatrix}n\\n_1\end{bmatrix}_q=[r]_{q^{n_1}}^{-1}\begin{bmatrix}n_1r\\n_1\end{bmatrix}_q,
\]
we obtain
\[
\dim(\langle\Delta\rangle^{K_N})=[r!]_{q^{n_1}}^{-1}\begin{bmatrix}n\\n_1,\dots,n_1\end{bmatrix}_q(q^{N-1})^{\frac{n_1(n_1-1)r(r-1)}{2}}\dim(\rho^{K_N})^r.
\]
The final asymptotic is by Proposition~\ref{cusp-asymp}.
\end{proof}
\begin{comment}
\begin{rmk}
In fact, since the Grothendieck ring $\mathcal R$ is generated by representations of the form $\langle\Delta\rangle$ (see \cite[Cor~7.5]{BZ2}), a combination of Lemma~\ref{induced-dim} and Lemma~\ref{exact-seq} gives an explicit formula for $G_\pi(X)$, for any representation $\pi$, in terms of $G_\rho$ for supercuspidal $\rho$. Such a formula is calculated explicitly in a special case in Proposition~\ref{example-prop}.
\end{rmk}
\end{comment}
\begin{rmk}
In the notation of the remark after Lemma~\ref{exact-seq}, Proposition~\ref{delta-dim} becomes:
\[
\mathscr I_{\langle\Delta\rangle}(X)=[r!]_{q^{n_1}}^{-1}X^{-\frac{r-1}2n}\mathscr I_\rho(X)^r.
\]
\end{rmk}
As a corollary, we generalize a result of Howe \cite{howe} that $G_\rho(X)\in{\mathbb Z}[X]$, for $\rho$ supercuspidal:
\begin{cor}\label{int-coeff}
Let $\pi$ be an arbitrary admissible representation of ${\mathrm{GL}}_n(F)$. Then, the growth polynomial $G_\pi(X)$ has integer coefficients.
\end{cor}
\begin{proof}
The statement holds for $\pi$ supercuspidal, by Corollary~\ref{fixed-poly} and \cite[Thm~3]{howe}, which states $c_{\mathcal O}(\pi)\in{\mathbb Z}$ for any nilpotent orbit ${\mathcal O}$.
The general case follows from Proposition~\ref{delta-dim}, Lemma~\ref{induced-dim}, and the fact that the Grothendieck ring $\mathcal R$ of representations of ${\mathrm{GL}}$ is generated as a ring by representations of the form $\langle\Delta\rangle$ (see \cite[Cor~7.5]{BZ2}). That $[r!]_{q^{n_1}}^{-1}\begin{bmatrix}n\\n_1,\dots,n_1\end{bmatrix}_q$ in Proposition~\ref{delta-dim} is an integer is more readily seen from the recursion~\eqref{integer-coeff-eq}.
\end{proof}
\begin{rmk}
The weaker statement, that $G_\pi(X)\in{\mathbb Q}[X]$, is immediate since $\dim(\pi^{K_N})$ is always an integer.
\end{rmk}
\subsection{Growth polynomial of arbitrary admissible representations} Although Proposition~\ref{delta-dim} allows for the calculation of growth polynomials for arbitrary representations, explicitly expressing representations in terms of $\langle\Delta\rangle$ is difficult in general. Nonetheless, Theorem~\ref{main-thm} calculates the leading term in terms of the Bernstein-Zelevinsky classification.
The following lemma allows us to relate the leading term of the growth polynomial of the representation $\langle a\rangle$ to that of the easier-understood representation $\pi(a)$.
\begin{lemma}\label{main-lemma}
For any multiset $a$ of segments,
\[
\dim(\langle a\rangle^{K_N})=(1+o(1))\dim(\pi(a)^{K_N})
\]
so in particular, $\dim_{GK}(\pi(a))=\dim_{GK}(\langle a\rangle)$.
\end{lemma}
\begin{proof}
First of all, we claim that if $a<b$ (i.e., $a$ can be obtained from $b$ via a chain of elementary operations) then $\dim_{GK}\pi(a)<\dim_{GK}\pi(b)$. Indeed, by Lemma~\ref{induced-dim} it suffices to check that for linked segments $\Delta$ and $\Delta'$,
\[
\dim(\langle\Delta\cap\Delta'\rangle\times\langle\Delta\cup\Delta'\rangle)<\dim(\langle\Delta\rangle\times\langle\Delta'\rangle).
\]
Let $\Delta=[\rho,\nu^{r-1}\rho]$ and $\Delta'=[\nu^s\rho,\nu^{s+t-1}\rho]$ for some $0<s\le r<s+t$ and $\rho$ a supercuspidal representation of ${\mathrm{GL}}_{n_1}(F)$. We hope to show
\[
\dim_{GK}(\langle\Delta\cap\Delta'\rangle\times\langle\Delta\cup\Delta'\rangle)=(t+s)(r-s)n_1^2+((t+s)^2+(r-s)^2)\frac{n_1(n_1-1)}2.
\]
is less than
\[
\dim_{GK}(\langle\Delta\rangle\times\langle\Delta'\rangle)=rtn_1^2+(r^2+t^2)\frac{n_1(n_1-1)}2.
\]
This reduces to the fact that $(t+s)^2+(t-s)^2>r^2+t^2$, i.e., the convexity of $x^2$.
Now by \cite[Thm~7.1]{BZ2}, in the Grothendieck ring $\mathcal R$ of representations of ${\mathrm{GL}}$ we have
\[
\pi(a)=\langle a\rangle+\sum_{b<a}m(b;a)\langle b\rangle
\]
for some nonnegative integers $m(b;a)$. For $b<a$ we have $\dim_{GK}\langle b\rangle\le \dim_{GK}\pi(b)<\dim_{GK}\pi(a)$, so
\[
\dim(\pi(a)^{K_N})=\dim(\langle a\rangle^{K_N})+\sum_{b<a}m(b;a)\dim(\langle b\rangle^{K_N})=(1+o(1))\dim(\langle a\rangle^{K_N}).\qedhere
\]\end{proof}
We finally arrive at our main theorem:
\begin{thm}\label{main-thm}
Let $\pi$ be an arbitrary admissible irreducible representation of ${\mathrm{GL}}_n(F)$. Then, there exists a multisegment $a=\{\Delta_1,\dots,\Delta_m\}$ such that $\rho=\langle a\rangle$, where $\Delta_i=[\rho_i,\nu^{r_i-1}\rho_i]$ for supercuspidal representations $\rho_i$ of ${\mathrm{GL}}_{n_i}(F)$. Then,
\[
G_\pi(X)=\frac{[n!]_q}{[r_1!]_{q^{n_1}}\cdots[r_m!]_{q^{n_m}}}X^{\frac12(n^2-n_1r_1^2-\cdots-n_mr_m^2)}+(\text{lower order terms}),
\]
so in particular, $\dim_{GK}(\pi)=\frac12(n^2-n_1r_1^2-\cdots-n_mr_m^2)$.
\end{thm}
\begin{proof}
Combining Lemma~\ref{induced-dim}, Lemma~\ref{main-lemma}, and Proposition~\ref{delta-dim}, we obtain:
\begin{align*}
\dim(\pi^{K_N})&=(1+o(1))\dim((\langle\Delta_1\rangle\times\cdots\times\langle\Delta_m\rangle)^{K_N})\\
&=(1+o(1))\begin{bmatrix}n\\n_1r_1,\dots,n_mr_m\end{bmatrix}_q(q^{N-1})^{\frac{n^2-n_1^2r_1^2-\cdots-n_m^2r_m^2}2}\prod_{i=1}^m\dim(\langle\Delta_i\rangle^{K_N})\\
&=(1+o(1))\frac{[n!]_q}{[r_1!]_{q^{n_1}}\cdots[r_m!]_{q^{n_m}}}(q^{N-1})^{\frac12(n^2-n_1r_1^2-\cdots-n_mr_m^2)}.\qedhere
\end{align*}
\end{proof}
\begin{example}\label{bz-example}
Consider representations with support $\{\rho,\nu\rho,\nu\rho,\nu^2\rho\}$, as in \cite[Ex~11.4]{BZ2}, with $\rho$ a supercuspidal representation of ${\mathrm{GL}}_{n_1}(F)$. Equivalently, they are irreducible subquotients of $\rho\times\nu\rho\times\nu\rho\times\nu^2\rho$ \cite[Thm~7.1]{BZ2}. Then, the Hasse diagram for multisegments supported on it is (where the partial order is induced from elementary operations):
\begin{center}\begin{tikzpicture}[scale=.7]
\node (max) at (0,2) {$\{\rho,\nu\rho,\nu\rho,\nu^2\rho\}$};
\node (l) at (-3,0) {$\{[\rho,\nu\rho],\nu\rho,\nu^2\rho\}$};
\node (r) at (3,0) {$\{\rho,\nu\rho,[\nu\rho,\nu^2\rho]\}$};
\node (zero) at (0,-2) {$\{[\rho,\nu\rho],[\nu\rho,\nu^2\rho]\}$};
\node (min) at (0,-4) {$\{[\rho,\nu^2\rho],\nu\rho\}$};
\draw (l)--(max)--(r);
\draw (l)--(zero)--(r);
\draw (zero)--(min);
\end{tikzpicture}\end{center}
and the corresponding GK-dimensions of $\langle a\rangle$ (which, due to Lemma~\ref{main-lemma} and its proof respects the Hasse diagram) are:
\begin{center}\begin{tikzpicture}[scale=.7]
\node (max) at (0,2) {$8n_1^2-2n_1$};
\node (l) at (-3,0) {$8n_1^2-3n_1$};
\node (r) at (3,0) {$8n_1^2-3n_1$};
\node (zero) at (0,-2) {$8n_1^2-4n_1$};
\node (min) at (0,-4) {$8n_1^2-5n_1$.};
\draw (l)--(max)--(r);
\draw (l)--(zero)--(r);
\draw (zero)--(min);
\end{tikzpicture}\end{center}
\end{example}
\begin{example}[Generalized Steinberg representations]
For a supercuspidal representation $\rho$ of ${\mathrm{GL}}_{n_1}$ let $\Delta=[\rho,\nu^{r-1}\rho]$, and consider the generalized Steinberg representation $\langle\Delta\rangle^t=\langle\rho,\dots,\nu^{r-1}\rho\rangle$ of ${\mathrm{GL}}_n$. Then, by Theorem~\ref{main-thm},
\[
G_{\langle\Delta\rangle^t}(X)=[n!]_qX^{\frac12(n-1)n}+(\text{lower order terms}).
\]
Proposition~\ref{example-prop} gives a more explicit formula.
\end{example}
\begin{example}[Generic representations]
More generally, if $\pi$ is a generic representation of ${\mathrm{GL}}_n(F)$, then
\[
G_\pi(X)=[n!]_qX^{\frac12(n-1)n}+(\text{lower order terms}).
\]
so the asymptotic formula is independent of the supercuspidal support. In fact, generic representations are characterized by being maximal (with respect to the order induced by elementary operations) in the poset of representations with a fixed supercuspidal support \cite[Thm~9.7]{BZ2}. Thus, $\dim_{GK}(\pi)=\frac{n^2-n}2$ if and only if $\pi$ is generic, recovering Rodier's result \cite{rodier2}.
\end{example}
\section{More examples}\label{section-more-examples}
In this section, we will only consider $\rho$-rigid representations, for some fixed supercuspidal representation $\rho$:
\begin{defn}
An irreducible representation $\pi$ of ${\mathrm{GL}}_n(F)$ is \emph{$\rho$-rigid} for some supercuspidal representation $\rho$ if the supercuspidal support of $\pi$ only contains representations of the form $\nu^i\rho$ for some $i\in{\mathbb Z}$.
\end{defn}
Only considering such representations suffice, by \cite[Prop~8.6]{BZ2} and Lemma~\ref{induced-dim}.
An explicit formula for the growth polynomial is available in some cases, when the structure of the lattice of multisegments is simple:
\begin{prop}\label{example-prop}
Let a $\rho$-rigid multiset of segments $a$ with support of size $s$ be such that any two of its segments have an empty intersection, where $\rho$ is a supercuspidal representation of ${\mathrm{GL}}_{n_1}(F)$ (so $n=n_1s$). Then,
\[
G_{\langle a\rangle}(X)=\begin{bmatrix}n\\n_1,\dots,n_1\end{bmatrix}_qX^{\frac{s(s-1)}2n_1^2}G_\rho(X)^{s}\sum_{b\le a}(-1)^{s-|b|}\prod_{i=1}^m[r_i!]_{q^{n_1}}^{-1}X^{-\frac{r_i(r_i-1)}2n_1},
\]
where $b=\{\Delta_1,\dots,\Delta_m\}\le a$, with $\Delta_i$ a segment of size $r_i$.
\end{prop}
\begin{proof}
By \cite[Prop~9.13]{BZ2}, in the Grothendieck ring we have the identity
\[
\langle a\rangle=\sum_{b\le a}(-1)^{|a|-|b|}\pi(b),
\]
so by Lemma~\ref{induced-dim} and Lemma~\ref{exact-seq} we have
\begin{align*}
G_{\langle a\rangle}(X)&=\sum_{b\le a}(-1)^{|a|-|b|}G_{\pi(b)}(X)\\
&=\sum_{b\le a}(-1)^{|a|-|b|}\begin{bmatrix}n\\r_1n_1,\dots,r_mn_m\end{bmatrix}_qX^{\frac{s^2-r_1^2-\dots-r_m^2}{2}n_1^2}\prod_{i=1}^mG_{\langle\Delta_i\rangle}(X).
\end{align*}
Now, applying Proposition~\ref{delta-dim} and simplifying gives the desired result.
\end{proof}
\begin{example}
Some example calculations are
\begin{align*}
G_{\langle[\rho,\nu\rho]\rangle^t}(X)&=\begin{bmatrix}2n_1\\n_1\end{bmatrix}_qX^{n_1^2}G_\rho(X)^2\{1-[2!]_{q^{n_1}}^{-1}X^{-n_1}\},\\
G_{\langle[\rho,\nu^2\rho]\rangle^t}(X)&=\begin{bmatrix}3n_1\\n_1,n_1,n_1\end{bmatrix}_qX^{3n_1^2}G_\rho(X)^3\{1-2[2!]_{q^{n_1}}^{-1}X^{-n_1}+[3!]_{q^{n_1}}^{-1}X^{-3n_1}\}.
\end{align*}
\end{example}
\begin{example}
In the notation of Proposition~\ref{example-prop}, when $n_1=1$ and $\rho=\chi$ is a character,
\[
G_{\langle a\rangle}(X)=\sum_{b\le a}(-1)^{s-|b|}\begin{bmatrix}s\\r_1,\dots,r_m\end{bmatrix}_qX^{\frac12(s^2-\sum_{i=1}^mr_i^2)}.
\]
\end{example}
\begin{rmk}
More generally, for a $\rho$-rigid multiset of segments $a$ with support of size $s$,
\[
G_{\langle a\rangle}(X)=Q(X)\cdot G_\rho(X)^s,
\]
for some polynomial $Q(X)$ determined by the structure of the poset of multisegments on the support of $a$ (as in Example~\ref{bz-example}). In Proposition~\ref{example-prop}, said poset is a Boolean lattice.
\end{rmk}
\section{General theory of local character expansions}\label{section-local-char}
Section~\ref{section-arbitrary} reduces the calculation of growth polynomials of representations to that of supercuspidal representations. To calculate the growth polynomials of supercuspidal representations, we use the \emph{local character expansion}, which tells us the germ of a character.
Throughout this section, \emph{assume $F$ has characteristic zero}. Let $G$ be the group of $F$-rational points of a connected reductive algebraic $F$-group (e.g., ${\mathrm{GL}}_n(F)$ or ${\mathrm{GL}}_m(D)$, for a division algebra $D$ over $F$.) Also let ${\mathfrak g}$ be its Lie algebra, so there is an exponential map
\[
\exp_G\colon{\mathfrak g}_0\to G_1,
\]
for some neighborhood ${\mathfrak g}_0$ of $0\in{\mathfrak g}$ and neighborhood $G_1$ of $1\in G$. In fact, $\exp_G$ is bijective for sufficiently small neighborhoods ${\mathfrak g}_0\ni0$ and $G_1\ni1$.
\begin{example}
For $G={\mathrm{GL}}_n(F)$ (resp., ${\mathrm{GL}}_m(D)$), we have ${\mathfrak g}=M_n(F)$ (resp., $M_m(D)$), and for ${\mathfrak g}_0\colonequals M_n(p{\mathscr O}_F)$ (resp., $M_m(p{\mathscr O}_D)$) and $G_1\colonequals 1+{\mathfrak g}_0$, there is an isomorphism of groups
\begin{align*}
{\mathfrak g}_0&\to G_1\\
X&\mapsto 1+X+\frac{X^2}2+\cdots.
\end{align*}
Unlike in the archimedean case, $\sum_{n\ge0}\frac{X^n}{n!}$ does not converge on the entire Lie algebra.
\end{example}
Let $Ch_\pi$ denote the character associated to an admissible representation $(\pi,V)$ of $G$. That is, for $f\in C_c^\infty(G)$ a compactly supported function on $G$,
\[
\mathrm{tr}\,\pi(f)=\int_GCh_\pi(g)f(g)dg,
\]
where $\pi(f)=\int_G f(g)\pi(g)dg\colon V\to V$ has finite rank so the trace is well-defined (restrict to a finite-dimensional subspace $\pi(f)(V)\subset W$ such that $\pi(f)(W)\subset W$). Note that $\pi(g)\colon V\to V$ for $g\in G$ need not have finite rank so its ``literal" trace is ill-defined.
For a compactly supported function $f\in C_c^\infty({\mathfrak g})$, define the \emph{Fourier transform} as
\[
\widehat f(X)\colonequals \int_{\mathfrak g} f(Y)\cdot\psi(B(X,Y))dY,
\]
where $B(-,-)$ is the Killing form on ${\mathfrak g}$ and $\psi$ is the fixed character of $F$ as defined in Section~\ref{section-terms}.
\begin{defn}
Let ${\mathcal N}_G$ be the set of nilpotent $\mathrm{Ad}\ G$-orbits in ${\mathfrak g}$, called the nilpotent orbits of $G$. It is partially ordered with respect to the order ${\mathcal O}_1\le{\mathcal O}_2$ when $\overline{{\mathcal O}}_1\subseteq \overline{{\mathcal O}}_2$, where $\overline{{\mathcal O}}_1\subset G$ is the Zariski closure of $\overline{{\mathcal O}}_1$.
\end{defn}
\begin{example}\label{gln-nilp}
For $G={\mathrm{GL}}_n(F)$, there is a 1-1 correspondence between the partially ordered sets:
\begin{enumerate}
\item\label{item1} ${\mathcal N}_G$, the set of nilpotent orbits of $G$;
\item\label{item2} equivalence classes of parabolic subgroups $P\subset G$; and
\item\label{item3} partitions of $n$,
\end{enumerate}
where the correspondence between \eqref{item1} and \eqref{item2} reverses order.
Indeed, for ${\mathcal O}\in{\mathcal N}_G$ let $x\in{\mathcal O}$ and consider the parabolic subgroup fixing the flag defined by $x$. Conversely, for any parabolic $P\subset G$, let $N\subset P$ be the unipotent radical and let $\mathfrak n\subset{\mathfrak g}$ be its Lie algebra, so $N=I_n+\mathfrak n$. Then, consider the unique nilpotent orbit ${\mathcal O}$ such that ${\mathcal O}\cap\mathfrak n\subset\mathfrak n$ is open. For each partition $\lambda$ of $n$, denote by ${\mathcal O}_\lambda\in{\mathcal N}_G$ the corresponding nilpotent orbit. Finally, the correspondence between \eqref{item2} and \eqref{item3} is demonstrated in the beginning of Section~\ref{section-arbitrary}.There is a similar correspondence for $G'={\mathrm{GL}}_m(D)$.
Note that when $\lambda$ is the partition $n=n_1+\cdots+n_r$, then for $N_\lambda\subset P_\lambda$ the nilpotent radical,
\[
\dim({\mathcal O}_\lambda)=2\dim(N_\lambda)=\sum_{i\ne j}n_in_j=n^2-n_1^2-\cdots-n_r^2.
\]
\end{example}
\begin{lemma}
Any nilpotent orbit ${\mathcal O}$ in $G$ has a $G$-invariant Radon measure $\mu_{\mathcal O}$. Moreover, letting ${\widehat\mu}_{\mathcal O}(f)\colonequals \mu_{\mathcal O}(\widehat f)$ for $f\in C_c^\infty({\mathfrak g})$ (where $\widehat f$ denotes the Fourier transform with respect to the self-dual measure), the distribution $\widehat\mu_{\mathcal O}$ is represented by a locally integrable function, again denoted $\widehat\mu_{\mathcal O}$.
\end{lemma}
\begin{example}
For $f\in C_c^\infty({\mathfrak g})$ we have $\mu_{\{0\}}(f)=f(0)$, so $\widehat\mu_{\{0\}}=1$.
\end{example}
Now, recall Howe and Harish-Chandra's local character expansion \cite{harish-chandra}:
\begin{prop}\label{howe-local-character}
Let $(\pi,V)$ be an irreducible admissible representation of $G$. Then, there exists a neighborhood $U$ of $0\in{\mathfrak g}$ such that for regular $X\in U$,
\[
Ch_\pi(\exp(X))=\sum_{{\mathcal O}}c_{\mathcal O}(\pi)\widehat{\mu}_{\mathcal O}(X),
\]
or, equivalently, for a function $f$ supported on $\exp(U)$,
\[
\mathrm{tr}\,\pi(f)=\sum_{\mathcal O} c_{\mathcal O}(\pi)\mu_{\mathcal O}(\widehat{f\circ\exp}),
\]
where the sum runs over nilpotent orbits of ${\mathfrak g}$, and the constant $c_{\mathcal O}(\pi)\in{\mathbb C}$ only depends on $\pi$ and ${\mathcal O}$.
\end{prop}
\begin{lemma}\label{induced-char}
Let ${\mathcal O}$ be a nilpotent orbit in ${\mathrm{GL}}_n(F)$ or ${\mathrm{GL}}_m(D)$, and let $P$ be the corresponding parabolic subgroup. Then, for $X\in{\mathfrak g}$ regular in a sufficiently small neighborhood of $0$,
\[
\widehat\mu_{\mathcal O}(X)=Ch_{\Ind_P^G(1_P)}(\exp X).
\]
In particular, for $f\in C_c^\infty({\mathfrak g})$ supported on a sufficiently small neighborhood of $0$,
\[
\mu_{\mathcal O}(\widehat f)=\mathrm{tr} \Ind_P^G(1_P)(f).
\]
\end{lemma}
\begin{proof}
For ${\mathrm{GL}}_n(F)$, this is \cite[Lem~5]{howe}, and the same argument works for ${\mathrm{GL}}_m(D)$.
\end{proof}
Let $(\pi,V)$ be an irreducible admissible representation of ${\mathrm{GL}}_n(F)$. Then, we may express $\dim(\pi^{K_N})$ as a polynomial of $q^{N-1}$, as in \cite{savin}:
\begin{cor}\label{fixed-poly} Let $\pi$ be an irreducible admissible representation of ${\mathrm{GL}}_n(F)$. Then,
\[
G_\pi(X)=\sum_{{\mathcal O}\in{\mathcal N}_G} c_{\mathcal O}(\pi) \begin{bmatrix}n\\n_1,\dots,n_r\end{bmatrix}_q X^{\frac12\dim{\mathcal O}},
\]
where $n=n_1+\cdots+n_r$ is the partition corresponding to ${\mathcal O}$ under the identification in Example~\ref{gln-nilp}.
\end{cor}
\begin{proof}
Let $V$ be the underlying vector space for $\pi$. Let $e_{K_N}$ be the normalized characteristic function, i.e.,
\[
e_{K_N}(x)=\begin{cases}\frac1{\mathrm{vol}(K_N)}&\text{if }x\in K_N\\0&\text{if }x\notin K_N.\end{cases}
\]
Then, $\pi(e_{K_N})\colon\pi\to\pi$ is a projection to $\pi^{K_N}$, so
\[
\dim(\pi^{K_N})=\mathrm{tr}\,\pi(e_{K_N}).
\]
Now by Proposition~\ref{howe-local-character} and Lemma~\ref{induced-char}, for any integer $N\gg0$,
\begin{align*}
\mathrm{tr}\,\pi(e_{K_N})&=\frac1{\mathrm{vol}(K_N)}\int_{K_N}Ch_\pi(g)dg\\
&=\sum_{\mathcal O} c_{\mathcal O}(\pi)\frac1{\mathrm{vol}(K_N)}\int_{K_N}Ch_{\Ind_P^G(1)}(g)dg\\
&=\sum_{\mathcal O} c_{\mathcal O}\mathrm{tr}\,(\Ind_P^G(1_P))(e_{K_N}).
\end{align*}
Now, by Lemma~\ref{induced-dim} and Example~\ref{gln-nilp},
\begin{align*}
\mathrm{tr}\,(\Ind_P^G(1_P))(e_{K_N})&=\dim(\Ind_P^G(1_P)^{K_N})\\%&=|P\backslash G/K_N|\\
&=\begin{bmatrix}n\\n_1,\dots,n_r\end{bmatrix}_qq^{\frac12\dim{\mathcal O}(N-1)}.\qedhere
\end{align*}
\end{proof}
\begin{rmk}
More generally, let $(\pi,V)$ be an admissible irreducible representation of ${\mathrm{GL}}_n(F)$, let $N\gg0$, and let $\varphi$ be an arbitrary irreducible representation of $K_N$. Then, by \cite[Prop~4.4]{Bush-Henn},
\[
e_\varphi(x)\colonequals\begin{cases}\frac{\dim\varphi}{\mathrm{vol}(K_N)}\mathrm{tr}\, (\varphi(x^{-1}))&\text{if }x\in K_N\\
0&\text{if }x\notin K_N\end{cases}
\]
is such that $\pi(e_\varphi)\colon V\to V$ is a projection to the $\varphi$-isotypic component, so a similar argument shows
\[
\dim(\hom_{K_N}(\varphi,\pi|_{K_N}))=\sum_{{\mathcal O}\in{\mathcal N}_G} c_{\mathcal O}(\pi) \dim(\varphi^{K_N\cap P})\begin{bmatrix}n\\n_1,\dots,n_r\end{bmatrix}_qq^{\frac12\dim{\mathcal O}(N-1)}.
\]
Thus, the coefficients $c_{\mathcal O}$ completely determine the structure of the $K_N$-representation $\pi|_{K_N}$.
\end{rmk}
It is known that:
\begin{prop}\label{cusp-asymp} For $\rho$ a supercuspidal representation of ${\mathrm{GL}}_n(F)$,
\[
G_\rho(X)=[n!]_qX^{\frac{n(n-1)}2}+(\text{lower order terms}).
\]
\end{prop}
\begin{proof}
Follows from \cite[Thm~3]{howe}, together with Corollary~\ref{fixed-poly}.
\end{proof}
Moreover, in for certain supercuspidal representations, the entire growth polynomial is known \cite[Prop~5.3]{murnaghan}:
\begin{prop}\label{murn}
Let $\rho$ be the supercuspidal representation corresponding to the generic character $(E/F,\theta)$ where $\theta=(\chi\circ N_{E/F})\phi$ for some generic character $\phi$ of $E$ over $F$, where $\phi$ has level $j$ (i.e., $j$ is the minimal integer such that $\phi$ is trivial on $1+{\mathfrak p}_E^j$). Then,
\begin{enumerate}
\item If $E/F$ is unramified,
\[
G_\rho(X)=\sum_{\mathcal O}(-1)^{n+r}n(r-1)!w_{\mathcal O}^{-1}q^{(n^2-n-\dim{\mathcal O})j/2}\begin{bmatrix}n\\n_1,\dots,n_r\end{bmatrix}_qX^{\frac12\dim{\mathcal O}}
\]
\item If $E/F$ is totally ramified,
\[
G_\rho(X)=\sum_{\mathcal O}(-1)^{n+r}(r-1)!w_{\mathcal O}^{-1}\frac q{q-1}(r-\sum_{i=1}^rq^{-n_i})q^{\frac{(n^2-n-\dim{\mathcal O})j}{2n}}\begin{bmatrix}n\\n_1,\dots,n_r\end{bmatrix}_qX^{\frac12\dim{\mathcal O}},
\]
\end{enumerate}
where $w_{\mathcal O}$ is the number of permutations of $r$ letters which fix the $r$-tuple $(n_1,\dots,n_r)$.
\end{prop}
\begin{proof}
Follows from \cite[Prop~5.3]{murnaghan}, together with Corollary~\ref{fixed-poly}.
\end{proof}
\begin{rmk}
For ${\mathrm{GL}}_\ell(F)$ with $\ell$ prime, this result suffices to give an explicit formula for all supercuspidal representations.
\end{rmk}
\section{Under Langlands functoriality}\label{section-functoriality}
Again, throughout this section, \emph{assume $F$ has characteristic zero}.
\subsection{Jacquet-Langlands correspondence}
Recall the Jacquet-Langlands correspondence \cite{JL}:
\begin{prop}\label{jl}
Let $D$ be a division algebra of dimension $d^2$ over $F$, and let $E^2({\mathrm{GL}}_n(F))$ (resp., $E^2({\mathrm{GL}}_m(D))$) be the set of irreducible square-integrable representations of ${\mathrm{GL}}_n(F)$ (resp, ${\mathrm{GL}}_m(D)$), where $n=md$. Then, there exists a bijection
\begin{align*}
E^2({\mathrm{GL}}_n(F))&\to E^2({\mathrm{GL}}_m(D))\\
\pi&\mapsto \pi_D
\end{align*}
such that for $g'\in{\mathrm{GL}}_m(D)$ a regular element, whose conjugacy class corresponds to a regular element $g\in{\mathrm{GL}}_n(F)$, then
\[
(-1)^mCh_{\pi_D}(g')=(-1)^nCh_\pi(g).
\]
\end{prop}
Now, we have:
\begin{prop}\label{j-l-formula}
Let $\pi\in E^2({\mathrm{GL}}_n(F))$ be an irreducible square-integrable representation, corresponding to $\pi_D\in E^2({\mathrm{GL}}_m(D))$ under the Jacquet-Langlands correspondence. Then
\[
(-1)^mc_{\mathcal O}(\pi)=(-1)^nc_{{\mathcal O}'}(\pi_D),
\]
where ${\mathcal O}'\subseteq{\mathcal O}$ are nilpotent orbits of $G'$ and $G$, respectively (via identifying the Lie algebras of $G'$ and $G$, tensored with $\overline F$).
\end{prop}
\begin{proof
Let $m=m_1+\dots+m_r$ be an arbitrary partition of $m$, and consider a block diagonal matrix $g=(g_1,\dots,g_r)\in{\mathrm{GL}}_n(F)$ where $g_i\in{\mathrm{GL}}_{dm_i}(F)$ are elliptic elements (i.e., has an irreducible characteristic polynomial) and the minimal polynomials of $g_i$ are coprime. Then $g$ is regular semisimple, corresponding to the diagonal matrix $g'=(g_1',\dots,g_r')\in{\mathrm{GL}}_m(D)$ where each $g_i'\in D_{dm_i}^\times\subset{\mathrm{GL}}_{m_i}(D)$ is similar to $g_i\in{\mathrm{GL}}_{dm_i}(F)$, where $D_{dm_i}\supset D$ is a division algebra of dimension $d^2m_i^2$ over $F$.
Then, by Proposition~\ref{howe-local-character} and Lemma~\ref{induced-char}, when $g$ as above is sufficiently close to $1$,
\[
Ch_\pi(g)=\sum_{{\mathcal O}}c_{\mathcal O}(\pi)Ch_{\Ind_P^G(1)}(g).
\]
Here by \cite[Thm~3]{dijk} the character $Ch_{\Ind_P^G(1)}(g)$ is locally supported on the union of the conjugates of $P$, and so $Ch_{\Ind_P^G(1)}(g)=0$ for $P$ not corresponding to a partition of the form $a_1d+\cdots+a_id$.
Let ${\mathcal N}_G^d$ be the set of such nilpotent orbits, i.e., those corresponding to partitions with all parts a multiple of $d$. There is a bijection ${\mathcal N}_{G'}\simeq{\mathcal N}_G^d$, sending ${\mathcal O}'\in{\mathcal N}_{G'}$ to the unique nilpotent orbit ${\mathcal O}$ of $G$ such that ${\mathcal O}'\otimes\overline F\subseteq{\mathcal O}\otimes\overline F$ (on the level of partitions, $a_1+\cdots+a_i\mapsto a_1d+\cdots+a_id$).
Now, Proposition~\ref{jl} gives the equality
\begin{equation}\label{jl-eq}
(-1)^m\sum_{{\mathcal O}\in\mathcal N_G^d}c_{\mathcal O}(\pi)Ch_{\Ind_P^G(1)}(g)=(-1)^n\sum_{{\mathcal O}'\in\mathcal N_{G'}}c_{{\mathcal O}'}(\pi_D)Ch_{\Ind_P^G(1)}(g),
\end{equation}
since by Lemma~\ref{induced-char},
\[
\widehat\mu_{{\mathcal O}'}(g')=Ch_{\Ind_{P'}^{G'}(1_{P'})}(g')=Ch_{\Ind_P^G(1_P)}(g).
\]
where $P'\colonequals (P\otimes\overline F)\cap G'\subset G'$. Moreover, for parabolic subgroups $P_1\subsetneq P_2$ of $G$ we have $\supp(Ch_{\Ind_{P_1}^G(1)})\subsetneq\supp(Ch_{\Ind_{P_2}^G(1)})$ in a neighborhood of $1\in G$, so the characters $\{Ch_{\Ind_P^G(1)}:P\subset G\text{ parabolic}\}$ are linearly independent. Thus, equation~\ref{jl-eq} gives \[(-1)^mc_{\mathcal O}(\pi)=(-1)^nc_{{\mathcal O}'}(\pi_D).\qedhere\]
\end{proof}
\begin{comment}[slight elaboration: {\color{red}If ${\mathcal O}_\lambda$ is the nilpotent orbit corresponding to $m=m_1+\cdots+m_r$ then we further have
\[
(-1)^m\sum_{{\mathcal O}\ge{\mathcal O}_\lambda}c_{\mathcal O}(\pi)Ch_P(g)=(-1)^n\sum_{{\mathcal O}\ge{\mathcal O}_\lambda}c_{{\mathcal O}'}(\pi_D)Ch_{P}(g),
\]
so by induction (from top to bottom), $(-1)^mc_{\mathcal O}(\pi)=(-1)^nc_{\mathcal O}(\pi_D)$.}]\end{comment}
\begin{cor}\label{const-term}
Let $\pi\in E^2({\mathrm{GL}}_n(F))$ be an irreducible square-integrable representation, and let $\pi_D\in E^2(D^\times)$ be the corresponding representation under the Jacquet-Langlands correspondence, where $D$ is a $n^2$-dimensional division algebra. Then the growth polynomial $G_\pi(X)$ has constant term $(-1)^{n-1}\dim(\pi_D)$.
\end{cor}
\begin{proof}
Since $D^\times$ is compact modulo center, the representation $\pi_D$ is finite-dimensional. Thus, $c_0(\pi_D)=\dim(\pi_D)$, and so $c_0(\pi)=(-1)^{n-1}c_0(\pi_D)=(-1)^{n-1}\dim(\pi_D)$.
\end{proof}
The combination of Proposition~\ref{cusp-asymp} and Corollary~\ref{const-term} allows the calculation for the polynomial $\dim(\pi^{K_N})$ for representations $\pi$ of ${\mathrm{GL}}_2(F)$:
\begin{example}\label{gl2-dim}
When $n=2$, \cite[Section~56]{Bush-Henn} provides an explicit description of the Jacquet-Langlands correspondence for minimal supercuspidal representations $\rho$ of ${\mathrm{GL}}_2$. Only treating these suffices since arbitrary supercuspidal representations are twists of minimal supercuspidal representations.
\begin{itemize}
\item If $\rho$ has level zero, then $\rho_D$ is induced from a character of $E^\times U_D^1$, so
\[
G_\rho(X)=(q+1)X-2.
\]
\end{itemize}
Otherwise, there is a simple stratum $({\mathfrak A}, n,\alpha)$ such that $\rho$ contains the character $\psi_\alpha$ of $U_{\mathfrak A}^{[n/2]+1}$, so $\rho\cong\cInd_J^G\Lambda$ where $J=E^\times U_{\mathfrak A}^{[(n+1)/2]}$ and $\Lambda$ is an irreducible representation on $J$:
\begin{itemize}
\item If $e_{\mathfrak A}=2$ then $\rho_D$ is induced from a character on $E^\times U_D^{[n/2]+1}$, so
\[
G_\rho(X)=(q+1)X-|E^\times U_D^{[n/2]+1}\backslash D^\times|.
\]
\item If $e_{\mathfrak A}=1$ and $n$ is odd then $\rho_D$ is induced from a $q$-dimensional representation of $E^\times U_D^n$, so
\[
G_\rho(X)=(q+1)X-q|E^\times U_D^n\backslash D^\times|,
\]
\item If $e_{\mathfrak A}=1$ and $n$ is even then $\rho_D$ is induced from a character of $E^\times U_D^{n+1}$, so
\[
G_\rho(X)=(q+1)X-|E^\times U_D^{n+1}\backslash D^\times|.
\]
\end{itemize}
In each of these cases, we have:
\begin{align*}
|E^\times U_D^m\backslash D^\times|&=|\langle\varpi_E\rangle\backslash\langle\varpi_D\rangle|\cdot|k_E^\times\backslash k_D^\times|\cdot|U_E^1U_D^m\backslash U_D^1|\\
&=\frac2e\cdot \frac{q^2-1}{q^{2/e}-1}|U_E^1U_D^m\backslash U_D^1|.
\end{align*}
Now, there is an exact sequence
\[
1\to (U_E^1\cap U_D^m)\backslash U_E^1\cong U_D^m\backslash U_E^1U_D^m\to U_D^m\backslash U_D^1\to U_E^1U_D^m\backslash U_D^1\to 1,
\]
where $m\ge1$. Noting that $U_E^1\cap U_D^m=U_E^{1+[(m-1)e/2]}$, we obtain
\[
|U_E^1U_D^m\backslash U_D^1|=\frac{|U_D^m\backslash U_D^1|}{|U_E^{1+[(m-1)e/2]}\backslash U_E^1|}=q^{2(m-1)-\frac2e[\frac e2(m-1)]}.\]
Thus, combining everything gives
\[
G_\rho(X)=\begin{cases}
(q+1)X-2&\text{if }\rho\text{ has level zero}\\
(q+1)X-(q+1)q^{[n/2]}&\text{if }e_{\mathfrak A}=2\\
(q+1)X-2q^n&\text{if }e_{\mathfrak A}=1.
\end{cases}
\]
\end{example}
\subsection{Base change} Another instance of Langlands functoriality is base change and automorphic induction. For a finite extension $E/F$ of degree $d$, base change is given by the commutative diagram
\begin{equation}\label{eq-base-change} \begin{tikzcd}
\mathbf{Irr}({\mathrm{GL}}_n(F)) \arrow{r}{\mathrm{rec}_F} \arrow[swap,dashed]{d}{\mathrm{BC}_{E/F}} & \mathcal G_n(F) \arrow{d}{\mathrm{Res}^{\mathcal W_F}_{\mathcal W_E}} \\%
\mathbf{Irr}({\mathrm{GL}}_n(E)) \arrow{r}{\mathrm{rec}_E}& \mathcal G_n(E),
\end{tikzcd}\end{equation}
and automorphic induction is given by
\[ \begin{tikzcd}
\mathbf{Irr}({\mathrm{GL}}_n(E)) \arrow{r}{\mathrm{rec}_E} \arrow[swap,dashed]{d}{\mathrm{AI}_{E/F}} & \mathcal G_n(E) \arrow{d}{\mathrm{Ind}^{\mathcal W_F}_{\mathcal W_E}} \\%
\mathbf{Irr}({\mathrm{GL}}_{nd}(F)) \arrow{r}{\mathrm{rec}_F}& \mathcal G_{nd}(F),
\end{tikzcd}\]
where $\mathrm{rec}_F$ denotes the local Langlands correspondence. When $E/F$ is a cyclic extension of prime degree, $\mathrm{BC}_{E/F}$ has a characterization in terms of characters \cite[Thm~6.2]{AC}:
\begin{defn}
Let $E/F$ be a cyclic extension of prime degree, and let $\tau\in\Gal(E/F)$ be a generator.
Moreover, let $(\Pi,V)$ be an irreducible representation of ${\mathrm{GL}}_n(E)$ such that $\Pi\cong\Pi^\tau$, where $\Pi^\tau(g)\colonequals\Pi(\tau(g))$ for $g\in{\mathrm{GL}}_n(E)$. Let $I_\tau\colon\Pi\to\Pi^\tau$ be such an isomorphism, appropriately normalized (require the Whittaker functional to be fixed). Then there exists a locally constant function $Ch_{\Pi,\tau}$ defined on ${\mathrm{GL}}_n(E)^{\tau-reg}$ such that
\[
\mathrm{tr}(\pi(f)\circ I_\tau)=\int_{{\mathrm{GL}}_n(E)^{\tau-reg}}f(g)Ch_{\Pi,\tau}(g)dg,
\]
where ${\mathrm{GL}}_n(E)^{\tau-reg}$ is the set of $g\in{\mathrm{GL}}_n(E)$ such that $g\tau(g)\cdots\tau^{d-1}(g)$ is regular.
\end{defn}
\begin{prop}[Shintani character formula]\label{shintani-char}
Let $E/F$ be a cyclic extension of prime degree $d$, and let $\tau\in\Gal(E/F)$ be a generator. Then for $\pi$ an irreducible tempered representation of ${\mathrm{GL}}_n(F)$,
\[
Ch_{\mathrm{BC}_{E/F}(\pi),\tau}(g)=Ch_\pi(\mathcal Ng),
\]
for any $g\in{\mathrm{GL}}_n(E)^{\tau-reg}$, where $\mathcal Ng\in{\mathrm{GL}}_n(F)$ is the norm of $g$, i.e., the unique (up to conjugacy) element of ${\mathrm{GL}}_n(F)$ conjugate to $g\tau(g)\cdots\tau^{d-1}(g)$ (which is regular by definition).
\end{prop}
Shintani's character formula gives another method to calculate the growth polynomial:
\begin{prop}\label{twisted-char}
Let $E/F$ be a cyclic extension of prime degree $d$, and let $\theta$ be a generic character of $E/F$ (i.e., $\theta\ne\theta^\tau$). Then for an integer $N\gg0$,
\[
G_{\mathrm{AI}_{E/F}\theta}(q^{[N/e]-1})=\mathrm{tr}(I_\tau|_{V^{K_N}})
\]
where $V\colonequals\theta\times\theta^\tau\times \cdots\times\theta^{\tau^{d-1}}$.
\end{prop}
\begin{proof}
Let $\pi=\mathrm{AI}_{E/F}(\theta)$ and $\Pi=\mathrm{BC}_{E/F}(\pi)$. Then, by Mackey's formula,
\begin{align*}
\mathrm{rec}_E(\Pi)&=(\Ind_{\mathcal W_E}^{\mathcal W_F}\mathrm{rec}_E\theta)|_{\mathcal W_E}\\
&=\bigoplus_{\sigma\in\mathcal W_E/\mathcal W_F}\mathrm{rec}_E\theta^\sigma\\
&=\mathrm{rec}_E\theta\oplus\mathrm{rec}_E\theta^\tau\oplus \cdots\oplus\mathrm{rec}_E\theta^{\tau^{d-1}},
\end{align*}
so that $\Pi=\theta\times\cdots\times\theta^{\tau^{d-1}}$. Now, for $N\gg0$, by Proposition~\ref{shintani-char} and Lemma~\ref{induced-char},
\begin{align*}
\mathrm{tr}(I_\tau|_{\Pi^{K_N}})&=\frac1{\mathrm{vol}(K_N(E))}\int_{K_N(E)^{\tau-reg}}Ch_{\Pi,\tau}(g)dg\\
&=\frac1{\mathrm{vol}(K_N(E))}\int_{K_N(E)^{\tau-reg}}Ch_{\pi}(\mathcal Ng)dg\\
&=\sum_{\mathcal O} c_{\mathcal O}(\pi)\frac1{\mathrm{vol}(K_N(E))}\int_{K_N(E)^{\tau-reg}}Ch_{\Ind_P^G(1)}(\mathcal Ng)dg,
\end{align*}
where the use of Lemma~\ref{induced-char} is justified because $\mathcal Ng\in{\mathrm{GL}}_n(F)$ is regular and lies in a small neighborhood of $1$.
We may safely replace $\Ind_P^G(1)$ with $\Ind_P^G(\chi_1\boxtimes\cdots\boxtimes\chi_r)$ for unitary characters $\chi_i$ such that $\chi_i\circ N_{E/F}$ are distinct, since the germs of their characters match. But then (by following the diagram~\eqref{eq-base-change}), the base change is \[\mathrm{BC}_{E/F}(\Ind_P^G(\chi_1\boxtimes\cdots\boxtimes\chi_r))=\Ind_{P(E)}^{G(E)}(\chi_1\circ N_{E/F}\boxtimes\cdots \boxtimes\chi_r\circ N_{E/F}),\] so by the same argument as above,
\begin{align*}
\frac1{\mathrm{vol}(K_N(E))}\int_{K_N(E)^{\tau-reg}}Ch_{\Ind_P^G(1)}(\mathcal Ng)dg&=\mathrm{tr}\big(I_\tau|_{\Ind_{P(E)}^{G(E)}(\chi_1\circ N_{E/F}\boxtimes\cdots\boxtimes \chi_r\circ N_{E/F})^{K_N}}\big).
\end{align*}
Vectors of the invariant space $\Ind_{P(E)}^{G(E)}(\chi_1\circ N_{E/F}\boxtimes\cdots \boxtimes\chi_r\circ N_{E/F})^{K_N}$ can be viewed as functions on $P(E)\backslash G(E)/K_N(E)$, and $\tau$ acts on it by the Galois action, so the trace is the size of the invariant elements
\[
(P(E)\backslash G(E)/K_N(E))^\tau=P(F)\backslash G(F)/(K_N(E)\cap G(F))=P(F)\backslash G(F)/K_{[N/e]}(F),
\]
so altogether, by Corollary~\ref{fixed-poly},
\begin{align*}
\mathrm{tr}(I_\tau|_{\Pi^{K_N}})&=\sum_{\mathcal O} c_{\mathcal O}(\pi)|P(F)\backslash G(F)/K_{[N/e]}(F)|\\
&=G_{\mathrm{AI}_{E/F}\theta}(q^{[N/e]-1}).\qedhere
\end{align*}
\end{proof}
We will give an alternative proof of the $n=2$ case of Proposition~\ref{murn}~(1), as well as Example~\ref{gl2-dim}:
\begin{prop}
Let $E/F$ be an unramified degree $2$ extension and let $\theta$ be a generic character of $E/F$, such that $\theta^\tau\theta^{-1}$ has level $\ell$ (i.e., $\ell$ is the minimal integer such that $\theta^\tau\theta^{-1}$ is trivial on $1+{\mathfrak p}_E^{\ell+1}$). Then,
\[
G_{\mathrm{AI}_{E/F}(\theta)}(X)=(q+1)X-2q^\ell.
\]
\end{prop}
\begin{proof}
Let $E=F(\sqrt\alpha)$ with $v_F(\alpha)=0$ and let $\varpi\in{\mathscr O}_F$ be the generator of the maximal ideal ${\mathfrak p}_F$. By Proposition~\ref{twisted-char}, it suffices to calculate the trace of $I_\tau$ on $(\theta\times\theta^\tau)^{K_N(E)}$. By \cite[Prop~4.5.6]{bump} it is given by
\begin{align*}
q^{-\ell}I_\tau\colon\nInd_{B(E)}^{G(E)}(\theta\boxtimes\theta^\tau)&\to \nInd_{B(E)}^{G(E)}(\theta\boxtimes\theta^\tau)^\tau\\
f&\mapsto\bigg(g\mapsto\lim_{s\to0^-} \int_Ef\big(\begin{pmatrix}&1\\1&x\end{pmatrix}\tau(g)\big)|x|^sdx\bigg).
\end{align*}
For $g_0\in{\mathrm{GL}}_2(E)$, define the twisted indicator function
\[
{\mathbf 1}_{g_0}(g)\colonequals\begin{cases}(\theta(\frac12)\boxtimes\theta^\tau(-\frac12))(b)&\text{if }g=bg_0u\text{ for some }b\in B(E),u\in K_N(E)\\
0&\text{if }g\notin B(E)g_0K_N(E),\end{cases}
\]
so that $\{{\mathbf 1}_{g_0}\}_{g_0\in B(E)\backslash{\mathrm{GL}}_2(E)/K_N(E)}$ is a basis for $(\theta\times\theta^\tau)^{K_N(E)}$. Note that by Iwasawa decomposition we may take $g_0\in K_0(E)$, in which case ${\mathbf 1}_{g_0}$ is well-defined for $N>\ell$. Thus,
\begin{equation}\label{trace-integral}
\mathrm{tr}(q^{-\ell}I_\tau|_{(\theta\times\theta^\tau)^{K_N}})=\sum_{g_0\in B(E)\backslash{\mathrm{GL}}_2(E)/K_N(E)}\lim_{s\to 0^-}\int_E{\mathbf 1}_{g_0}\big(\begin{pmatrix}&1\\1&x\end{pmatrix}\tau(g_0)\big)|x|^sdx.
\end{equation}
Moreover, for arbitrary $k\in K_0(F)$ we have
\[
{\mathbf 1}_{g_0}\big(\begin{pmatrix}&1\\1&x\end{pmatrix}\tau(g_0)\big)={\mathbf 1}_{g_0k}\big(\begin{pmatrix}&1\\1&x\end{pmatrix}\tau(g_0k)\big),
\]
since if $\begin{pmatrix}&1\\1&x\end{pmatrix}\tau(g_0)=bg_0u$ then $\begin{pmatrix}&1\\1&x\end{pmatrix}\tau(g_0k)=bg_0k(k^{-1}uk)$, where $k^{-1}uk\in K_N(E)$. Thus, the sum in equation~\eqref{trace-integral} can be combined by $K_0(F)$-orbits. Those orbits are represented by
\[
\big\{I_2,\begin{pmatrix}1\\\varpi^i\sqrt{\alpha}&1\end{pmatrix}:0\le i<N\big\},
\]
where the orbit of $I_2$ has size $q^{N-1}(q+1)$ by Lemma~\ref{parabolic-size}, and the orbit of $\begin{pmatrix}1\\\varpi^i\sqrt{\alpha}&1\end{pmatrix}$ has size $(q^2-q)q^{2(N-1)}$ for $i=0$ and $(q^2-1)q^{2(N-1)-i}$ for $i>0$. What is left is to calculate the integral in equation~\ref{trace-integral} for each of these choices of $g_0$.
When $g_0=I_2$, then $\begin{pmatrix}&1\\1&x\end{pmatrix}\in BK_N$ exactly when $v_E(x)\le -N$, in which case $b=\begin{pmatrix}-x^{-1}&*\\&x\end{pmatrix}$ and so the desired integral is
\[
\lim_{s\to 0^-}\int_{v(x)\le -N}\theta(-x^{-1})\theta^\tau(x)|x|^sdx.
\]
The change of variables $x\mapsto ux$ for $u\in{\mathscr O}_E^\times$ such that $\theta(\tau(u))\ne\theta(u)$ shows the integral is $0$.
Next, when $g_0=\begin{pmatrix}1\\\varpi^i\sqrt{\alpha}&1\end{pmatrix}$, we have
\[
\begin{pmatrix}&1\\1&x\end{pmatrix}\begin{pmatrix}1\\-\varpi^i\sqrt{\alpha}&1\end{pmatrix}=\begin{pmatrix}*&*\\1-x\varpi^i\sqrt\alpha&x\end{pmatrix}\in Bg_0K_N
\]
exactly when $(1-x\varpi^i\alpha)x^{-1}\equiv\varpi^i\alpha\pmod{{\mathfrak p}_E^N}$, i.e., $x\in\frac1{2\varpi^i\sqrt\alpha}(1+{\mathfrak p}_E^{N-i})$. Thus, the integral is (using that the measure $\frac{dx}{|x|}$ is a Haar measure for $E^\times$):
\[
\int_{\frac1{2\varpi^i\sqrt\alpha}(1+{\mathfrak p}_E^{N-i})}\theta(-x^{-1})\theta^\tau(x)\frac{dx}{|x|}=\int_{1+{\mathfrak p}_E^{N-i}}\theta(x^{-1})\theta^\tau(x)\frac{dx}{|x|}.
\]
If $N-i\ge\ell+1$ then $\theta(x^{-1})\theta^\tau(x)=1$ on $1+{\mathfrak p}_E^{N-i}$, so the integral is simply $\int_{1+{\mathfrak p}^{N-i}}\frac{dx}{|x|}=q_E^{(i-N)+\frac12}=q^{2(i-N)+1}$. On the other hand, if $N-i\le\ell$ then looking at the change of variables $x\mapsto ux$ for $u\in 1+{\mathfrak p}_E^\ell$ such that $\theta(\tau(u))\ne\theta(u)$ shows the integral is $0$. Thus, in conclusion,
\begin{align*}
q^{-\ell}\mathrm{tr}(I_\tau|_{(\theta\times\theta^\tau)^{K_N}})&=(q^2-q)q^{2(N-1)}q^{-2N+1}+\sum_{i=0}^{N-\ell-1}(q^2-1)q^{2(N-1)-i}q^{2(i-N)+1}\\
&=(q+1)q^{N-\ell-1}-2.\qedhere
\end{align*}
\end{proof}
\begin{rmk}
More generally, let $\rho\in\mathbf{Irr}({\mathrm{GL}}_n(E))$ be supercuspidal and consider $\pi=\mathrm{AI}_{E/F}\rho$. Then, the same arguments show the growth polynomial $G_\pi$ is related to the trace of the intertwining integral $\rho\times\cdots\times\rho^{\tau^{d-1}}\to(\rho\times\cdots\times\rho^{\tau^{d-1}})^\tau$. Moreover, the trace is given by a sum over the double coset $P_{n+\cdots+n}(E)\backslash{\mathrm{GL}}_{nd}(E)/K_N(E)K_0(F)$, the $K_0(F)$-orbits in the partial flag variety over ${\mathscr O}_E/{\mathfrak p}_E^N$. Looking more closely at the combinatorial structure of such double cosets is expected to give new results.
\end{rmk}
\section*{Acknowledgement}
The author thanks Hao Peng for mentoring this project and giving helpful advice. We thank Zhiyu Zhang for suggesting this project. We thank Prof. Wei Zhang and Prof. Ju-Lee Kim for giving us some relevant references and pointing us in the right direction. We thank Dr. Max Gurevich for pointing out an error in the original argument for Proposition~\ref{delta-dim}. We thank Prof. Marie-France Vign\'{e}ras and Prof. Guy Henniart for helpful discussions, as well as fixing an error in Remark~\ref{gl2-dim}.
Finally, we thank Prof. David Jerison and Prof. Ankur Moitra for organizing the SPUR program and helping us through insightful discussions.
|
\section{Synonyms}
Robotics in the Arctic;
Robotics in Antarctica;
Robotics in polar regions;
Robotics in subarctic areas;
Robotics in the cryosphere;
Robotics in cold environments.
\section{Definitions}
The terms \emph{robotics in snow and ice} refers to robotic systems being studied, developed, and used in areas where water can be found in its solid state.
This specialized branch of field robotics investigates the impact of extreme conditions related to cold environments on autonomous vehicles.
\section{Overview}
Regions of Earth where water can be found in its solid form are within the realm of the cryosphere.
As these regions required humans to be protected against the cold, which reduces mobility and operational time, robots have been developed to collect data or act remotely on the environment on their behalf.
Robotic systems deployed in the cryosphere need to be robust to subzero temperature while being challenged by a large spectrum of precipitation and terrain.
On the ground, sub-domains of the cryosphere include snow cover, freshwater ice, frozen ground, sea ice, glaciers, ice caps, and ice sheet.
In the air, solid water can take the form of freezing rain, hail, pellets, snow, and everything in between.
The combination of air temperature, precipitation, and ground conditions makes it very challenging for robotic systems to accomplish a given task as their visibility, maneuverability, and operation time will be impacted.
Broadly speaking, these conditions have pushed forward the research of novel solutions related to perception algorithms, control algorithms, locomotion designs, physical integrity, and the energy management of robots.
Originally a testbed for space exploration, applications extended rapidly to fundamental Earth sciences, transportation, and forestry.
While research related to \emph{robotics in snow and ice} is still in its infancy, it is a clear fertile ground for discoveries related to remote locations and the robustness of robots.
\subsection{Impact of the cryosphere on robotics}
Robotic systems must be well adapted for the specificities of locations where they need to achieve a task.
Cold environments are often categorized by their latitudes (e.g., Arctic, Antarctic, polar regions, subarctic) and elevation.
Instead of these two axes, we will be using directly sub-domains of the cryosphere to explain key elements of \emph{robotics in snow and ice}.
\autoref{fig:zones} relates latitudes and elevation to these sub-domains while highlighting key robots that were deployed in the cryosphere.
The historical context in which these robots were deployed will be described in later sections.
\begin{figure*}[htbp]
\begin{center}
\includegraphics[width=\textwidth,trim={2px 0 0 3px},clip]{figures/zones.pdf}
\caption{Representation of sub-domains of the cryosphere as a function of latitude and elevation.
Light blue represents water and labels with dashed lines highlight where key robots were deployed in different sub-domains.
The size of vegetation goes down with latitude and elevation.
}
\label{fig:zones}
\end{center}
\end{figure*}
The most common sub-domain of the cryosphere is snow covers.
They happen in any place where snow has time to accumulate onto the ground before melting.
The texture of snow cover is complex and highly varied (e.g., fresh powder, layered ice crusts, and compacted by winds) and constantly evolving.
Navigation through fresh snowfield is challenging due to deep sinkage, snow resistance, traction loss, and ingestion of snow into the drive mechanism~\citep{Lever2009}.
As shown in \autoref{fig:deep_snow}, specialized designs must be considered to ensure that a vehicle can rise over a snowbank when driving off-road.
Even with specialized locomotion designs, tight turns in soft snow can result in traction loss and wheel sinkage causing the robot to stall~\citep{Stansbury2004}.
On the perception side, snow covers are highly reflective and produce few textures, which is challenging for vision-based algorithms.
For example, early field deployments in Antarctica could not rely on stereo vision because of missing texture causing disparity matching to fail~\citep{Moorehead1999}.
Moreover, camera autoexposure algorithms tend to struggle on sunny days as the luminosity extends beyond the typical dynamic range of the photosensor, the texture used for feature extraction in localization algorithms is washed out, the ground constantly changes with winds, and images using auto-white balance will shift toward the blue spectrum~\citep{Paton2017}.
Even when specialized algorithms are used to enhance images, the number of features will be significantly lower than that of a typical outdoor scene~\citep{Williams2009}.
\begin{figure}[hbtp]
\begin{center}
\includegraphics[width=\columnwidth]{figures/deep_snow.jpg}
\caption{A \SI{500}{\kg} robot navigating in a subarctic region.
Specialized tracks allow the robot to navigate over deep and soft snow.}
\label{fig:deep_snow}
\end{center}
\end{figure}
Another sub-domain to consider is waterbodies, whether they are fresh or salty, as they have unusual dynamics that robotic systems must be aware of.
For example, water around \SI{4}{\celsius} is denser than the water around \SI{0}{\celsius}.
This density variation results in shallow ponds freezing completely, while lakes and rivers will produce a large frozen crust protecting liquid water maintained at around \SI{4}{\celsius}.
A key application affected by this phenomenon is good transportation in northern regions.
This kind of transport relies on ice roads, which are mostly crossing frozen lakes and rivers.
Advancement in autonomous transportation has the potential of reducing transportation risks, but robots must be conceived with an understanding of ice formation.
Sea ice follows the same process, but the crust will tend to break under external forces, such as tides and winds.
This sea ice must not be misclassified with icebergs, which also float on oceans, but were generally part of a large ice shelf that broke before drifting away.
The density of icebergs is so high that even icebreakers must avoid them during navigation.
Moreover, icebergs are an important threat to permanent installations on water and to ship navigation in polar regions~\citep{Zhou2019}.
Any long-range autonomous navigation on water would need to monitor icebergs to avoid them.
Progressing toward the poles, we find more sub-domains of the cryosphere, such as ice shelves and glaciers, with the most prominent being ice sheets.
Greenland and Antarctica are the only ice sheets existing on Earth.
The Antarctic ice sheet is subject to extreme cold conditions, requiring a robot to be operational at \SI{-40}{\celsius} for summer deployments, a temperature well below ratings of typical electronics~\citep{Hoffman2019}.
In combination with wind speed exceeding \SI{20}{\m \per \second}, cold will induce stress on the main body of a robotic platform~\citep{Hoffman2019}, affect manoeuvrability~\citep{Shillcutt1999}, crack seals relying on glue or tape letting snow inside~ \citep{Akers2004}, and any supporting equipment relying on small batteries will have a noticeable reduced lifespan~\citep{Stansbury2004}.
In terms of traversability, ice sheets and glaciers are equivalent to large plains subject to two specific hazards.
Firstly, hard snow will be eroded by winds creating waves-shaped ridges, sometimes referred to as sastrugi, which vary from a few centimetres to two metres high.
These obstacles can be high enough to flip a robot going downward or hard to overcome upward due to reduced traction.
Because of their shapes and lack of colours, sastrugi are notoriously hard to detect~\citep{Gifford2009}.
Moreover, these open areas have few structures limiting winds to gain speed, which will compact the snow and even carve blue-ice fields~\citep{Foessel1999}.
The second major hazard is large crevasses that can be hidden under thin layers of snow.
These crevasses threaten humans and vehicles navigating in the area.
Manual monitoring of these deep fractures is typically done using ground-penetrating radars.
Surveying hidden crevasses is another beneficial application for robots in these environments as it is dangerous and slow for humans to detect and mark their location~\citep{Trautmann2009,Lever2013}.
In terms of localization, the lack of vegetation produces large snow-covered plains, where the most useful information for visual-based localization is situated on the skyline~\citep{Barfoot2011}.
Having feature points on the horizon makes it more challenging to estimate the linear motion of a vehicle~\citep{Paton2017}.
Moreover, when navigating at a high latitude, the sun may move following the horizon for a long period.
This kind of motion of the sun cast long and rapidly moving shadows during the day, thus limiting the ability to relocalize in a given environment~\citep{Barfoot2010,Paton2017}.
Ice caps are also permanent ice anchored on the ground but cover a smaller area than ice sheets.
They can mostly be found on top of mountains, but are not limited to high elevation.
With the rapid development of \acp{UAV} combined with advances in photogrammetry, mountains and their snowpack can be surveyed at a faster pace.
However, \citet{Revuelto2021} observed that shadows cast by mountains impact greatly the 3D reconstruction, especially during winter, where the sun has lower incidence angles.
Overall, different sub-domains of the cryosphere are extremely challenging for robotic systems and must be well understood when expecting safe autonomous behaviours.
\subsection{Impact of precipitation on robotics}
The most famous type of precipitation associated with the cryosphere is snow.
In reality, the interaction of cold air and water is complicated and can produce a continuous spectrum of types of precipitation, such as the ones depicted in \autoref{fig:precipitation}.
In continental climates, it is common to see a mix of freezing rain, pellets, and snow during the same storm, especially close to spring and autumn, which can add challenges to perception algorithms.
\begin{figure*}[hbtp]
\begin{center}
\includegraphics[width=\textwidth,trim={0 2px 0 0},clip]{figures/precipitation.pdf}
\caption{Representation of the continuous transition between different types of precipitation from rain to snow.
Yellow represents warm air and white is cold air.
Different proportions of cold and warm air will make the transition between rain, hail, freezing rain, pellet, and snow.
While rain can be drained by the ground, solid precipitation will accumulate on the ground, thus changing rapidly the topology of the surroundings.
This dynamic topology combined with low visibility poses major challenges to robotic systems.}
\label{fig:precipitation}
\end{center}
\end{figure*}
This variety of precipitation is an open problem for the localization of autonomous cars~\citep{Pitropov2020}.
Pellets and snow are known to degrade depth perception by adding noise to stereo and lidar sensing~\citep{Foessel1999}.
For example, a snowflake can trigger a lidar reading at up to \SI{20}{\m}~\citep{Charron2018}.
Millimetre-wave radar is less subjected to dense precipitation, but can still be occluded by wet snow sticking to the sensor optic~\citep{Hong2020}.
Moreover, the potential size of hail can put the integrity of the robot at risk and can cause physical damage to light structures.
In general, freezing rain is known to be dangerous for infrastructure, which can collapse under the weight of the ice rapidly accumulating.
More related to robotics, a thick layer of ice can break antennas used for communication with a robot, attenuate \ac{GPS} signals used for localization, and hugely limit the efficiency of solar panels.
Moreover, the newly created slippery surfaces will perturb path tracking and control algorithms.
Finally, whiteout conditions happen with a combination of snow and high wind or snow and fog.
Sustained high winds combined with snow are called a blizzard, during which it becomes uncertain whether the snow is rising from the ground or falling from the sky.
Throughout a whiteout, such as the one shown in \autoref{fig:whiteout}, light is completely diffused (i.e., not producing any shadows) rendering the ground and the sky indistinguishable.
These conditions reduce considerably the number of features that can be used to localize the robot.
It is not unusual to observe even humans feeling nauseous or subject to dysequilibrium when moving in a whiteout because of conflicting sensory stimulus~\citep{Hausler1995}.
Pilots will typically not fly in these conditions causing seriously delays~\citep{Williams2010} or even cancellations of robotics field deployments~\citep{Bonanno2003}.
Meteorological events, especially in cold regions, are diverse and have a direct impact on autonomous navigation safety.
Any long-term tasks will have to plan for precipitation sporadically limiting the perception of a robot in the cryosphere.
\begin{figure}[hbtp]
\begin{center}
\includegraphics[width=\columnwidth]{figures/whiteout.pdf}
\caption{Example of a perceptual challenge during a whiteout caused by fog and light snow on a mountain.
The lack of distinction between the ground and the sky combined with diffused lights pushes to the limit algorithms aiming at scene interpretation or localization. }
\label{fig:whiteout}
\end{center}
\end{figure}
\section{Key Research Findings}
The current scientific literature is very sparse when it comes to field testing in ice and snow.
Nonetheless, pioneering works that helped gather critical field knowledge about deploying robots in harsh and cold conditions are highlighted here.
Moreover, robots described in this section can also be found in \autoref{fig:zones}, showing how they spread over different sub-domains of the cryosphere.
In December 1992, an eight-legged robot named \emph{Dante} was deployed on the edge of an active volcano in Antarctica~\citep{Wettergreen1993}.
Although its descent was interrupted at an early stage, valuable lessons were learned when deploying this \SI{400}{\kg} walking robot on Mount Erebus and opened new research opportunities for remote exploration missions.
Two years later, a new version named \emph{Dante~II} managed to rappel in another active volcano, Mount Spurr, Alaska~\citep{Bares1999}.
The robot could walk through a mix of snow, ice, rocks, and ashes to transmit images from the crater to volcanologists located in a remote control station.
Developed for more level surfaces, \emph{Nomad} was first validated on an ice shelf close to Patriot Hills Base Camp, Antarctica, in November 1998~\citep{Moorehead1999} before being sent to the periphery of a glacier of the same continent two years later~\citep{Apostolopoulos2000}.
Nomad drove \SI{10.3}{\km} in the harsh weather of Antarctica and could autonomously discover meteorites during its second deployment.
Often forgotten in the literature, modified \emph{Snowcats} using cameras as a guiding system was tested in the Italian Alps~\citep{Broggi2002}.
A Snowcat is a truck-sized vehicle equipped with a pair of tracks and is common on ski slopes to prepare the trails.
In 2002, the modified vehicles were sent to Zucchelli Station, Antarctica, but the weather conditions were too critical to allow further testing~\citep{Bonanno2003}.
This unfortunate event highlights the challenge of field testing in remote sub-domains of the cryosphere.
Moved solely by the wind, the \emph{Tumbleweed Polar Rover} was first tested on the ice sheet of Greenland~\citep{Behar2004} to be finally sent from the Amundsen-Scott South Pole Station, Antarctica, in 2004~\citep{JPL2004}.
The rover traveled around \SI{130}{\km} while transmitting information about its coordinates, temperature, and air pressure through a satellite link.
Following a similar validation sequence, \emph{MARVIN 1} started to experiment with polar conditions in the summer of 2003 in Greenland, before \emph{MARVIN 2} was deployed in Antarctica three years later~\citep{Gifford2009}.
In both cases, the robot demonstrated early potential for seismic and radar remote sensing of the ice sheet.
Meanwhile, \emph{Cool Robot} and its upgraded version \emph{Yeti} were deployed from 2005 to 2011, also on both ice sheets.
These battery-powered robots traveled over \SI{600}{\km} in three consecutive field seasons.
This line of research culminated with deployment at the Old Pole, Antarctica, where ground-penetrating radar was used to survey the remains of an old camp~\citep{Williams2014}.
Finally, during the summer of 2009, a six-wheeled robot with an articulated chassis, named \emph{ROC6}, was deployed near the Haughton Crater, in the Canadian High Arctic~\citep{Barfoot2011}.
Although located at a very high latitude on Devon Island, this deployment was done on exposed permafrost, considered a polar desert, where ground-penetrating radar could be used to survey subsurface structures.
A handful of key observations are recurrent from these deployments.
Except for the Snowcats, all of these deployments were used to push the boundaries of robotics on one side and to validate extraterrestrial exploration solutions on the other.
Both ice sheets were confirmed to have interesting properties common to the frozen Europa moon~\cite{Gifford2009}, while polar deserts share similarities with the surface of Mars~\citep{Barfoot2011}.
Moreover, most of these deployments were done in large open spaces, where \ac{GPS} navigation was meeting the overall navigation accuracy of their missions.
Lastly, deploying a complex system in remote and harsh locations is costly and requires a high level of planning to fulfill scientific goals.
Equipment needs to be shipped well in advance~\citep{Barfoot2010}, which puts pressure on the storage capability of fragile equipment in cold.
Moreover, the lack of infrastructure limits repair and modification to a minimum~\citep{Morad2020}.
\section{Examples of Application}
Beyond mockups for planetary exploration, robots being robust to snow and ice can support many applications.
Robotics can support Earth sciences in remote locations, where manual sampling of the environment can be tedious and sometimes dangerous.
Surveys using autonomous vehicles can bring higher spatiotemporal resolution and support many interdisciplinary sciences.
Past efforts were related to volcanology~\citep{Wettergreen1995}, mineralogy~\citep{Apostolopoulos2000}, atmospheric science~\citep{Behar2004}, seismology~\citep{Gifford2009}, glaciology~\citep{Ray2020}, and geomorphology~\citep{Barfoot2011} to name few.
Being able to move between two stations autonomously can support good transportation in remote locations.
Ice roads are often used to travel in winter, where building permanent infrastructure would be too expensive.
These roads, along with the main route between McMurdo Station and the Amundsen–Scott South Pole Station, need to be inspected regularly for crevasses or ice thickness.
Autonomous vehicles already demonstrated the potential to reduce risks associated with such tasks~\citep{Lever2013}.
Moreover, snow also needs to be regularly removed from roads to be accessible by less specialized vehicles.
Snow removal is an essential service in many inhabited areas with large snow cover and can be costly.
Snow removal is also a time-critical operation for many airports.
Also relying on frozen ground to avoid sinkage of the heavy machinery, forestry vehicles often operate during winter.
To sum up, applications related to snow and ice are not limited to expeditions in remote locations to show befits of automation.
\section{Future Directions for Research}
Self-driving cars are taking a larger place in the public sphere and are expected on our roads shortly.
Although great technological advances have occurred during the last decade, autonomous driving during winter still raises reliability concerns.
More effort will be needed to make self-driving cars safe to use in continental climates and datasets on this topic are on the rise~\citep{Pavlov2019,Pitropov2020,Barnes2020}.
It is interesting to point out that, up to now, no autonomous navigation algorithm was demonstrated during whiteout conditions.
Along with low visibility snowstorms, extreme meteorological events are difficult to plan for in a scientific experiment, thus making it difficult to validate the safety of robots in these conditions.
As it is a source of major scientific funding, missions preparing for space exploration are expected to continue~\citep{Reid2020}.
In the face of climate change, building finer models of the environment by using autonomous platforms to carry Earth science sensors in a harsh environment will also remain and even gain in importance.
Finally, roboticists will progress to extend autonomy solutions for different parts of the cryosphere.
For example, a research program named SNOW (Self-driving Navigation Optimized for Winter) investigates autonomous navigation solutions in subarctic forests \citep{Baril2020,Baril2022}.
Another interesting initiative is the 3D reconstruction of icebergs using the autonomous underwater vehicle \emph{Scolum}~\citep{Zhou2019}.
Finally, although early deployment in remote areas of the cryosphere relied on \acp{UGV}, \acp{UAV} were recently used to reconstruct snow-covered alpine mountains~\citep{Revuelto2021}.
Another interesting direction is the exploitation of the \emph{midnight sun} to extend the mission time of a solar-powered \ac{UAV} to monitor the calving of glaciers~\citep{Jouvet2018}.
At a higher level, four key challenges need to be addressed to enhance autonomy in snow and ice.
First, deployment in a remote location puts pressure on long-range communication in a context where there is little infrastructure to support deployment.
Second, energy economy or energy harvesting solutions need to be put forward for long-term autonomous missions.
Currently, gas engines may interfere with environmental sampling, while battery performances degrade rapidly in cold.
Third, solutions for localization in rapidly changing environments and high precipitation need to be improved.
For example, snowbanks swiftly move with the wind, snow covers heavily deform under the weight of vehicles, and blizzard drastically degrades the perception capability of a robot.
Finally, mobility on ice and snow is notoriously difficult.
As an illustration, turning on the spot in deep snow will often stall a robot resulting in a vehicle digging itself even more as controllers try to compensate for an increasing tracking error.
Moreover, path tracking algorithms need to cope with a larger spectrum of ground friction levels than what is typically tested in dry conditions.
At a micro level, the ground can be a variable composition of gravels, ice, and snow continuously changing the traction force of the robot.
Surely, the presented research trends combined with these ongoing challenges make \emph{robotics in snow and ice} an exciting research field for many years to come as potential applications evolve.
There is increasing demand to better understand rapid climate change, which will require higher spatiotemporal data to be collected in the cryosphere.
Moreover, there is rapidly growing investment in space ranging from planetary exploration to building habitats, bringing more attention to Earth testbeds.
\Urlmuskip=0mu plus 1mu\relax
\def\do\/\do-{\do\/\do-}
\bibliographystyle{spbasic}
|
\section{Spin- and chirality-related quantities}
In the simulation, we identify magnetic phases from the spin structure factor, magnetization, spin scalar chirality, and skyrmion number in each sublattice.
The spin structure factor for the sublattice $\alpha=\mathrm{A},\mathrm{B}$ is defined as
\begin{align}
S^\mu(\alpha,\bm{q}) = \left\langle\frac{1}{N}\sum_{j,k}S^\mu_{\alpha j}S^\mu_{\alpha k} e^{i\bm{q}\cdot(\bm{R}_{\alpha j}-\bm{R}_{\alpha k})}\right\rangle,
\end{align}
where $\mu=x,y,z$, $\bm{R}_{\alpha j}$ is the position vector at site $j$ on the sublattice $\alpha$, $N$ is the number of the unit cell, and $\langle\cdots\rangle$ is the average over the Monte Carlo samples.
We also calculate the in-plane spin structure factor, $S^\perp(\alpha,\bm{q})=S^x(\alpha,\bm{q})+S^y(\alpha,\bm{q})$.
The magnetization for the sablattice $\alpha$ is defined as
\begin{align}
M_\alpha = \left\langle\frac{1}{N}\sum_{j}S^z_{\alpha j}\right\rangle.
\end{align}
The local spin scalar chirality of the triangle on the sublattice $\alpha$ is defined as
\begin{align}
\chi_{\alpha \bm{r}}= \bm{S}_{\alpha j}\cdot( \bm{S}_{\alpha k}\times \bm{S}_{\alpha l}),
\end{align}
where the position vector $\bm{r}$ represents the triangle center and the triangle consists of $(j,k,l)$ sites labeled in the counterclockwise order.
The uniform spin scalar chirality of the sublattice $\alpha$ is given by
\begin{align}
\chi_\alpha^\mathrm{sc} = \left\langle \frac{1}{N}\sum_{\bm{r}} \chi_{\alpha\bm{r}} \right\rangle.
\end{align}
The skyrmion density at the triangle $\bm{r}$ on the sublattice $\alpha$~\cite{BERG1981412} is defined as
\begin{align}
\tan\left(\frac{\Omega_{\alpha \bm{r}}}{2}\right) = \left[\frac{\bm{S}_{\alpha j}\cdot( \bm{S}_{\alpha k}\times \bm{S}_{\alpha l})}{1+\bm{S}_{\alpha j}\cdot\bm{S}_{\alpha k}+\bm{S}_{\alpha k}\cdot\bm{S}_{\alpha l}+\bm{S}_{\alpha l}\cdot\bm{S}_{\alpha j}} \right].
\end{align}
Then, the skyrmion number for the sublattice $\alpha$ is given by
\begin{align}
\label{eq:Nsk}
N^\alpha_\mathrm{sk} = \frac{1}{4\pi N_\mathrm{m}}\left\langle \sum_{\bm{r}} \Omega_{\alpha \bm{r}} \right\rangle,
\end{align}
where $N_\mathrm{m}$ is the number of the magnetic unit cell.
\section{Trivial phases I-V}
In Fig.~3, we find the field-induced trivial phases denoted as I-V with $N^\mathrm{tot}_\mathrm{sk}=N^\mathrm{stagg}_\mathrm{sk}=0$.
We show their spin configurations, chirality configurations, and spin structure factors in Fig.~\ref{fig:3q}.
Since all the phases exhibit the triple-$Q$ peaks at $\bm{Q}_1=(0,\pi/3)$, $\bm{Q}_2=(-\sqrt{3}\pi/6,-\pi/6)$, and $\bm{Q}_3=(\sqrt{3}\pi/6,-\pi/6)$ in the spin structure factor, they are regarded as the triple-$Q$ states.
The trivial spin configurations I and II are characterized by the double-$Q$ in-plane spin structure factor and the single-$Q$ out-of-plane one, as shown in Figs.~\ref{fig:3q}(a) and \ref{fig:3q}(b), respectively, where the peak positions depend on the sublattices.
The trivial state I is characterized by the sublattice-dependent in-plane spin structure factor and sublattice-independent our-of-plane one: $S^\perp({\rm A},\bm{Q}_1)>S^\perp({\rm A},\bm{Q}_2)$ and $S^z({\rm A},\bm{Q}_3)$ for the sublattice A and $S^\perp({\rm B},\bm{Q}_2)>S^\perp({\rm B},\bm{Q}_1)$ and $S^z({\rm B},\bm{Q}_3)$ for the sublattice B.
The trivial phase II shows the sublattice-dependent in-plane and out-of-plane spin structure factors: $S^\perp({\rm A},\bm{Q}_3)>S^\perp({\rm A},\bm{Q}_2)$ and $S^z({\rm A},\bm{Q}_1)$ for the sublattice A and $S^\perp({\rm B},\bm{Q}_1)>S^\perp({\rm B},\bm{Q}_2)$ and $S^z({\rm B},\bm{Q}_3)$ for the sublattice B.
Meanwhile, the trivial spin configurations III-V show the sublattice-independent in-plane and out-of-plane spin structure factors:
$S^\perp(\alpha,\bm{Q}_1)=S^\perp(\alpha,\bm{Q}_2)$ and $S^z(\alpha,\bm{Q}_3)$ for $\alpha=\mathrm{A},\mathrm{B}$ in the phase III,
$S^{\perp(z)}(\alpha,\bm{Q}_1)=S^{\perp(z)}(\alpha,\bm{Q}_2)=S^{\perp(z)}(\alpha,\bm{Q}_3)$ for $\alpha=\mathrm{A},\mathrm{B}$ in the phase IV,
and $S^\perp(\alpha,\bm{Q}_2)>S^\perp(\alpha,\bm{Q}_1)=S^\perp(\alpha,\bm{Q}_3)$ and $S^z(\alpha,\bm{Q}_1-\bm{Q}_3)>S^z(\alpha,\bm{Q}_2)$ for $\alpha=\mathrm{A},\mathrm{B}$ in the phase V.
\begin{figure*}[h!]
\begin{center}
\includegraphics[width=1.0\hsize]{supplementfigS1.pdf}
\caption{
\label{fig:3q}
First (fourth) column: Spin ($\bm{S}_{\alpha i}$) and chirality ($\chi_{\alpha \bm{r}}$) configurations on the sublattice A (B) of
(a) the trivial state I at $\Theta=\pi/24$ and $H=0.3$,
(b) the trivial state II at $\Theta=\pi/24$ and $H=1$,
(c) the trivial state III at $\Theta=5\pi/48$ and $H=1$,
(d) the trivial state IV at $\Theta=\pi/6$ and $H=1$,
and (e) the trivial state V at $\Theta=\pi/4$ and $H=1$.
The arrows, contours of arrows, and contours of circles show the $xy$ spin component, $z$ spin component, and spin scalar chirality, respectively.
Second and third (fifth and sixth) columns: The in-plane [$S^\perp(\alpha,\bm{q})$] and out-of-plane [$S^z(\alpha,\bm{q})$] spin structure factors for the sublattice A (B) in momentum space, respectively.
The circle and square highlight the dominant and subdominant peaks, respectively.
The hexagons with a solid line show the first Brillouin zone.
The $\bm{q}=\bm{0}$ component is removed for better visibility.
}
\end{center}
\end{figure*}
\section{Skyrmion crystals in the models with different form factors}
We consider the effective spin model with different form factors for the bilinear and biquadratic interactions, which is given by
\begin{align}
\mathcal{H}^\mathrm{eff}&=-2J\sum_{\eta}\sum_{\alpha,\beta}X^{\alpha\beta}\bm{S}_{\alpha \bm{Q}_\eta}\cdot\bm{S}_{\beta -\bm{Q}_\eta} \nonumber\\
&+2\frac{K}{N}\sum_{\eta}\left(\sum_{\alpha,\beta}X^{\alpha\beta}_\mathrm{K}\bm{S}_{\alpha \bm{Q}_\eta}\cdot\bm{S}_{\beta -\bm{Q}_\eta}\right)^2 \nonumber \\
&-H\sum_{\alpha,i}S_{\alpha i}^z.
\end{align}
Here, $X^{\alpha\beta}$ and $X^{\alpha\beta}_\mathrm{K}$ are the form factors for the RKKY interaction and biquadratic interaction, respectively.
Although we consider the case of $X^{\alpha\beta}=X^{\alpha\beta}_\mathrm{K}$ in the main text for simplicity, the form factors are usually different.
In this section, we show that the SkX phases in Fig.~3 also apper even for $X^{\alpha\beta}\neq X^{\alpha\beta}_\mathrm{K}$.
Specifically, we consider three different $X^{\alpha\beta}_\mathrm{K}$ while fixing $X^{\rm AA}\equiv \cos^2\Theta$ and $X^{\rm AB}\equiv -\sin^2\Theta$ to cover various situations:
(i) $X^{\rm AA}_\mathrm{K}\equiv \cos^2\Theta$ and $X^{\rm AB}_\mathrm{K}\equiv \sin^2\Theta$,
(ii) $X^{\rm AA}_\mathrm{K}\equiv \sin^2\Theta$ and $X^{\rm AB}_\mathrm{K}\equiv -\cos^2\Theta$,
and (iii) $X^{\rm AA}_\mathrm{K}\equiv \sin^2\Theta$ and $X^{\rm AB}_\mathrm{K}\equiv \cos^2\Theta$.
By following the same manner of the simulation in the main text, we find the FC SkX I, AFC SkXs I and II, and FerriC SkX in various $\Theta$ and $H$.
For example, we show critical magnetic fields for multiple topological transitions in Table~\ref{tab:Hc}, where the AFC SkX II appears at low fields including $H=0$, the FerriC SkX appears above $H_{{\rm c} 1}$, the AFC SkX I appears above $H_{{\rm c} 2}$, the FC SkX I appears above $H_{{\rm c} 3}$, and other trivial states appear above $H_{{\rm c} 4}$ while increasing $H$.
In this way, a variety of the SkXs discussed in the main text are ubiquitously stabilized in the effective spin model with the antiferromagnetic intersublattice RKKY interaction and the positive biquadratic interaction irrespective of the form factor $X^{\alpha\beta}_\mathrm{K}$.
\begin{table}[h!]
\caption{\label{tab:Hc}
Critical magnetic fields $H_{{\rm c} 1}$, $H_{{\rm c} 2}$, $H_{{\rm c} 3}$ and $H_{{\rm c} 4}$ for multiple topological transitions at specific $\Theta$.
Here, we find the AFC SkX II at $0\le H < H_{{\rm c} 1}$, the FerriC SkX at $ H_{{\rm c} 1}< H < H_{{\rm c} 2}$, the AFC SkX I at $ H_{{\rm c} 2}< H < H_{{\rm c} 3}$, and the FC SkX at $ H_{{\rm c} 3}< H < H_{{\rm c} 4}$.
}
\begin{ruledtabular}
\begin{tabular}{llllll}
& $\Theta$ & $H_{{\rm c}1}$ & $H_{{\rm c}2}$ & $H_{{\rm c}3}$ & $H_{{\rm c}4}$ \\ \hline
case (i) & $5\pi /48$ & 0.3875 & 0.4125 & 0.4825 & 0.8625 \\
case (ii) & $\pi /6$ & 0.2625 & 0.3375 & 0.4625 & 0.9125\\
case (iii) & $7\pi /48$ & 0.3375 & 0.3625 & 0.5375 & 0.9625
\end{tabular}
\end{ruledtabular}
\end{table}
|
\section{Some appendix section}
\fromplm{TODO}
\section{Introduction}
\label{sec:intro}
Simulations have long been used to understand emergent phenomena in complex many-body systems. Special-purpose analog simulators trade-off the generality of digital implementations for either scalability or access to regimes challenging for digital computers. In this regard, photonic analog simulators~\cite{Peruzzo2010,Lahini2009,Screiber2012,Lin2016,Harris2017,Lin2018,Muniz2019,Wang2020} complement developments in platforms like superconducting circuits~\cite{Hung2021,Karamlou2022} and ultracold atoms~\cite{An2017,Periwal2021} by enabling, in principle, extremely large-scale simulations. Photonic simulation has a long history and has led to the discovery of a variety of phenomena challenging to realize in conventional condensed-matter systems, such as topological phase transitions~\cite{Ozawa2019,Lustig2019,Yang2019,Dutt2020,Lustig2021,Wang2021,Leefmans2022} and non-Hermitian exceptional points~\cite{Regensburger2012,Eichelkraut2013,Hodaei2017,Weidemann2021,Xia2021}, which in turn has led to new photonic devices with applications far beyond basic physical science~\cite{Bandres2018,Solnyshkov2018,Hokmabadi2019,Zeng2020}.
Although telecommunication technologies routinely utilize the high-bandwidth inherent to optics, harnessing the frequency parallelism of light for large-scale analog simulation has largely remained unexplored. One promising approach is to implement synthetic frequency dimensions~\cite{Bersch2009,Schwartz2013,Ozawa2016,Yuan2016,Bell2017,Qin2018,Dutt2019,Hu2020,Dutt2020,Li2021,Chen2021,Wang2021,YuanDutt2021,Zhao2021,Balcytis2022,Dutt2022}, in which optical frequency modes are mapped to lattice sites to perform bosonic analog simulations. Simulators using synthetic frequency dimensions have been shown to be versatile, implementing synthetic electric and magnetic fields~\cite{Bersch2009,Miyake2013,Yuan2015,Yuan2016,Qin2018,QinYuan2018,Peterson2019,Lee2020,DErrico2021,Li2021,Chen2021,Englebert2021}, non-Hermitian coupling~\cite{WangDutt2021,Wang2021}, and nonlinear interactions~\cite{Englebert2021,Tusnin2020}. So far, however, these have suffered from restricted system sizes, limited programmability, or lacked techniques for probing the underlying dynamics.
Here, we demonstrate a frequency-mode-based platform that can simulate transport of arbitrary excitations in planar and non-planar optical lattices with up to 100,000 sites -- orders of magnitude greater than achieved previously in programmable simulators~\cite{YuanDutt2021}. By pursuing a dense spectrum with MHz mode spacing, we leverage developments in both optical frequency combs and high-frequency optoelectronics to manipulate and probe a large number of optical frequency modes in a ring cavity. These technologies enable arbitrary frequency encoding of input states, highly programmable lattices, and tools to read out the dynamics of our simulator.
The class of Hamiltonians that our system is able to simulate is
\begin{equation}
H = \sum_{i < j} J_{i-j} a^{\dagger}_i a_j + \text{H.c.}
\label{eq:hamiltonian}
\end{equation}
Hamiltonians in this class describe non-interacting bosons on lattices with translational invariance. $a^{\dagger}_l$ and $a_l$ are, respectively, the bosonic creation and annihilation operators for the $l$th lattice site. The lattice geometries are defined by the tunneling rates $J_k = \vert J_k \vert e^{i\phi_k}$, which encode the coupling of sites a distance $k$ apart. In the synthetic-frequency-dimensions approach \cite{YuanDutt2021} that we adopt, the lattice operators $a_l$, $a_l^{\dagger}$ are associated with the $l$th frequency mode of an optical cavity; the cavity modes are spaced apart in frequency by $\Omega$ (the free spectral range of the cavity). The tunneling rates between sites $\left( J_k \right)_{k=1,2,3,\ldots}$ are physically realized using a phase modulator within the optical cavity (Fig.~\ref{fig:figure1}A) -- intuitively, the modulator creates optical sidebands at the frequencies contained in the modulation signal $v(t)$ and it is these sidebands that can cause coupling between cavity modes. By setting the frequency components in $v(t)$ appropriately, different lists of tunneling amplitudes $\left( J_k \right)$ can be programmed (Fig.~\ref{fig:figure1}A), which in turn realize different lattice geometries (Fig.~\ref{fig:figure1}C). In our simulator, there is an additional term in the Hamiltonian accounting for the gain/loss balance, but this is kept close to zero (see Supplementary Section~\ref{sec:setup_modeling}).
The main goal of the photonic simulator we present in this work is to be able study the transport of a variety of input excitations in any Hamiltonian in the class defined by Eq.~\eqref{eq:hamiltonian}, where the complex parameters $J_k$ can be programmed arbitrarily -- this allows us to study a diversity of different lattices, including ones that are multidimensional.
\begin{figure*}[h]
\centering
\includegraphics[width=.95\textwidth]{ Figure1_v2.pdf}
\caption{ \textbf{ Simulations of large-scale bosonic transport with programmable photonic simulator. } (A) Dynamic modulation of a fiber ring resonator couples frequency components of the intracavity field in each roundtrip, represented in the basis of the frequency modes as $U = F^{\dagger} e^{iV} F$. Here, $V \propto \text{diag} (\vec{v})$ is the diagonal voltage operator defining the modulation signal $\vec{v} = (v(t_1),v(t_2),\ldots,v(1/\Omega))$, and $F$ is the discrete Fourier transform. The components of the voltage modulation define the coupling (bottom). By modulating at multiples of the mode spacing ($\Omega$), we only couple long lived modes of the cavity, allowing for injected signals to propagate in frequency for many multiples of the roundtrip time. (B) Engineered long-range coupling maps the one-dimensional spectrum to $L \times L$ two-dimensional lattice with twisted boundary conditions. As $L$ grows large, the lattice approaches a smooth 2D plane. (C) A set of voltage signals defining lattices in 2 and 3 dimensions (top; see main text and Supplementary~Fig.~\ref{fig:modulation_schemes} for details), and the response to a single frequency drive (single-site) for a twisted 2D square lattice with over 20,000 lattice sites compared with tight-binding simulations of a 2D square lattice (bottom). }
\label{fig:figure1}
\end{figure*}
\begin{figure*}[h]
\centering
\floatbox[{\capbeside\thisfloatsetup{capbesideposition={right,top},capbesidewidth=8cm}}]{figure}[\FBwidth]
{
\captionsetup{justification=raggedright}
\caption{
\textbf{Optical band-structure measurements of 2D and 3D lattices. } (A) The band structure for a twisted 2D square lattice is measured from the time-domain response of the cavity to scanning single-frequency injection as a function of the detuning $\Delta$~\cite{Dutt2019}, here demonstrated using a toy example with linear lattice size $L = 3$. This time-trace output is sliced up into chunks of length $L$, allowing the reconstruction of a full 2D band structure measured in a single shot (see Supplementary Section~\ref{sec:supp_bs}). (B) Reconstructed band structure for a 2D square lattice with large $L$, comparing analytic (left) with experimental results (right) for $L = 100$. As $L$ grows large, the effect of the twisted boundary condition in the band structure becomes negligible, and the measured band structure approximates that of a regular 2D square lattice. (C) Data of the full bandstructure plotted along slices that connect special points of the Brillouin zone, compared with analytic results for a true 2D square lattice (black). These points, highlighted in the bottom-left, denote locations in momentum space with high symmetry. The density of states $g(E)$ is directly measured by summing the time-domain response (right). Band structures and density of states for (D) 2D triangular lattice ($L = 100$), (E) 3D square lattice ($L = 28$), and (F) 3D hexagonal lattice ($L = 28$).
}\label{fig:figure2}}
{
\includegraphics[width=0.45\textwidth]{Figure2_v5.pdf}
}
\end{figure*}
The Hamiltonian in Eq.~\eqref{eq:hamiltonian} describes a 1D lattice (with nonlocal couplings) but we can implement effective higher-dimensional lattices by suitably programming the couplings $\left( J_k \right)$ to reflect the local geometry. For example, an effective 2D square lattice can be realized by coupling nearest-neighbors and $L$th nearest-neighbors, i.e., $\left( J_k \right)_{k=1,2,3,\ldots} = \left( J_1,0,\ldots,0,J_L,0,\ldots \right)$ (Fig.~\ref{fig:figure1}B). While this effectively produces a 2D lattice with a twisted boundary condition~\cite{noteforTBCs}, a local excitation with finite lifetime will be insensitive to this twist for a sufficiently large system size. The vanishing effects of the twisted boundary condition can be seen in the steady-state response to a single-site excitation in the comparison shown in Fig.~\ref{fig:figure1}C between our experiment and simulations of a true 2D tight-binding lattice.
The frequency-multiplexed platform has a convenient encoding of reciprocal space for lattice systems. In the mapping from lattice sites to frequency modes, time maps to momentum~\cite{Dutt2019}. In the same vein, since the Fourier components of the modulation signal define the connectivity, the modulation signal in the time domain defines the band structure. For a 1D lattice, this correspondence is exact: the modulation signal $v(t) = -V_0 \cos(\Omega t)$ couples nearest-neighbor modes, implementing a 1D tight-binding chain with band structure $E(k) = -J \cos (ka)$. Here the lattice spacing $a$ is identified with the mode spacing $\Omega$, and momentum $k$ with time $t$. More generally, the action of phase modulation on the frequency modes can be expressed as a unitary operator $U = F^{\dagger} e^{i V} F$, where $F$ is the discrete Fourier transform, and $V \propto \text{diag}(\vec{v})$ is a diagonal matrix whose values are proportional to the voltage signal $\vec{v} = (v(t_1),v(t_2),\ldots,v(1/\Omega))$. In our simulator, the operator $U$ implements the time evolution defined by the Hamiltonian in Eq.~\eqref{eq:hamiltonian}. Thus, the modulation signal $v(t)$ defines the time evolution in a diagonal basis, and therefore encodes energy eigenvalues of the lattice, i.e., the band structure. As a consequence, this permits us to encode arbitrary lattices that have with a single-band band structure. Additionally, it provides us direct access to momentum-space lattice measurements \cite{Dutt2019}.
\begin{figure*}[h]
\centering
\includegraphics[width=.9\textwidth]{ Figure3_v5.pdf}
\caption{ \textbf{Input state preparation. } (A) Scheme for preparing arbitrary input states: a single frequency tone is modulated with a signal encoding both amplitude and phase of a given state, producing symmetric sidebands (orange spectrum). The initial tone and the unwanted sideband is then rejected with a bandpass filter (red envelope), leaving only the positive sidebands which are sent into the cavity. (B) Experimental measurements of input states for increasing number of modes programmed in the input signal for: a standing wavepacket, an angular wave enveloped with a Gaussian centered at zero, higher angular state enveloped with an offset radial Gaussian, and the Cornell University logo. The steady state outputs of these states for a 2D $L \times L$ square lattice are shown to the right, along with comparision with theory. (C) We excite momentum eigenstates of a 2D square lattice with momenta in various directions enveloped with a Gaussian. Left shows the representation of the input state in momentum space $\vec{k} = (k_x,k_y)$, right shows the experimental steady state in position space $\vec{x} = (x,y)$. Here, local momentum eigenstates are continuously excited at the center, and propagate with a well-defined momentum before decaying. }
\label{fig:figure3}
\end{figure*}
\begin{figure}[h!]
\centering
\includegraphics[width=.95\textwidth]{ Figure4_v2.pdf}
\caption{ \textbf{Time-reversal-symmetry breaking in a 2D triangular lattice due to an effective gauge field. } (A) Complex hopping terms induces a nonzero local magnetic flux within a plaquette of a triangular lattice. (B) The introduction of the magnetic field breaks time reversal symmetry, as can be seen in the asymmetry of the K and K’ points in the band structure after performing Pierelis substitution (bottom). (C) Measured steady state spectral response due to a single-site injection under the influence of the synthetic local magnetic field. The presence of the synthetic field leads to a departure from 6-fold symmetry to 3-fold symmetry in the transport. Experimental data (left) is compared with simulations (right).}
\label{fig:figure4}
\end{figure}
To extend the above momentum-to-time analogy to 2D and 3D, we require the number of modes to grow large enough that the finite-size effects from the twisted boundary condition vanish. For small $L$, slices of the band structure along the slow axis (i.e., the axis corresponding to transport along nearest-neighbors) suffer from an asymmetry near the boundaries of the Brillouin zone (Fig.~\ref{fig:figure2}A). This is due to the twisted boundary conditions, which makes the two directions no longer independent, since $L$ hops along $\Omega$ reaches will reach the same position as a single hop along $L\Omega$. Concretely, the asymmetry in the band structure can be seen by comparing the two-tone signal we use to generate a 2D lattice, $v(t) = -2V_0 \cos (\Omega t) - 2V_0 \cos( L\Omega t)$, and a true 2D tight-binding lattice with nearest-neighbor hopping, which has a band structure $E(\vec{k}) = -2J \cos(k_x a) - 2J \cos(k_y a)$. The latter has two independent reciprocal lattice vectors, $k_x$ and $k_y$. For $L \gg 1$ however, we can rely on a separation of timescales and treat $\Omega' = L\Omega$ as an effective independent degree of freedom. This approach can be readily extended to higher-dimensional lattices, e.g. for a 3D square lattic, $\Omega$, $L\Omega$, and $L^2\Omega$ are the independent degrees of freedom.
Fig.~\ref{fig:figure2}A outlines how we use the methods introduced in Ref.~\cite{Dutt2019} to measure the band structure of 2D lattice in a single-shot, then slice up the measured band structure in periods of $T_{\text{fast}} = 1/L\Omega$ to reconstruct the 2D full band structure. See Supplementary Section~\ref{sec:supp_bs} for full details on this reconstruction. As $L \rightarrow \infty$, the band structure of our effective 2D square lattice approaches that of a regular 2D square lattice, as seen when comparing Figs.~\ref{fig:figure2}A and B. Slices through high symmetry points of the full band structure are shown in Fig.~\ref{fig:figure2}C-F for a 2D square, 2D triangular, a 3D simple cubic, and a 3D hexagonal lattice, along with the respective density of states for each. Theoretical curves for ordinary tight-binding lattices are shown in black.
High-bandwidth telecommunications optoelectronics enable the study of transport in our platform for arbitrary input states. Our scheme is enabled by 12-GHz electro-optic modulation, summarized schematically in Fig.~\ref{fig:figure3}A. This technique allows us to specify the amplitude and phases of input excitations for up to around 4000 lattice sites, limited primarily by a bandpass filter used in our implementation (see Supplementary Section~\ref{sec:input_state_prep} for details). Figure~\ref{fig:figure3}B shows experimental measurements of various input states, including standing wavepacket eigenstates, angular wavepackets and a Cornell `C'. The right column displays their respective steady-state response. With full control over both amplitude and phase, we are able to excite states with net momentum. Fig.~\ref{fig:figure3}C shows the steady state response of momentum eigenstates of a 2D square lattice enveloped with a Gaussian for a discrete set of nonzero input momenta. The momentum of each input state is labeled by its respective momentum distribution within the Brillouin zone, shown in the left column of Fig.~\ref{fig:figure3}C.
\begin{figure*}[h]
\centering
\includegraphics{ Figure5.pdf}
\caption{\textbf{Simulations of bosonic transport in a tree-like geometry with a graph comprising over 100,000 sites.} (A) Non-equilibrium correlation measurements for a 1D chain with non-local interactions, characterized by the degree of locality $s$ (see Eq. \eqref{eq:tree}). As $s$ is tuned from $-1$ (left) to $+1$ (right), the correlations transition from locally decaying to treelike \cite{Periwal2021}. Top row is experimental data, bottom row is theory. The lattice cartoons on the left and right schematically show the coupling form for a single lattice site (position 0). (B) Optical spectrum measurement for response to a single site injection with $s = 0.5$, showing measurable steady-state population in $>$ 100,000 lattice sites. (C-E) RF spectrum measurement showing the lattice occupation with single-site resolution of the zoom-in in the full optical spectrum, comparing with simulations for windows of 20 GHz (C), 500 MHz (D), and 100 MHz (E). Theoretical results have been inverted for clarity.}
\label{fig:figure5}
\end{figure*}
By programming the phases and detunings of the coupling Hamiltonian (Eq.~\eqref{eq:hamiltonian}), we implemented synthetic magnetic and electric fields respectively (see Supplementary Section~\ref{sec:supp_results} for measurements for synthetic electric fields) \cite{Lee2020,DErrico2021,Li2021,Chen2021,Bersch2009,Yuan2016,Qin2018,QinYuan2018,Miyake2013,Yuan2015,Peterson2019}. Figure~\ref{fig:figure4} shows the effect of a synthetic gauge field applied to a 2D triangular lattice, giving rise to a global zero magnetic field, but nonzero local magnetic field. By adding a relative phase along nearest neighbor hoppings, a nonzero magnetic flux going around each plaquette of the triangular lattice is induced. Shown in Fig.~\ref{fig:figure4}B, the addition of this field breaks time-reversal symmetry, which for the triangular and honeycomb lattices, maps the $K$ to $K'$ points \cite{Haldane1988}. This results in a reduction of a 6-fold symmetry to a 3-fold symmetry in the transport of injected light, where propagation of light is prohibited in certain directions, as shown in the heterodyne measurements of the steady-state density in a triangular geometry with 10,000 lattice sites (Fig.~\ref{fig:figure4}). This time-reversal-symmetry breaking with local non-zero fields is one key ingredient in observations of the quantum valley Hall effect seen in honeycomb lattices~\cite{Haldane1988,Mak2014,JimnezGaln2020}.
In addition to lattices found in traditional condensed-matter systems, our photonic simulator is capable of simulating systems not realizable in crystalline materials. Systems with non-planar connectivity are particularly interesting given their realization in solid state systems are impractical, yet contain rich physics.
~\textcite{Periwal2021} recently experimentally demonstrated a simulation of a graph with exotic long-range connectivity given by
\begin{equation}
J_{i - j} \propto \begin{cases}
\vert i - j \vert ^s & \vert i - j \vert = 2^n, n \in \mathbb{Z}^+\\
0 & \text{otherwise}.
\end{cases}
\label{eq:tree}
\end{equation}
This describes a system that can be continuously changed, using the parameter $s$, from an Archimedean-geometry regime in which correlations between sites decay with Archimedean distance $\vert i - j \vert $, to a non-Archimedean-geometry regime in which the correlations between sites have a treelike structure. The hierarchical geometry of this treelike system is a toy model for $p$-adic AdS/CFT correspondence~\cite{Bentsen2019} studied
previously using atomic ensembles in an optical lattice with 16 sites~\cite{Periwal2021}. We experimentally show this transition in Fig.~\ref{fig:figure5}A in the measurements of correlations of the lattice as $s$ is tuned (see Supplementary~Section~\ref{sec:supp_results} for details on the correlation measurements and another example with a 1D lattice). Near the transition, at $s = 0.5$, the lattice exhibits both strong local and nonlocal connectivity, resulting in dense yet extremely large lattices, shown in both optical spectrum measurements in Fig.~\ref{fig:figure5}B as well as RF spectrum measurements in Fig.~\ref{fig:figure5}C.
\section{Discussion}
\label{sec:discussion}
While some graphs, such as the tree-like example depicted in Fig.~\ref{fig:figure5}, result in steady-state occupations that span 100,000 or more lattice sites, quantifying the absolute size of our simulator requires some nuance. On one hand, based on the dispersion and bandwidth of the elements inside the cavity, we believe the lattices we simulate span several THz, corresponding to millions of lattice sites. On the other hand, as in real systems, local excitations in locally connected lattices will rarely explore most of this large lattice before their amplitude decays below the noise floor of our detectors. In these lattices, the steady-state response of a local injection is detectable at most $10^4$ lattice sites in the vicinity of the injected wavepacket.
The examples here cover only a narrow swath of the bosonic physics that can be simulated with frequency-domain coupling of photonic modes. By optimization of intracavity dispersion, loss, and power, observing dynamics on lattices spanning millions or more sites is feasible. By adding intracavity amplitude modulators, non-Hermitian Hamiltonians can be realized~\cite{Wang2021}. By adding multiple spatial modes or coupled cavities, multi-band physics~\cite{Dutt2020}, defects, and hard lattice edges may be simulated~\cite{Dutt2022}, enabling the study of much more complex topological phenomena. By varying modulations over multiple cavity periods, dynamical, non-equilibrium phases~\cite{Battiston2021,Disa2021} and other transient phenomena may be realized. In unmodulated cavities, Kerr nonlinearities give rise to locked combs defined by dissipative cavity solitons~\cite{Kippenberg2018}. Intermodal interactions lead to four-mode interactions~\cite{Tusnin2020,Englebert2021}, which could be systematically programmed via reconfigurable intracavity dispersion, spectral loss, and/or additional mode families~\cite{Wright2017,Karpov2019}. Such platforms would allow new regimes of intricately tailored frequency combs, and would permit photonic frequency-mode simulators to explore the physics of models with higher-order interactions, a regime which remains a challenging frontier of network science~\cite{Battiston2021}. Photonic simulators have been demonstrated to be robust platforms for exploring condensed-matter phenomena, some of which had previously been inaccessible. Frequency-domain photonic simulators benefit from programmability and the potential for large scale, with near-term prospects to scale to sizes that would be impractical to simulate with conventional computers, and in turn may enable the study of emergent behavior that cannot currently be explored in any setting.
\section*{Data and code availability}
All data generated and code used in this work is
available at: \url{https://doi.org/10.5281/zenodo.6959554}
\section*{Author Contributions}
A.S., L.G.W. and P.L.M. developed the concept. A.S. and L.G.W. built the experimental setup, with early contributions from H.K.D. A.S. performed the experiments, the data analysis, and the numerical simulations (theory). P.F.W. performed experimental data collection. A.S., L.G.W. and P.L.M. wrote the manuscript. L.G.W. and P.L.M. supervised the project.
\section*{Acknowledgements}
P.L.M. gratefully acknowledges financial support from a David and Lucile Packard Foundation Fellowship, and also acknowledges membership of the CIFAR Quantum Information Science Program as an Azrieli Global Scholar. The authors wish to thank NTT Research for their financial and technical support. Portions of this work were supported by the National Science Foundation (award CCF-1918549). We acknowledge helpful discussions with David Hathcock, Erich Mueller, Sridhar Prabhu, Eliott Rosenberg, and members of the NTT PHI Lab / NSF Expeditions research collaboration. We also thank Avik Dutt for helpful discussions and for feedback on a draft of the manuscript. We thank Michael Buttolph for assistance with fiber splicing, and Valene Tjong for contributing instrumentation-control code.
|
\section{Introduction}
Instructors frequently look for ways to support their students and improve their performance. With access to online learning environments, instructors can gather feedback in a larger scale setting. Since optimal instructional designs and scaffolds may not be known ahead of time, multiple possibilities can be tested using Uniform Random (UR) A/B experiments, also known as randomized control trials. In a UR A/B experiment, students are uniformly assigned to the different conditions that an instructor or researcher would like to test to learn about their relative effectiveness.
One challenge of this approach is how to use data more rapidly to help current students. To mitigate this, we can aim to maximize total learning by having most students being subject to the more effective conditions as they become known.
Adaptive randomization is an effective strategy for assigning more students to the current most optimal condition, while retaining the ability to test other conditions. We use a Multi-Armed Bandit (MAB) algorithm that uses machine learning to increase the number of students assigned to the current most effective condition (or \textit{arm}) \cite{lomas2016interface}, \cite{williams2018enhancing}. MAB are commonly used for rapid use of data in different areas such as marketing to optimize the benefits of the users and balance exploration vs. exploitation \cite{lomas2016interface}, \cite{rafferty2018bandit}. For this study, we used Thompson Sampling (TS), a probability matching algorithm, where the probability of assignment is proportional to the probability that the arm is optimal \cite{lomas2016interface}.
In this paper, we present a real-world experiment to illustrate the benefits and limitations of using UR A/B experiments and TS in educational settings. First, we use UR A/B experiments to evaluate different versions of emails in a homework reminder intervention to determine if a more effective version can be identified. We then compare the results of UR A/B experiments with the TS results to study its performance and benefits. Our experimental design allows us to compare classical balanced A/B comparisons side-by-side with a TS adaptive experiment to evaluate the trade-offs of using each of these methods.
\section{Multi-Armed Bandit (MAB) Algorithms}
To optimize the experience of students, we use the TS algorithm designed to solve MAB problems, useful for adaptively assigning participants to conditions.
\cite{williams2018enhancing}
The stochastic MAB problem is the problem of sequentially choosing from a discrete set of actions to maximize cumulative reward, where a reward is some measure of the effectiveness of the chosen action (arm). In this paper, we focus on the MAB problem with binary rewards. More precisely, we choose between $K$ versions (arms), and we denote the choice of action at step $t$ of the experiment by $x_t$. Assuming we choose the $k-{th}$ arm, where $k \in \{1, 2, \ldots, K\}$, then $x_t = k$, and we receive a reward $r_t$ with probability $p_k$.
TS shows strong empirical performance in maximizing the cumulative reward \cite{sutton2018reinforcement}. TS is a Bayesian algorithm that maintains a posterior distribution over each reward $p_k$. In our case, we use a Beta prior with parameters $\alpha_k$ and $\beta_k$. Arms are chosen by sampling values from the posteriors over each arm, and choosing the arm corresponding to the highest sample drawn. The posterior distribution is then updated based on the chosen action $x_t=k$ and observed reward $r_t$. We use a uniform prior for each arm, $\alpha_k=\beta_k=1$, for all $k \in \{1, 2 ,..., K\}$.
\[(\alpha_k, \beta_k) \leftarrow
\begin{cases}
(\alpha_k, \beta_k) & \text{if } x_t \neq k\\
(\alpha_k, \beta_k) + (r_t, 1 - r_t) & \text{if } x_t = k\\
\end{cases}\]
\section{Methods: Traditional and Adaptive Experimentation}
For three consecutive weeks, we tested four different versions of the emails to investigate which might be more effective in leading students to click on the homework link appended in the email. To evaluate how TS adapts the distribution of students to each email version, we sent the messages in four different batches on consecutive days of the week (Tuesday to Friday).
Using UR, each of the four email reminders has the same probability of being assigned to a student. For the TS algorithm, the probability of assignment of the email reminders version is proportional to the \textit{reward} (in our case, click rate) in all the previous batches, which is updated after each batch.
For our interventions, we used two different variations of the TS algorithm. For Weeks 1 and 2, we used a UR-TS Hybrid where the TS updating of the probability that an arm has the highest click rate used data from the UR participants too. This hybrid is called $\epsilon$[0.5]-TS, because with epsilon (epsilon = 0.5) probability, arms are assigned using UR. This takes inspiration from past algorithms like epsilon-greedy \cite{watkins1989learning} and top-two TS \cite{russo2016simple}. This is interesting to investigate because scientists may want to get the benefits of obtaining data under UR (in case TS introduces biases \cite{rafferty2019statistical}) for analysis, while also then using that data to \textit{improve} the performance of the TS algorithm. For Week 3, we applied traditional TS that did not use the data from the UR assignment
\section{Analysis and Results}
\ignore{
\begin{table}[t]
\centering
\begin{tabular}{c|c c}
& Click Rate (1) & Click Rate (2)\\
\hline
Intercept & 0.1947** (0.0228) & 0.2265** (0.0307)\\
Arm 2 & -0.0021 (0.0314) & -0.0504 (0.0464)\\
Arm 3 & 0.0092 (0.0313) & -0.0324 (0.0443)\\
Arm 4 & -0.0109 (0.0323) & -0.0445 (0.0465)\\
\hline
Weekly Effects & Yes & Yes\\
Participant Effects & No & Yes\\
No. Observations & 1278 & 1278
\end{tabular}
\caption{Table entries represent the overall effect of receiving an email reminder on the propensity to click the homework link using a panel data regression estimated on data from the UR group for the three weeks of interventions. The outcome variable in each column represents if the student clicked on the email link for the version they received. The regressor in each scenario is an indicator of whether a student clicked on the link. The Weekly Effects and Participant Effects rows present whether time (week) and participant (student) fixed effects (indicator variables) were included in the regression model. The standard error of the treatment effect estimate is presented in parenthesis. Significance levels: * $p < 0.05$, ** $p < 0.01$, z-test.}
\label{tab:panreg}
\end{table}
}
Using a panel regression with week fixed effects (i.e., include indicators for Week2-$\epsilon$[0.5]-TS and Week3-TS), we can aggregate the uniform portion of the experiment for the three weeks to evaluate the effects of the four arms on student click rate. We find that the average click rate in our sample across the three weeks is around $19\%.$ All four arms had click rates are within 2\% of each other and their difference is not statistically significant regardless of time and participant effects. The lack of an optimal arm suggests that all students were assigned to fairly similar treatments. These results are robust to also including student fixed effects within the panel regression model.
For both of the later weeks, $\epsilon$[0.5]-TS behaves similarly, favouring one particular arm. Some arms are assigned a substantially higher number of students and the others far fewer, but the seemingly favoured arm varies across the two weeks and is not consistent in Week2-$\epsilon$[0.5]-TS. In Table \ref{tab:ts_reward}, we show the cumulative reward per arm and per batch, which influenced the probability of assignment. Even though the probability of assignment aligns with the click rates from the previous batch, the algorithm drastically shifts to the arm with the highest reward and leaves minimal room for exploration in batch 3. As the click rates were not consistent across the different batches, we are unable to identify the presence of the most efficient arm, which is reflected on the cumulative rewards in Table \ref{tab:ts_reward}, and could be a result of there being minimal differences between arms.
\begin{table}[!t]
\centering
\begin{tabular}{c c|c c c c c c c c}
& & \multicolumn{2}{c}{Arm 1} & \multicolumn{2}{c}{Arm 2} & \multicolumn{2}{c}{Arm 3} & \multicolumn{2}{c}{Arm 4} \\
& & CCR & PA & CCR & PA & CCR & PA & CCR & PA\\
\hline
\multirow{4}{*}{\shortstack{Week1-$\epsilon$[0.5]-TS}} & Batch 1 & 0.200 & 0.117 & 0.277 & \textbf{0.659} & 0.212 & 0.177 & 0.167 & 0.047\\
& Batch 2 & 0.219 & \textbf{0.466} & 0.22 & 0.443 & 0.149 & 0.040 & 0.152 & 0.052\\
& Batch 3 & 0.206 & 0.434 & 0.209 & \textbf{0.452} & 0.137 & 0.017 & 0.163 & 0.097\\
& Batch 4 & 0.163 & 0.077 & 0.205 & \textbf{0.697} & 0.161 & 0.104 & 0.162 & 0.122\\
\hline
\multirow{4}{*}{\shortstack{Week2-$\epsilon$[0.5]-TS}} & Batch 1 & 0.213 & 0.116 & 0.086 & 0.000 & 0.246 & 0.225 & 0.300 & \textbf{0.659}\\
& Batch 2 & 0.198 & 0.082 & 0.14 & 0.004 & 0.244 & 0.311 & 0.266 & \textbf{0.603}\\
& Batch 3 & 0.197 & 0.065 & 0.186 & 0.041 & 0.261 & \textbf{0.666} & 0.235 & 0.229\\
& Batch 4 & 0.194 & 0.052 & 0.209 & 0.109 & 0.249 & \textbf{0.505} & 0.240 & 0.334\\
\hline
\multirow{4}{*}{\shortstack{Week3-TS}} & Batch 1 & 0.231 & \textbf{0.477} & 0.153 & 0.056 & 0.22 & 0.389 & 0.157 & 0.078\\
& Batch 2 & 0.233 & \textbf{0.926} & 0.137 & 0.030 & 0.135 & 0.033 & 0.120 & 0.011\\
& Batch 3 & 0.183 & \textbf{0.510} & 0.129 & 0.039 & 0.133 & 0.070 & 0.174 & 0.381\\
& Batch 4 & 0.181 & 0.405 & 0.152 & 0.086 & 0.144 & 0.076 & 0.181 & \textbf{0.433}\\
\end{tabular}
\caption{The Table presents the cumulative click rate (CCR) and the probability of assignment (PA) for each arm in the next batch for the three weeks of interventions. The arms represent a version of the message. Each batch represents a separate day of message deployments: Monday (Batch 1), Tuesday (Batch 2), Wednesday (Batch 3), and Thursday (Batch 4). The cumulative click rate of each arm represents the percentage of people who received the message that clicked accumulated up to and including that batch. The probability of assignment is the probability that a particular arm will be chosen in the next batch calculated using 1000000 Monte Carlo simulations. Bold values highlight the arm with the highest probability of assignment for each batch on every week. }
\label{tab:ts_reward}
\end{table}
\ignore{
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.17]{figures/ts_assignments.png}
\caption{The Figure shows the TS assignment distribution for the three weeks of the intervention. Each arm represent a version of the message. Each batch represents a separate day of message deployments: Tuesday (Batch 1), Wednesday (Batch 2), Thursday (Batch 3), and Friday (Batch 4).}
\label{fig:ts_assign}
\end{figure}
}
\section{Discussion \& Limitations}
We present a case study of adaptive experimentation, using the Thompson Sampling MAB algorithm, when a difference between arms is not observed (i.e., the arms/conditions are equally effective), according to a traditional Uniform Random experiment. We illustrate how TS may randomly favour an arm, even when giving this arm more frequently has no consequences for participants. The TS algorithm minimizes regret --- it aims to keep participants from being assigned to sub-optimal arms --- so in the case where arms are equivalent, one could argue that any of them could be presented. However, this can be problematic for scientific inference and statistical analysis \cite{rafferty2019statistical}.
One limitation is that there might have been unobserved confounding variables that caused the click rates to change in one particular batch (e.g., students had an assignment due or a test), which will also affect the algorithm – this is one concern to keep in mind in applying MAB algorithms for adaptive experimentation.
\section{Conclusion \& Future Work}
This paper provides an example of a real-world intervention using adaptive experiments, which can help instructors and researchers to use the results from experiments to more rapidly benefit students. We illustrate an instance of conducting such experiments using the TS algorithm, where the results suggest there is no difference between arms/conditions. We hope that this paper provides a first step for instructors and researchers to investigate adaptive experimentation in education. Future work can explore how the algorithm behaves in a wider variety of scenarios, such as different batch sizes and structure, and alternative differences between the arm/conditions.
\section{Acknowledgements}
This work was partially supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) (\#RGPIN-2019-06968), as well as by the Office of Naval Research (ONR) (\#N00014-21-1-2576).
\bibliographystyle{splncs04}
|
\section*{Acknowledgments}
\radd{The authors are grateful to the reviewers whose comments improved this paper.}%
\mpadd{
The authors also thank
P.~Bonofiglo,
N.~Dreval,
N.~Gorelenkov,
W.W.~Heidbrink,
and
Y.~Todo
for fruitful discussions.
}
This work was supported by US DOE grants DE-SC0014264, DE-AC02-09CH11466, DE-SC0020412, and DE-SC0020337,
as well as the Brazilian agency FAPESP Project 2011/50773-0.
This work has been carried out within the framework of the EUROfusion Consortium, funded by the European Union via the Euratom Research and Training Programme (Grant Agreement No 101052200 -- EUROfusion). Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Commission. Neither the European Union nor the European Commission can be held responsible for them.
This research used resources of the National Energy Research Scientific Computing Center (NERSC), a U.S. Department of Energy Office of Science User Facility located at Lawrence Berkeley National Laboratory, operated under Contract No. DE-AC02-05CH11231 using NERSC award FES-ERCAP20598.
\subsubsection{#1}\hfill
\newcommand{\add}[1]{\textcolor{red}{#1}}
\renewcommand{\add}[1]{#1\unskip}
\newcommand{\sub}[1]{\sout{#1}}
\renewcommand{\sub}[1]{\unskip}
\newcommand{\mpadd}[1]{\textcolor{red}{#1}}
\renewcommand{\mpadd}[1]{#1\unskip}
\newcommand{\mpsub}[1]{\sout{#1}}
\renewcommand{\mpsub}[1]{\unskip}
\newcommand{\nfadd}[1]{\textcolor{red}{#1}}
\renewcommand{\nfadd}[1]{#1\unskip}
\newcommand{\nfsub}[1]{\sout{#1}}
\renewcommand{\nfsub}[1]{\unskip}
\newcommand{\radd}[1]{\textcolor{red}{#1}}
\renewcommand{\radd}[1]{#1\unskip}
\newcommand{\rsub}[1]{\sout{#1}}
\renewcommand{\rsub}[1]{\unskip}
\newcommand{\tadd}[1]{\textcolor{red}{#1}}
\renewcommand{\tadd}[1]{#1\unskip}
\newcommand{\tsub}[1]{\sout{#1}}
\renewcommand{\tsub}[1]{\unskip}
\newcommand{\nadd}[1]{\textcolor{red}{#1}}
\renewcommand{\nadd}[1]{#1\unskip}
\newcommand{\nsub}[1]{\sout{#1}}
\renewcommand{\nsub}[1]{\unskip}
\newcommand{4768}{4768}
\newcommand{Alfv\'{e}n\xspace}{Alfv\'{e}n\xspace}
\newcommand{\Alfven Eigenmode\xspace}{Alfv\'{e}n\xspace Eigenmode\xspace}
\newcommand{\Alfven Eigenmodes\xspace}{Alfv\'{e}n\xspace Eigenmodes\xspace}
\renewcommand{\AE}{AE\xspace}
\newcommand{AEs\xspace}{AEs\xspace}
\newcommand{TAE\xspace}{TAE\xspace}
\newcommand{TAEs\xspace}{TAEs\xspace}
\newcommand{GAE\xspace}{GAE\xspace}
\newcommand{GAEs\xspace}{GAEs\xspace}
\newcommand{EAE\xspace}{EAE\xspace}
\newcommand{EAEs\xspace}{EAEs\xspace}
\newcommand{RSAE\xspace}{RSAE\xspace}
\newcommand{RSAEs\xspace}{RSAEs\xspace}
\newcommand{EPM\xspace}{EPM\xspace}
\newcommand{EPMs\xspace}{EPMs\xspace}
\newcommand{SAW\xspace}{SAW\xspace}
\newcommand{SAWs\xspace}{SAWs\xspace}
\newcommand{\AlfvenEigenmode Active Diagnostic\xspace}{\Alfven Eigenmode\xspace Active Diagnostic\xspace}
\newcommand{AEAD\xspace}{AEAD\xspace}
\newcommand{\AE antenna\xspace}{\AE antenna\xspace}
\newcommand{\red{[CITE]}\xspace}{\red{[CITE]}\xspace}
\newcommand{limiter\xspace}{limiter\xspace}
\newcommand{inner limiter\xspace}{inner limiter\xspace}
\newcommand{outer limiter\xspace}{outer limiter\xspace}
\newcommand{X-point\xspace}{X-point\xspace}
\renewcommand{\etal}{\emph{et al}\xspace}
\renewcommand{\NF}{\emph{Nucl. Fusion}\xspace}
\newcommand{EP\xspace}{EP\xspace}
\newcommand{EPs\xspace}{EPs\xspace}
\newcommand{FI\xspace}{FI\xspace}
\newcommand{FIs\xspace}{FIs\xspace}
\newcommand{RF\xspace}{RF\xspace}
\newcommand{deuterium\xspace}{deuterium\xspace}
\newcommand{hydrogen\xspace}{hydrogen\xspace}
\newcommand{helium\xspace}{helium\xspace}
\renewcommand{\H}{\mathrm{H}}
\newcommand{^{3}\mathrm{He}}{^{3}\mathrm{He}}
\newcommand{^{4}\mathrm{He}}{^{4}\mathrm{He}}
\newcommand{MISHKA\xspace}{MISHKA\xspace}
\newcommand{EFIT\xspace}{EFIT\xspace}
\newcommand{EFTP\xspace}{EFTP\xspace}
\newcommand{EFTF\xspace}{EFTF\xspace}
\newcommand{HELENA\xspace}{HELENA\xspace}
\newcommand{CSCAS\xspace}{CSCAS\xspace}
\newcommand{CASTOR\xspace}{CASTOR\xspace}
\newcommand{JOREK\xspace}{JOREK\xspace}
\newcommand{pdf\xspace}{pdf\xspace}
\newcommand{pdfs\xspace}{pdfs\xspace}
\newcommand{SparSpec\xspace}{SparSpec\xspace}
\newcommand{versus\xspace}%{vs\xspace}{versus\xspace
\newcommand{C-Mod\xspace}{C-Mod\xspace}
\newcommand{Alcator \CMod}{Alcator C-Mod\xspace}
\newcommand{plasma-antenna\xspace}{plasma-antenna\xspace}
\renewcommand{plasma-antenna\xspace}{PA\xspace}
\newcommand{\pa}{plasma-antenna\xspace}
\newcommand{low-$f$\xspace}{low-$f$\xspace}
\newcommand{high-$f$\xspace}{high-$f$\xspace}
\newcommand{HWHM\xspace}{HWHM\xspace}
\newcommand{NOVA-K\xspace}{NOVA-K\xspace}
\newcommand{NBI\xspace}{NBI\xspace}
\newcommand{(\mathrm{D}_\mathrm{NBI})}{(\mathrm{D}_\mathrm{NBI})}
\newcommand{ICRH\xspace}{ICRH\xspace}
\newcommand{TRANSP\xspace}{TRANSP\xspace}
\newcommand{H-mode\xspace}{H-mode\xspace}
\newcommand{L-mode\xspace}{L-mode\xspace}
\newcommand{M-mode\xspace}{M-mode\xspace}
\newcommand{MEGA\xspace}{MEGA\xspace}
\newcommand{TORIC\xspace}{TORIC\xspace}
\newcommand{NUBEAM\xspace}{NUBEAM\xspace}
\newcommand{FIDF\xspace}{FIDF\xspace}
\newcommand{SXR\xspace}{SXR\xspace}
\newcommand{ALCON\xspace}{ALCON\xspace}
\newcommand{ORBIT\xspace}{ORBIT\xspace}
\newcommand{\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}}{\beta_\mathrm{fi}
\newcommand{{\sim}}{{\sim}}
\newcommand{{-}}{{-}}
\newcommand{\omega_0}{\omega_0}
\newcommand{f}{f}
\newcommand{f_0}{f_0}
\newcommand{\gamma/\wo}{\gamma/\omega_0}
\newcommand{\gamma}{\gamma}
\newcommand{\g_{\mathrm{low-}f}}{\gamma_\mathrm{low}}
\newcommand{\g_{\mathrm{high-}f}}{\gamma_\mathrm{high}}
\newcommand{\Delta(\gamma/\wo)}{\Delta(\gamma/\omega_0)}
\newcommand{n}{n}
\newcommand{m}{m}
\newcommand{M}{M}
\newcommand{\vert\n\vert}{\vertn\vert}
\newcommand{q}{q}
\newcommand{q_0}{q_0}
\newcommand{q_{95}}{q_{95}}
\newcommand{n_\mathrm{e95}}%{n_{95}}{n_\mathrm{e95}
\newcommand{\nabla\ln\nnf}%{\nabla\nnf}{\nabla\lnn_\mathrm{e95}}%{n_{95}
\newcommand{\nabla\nnf}{\nablan_\mathrm{e95}}%{n_{95}}
\newcommand{P_{\mathrm{\NBI}}}{P_{\mathrm{NBI\xspace}}}
\newcommand{P_{\mathrm{RF}}}{P_{\mathrm{RF}}}
\newcommand{P_{\mathrm{\ICRH}}}{P_{\mathrm{ICRH\xspace}}}
\newcommand{\mathrm{N_{bin}}}%{N_\mathrm{bin}}{\mathrm{N_{bin}}
\newcommand{\mathrm{N_{tot}}}%{N_\mathrm{tot}}{\mathrm{N_{tot}}
\newcommand{\f_\mathrm{TAE}}{f_\mathrm{TAE}}
\newcommand{\f_\mathrm{\MISHKA}}{f_\mathrm{MISHKA\xspace}}
\newcommand{B_0}{B_0}
\newcommand{R_0}{R_0}
\newcommand{\mu_0}{\mu_0}
\newcommand{m_\mathrm{D}}{m_\mathrm{D}}
\newcommand{m_\mathrm{H}}{m_\mathrm{H}}
\renewcommand{\ne}{n_{\mathrm{e}}}
\newcommand{n_{\mathrm{e}0}}{n_{\mathrm{e}0}}
\renewcommand{\ni}{n_{\mathrm{i}}}
\newcommand{T_{\mathrm{e}}}{T_{\mathrm{e}}}
\newcommand{T_{\mathrm{e}0}}{T_{\mathrm{e}0}}
\newcommand{T_{\mathrm{i}}}{T_{\mathrm{i}}}
\newcommand{A_\mathrm{eff}}{A_\mathrm{eff}}
\newcommand{m_\mathrm{eff}}{m_\mathrm{eff}}
\newcommand{R^2}{R^2}
\newcommand{I_\mathrm{ant}}{I_\mathrm{ant}}
\newcommand{\chi^2}{\chi^2}
\newcommand{N}{N}
\newcommand{\n_\mathrm{max}}{n_\mathrm{max}}
\newcommand{\n_0}%{\n_\mathrm{res}}{n_0
\newcommand{\n_\mathrm{ant}}{n_\mathrm{ant}}
\newcommand{X}{X}
\newcommand{A}{A}
\newcommand{I_\mathrm{p}}{I_\mathrm{p}}
\newcommand{d_\mathrm{sep}}%{d_\mathrm{LCFS}}{d_\mathrm{sep}
\newcommand{\dlfs}{d_\mathrm{sep}}%{d_\mathrm{LCFS}}
\newcommand{s_{95}}{s_{95}}
\newcommand{\kappa}{\kappa}
\newcommand{\delta_u}{\delta_u}
\newcommand{\delta_l}{\delta_l}
\newcommand{\beta_\mathrm{N}}{\beta_\mathrm{N}}
\newcommand{\ell_\mathrm{i}}{\ell_\mathrm{i}}
\newcommand{f_\mathrm{ELM}}{f_\mathrm{ELM}}
\newcommand{v_\mathrm{A}}{v_\mathrm{A}}
\newcommand{\vA}{v_\mathrm{A}}
\newcommand{\wo\Ro/\vA}{\omega_0R_0/v_\mathrm{A}}
\newcommand{\Delta(\eign)}{\Delta(\wo\Ro/\vA)}
\renewcommand{\d}{d}
\newcommand{\hat{\eta}}{\hat{\eta}}
\newcommand{s}{s}
\newcommand{s_\mathrm{ant}}{s_\mathrm{ant}}
\newcommand{s_\mathrm{wall}}{s_\mathrm{wall}}
\renewcommand{\t}{t}
\newcommand{\psi_\mathrm{N}}{\psi_\mathrm{N}}
\newcommand{\sqrt{\psin}}{\sqrt{\psi_\mathrm{N}}}
\newcommand{\rho_\mathrm{pol}}{\rho_\mathrm{pol}}
\newcommand{r_\mathrm{w}}{r_\mathrm{w}}
\newcommand{v_\parallel}{v_\parallel}
\newcommand{v_0}{v_0}
\newcommand{p_\mathrm{FI}}{p_\mathrm{FI}}
\newcommand{\omega_\mathrm{*FI}}{\omega_\mathrm{*FI}}
\newcommand{n_\mathrm{He3}}%{n_{\Hethree}}{n_\mathrm{He3}
\newcommand{v_\mathrm{te}}{v_\mathrm{te}}
\newcommand{v_\mathrm{ti}}{v_\mathrm{ti}}
\newcommand{0.49\columnwidth}%{0.49\textwidth}{0.49\columnwidth
\newcommand{0.325\columnwidth}%{0.49\textwidth}{0.325\columnwidth
\newcommand{\SI}[2]{#1~\mathrm{#2}}
\newcommand{\mathrm{d}}{\mathrm{d}}
\newcommand{\deriv}[2]{\frac{\mathrm{d} #1}{\mathrm{d} #2}}
\newcommand{\JPN}[1]{\mathrm{JPN}~#1}
\newcommand{\N}[1]{(\mathrm{N_{tot}}}%{N_\mathrm{tot} = #1)
\newcommand{\abs}[1]{\vert #1 \vert}
\newcommand{\norm}[1]{\vert\vert #1 \vert\vert}
\newcommand{\expo}[1]{\exp\left( #1 \right)}
\newcommand{\n_i}{n_i}
\newcommand{\n_j}{n_j}
\newcommand{\n_k}{n_k}
\newcommand{\phi_k}{\phi_k}
\newcommand{\theta_i}{\theta_i}
\newcommand{\theta_j}{\theta_j}
\newcommand{\theta_k}{\theta_k}
\newcommand{\nmax}%{N_\mathrm{max}}{\n_\mathrm{max}
\newcommand{\n^*}%{N_\mathrm{p}}{n^*
\newcommand{\mod}[1]{\mathrm{mod}\left( #1,2\pi\right)}
\newcommand{\atantwo}[1]{\mathrm{atan2}\left(#1\right)}
\renewcommand{\xi}{x_i}
\newcommand{y_i}{y_i}
\newcommand{x_j}{x_j}
\newcommand{y_j}{y_j}
\newcommand{\floor}[1]{\lfloor#1\rfloor}
\newcommand{m_k}{m_k}
\renewcommand{\vec}[1]{\mathbf{#1}}
\newcommand{\vec{\phi}}{\vec{\phi}}
\newcommand{\vec{x}}{\vec{x}}
\newcommand{\vec{y}}{\vec{y}}
\section{Experimental measurements of \AE stability}\label{sec:experiment}
In this section, we report the novel experimental observation of stable and unstable AEs\xspace \emph{simultaneously} measured in JPN~94700. This pulse is part of the three-ion-heating scenario development experiments at JET \cite{Kazakov2020}; in fact, extensive analyses of the following, similar plasma discharge, JPN~94701, are reported in \cite{Kazakov2021}. Experimental results are presented here, and further kinetic-MHD simulations with NOVA-K\xspace \cite{Cheng1992,Fu1992,Gorelenkov1999} and MEGA\xspace \cite{Todo1998} are given in \cref{sec:novak,sec:mega}, respectively.
The time evolution of plasma parameters for JPN~94700 are shown in \cref{fig:params}. During the time range of interest, $t=\SI{7.5\mydash11}{s}$ (shaded), steady-state plasma parameters are $B_0 = \SI{3.7}{T}$, $I_\mathrm{p} = \SI{2.5}{MA}$, $n_{\mathrm{e}0} = \SI{6\times10^{19}}{m^{-3}}$, and $T_{\mathrm{e}0}=\SI{5.5\mydash7}{keV}$, although note the large sawtooth crashes at $t\approx\SI{8}{s}$ and $\SI{8.8}{s}$. Auxiliary heating powers are Neutral Beam Injection (NBI\xspace) of $P_{\mathrm{\NBI}} = \SI{6}{MW}$, with ${\sim}\SI{300}{ms}$ ``blips'' up to $\SI{8}{MW}$ at $t\approx\SI{9}{s}$ and $\SI{9.6}{s}$, and Ion Cyclotron Resonance Heating (ICRH\xspace) of $P_{\mathrm{\ICRH}} = \SI{4}{MW}$ stepping up to $\SI{6}{MW}$ at $t\approx\SI{8.5}{s}$. The concentration of $^{3}\mathrm{He}$ is $n_\mathrm{He3}/\ne \approx 27\%$ as part of the D-(DNBI)-$^{3}\mathrm{He}$ heating scenario \cite{Kazakov2021}.
\begin{figure}[h!]
\centering
\begin{subfigure}{0.49\columnwidth}%{0.49\textwidth}
\includegraphics[width=\textwidth]{params_94700_220112.pdf}
\caption{}
\label{fig:params}
\end{subfigure}
\begin{subfigure}{0.49\columnwidth}%{0.49\textwidth}
\includegraphics[width=\textwidth]{profiles_94700_220426.pdf}
\caption{}
\label{fig:profiles}
\end{subfigure}
\caption{(a)~Plasma parameters for JPN~94700: toroidal magnetic field (dashed), plasma current (solid), central electron density from Thomson Scattering (TS) and temperature from Electron Cyclotron Emission, and heating powers from NBI\xspace (solid) and ICRH\xspace (dashed). Both stable and unstable AEs\xspace were measured during the shaded time interval. (b)~Profiles at $t = \SI{9.9}{s}$: safety factor from EFIT\xspace \cite{Lao1985} constrained by pressure only (solid) and additionally polarimetry (dashed) \radd{with $q=1$ dotted}, electron density and temperature from TS, and rotation frequency from $^{3}\mathrm{He}$ charge exchange. Experimental data are shown as circles with uncertainties as error bars, while solid lines are fits to the data.}
\label{fig:params_and_profiles}
\end{figure}
Profiles of various plasma parameters are shown in \cref{fig:profiles} for one time of interest, $\t = \SI{9.9}{s}$. The electron density and temperature profiles, from Thomson scattering, match the fitted profiles from TRANSP\xspace \cite{TRANSP,Hawryluk1980,Ongena2012} well. Note here that an effective ion density profile, including the relatively high concentration of $^{3}\mathrm{He}$, is used in the simulations of the following sections. The rotation profile is obtained from $^{3}\mathrm{He}$ charge exchange spectroscopy.
The safety factor profile is solved iteratively with EFIT\xspace \cite{Lao1985}, including constraints from kinetic pressure and additionally polarimetry; both $q$-profiles are similar in \cref{fig:profiles}. It is important to note here that this can be one of the largest sources of uncertainty in the analysis below. The observation of Reverse Shear AEs\xspace (RSAEs\xspace), or Alfv\'{e}n\xspace Cascades, \radd{starting at $\t\approx\SI{8.9}{s}$} (see \cref{fig:spectrogram_and_pmn}) indicates that a non-monotonic $q$-profile exists for some time period \radd{following the sawtooth crash at $\t \approx \SI{8.8}{s}$}. \rsub{, and} Further constraints from Motional Stark Effect data have been used to identify \rsub{this} \radd{reverse shear} in similar plasmas \cite{Dreval2021}; however, such data were unfortunately not available for this pulse. \radd{That said, the upward-chirping RSAEs\xspace attain their maximum frequencies around $\t \approx \SI{9.9}{s}$, meaning that relaxation to a monotonic $q$-profile is plausible by this time. A central safety factor $q_0 \approx 0.8$ is consistent with the previously sawtoothing profile, but with fast ion stabilization.} As will be shown, good agreement between experiment and modeling is attained with the $q$-profile shown in \cref{fig:profiles}.
TAEs\xspace are destabilized by the FI\xspace population during $\t=\SI{7.5\mydash11}{s}$, as seen in the spectrogram of \cref{fig:spectrogram}. This is the Fourier decomposition of magnetic fluctuation data in both time and toroidal angle; thus, we can clearly identify coherent TAEs\xspace in time, frequency, and with a specific $n$-value. At first ($\t\approx\SI{7.5\mydash8}{s}$ and $P_{\mathrm{\ICRH}}=\SI{4}{MW}$), only $n=2,3$ TAEs\xspace are driven unstable, and then the modes quickly disappear with the sawtooth crash at $\t\approx\SI{8}{s}$. \sub{, presumably from the redistribution of FIs\xspace.} \tsub{This is most likely due to the disappearance of the $q=1$ surface in the plasma, which is the approximate location of the modes, as discussed below and in the next sections.} \tadd{This is likely due to the changing $q$-profile and/or redistribution of FIs\xspace.}
\begin{figure}[h!]
\centering
\begin{subfigure}{0.49\columnwidth}%{0.49\textwidth}
\includegraphics[width=\textwidth]{mode_analysis_94700_220112.png}
\caption{}
\label{fig:spectrogram}
\end{subfigure}
\begin{subfigure}{0.49\columnwidth}%{0.49\textwidth}
\includegraphics[width=\textwidth]{mode_analysis_IK_PMN_94700_220114.png}
\caption{}
\label{fig:pmn}
\end{subfigure}
\caption{Fourier decomposition of magnetics data with (a)~toroidal and (b)~poloidal mode number analysis for JPN~94700. Note that AEAD\xspace signals and ICRH\xspace beat frequencies have been (almost) subtracted from the background in (b).}
\label{fig:spectrogram_and_pmn}
\end{figure}
As ICRH\xspace increases to $P_{\mathrm{\ICRH}}=\SI{6}{MW}$, $n=2\mydash4$ TAEs\xspace are destabilized from $\t\approx\SI{8.3\mydash8.8}{s}$, again stabilizing with the sawtooth crash at $\t\approx\SI{8.8}{s}$. From $\t\approx\SI{9\mydash11}{s}$, the FIs\xspace destabilize $n=3-5$ TAEs\xspace and stabilize the sawteeth; \add{interestingly, note that the $n=2$ mode has now stabilized.} The $n=1$ and $6$ TAEs\xspace are also not observed in the spectrogram during this time interval at all. Furthermore, the frequency-spacing between mode numbers is $\Deltaf \approx \SI{10\mydash20}{kHz}$, \mpadd{which is slightly larger than the on-axis rotation frequency (see \cref{fig:profiles}); the additional shift could be explained by different mode locations within the TAE\xspace gap.} \mpsub{which is larger than the on-axis rotation measured from charge exchange (see \cref{fig:profiles}); thus, more than just the Doppler shift is at play here.}
\radd{From the magnetics data, the amplitudes of the $n=3\mydash5$ modes decrease by approximately an order of magnitude with each increasing $n$, and this is qualitatively seen in the spectrograms of \cref{fig:spectrogram_and_pmn}. However, deducing true mode amplitudes with edge diagnostics is difficult, especially with fields decreasing approximately as $\propto r^{-m}$. That said,} \rsub{Finally,} the $n=2,3$ TAEs\xspace are also observed as \mpadd{low-amplitude} fluctuations in the SXR\xspace array, localizing them \mpadd{approximately} within \rsub{$q<2$} \radd{$q\sim2$};%
\footnote{\label{fn:95683}\radd{Reflectometry data were not available for the pulse of interest, JPN~94700; however, those data from a similar discharge, JPN~95683, localized similar TAEs\xspace just outside $q\sim1$ \cite{Kazakov2021}.}}
$n=4,5$ TAEs\xspace are not seen in the SXR\xspace data, likely due to a too low \mpadd{signal-to-noise ratio.} \mpsub{of amplitudes.}
Throughout this discharge, the AEAD\xspace scans in frequency, \mpadd{sometimes resonantly exciting stable AEs\xspace.} \mpsub{attempting to resonantly excite \emph{stable} AEs\xspace, but not interacting with unstable AEs\xspace.} This scan can be seen in \cref{fig:spectrogram} as the triangular waveform. Interestingly, the predominant mode number identified by the magnetic probes for the AEAD\xspace perturbation is $n=5$, although all antennas have the same phase; therefore, power \emph{should} be injected into a spectrum of mode numbers peaked around $n=0$, with a higher likelihood of exciting even than odd $n$-numbers.
A poloidal mode number ($m$) analysis is also performed, with the resulting spectrogram shown in \cref{fig:pmn}. As with the $n$-calculation, this is a chi-square minimization of the phase difference between probes, iterating over all probes as the reference \cite{Tinguely2020}. \radd{We calculate the poloidal angle of each probe with respect to the magnetic axis from EFIT\xspace, but ignore the mode location in the evaluation.} \rsub{However,} We additionally include an inverse variance weighting, assuming that the standard deviation is the distance from the poloidal probe position to the magnetic axis. \mpadd{Furthermore, we restrict the poloidal harmonic range to $m \in [n,3n]$, assuming the mode is located within $q\leq3$ for computational efficiency. This assumption is valid for this analysis, as will be shown in the next sections, but is an overall limitation; thus, there could be possible effects from aliasing.}
While there are certainly larger uncertainties associated with this $m$-calculation, the results align with expectations: a mix of $m=4,5$ (brown, red) is seen for the $n=2$ TAE\xspace\radd{, and} $m=3$ (green) dominates for $n=3$; \radd{these are then located within $q<5/2$, agreeing approximately with the SXR\xspace data.} Many $m$-numbers appear for both $n=4$ and $5$, but $m=12$ (yellow) and $14$ (orange), respectively, might stand out as the clearest poloidal harmonics; \radd{this suggests localizations closer to $q\sim3$, although they could not be identified in the SXR\xspace signal.} Finally, note that both the AEAD\xspace and ICRH\xspace beat frequencies are (almost) subtracted from the background of this spectrogram; this was done only as part of a larger database analysis of unstable TAEs\xspace in JET.
The AEAD\xspace frequency scan is reproduced in \cref{fig:resonances}; here, the amplitude of the magnetic response (summed over all probes) is also shown, and two pairs of broad peaks are visible around $\t\approx\SI{9.5}{s}$ and $\SI{10.1}{s}$. These are identified as four measurements of the same stable \AE resonance, with resonant frequency $f_0\approx\SI{245}{kHz}$. This is higher than the unstable $n=5$ TAE\xspace in \cref{fig:spectrogram}, indicating that it could be a Doppler-shifted $n=6$ TAE\xspace. \rsub{Unfortunately, too few magnetic probes measured this mode to reliably evaluate its $n$-value.} \radd{Unfortunately, this mode could only be measured by a few magnetic probes; the limited number of probes and their toroidal separation did not enable a reliable determination of the mode’s toroidal mode number.} The \radd{(normalized)} damping rate measurements range from $-\gamma/\wo \approx 2\mydash3\%$, with relatively large uncertainties $\Delta(\gamma/\wo) \approx \pm 1\%$.%
\footnote{\radd{Note that $\gamma$ has units $\mathrm{s^{-1}}$, and $\omega_0$ has units $\mathrm{rad\,s^{-1}}$ and not $\mathrm{Hz}$.}}
A stable $n=6$ mode is \nfadd{expected} \nfsub{not unexpected} since the $n=5$ mode is \radd{low amplitude and could be} only marginally unstable, i.e. $\gamma/\wo \sim 0$
\begin{figure}[h!]
\centering
\includegraphics[width=0.49\columnwidth}%{0.49\textwidth]{resonances_94700_220112.pdf}
\caption{Stable \AE resonance measurements: the magnetic response amplitude (summed over all probes), AEAD\xspace (dashed) and resonant frequencies (circles), and normalized damping rates with uncertainties as error bars.}
\label{fig:resonances}
\end{figure}
The deuterium FI\xspace distribution function is calculated from TRANSP\xspace \cite{TRANSP,Hawryluk1980,Ongena2012}, using the NUBEAM\xspace \cite{NUBEAM} and TORIC\xspace \cite{TORIC} modules, and is shown in \cref{fig:fidf_both} as a function of radius, energy, and pitch. The NBI\xspace population is seen with a birth energy ${\sim}\SI{100}{keV}$, and the ICRH\xspace-accelerated tail extends to energies ${\sim}\SI{2.5}{MeV}$ with a dominant volume-averaged pitch $v_\parallel/v\approx0.5$. As will be discussed in the following sections, analytic fits of the FI\xspace distribution function are performed for both NOVA-K\xspace and MEGA\xspace simulations.
\begin{figure}[h!]
\centering
\begin{subfigure}{0.49\columnwidth}%{0.49\textwidth}
\includegraphics[width=\textwidth]{FIDF_94700_Evpsin_220126.pdf}
\caption{}
\label{fig:fidf_Evpsin}
\end{subfigure}
\begin{subfigure}{0.49\columnwidth}%{0.49\textwidth}
\includegraphics[width=\textwidth]{FIDF_94700_Evpitch_220126.pdf}
\caption{}
\label{fig:fidf_Evpitch}
\end{subfigure}
\caption{Deuterium FI\xspace distribution functions (logarithmic) from TRANSP\xspace for JPN~94700 and integrated over $\t = \SI{9.7\mydash9.9}{s}$: (a)~averaged over all pitches, and (b)~at $\sqrt{\psin} \approx 0.5$.}
\label{fig:fidf_both}
\end{figure}
\section{Introduction}\label{sec:intro}
\Alfven Eigenmodes\xspace (AEs\xspace), destabilized by fast ions (FIs\xspace) in tokamak plasmas, have been the subject of intense investigation due to their potential to enhance the transport and thus degrade the confinement of said FIs\xspace \cite{Fasoli2007}. This should be avoided in future fusion devices, like ITER and SPARC, which aim to maximize confinement and fusion power.
Unstable AEs\xspace are often easily observable in high-frequency fluctuation data, e.g. from passive magnetic probes, interferometry, reflectometry, soft x-ray (SXR\xspace), and other signals. However, the measurement of \emph{stable} AEs\xspace requires a system like JET's \AlfvenEigenmode Active Diagnostic\xspace (AEAD\xspace) \cite{Fasoli1995}, comprising two in-vessel, toroidal arrays of four antennas each \cite{Panis2010}. The antennas are independently powered and phased by amplifiers with currents up to ${\sim}\SI{10}{A}$, frequencies ranging $\SI{25\mydash250}{kHz}$, and a toroidal mode number spectrum $\vert\n\vert<20$ \cite{Puglia2016}. When the AEAD\xspace's frequency scan intersects the resonant frequency $\omega_0=2\pif_0$ \radd{$\mathrm{[rad\,s^{-1}]}$} of a stable%
\footnote{\mpadd{Note that the AEAD\xspace \emph{can} interact with unstable AEs\xspace, but for sufficiently destabilized mode amplitudes, the growth rate cannot be straightforwardly determined from the response.}}
\AE, the plasma responds like a weakly damped, driven harmonic oscillator; the damping rate $\gamma$ \radd{$\mathrm{[s^{-1}]}$} and $n$-number are then deduced from the magnetic response measured by up to fourteen fast magnetic probes \cite{Tinguely2020}.
Many past studies have analyzed unstable AEs\xspace and compared with modeling efforts (see \cite{Wong1999,Heidbrink2008} and references therein). It has been found \cite{Taimourzadeh2019} that many codes can identify the same \AE and calculate similar contributions of \emph{drive} to the total growth rate. However, the evaluation of the contribution from \emph{damping} is more variable. Therefore, much work has gone into the analysis of stable AEs\xspace and their net damping rates, in both JET %
\cite{Fasoli1995,Borba2010,Tinguely2021,Tinguely2022} \radd{(and references therein)}
and Alcator C-Mod %
\cite{Snipes2004,Snipes2005,Snipes2006}%
. Still, there can be a disconnect of sorts between un/stable AEs\xspace if they are not measured under the same plasma conditions.
In this paper, we \tsub{present} \tadd{report (for the first time, to the authors' knowledge)} the \emph{simultaneous} measurements of three destabilized Toroidicity-induced \Alfven Eigenmodes\xspace (TAEs\xspace) and one stabilized TAE\xspace in a JET deuterium plasma discharge.\tsub{a \rsub{rather} unique event \radd{to the authors' knowledge.}}
Experimental values of the net growth rates, frequencies, and mode numbers and locations allow direct comparison with \AE stability codes. Agreement between experiment and models gives us greater confidence in the prediction of \AE stability for future devices; in particular, complementary analyses from the recent JET DT campaign \cite{Mailloux2022} will identify and validate the contribution of alphas to \AE drive.
\nadd{
We note here that such a measurement of concurrent unstable and stable AEs\xspace can be difficult to achieve in JET: First, the AEAD\xspace perturbation only resonates with a stable \AE if their frequencies match and the mode location is accessible. Then, the fast magnetics only measure the resonance if the signal is sufficiently strong and the damping rate not too large. In JET, the magnetic field strength is typically too high for FIs\xspace from Neutral Beam Injection (NBI\xspace) to drive AEs\xspace unstable; thus, AEs\xspace are usually only destabilized during Ion Cyclotron Resonance Heating (ICRH\xspace). However, the AEAD\xspace is not efficient during H-mode\xspace \cite{Tinguely2022}, so the AEAD\xspace measures only few stable AEs\xspace during ICRH\xspace (and NBI\xspace), regardless of unstable AEs\xspace being observed at the same time.
}
The outline of the rest of the paper is as follows: In \cref{sec:experiment}, the experimental measurements of both stable and unstable AEs\xspace are presented. \Cref{sec:novak,sec:mega} then show simulation results from hybrid kinetic-MHD codes NOVA-K\xspace \cite{Cheng1992,Fu1992,Gorelenkov1999} and MEGA\xspace \cite{Todo1998}, respectively. Finally, a discussion and summary is given in \cref{sec:summary}.
\section*{References}
\bibliographystyle{unsrt}
\section{Hybrid kinetic-MHD simulations with MEGA}\label{sec:mega}
Despite NOVA-K\xspace modeling successfully reproducing many of the characteristics of the observed TAEs\xspace (such as mode frequency and radial location), the simulations of the previous section predicted an unstable $n=2$ TAE\xspace at $\t=\SI{9.9}{s}$, while such a mode is not detected by any fluctuation diagnostic at that time.
On the other hand, large radiative and continuum damping are predicted for the $n=5$ mode, thus stabilizing it, yet the TAE\xspace is clearly observed in experiment.
To further investigate these features and \mpsub{try to} resolve these discrepancies, a more advanced hybrid kinetic-MHD code MEGA\xspace \cite{Todo1998} is employed. MEGA\xspace self-consistently solves the evolution of a nonlinear, resistive, full-MHD background and the dynamics of the FI\xspace population. The FIs\xspace are simulated using gyro-kinetic markers that include finite Larmor radius effects. Both the fluid plasma and FIs\xspace are coupled together by means of including the FI\xspace current density term in the MHD momentum equation applied in the particle-in-cell algorithm.
The model for the thermal plasma is derived by Hazeltine and Meiss in \cite{Hazeltine1985}, and its implementation includes diamagnetic drifts and plasma rotation, providing good accuracy of the simulated TAE\xspace frequency.
As with the NOVA-K\xspace simulations described in the previous section, the electron temperature, plasma rotation, and ion density profiles used in these simulations are taken from TRANSP\xspace (see \cref{fig:params_and_profiles}). As mentioned, the density profile is corrected to account for the $^{3}\mathrm{He}$ concentration, which will affect the simulated mode frequency.
For these simulations, the cylindrical grid used to solve the MHD equations has a resolution of $N_R \times N_\phi \times N_Z = 128 \times 192 \times 128$\radd{; this resolution in the poloidal plane has good convergence and has been used previously to investigate harmonics $m>10$ \cite{Todo2021}}. All modes above $n=6$ are filtered out, so a minimum of 32 grid points in the toroidal direction per harmonic is ensured. The simulation time step is 4\% of a gyro-period of a deuterium ion at the magnetic axis. For the dissipative parameters (resistivity $\eta$, viscosity $\nu$, and diffusivity $\chi$), normalized values $\eta/\mu_0 = \nu = \chi= 5\times10^{-7}v_A R_0$ are used, where $v_A$ is the Alfv\'{e}n\xspace velocity at the magnetic axis and $R_0$ is the major radius.
Approximately 6.2~million markers ($\mysim2$ per simulation grid point) are employed to resolve the FI\xspace population using the $\delta f$ method. A quasi-analytical%
\footnote{\mpadd{In MEGA\xspace, markers are uniformly distributed in real and velocity space, with weightings based on energy, normalized pitch angle, and poloidal flux.}}%
, anisotropic\radd{, single} slowing-down distribution is used, where the maximum energy is set at $E_\mathrm{max} = \SI{2}{MeV}$ \radd{and the critical velocity $v_c$ is radially dependent on $T_e(r)$}. In order to match the ICRH\xspace absorption layer with the turning point of the trapped FIs\xspace, the normalized pitch \rsub{angle} of the simulated FI\xspace distribution is centered at $\Lambda = \mu B_0/E = 0.8$\radd{, with a width of $\Delta\Lambda = 0.3$,} producing a real-space-integrated velocity space similar to that depicted in \cref{fig:fidf_Evpitch}.
A simulation is performed where the different harmonics grow linearly together and then saturate. \Cref{fig:mega_2}(a) shows the poloidal projection of the radial velocity perturbation associated with the $n=4$ instability. The toroidal harmonics $n=3\mydash5$ are found to be unstable, in agreement with the experiment. The poloidal harmonics of each toroidal harmonic are depicted in \cref{fig:mega_2}(b)\radd{, and good agreement is seen when comparing to the mode structures from NOVA-K\xspace in \cref{fig:novak_all}.}
The locations of these harmonics \rsub{agrees} \radd{are also consistent} with the observed locations in this and previous experiments
\begin{figure}[h!]
\centering
\includegraphics[width=\textwidth]{mega_2.pdf}
\caption{(a)~MEGA\xspace-simulated poloidal structure of the radial velocity perturbation of the unstable $n=4$ TAE\xspace. Radial profiles of the (b)~radial velocity and (c)~magnetic field perturbation for the unstable $n=3\mydash5$ modes, with the $q$-profile overlaid \radd{and surface $q = (n+1/2)/n$ marked}.}
\label{fig:mega_2}
\end{figure}
Next, another simulation is performed with twice the FI\xspace pressure from TRANSP\xspace, i.e. $2\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$, aiming to observe the toroidal harmonics that remained stable in the previous simulation. The evolution of these fields are Fourier decomposed and plotted together with Alfv\'{e}n\xspace continua in \cref{fig:mega_3}\radd{; frequencies are provided in \cref{tab:mega}}. \rsub{As expected} The modes $n=3-6$, as expected, can be classified as TAEs\xspace since they are located within the continuum gap calculated by the ALCON\xspace code \cite{Deng2012}. The frequency \radd{of the ALCON\xspace 1-D estimation of the Alfv\'{e}n\xspace continuum for} each harmonic is Doppler-shifted by the \nfadd{experimentally measured} toroidal rotation included in the simulation. \radd{One can observe that the $n=2$ instability strongly intersects the Alfv\'{e}n\xspace continuum; thus, this mode can be classified as an energetic particle mode (EPM\xspace)} \tadd{at the time of interest. However, this does not mean that the unstable $n=2$ mode observed experimentally earlier in time (see \cref{fig:spectrogram_and_pmn}, $\t\approx\SI{7.5-9}{s}$) is also an EPM\xspace.}
\begin{figure}[h!]
\centering
\includegraphics[width=\textwidth]{mega_3.pdf}
\caption{Fourier-decomposed radial velocity from MEGA\xspace together with the Alfv\'{e}n\xspace continuum calculated by the ALCON\xspace code \cite{Deng2012}. The modes are located within the TAE\xspace gap\radd{, except the $n=2$ EPM\xspace.}}
\label{fig:mega_3}
\end{figure}
\rsub{
Comparing \cref{fig:novak_all,fig:mega_2,fig:mega_3}, we see that MEGA\xspace's predicted mode structures are slightly more core-localized, within $\rho_\mathrm{pol}=\sqrt{\psin} \leq 0.6$, while those from NOVA-K\xspace span the mid-radius region $0.2 \leq \rho_\mathrm{pol} \leq 0.8$
However, the intersections with the Alfv\'{e}n\xspace continua appear similar for both codes \radd{except for the $n=2$ EPM\xspace}.
}
\radd{
Comparing \cref{fig:spectrogram_and_pmn,tab:novak,tab:mega}, we see that MEGA\xspace's predicted mode frequencies are too low compared to experimental values for $n=2\mydash4$, but closer than NOVA-K\xspace for $n=5,6$. In addition, it is interesting that MEGA\xspace does not find an $n=2$ TAE\xspace in the gap indicated by ALCON\xspace (see \cref{fig:mega_3}) around ${\sim}\SI{180}{kHz}$, while the EPM\xspace is destabilized instead.
}
Since MEGA\xspace reproduces the TAEs\xspace satisfactorily and finds $n=3\mydash5$ unstable as in the experiment, a scan in $\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$ is performed to further investigate the intrinsic damping associated with each mode and how sensitive the FI\xspace drive is to the applied FI\xspace pressure.
To disentangle these effects for each TAE\xspace, single-$n$ runs are performed, simulating only a fraction of the toroidal angle $\phi \in [0, 2\pi/n]$. The toroidal resolution is reduced to $N_\phi = 32$ for these runs. Seven values of $\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$ are simulated, ranging from the value inferred from TRANSP\xspace up to $2.25\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$. \Cref{fig:mega_4}(a) shows the temporal evolution of the energy of the $n=3$ TAE\xspace for these different values of FI\xspace pressure, while \cref{fig:mega_4}(b) shows the temporal evolution of the different toroidal harmonics simulated for the case with $2\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$.
\begin{figure}[h!]
\centering
\includegraphics[width=0.8\linewidth]{mega_4-eps-converted-to.pdf}
\caption{(a)~Temporal evolution of the $n=3$ TAE\xspace energy for different values of FI\xspace pressure $\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$ in MEGA\xspace. (b)~Temporal evolution of different toroidal harmonics for $2\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$. (c)~Growth rates as a function of $\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$: $n=3\mydash6$ have similar values of inferred damping (at $\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}=0$) while $n=2$ has the largest. The slope of $n=6$ is lower, indicating that the mode is stable due to poor FI\xspace drive. The stability of an $n=3$ TAE\xspace with increased resistivity ($10\eta$) is also included (dashed), producing a line almost parallel to the reference $n=3$ case, indicating that the resistivity primarily affects the mode damping, not drive.}
\label{fig:mega_4}
\end{figure}
The growth rate of the instability of each simulation is \mpadd{evaluated} \mpsub{measured} during the linear phase and plotted in \cref{fig:mega_4}(c) as a function of the FI\xspace pressure\mpadd{; specific values are also given in \cref{tab:mega}}. Note that $n=2$ and 6 are only observed for simulations with pressures ${\geq}1.75\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$ because those modes remain stable for lower values. A linear fit is performed to the growth rates of each toroidal mode number. The crossing point with the ordinate ($\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}=0$) indicates the negative growth rate of the instability in the case without FIs\xspace, thereby indirectly estimating the total\add{, or ``intrinsic,''} mode damping.
\begin{table}[h!]
\caption{\mpadd{\rsub{Normalized net damping} \radd{Total growth} rates (\%\radd{, damping $<0$}) calculated from MEGA\xspace for different values of the fast ion pressure $\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$ \radd{from TRANSP\xspace}}.}
\label{tab:mega}
\centering
\begin{tabular}{l c c c c c}
\hline
Toroidal mode number ($n$) & 2 & 3 & 4 & 5 & 6 \\
\radd{Frequency (kHz, lab frame)} & \radd{125} & \radd{170} & \radd{200} & \radd{230} & \radd{240} \\
\rsub{Damping $\gamma/\wo$ (\%)} \radd{Fast ion beta} & \multicolumn{5}{c}{\radd{Total growth rates $\gamma/\wo$ (\%)}} \\
\hline
$0.0\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$ & -17.9 \radd{$\pm$ 0.9} & -6.9 \radd{$\pm$ 0.4} & -8.0 \radd{$\pm$ 0.6} & -8.9 \radd{$\pm$ 0.4} & -4.7 \radd{$\pm$ 1.0} \\
$0.7\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$ & -7.5 \radd{$\pm$ 1.0} & 2.3 \radd{$\pm$ 0.5} & 1.7 \radd{$\pm$ 0.7} & \rsub{$\mysim0$} \radd{0.1~$\pm$~0.5 } & -1.2 \radd{$\pm$ 1.1} \\
$1.0\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$ & -3.0 \radd{$\pm$ 1.0} & 6.2 \radd{$\pm$ 0.5} & 5.9 \radd{$\pm$ 0.7} & 3.7 \radd{$\pm$ 0.5} & 0.3 \radd{$\pm$ 1.1} \\
\hline
\end{tabular}
\end{table}
It is interesting\radd{, although perhaps expected,} that the $n=3\mydash6$ TAEs\xspace have similar values of the intrinsic damping, i.e. $-\gamma/\wo \approx 5\%\mydash10\%$, while $n=2$ \radd{EPM\xspace} is almost twice as large. In view of these computational results, one could argue that $n=2$ and 6 \rsub{TAEs\xspace} \radd{modes} are stable in the experiment but for different reasons: the $n=2$ \rsub{mode} \radd{EPM\xspace} has a larger intrinsic damping \radd{due to the strong interaction with the SAW\xspace continuum,} while the $n=6$ \rsub{mode} \radd{TAE\xspace} has a lower FI\xspace drive, which is reflected by the lower slope in \cref{fig:mega_4}(c). \add{This is only partially found in the NOVA-K\xspace results (see \cref{tab:novak}): The total intrinsic damping is predicted to increase with $n$, in disagreement with MEGA\xspace. On the other hand, the drive of the $n=2\mydash4$ modes is about twice that of the $n=6$ mode, in agreement with MEGA\xspace, although the same applies to $n=5$, which is a discrepancy. Moreover, there is a factor of 1-10 difference in the intrinsic damping values predicted by NOVA-K\xspace and MEGA\xspace, but part of this may be due to the resistivity used in MEGA\xspace, discussed further below.}
At the experimental value of $\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$, the $n=3\mydash5$ TAEs\xspace are unstable and decreasing in their growth rates. This qualitatively agrees with decreasing (saturated) mode amplitudes observed by the magnetic probes in experiment (see \cref{fig:spectrogram_and_pmn}). The $n=2$ \rsub{TAE\xspace} \radd{mode} is heavily damped around ${\sim}\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$, also in agreement with experiment. At the TRANSP\xspace value of $\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$, the linear extrapolation of the $n=6$ TAE\xspace growth rate would suggest a marginally \emph{unstable} mode with $\gamma/\wo \approx +0.3\%$\radd{, although the uncertainty is significant $\mysim1\%$ (see \cref{tab:mega})}. To best match the measured net damping rate of $-\gamma/\wo \approx 2\% \pm 1\%$, we would need to decrease the FI\xspace pressure to $\mysim0.7\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$, at which point the $n=5$ TAE\xspace would be only marginally unstable, also in better agreement with experiment.
It is important to note that the validity of the estimated mode damping depends on the resistivity $\eta$ used in these simulations; however, a value larger than the Spitzer value was chosen to accommodate MEGA\xspace's high-resolution grid \cite{Bierwage2016}. To test the sensitivity to $\eta$, the $\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$ scan was run again for the $n=3$ harmonic using a resistivity ten times larger. The resulting growth rates are shown in \cref{fig:mega_4}(c). As expected, the inferred damping rate has increased by $\mysim50\%$,
but not enough to exceed the damping estimated for the $n=2$ \rsub{TAE\xspace} \radd{EPM\xspace}. The slope of the (dashed) line is very similar to the reference case, suggesting that $\eta$ has little effect on the FI\xspace drive. In reality, using the true (lower) resistivity would \emph{decrease} the net damping. Thus, while variations of $\eta$ might slightly change the inferred damping, the conclusion that the $n=2$ mode remains stable due to larger intrinsic damping and $n=6$ due to lower FI\xspace drive is still valid.
\section{Simulation}\label{sec:simulation}
\section{Kinetic-MHD simulations with NOVA-K}\label{sec:novak}
The kinetic-MHD code NOVA-K\xspace \cite{Cheng1992,Fu1992,Gorelenkov1999} is used to identify AEs\xspace and assess their stability at one time, $\t=\SI{9.9}{s}$. This time is selected because it occurs during a relatively steady-state period of the pulse (see \cref{fig:params}), i.e. after sawtooth stabilization, and in between the measurements of stable AEs\xspace by the AEAD\xspace (see \cref{fig:resonances}). The plasma profiles (see \cref{fig:profiles}) are used in the calculation of the Alfv\'{e}n\xspace continua and eigenmode structures for $n=2\mydash6$, with a subset shown in \cref{fig:novak_all}. Note that the TAE\xspace eigenfrequencies are in the \emph{lab} frame, as \nfadd{the experimental} rotation \nfadd{profile} is included in NOVA-K\xspace; they span $f \approx \SI{195\mydash225}{kHz}$ (see \cref{tab:novak}), which is only slightly lower than those observed experimentally (see \cref{fig:spectrogram_and_pmn}). This $\mysim10\%$ difference is likely the result of uncertainties in the density, safety factor, and rotation profiles; \mpadd{the adiabatic index used in NOVA-K\xspace could also play a role \cite{VanZeeland2016}.}
\newcommand{0.49\textwidth}{0.49\textwidth}
\begin{figure}[h!]
\centering
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=\textwidth]{novak_94700_n3_f195847_220113.pdf}
\caption{}
\label{fig:novak3}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=\textwidth]{novak_94700_n4_f206525_220119.pdf}
\caption{}
\label{fig:novak4}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=\textwidth]{novak_94700_n5_f215054_220113.pdf}
\caption{}
\label{fig:novak5}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=\textwidth]{novak_94700_n6_f224672_220113.pdf}
\caption{}
\label{fig:novak6}
\end{subfigure}
\caption{Continua (thin lines) and poloidal mode structure (solid, dot-dashed) from NOVA-K\xspace at $\t = \SI{9.9}{s}$ for $n = $~(a)~3, (b)~4, (c)~5, (d)~6. The frequency (lab-frame) is indicated by the horizontal dotted line. $\psi_\mathrm{N}$ is the normalized poloidal flux. While the ranges of poloidal harmonics ($m$) vary for each subplot, the colors and line styles are the same for all.}
\label{fig:novak_all}
\end{figure}
The poloidal mode structures \radd{from NOVA-K\xspace} for each $n$ \radd{(see \cref{fig:novak_all})} are primarily dominated by harmonics $m = n,n+1$, i.e. localizing the TAEs\xspace around $q\approx1$ and consistent with the SXR\xspace data, although the overall structures are quite global. \radd{From the NOVA-K\xspace results,} we may have expected $m=4$ \rsub{(instead of $m=3$)} to be identified \radd{as the dominant poloidal harmonic} for the $n=3$ TAE\xspace in \cref{fig:pmn}\radd{, but instead $m=3$ is found experimentally}. Similarly, the higher poloidal harmonics $m=12$ and $14$ for $n=4$ and $5$ TAEs\xspace, respectively, are not dominant in NOVA-K\xspace, but perhaps their closer location to the poloidal probes increases their detection probability \radd{ in experiment. For example, see the $m=14$ harmonic for the $n=5$ TAE\xspace in \cref{fig:novak5}.} \radd{Unfortunately, the SXR\xspace array can help localize these modes, identified by frequency, but their low signals and measurement by only a few SXR\xspace lines-of-sight make a mode number evaluation infeasible. This demonstrates the difficulty of measuring some properties of core modes from edge diagnostics.}
The stability of the $n=2\mydash6$ TAEs\xspace is computed including the effects of the FIs\xspace. In NOVA-K\xspace, an energy threshol
\footnote{Note that the threshold was varied among $\SI{125}{keV}$, $\SI{150}{keV}$, and $\SI{200}{keV}$ with little impact on the stability results.}
is used to discriminate the NBI\xspace and ICRH\xspace populations within the FI\xspace distribution function \radd{from TRANSP\xspace and NUBEAM\xspace}, which is important here because this three-ion-heating scenario specifically heats the DNBI ions. Then, for each radial position, the \radd{local} NBI\xspace population is fit with \radd{an anisotropic} slowing down distribution, while the ICRH\xspace tail is fit with an \radd{anisotropic equivalent} \rsub{average} temperature \radd{\cite{Snipes2005}}. Furthermore, the ICRH\xspace resonance layer is modeled as a Gaussian profile, in this case with its peak at the magnetic axis ($R_0 \approx \SI{3}{m}$) and width $\Delta R \approx \SI{0.24}{m}$, similar to that shown in \cite{Kazakov2021} \mpadd{(see Fig.~12a therein)}.
The breakdown of various drive and damping mechanisms is given in \cref{tab:novak}. Radiative damping is the largest contribution; this has also been identified \radd{in} other recent analyses of (asynchronous) unstable \cite{Nabais2018,Aslanyan2019} and stable \cite{Tinguely2021,Tinguely2022} AEs\xspace in JET.
\mpadd{While the implementation of radiative damping\radd{, which is averaged over the radial mode structure in NOVA-K\xspace,} comes from \cite{Berk1993,Fu1996}, \radd{Ref.}~\cite{Mett1992} can help us estimate the \emph{relative} uncertainty:
\radd{$\Delta(\gamma/\wo)_\mathrm{rad} \propto (1/3) (\gamma/\wo)_\mathrm{rad} \left(\mathrm{d} \ln T + \mathrm{d} \ln q \right)$}%
, which is $\mysim8\%$ assuming 10\% uncertainties in the plasma profiles.} %
Continuum damping can also be a significant contributor, which is clear from the eigenmodes' intersecting the continuum around $\sqrt{\psin} \approx 0.8$ in \cref{fig:novak_all}, \mpadd{although absolute uncertainties can be of order $\Delta(\gamma/\wo)\sim0.1\%$ \cite{Bowden2014}.}
\begin{table}[h!]
\caption{\rsub{Normalized damping} \radd{Growth} rates (\%\radd{, damping $<0$}) calculated from NOVA-K\xspace. The frequency is in the lab frame\mpadd{; the \emph{absolute} uncertainty in continuum damping is $\pm0.1\%$, while the \emph{relative} uncertainty in radiative damping is $\pm8\%$.} \mpsub{, and uncertainties of continuum damping are $\pm 0.1\%$.}
Totals are given \textbf{with} and \emph{without} the contributions from fast ions, which include finite Larmor radius effects.}
\label{tab:novak}
\centering
\begin{tabular}{l c c c c c}
\hline
Toroidal mode number ($n$) & 2 & 3 & 4 & 5 & 6 \\
Frequency (kHz, lab frame) & 179.9 & 195.9 & 206.5 & 215.1 & 224.7 \\
Damping\radd{/drive mechanism} \rsub{$\gamma/\wo$ (\%)} & \multicolumn{5}{c}{\radd{Growth rate $\gamma/\wo$ (\%)}} \\
\hline
Continuum & -0.06 & -0.03 & -0.77 & -0.33 & -0.31 \\
Radiative & -1.03 & -2.01 & -2.84 & -4.44 & -4.76 \\
Electron collisional & -0.32 & -0.39 & -0.04 & -0.04 & -0.01 \\
Electron Landau & -0.09 & -0.11 & -0.07 & -0.12 & -0.12 \\
Ion Landau & -0.02 & -0.01 & -0.04 & -0.01 & -0.01 \\
NBI\xspace fast ions & -0.07 & -0.12 & -0.11 & $\sim$0 & -0.05 \\
ICRH\xspace fast ions & 4.08 & 4.30 & 4.45 & 2.01 & 2.34 \\
\emph{Total (w/o fast ions)} & \emph{-1.53} & \emph{-2.54} & \emph{-3.76} & \emph{-4.94} & \emph{-5.22} \\
\textbf{Total (w/ fast ions)} & \textbf{2.48} & \textbf{1.64} & \textbf{0.58} & \textbf{-2.93} & \textbf{-2.92} \\
\hline
\end{tabular}
\end{table}
Due to the high magnetic field (and thus high Alfv\'{e}n\xspace speed), NBI\xspace ions actually Landau damp these modes \radd{in JET \cite{Borba2000}}, but ICRH\xspace FIs\xspace provide significant drive, $\gamma/\wo \approx 2\mydash4\%$. Thus, $n=2\mydash4$ TAEs\xspace are predicted by NOVA-K\xspace to be destabilized, with decreasing total growth rates, while $n=5,6$ TAEs\xspace are stabilized with similar total damping rates. The trend of decreasing drive with $n$ is at least consistent for $n=3\mydash5$ modes, although of course $n=5$ is \rsub{marginally} \emph{unstable} in experiment (see \cref{fig:spectrogram_and_pmn}). Additionally, the $n=2$ TAE\xspace is stable at $\t=\SI{9.9}{s}$ in experiment, not strongly driven. These are currently unresolved discrepancies. That said, the net damping rate $\gamma/\wo \approx -2.9\%$ for the $n=6$ TAE\xspace agrees within experimental uncertainties.
\section{Summary}\label{sec:summary}
In this work, we presented novel measurements of both unstable and stable TAEs\xspace observed simultaneously in a JET deuterium plasma, with steady-state parameters $B_0 = \SI{3.7}{T}$, $I_\mathrm{p} = \SI{2.5}{MA}$, $n_{\mathrm{e}0} = \SI{6\times10^{19}}{m^{-3}}$, and $T_{\mathrm{e}0}=\SI{5.5\mydash7}{keV}$ (see \cref{fig:params_and_profiles}). The three-ion-heating scheme (D-DNBI-27\% $^{3}\mathrm{He}$) was used in this discharge to accelerate ${\sim}\SI{100}{keV}$ NBI-deuterons to ${\sim}$MeV energies (see \cref{fig:fidf_both}) with a dominant pitch $v_\parallel/v \approx 0.5$. This fast ion (FI\xspace) population destabilized $n=2\mydash5$ TAEs\xspace, and their saturated mode amplitudes decreased with toroidal mode number $n$ (see \cref{fig:spectrogram_and_pmn}), from which decreasing FI\xspace drive was inferred. Synchronously, JET's in-vessel \AlfvenEigenmode Active Diagnostic\xspace (AEAD\xspace) resonantly excited a stable \AE with normalized damping rate $-\gamma/\wo\approx2\% \pm 1\%$ (see \cref{fig:resonances}). Its Doppler-shifted frequency, $f_0 \approx \SI{245}{kHz}$, indicated that it was likely an $n=6$ TAE\xspace.
Simulations with two hybrid kinetic-MHD codes, NOVA-K\xspace and MEGA\xspace, both identified AEs\xspace with similar eigenfrequencies, poloidal mode structures, and radial locations as those observed in experiment. NOVA-K\xspace's calculation of \AE stability demonstrated good agreement with the $n=3$, 4, and 6 TAEs\xspace (see \cref{fig:novak_all}), matching the experimentally measured damping rate of the $n=6$ mode within uncertainties and identifying radiative damping as the dominant mechanism (see \cref{tab:novak}). However, there were two major discrepancies between NOVA-K\xspace and experiment: the stable $n=2$ TAE\xspace, i.e. not seen in fluctuation data, was predicted by NOVA-K\xspace to be strongly driven, while the marginally unstable $n=5$ TAE\xspace was predicted to be strongly damped.
Improved agreement was then obtained with MEGA\xspace for all modes: the unstable $n=3\mydash5$ and stable $n=2,6$ \rsub{TAEs\xspace} \radd{modes} (see \cref{fig:mega_2,fig:mega_3}). By scanning the FI\xspace pressure $\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$ in MEGA\xspace, \radd{an $n=2$ EPM\xspace and $n=6$ TAE\xspace were destabilized and identified.} The intrinsic damping was extrapolated to $\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}=0$ (see \cref{fig:mega_4,tab:mega}), and the $n=2,6$ \rsub{TAEs\xspace} \radd{modes} were found to be stabilized for two different reasons: the former \radd{EPM\xspace} by higher intrinsic damping \radd{due to strong interaction with the Alfv\'{e}n\xspace continuum}, and the latter TAE\xspace by lower fast ion drive. Yet the best agreement of the MEGA\xspace growth rates with experiment was not found at the value of $\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$ inferred from TRANSP\xspace, but rather at $\mysim0.7\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$.
These disagreements between experiment and simulation still need to be resolved. One resolution could be found by manipulating the various inputs within experimental error bars, and that exploration is left for future work. \radd{The codes also include different physics (see \cite{Taimourzadeh2019} and others for further details): NOVA-K\xspace is a perturbative eigenvalue code with some contributions to the linear growth/damping rate calculated analytically, whereas MEGA\xspace is a first-principle, resistive, fully nonlinear initial value code. The benefit of NOVA-K\xspace is the identification of various drive and damping mechanisms and their uncertainties, whereas MEGA\xspace gives only the total - though perhaps more accurate - growth rate.}
The codes may also need to improve their stability calculations of intrinsic damping and drive. These data, as well as large databases of stable AEs\xspace (see \cite{Tinguely2020,Tinguely2021,Tinguely2022} and references therein), could help guide this effort. Nevertheless, the trends of decreasing growth rate for $n=3\mydash6$ TAEs\xspace in both MEGA\xspace and NOVA-K\xspace is promising, and so is the relatively good match with the experimentally measured damping rate. \nfadd{Preliminary results from the orbit-following code ORBIT\xspace \cite{ORBIT} indicate some strong interactions of the TAEs\xspace with trapped fast ions, and this should be looked into further.} Additional upcoming work will include similar analyses of \AE stability in DT plasmas from the recent JET 2021 campaign, during which the AEAD\xspace actively excited \rsub{$\mysim1000$} \radd{over 2000} stable AEs\xspace.
\mpadd{
Looking toward future tokamaks, like ITER and SPARC, a similar modeling strategy could be used \nfadd{to} predict and optimize \AE stability: Multiple codes, e.g. NOVA-K\xspace and MEGA\xspace, should be used to identify AEs\xspace (and other FI\xspace-driven modes), and a cross-check performed among them to confirm similar mode frequencies, structures, locations, etc. The findings of this work suggest that scanning the FI\xspace pressure ${\pm}30\%$ may be required (at least in MEGA\xspace) to best match reality, although $+30\%$ may be too conservative \radd{and uncertainties must be considered}. Both the MEGA\xspace $\beta_\mathrm{fi}}%{\beta_\mathrm{ep}}{\beta_\mathrm{\FI}$-scan and NOVA-K\xspace breakdown can then identify the dominant contributions to damping and drive, and various parameters could be tuned to help stabilize those modes, e.g. plasma elongation or ICRH\xspace resonance location.
}
|
\section{Supplementary Material}
Figures~\ref{fig1}(a)-(b) show photos of
co-aligned $\kappa$-(ET)$_2$Cu$_{2}$(CN)$_3$ ($\kappa$-CN) crystals
and the typical shape of the $\kappa$-CN crystal.
Owing to the monoclinic symmetry $P2_{1}/C$,
with a monoclinic angle 113.42~deg and lattice parameters of
$a=16.117$~\AA, $b=8.586$~\AA, and $c=13.397$~\AA~\cite{Geiser91},
the $a$-axis is not perpendicular to the $bc$ plane,
and the crystals mounted on the aluminum plate
contain two choices of the structural domains,
according to their tilted $a$-axes [Fig.~\ref{fig1} (c)].
Phonon signals were measured along the $b^{*}$($=c\times a$) axis
at $\mbox{\boldmath $Q$}$=(060),
which is common to the two tilted $a$-axes.
In each ET layer, the long ET molecular axes are inclined approximately $30^{\circ}$
with respect to the layer normal.
In $\kappa$-CN with a monoclinic structure,
ET layers of the same orientation are stacked with an anion layer in between [see Fig.~\ref{fig1}(d)].
In contrast, two adjacent ET layers with opposite orientations are stacked alternately
in $\kappa$-(ET)$_{2}$Cu[N(CN)$_{2}$]Cl ($\kappa$-Cl)
with an orthorhombic structure~\cite{Geiser91_PhysC}.
Note that the differences in the mode energies revealed
for $\kappa$-CN and $\kappa$-Cl are likely
due to differences in the crystal structure, resulting in a somewhat harder lattice for $\kappa$-CN.
Given that in $\kappa$-CN, the ET dimer is stacked
in the same orientation and the anion layer has a closely packed structure
[Figs.~\ref{fig1}(d) and (e)],
the phonon modes of $\kappa$-CN
are expected to be harder compared to those of $\kappa$-Cl.
\begin{figure}
\includegraphics[keepaspectratio=true,width=10cm,clip]{k-CN_suppl_Fig1_submit.eps}
\caption{\label{fig1} (color online)
Images of (a) co-aligned $\kappa$-(ET)$_2$Cu$_{2}$(CN)$_3$ ($\kappa$-CN) crystals
and (b) the typical shape of the $\kappa$-CN crystal with a flat surface of the $bc$-plane.
Crystal structure of $\kappa$-CN : (c) side view of the layered structure.
Side view of the layered structure for (d) $\kappa$-CN and
(e) $\kappa$-(ET)$_{2}$Cu[N(CN)$_{2}$]Cl.
}
\end{figure}
Figures~\ref{fig2} show rocking curves of (021) and (012) Bragg peaks
for the co-aligned $\kappa$-CN crystals.
Most of the crystals are co-aligned within 10 degrees for both Bragg peaks.
Although the overall mosaicity of the sample
is larger than the instrumental resolution ($\sim1$~deg),
the main conclusion of this study is not affected
because the phonon mode at $E=4.7$~meV
is an optical mode with no dispersion near $\Gamma$-points.
\begin{figure}
\includegraphics[keepaspectratio=true,width=10.0cm,clip]{k-CN_suppl_Fig2_submit.eps}
\caption{\label{fig2}
Rocking curves of (a) (021) and (b) (012) Bragg peaks
for the co-aligned $\kappa$-CN crystals obtained using the IN3 spectrometer.
}
\end{figure}
Figure~\ref{fig3} shows the all temperature dependence
for constant-Q scans at (060) for
$\kappa$-(BEDT-TTF)$_{2}$Cu$_{2}$(CN)$_{3}$ ($\kappa$-CN),
some of which are not shown in the manuscript.
The phonon spectra for $\kappa$-CN are broad above $T=6$~K.
This is in contrast to the phonon spectra of $\kappa$-(BEDT-TTF)$_{2}$Cu[N(CN)$_{2}$]Cl
($\kappa$-Cl) showing recovery of sharp phonon peaks at $T=100$~K,
as shown in Fig.~\ref{fig4}.
The differences in phonon damping between $\kappa$-CN
and $\kappa$-Cl could be related to the itinerancy of
$\pi$-electrons.
\begin{figure}
\includegraphics[keepaspectratio=true,width=13.0cm,clip]{k-CN_suppl_Fig3_submit.eps}
\caption{\label{fig3}
Temperature dependence of constant-Q scans at (060).
The solid lines are fits to damped
harmonic oscillator functions at $E\sim~2.0$, 2.9, 3.7, 4.7, 5.7 and 7.1~meV
convolved with the instrumental resolution.
The dashed lines show the peak profiles for each mode.
The breathing mode of the ET dimer at 4.7~meV is represented by a red hatch.
The horizontal bars represent instrumental energy resolution at $E=4.7$~meV.
}
\end{figure}
\begin{figure*}
\includegraphics[keepaspectratio=true,width=10.0cm,clip]{k-CN_suppl_Fig4_submit.eps}
\caption{\label{fig4}
Phonon spectra at $T=100$~K for $\kappa$-CN and $\kappa$-Cl~\cite{Matsuura19}
which were measured at $Q=$(603) and (060), respectively.
}
\end{figure*}
Figure~\ref{fig5} shows the temperature dependence
of the electric resistivity $\rho$ for $\kappa$-CN and $\kappa$-Cl~\cite{Matsuura19}.
Although a hydrogenated sample was used for $\kappa$-CN,
deuterization did not affect on the resistivity in the case of $\kappa$-CN
since $\kappa$-CN is deep in the dimer Mott-insulator phase.
$\rho$ for $\kappa$-CN shows semiconducting behavior
for all temperatures below $T=300$~K.
However, $\rho$ for $\kappa$-Cl shows
a rapid increase below $T_\mathrm{INS}\sim50-60$~K
below which the charge gap opens in optical conductivity.
\begin{figure}
\includegraphics[keepaspectratio=true,width=10cm,clip]{k-CN_suppl_Fig5_submit.eps}
\caption{\label{fig5}
Temperature dependence of the electric resistivity $\rho$
for $\kappa$-CN and $\kappa$-Cl.
A deuterized sample was used for $\kappa$-Cl
whereas a hydrogenated sample was used for $\kappa$-CN.
}
\end{figure}
As the temperature decreases,
$\kappa$-CN approaches the CO phase boundary,
and the energy of the intra-dimer charge fluctuation exhibits
softening (Fig.~\ref{fig7})~\cite{Naka13}.
In the theoretical phase diagram for $c$-CO [Fig.~3(a) in the main text],
reentrant behavior of the CO phase boundary is observed
in the AFM/QSL phase
owing to the competition between the exclusive AFM and CO phases.
In the AFM/QSL phase,
$\omega_\mathrm{charge}$ becomes harder as
it shifts away from the CO phase boundary (Fig.~\ref{fig7}).
Thus, the coupling between the charge and lattice is expected to
be weak below $T^{*}$ because $\omega_\mathrm{charge}$ and
$\omega_\mathrm{phonon}$ are no longer close to each other.
\begin{figure}
\includegraphics[keepaspectratio=true,width=10cm,clip]{k-CN_suppl_Fig7_submit.eps}
\caption{\label{fig7}
Schematic diagram of the
temperature dependence of the charge
fluctuations $\omega_\mathrm{charge}$
and coupled phonon frequency $\omega_\mathrm{charge}$.
}
\end{figure}
The crystal structures in Figs.~\ref{fig1} were produced by VESTA software\cite{Momma_11}.
|
\section{Introduction}
\label{Sec:introduction}
The LHCb collaboration continues its exploration of hadronic interactions as revealed by electroweak decays of heavy hadrons, recently announcing the discovery of a signal in the $J/\psi\, \Lambda$ mass spectrum of $B^-\to J/\psi \, \Lambda\, \bar{p}$~\cite{ChenParticleZoo}. The mass and width of $P_{\psi s}^\Lambda(4338)$ are \begin{align}M&= 4338.2 \pm 0.7~\textrm{ MeV},\\ \Gamma &= 7.0 \pm 1.2~\textrm{ MeV},\end{align}
and $J^P= 1/2^-$ quantum numbers are preferred.
Because of its proximity to $\Xi_c\bar D$ threshold, a molecular interpretation of $P_{\psi s}^\Lambda(4338)$ has been proposed~\cite{Karliner:2022erb,Wang:2022mxy,Yan:2022wuz}. Molecules with $\Xi_c\bar D$ constituents have been predicted in a wide range of models, typically assuming a binding interaction due to boson exchange, or effective field theory constrained by heavy quark symmetry \cite{Wang:2019nvm,Xiao:2019gjd,Peng:2019wys,Peng:2020hql,Chen:2020uif,Dong:2021juy,Yan:2021nio,Chen:2021cfl,Chen:2020kco,Liu:2020hcv,Zhu:2021lhd,Xiao:2021rgp,Chen:2022onm,Wang:2019nvm}. In such models, whether or a not a particular state binds is ultimately determined by one or more parameters which are fit to data, such as the form factor cut-off, or contact terms attributed to unknown short-distance physics. For this reason, the molecular approach is only robust to the extent that it can simultaneously describe several different states with the same fit parameters. With respect to this criterion, the molecular scenario for $P_{\psi s}^\Lambda(4338)$ runs into problems.
If $P_{\psi s}^\Lambda(4338)$ is a $\Xi_c\bar D$ molecule ($1/2^-$), it could potentially have $\Xi_c\bar D^*$ partners ($1/2^-$ and $3/2^-$). From heavy quark symmetry, the potentials in all three channels are identical,
\begin{align}
V(\Xi_c\bar D,1/2^-)=V(\Xi_c\bar D^*,1/2^-)=V(\Xi_c\bar D^*,3/2^-),\label{eq:constraint}
\end{align}
which implies the three states should have comparable binding energies, an expectation which is confirmed in a one-boson model calculation \cite{Wang:2022mxy}. The $P_{cs}(4459)$ state observed at LHCb~\cite{Aaij:2020gdg} is a candidate for a $\Xi_c\bar D^*$ molecule~\cite{Peng:2020hql,Chen:2020kco,Liu:2020hcv,Zhu:2021lhd,Xiao:2021rgp}, but if it is the partner of $P_{\psi s}^\Lambda(4338)$ as a $\Xi_c\bar D$ molecule, it implies a drastic violation of heavy-quark symmetry, since its binding energy is so large (19~MeV) compared to $P_{\psi s}^\Lambda(4338)$ (which is at threshold).
Similarly, the hypothesis that $P_{cs}(4459)$ consists of two states \cite{Aaij:2020gdg}, interpreted as $ \Xi_c\bar D^*$ molecules ($1/2^-$ and $3/2^-$) \cite{Karliner:2022erb,Wang:2022mxy}, is problematic, because the 13~MeV mass splitting contradicts the above expectation from heavy-quark symmetry. Although some level of splitting is expected in models (for example due to coupled-channel effects)~\cite{Wang:2019nvm,Peng:2020hql}, it remains to be seen whether such models can obtain the right level of splitting, and also reconcile the significant $\Xi_c\bar D^*$ binding with the lack of binding in $\Xi_c\bar D$. We note that in the model of ref.~\cite{Yan:2022wuz}, there is no region of parameter space in which both $P_{\psi s}^\Lambda(4338)$ and two $P_{cs}(4459)$ states can be accommodated.
Refs.~\cite{Karliner:2022erb,Wang:2022mxy} have also argued for an analogy between $P_{\psi s}^\Lambda(4338)$ and $P_{cs}(4459)$ (as $\Xi_c\bar D\*$ molecules), and $P_c(4312)$, $P_c(4440)$ and $P_c(4457)$ \cite{Aaij:2015tga,Aaij:2019vzc,Aaij:2016phn} (as $\Sigma_c\bar D\*$ molecules). The analogy is misleading, considering that the corresponding potentials are neither related by $SU(3)$ flavor (as $\Xi_c$ and $\Sigma_c$ belong to different flavor multiplets), nor heavy-quark spin symmetry (as $\Xi_c$ and $\Sigma_c$ have different light quark spins) \cite{Peng:2020hql}. Indeed, heavy quark symmetry implies a completely different pattern of states in $\Sigma_c\bar D\*$ systems (where the potentials are spin-dependent~\cite{Liu:2019tjn,Liu:2019zvb,Du:2019pij,Sakai:2019qph,Valderrama:2019chc,Xiao:2019aya,Peng:2020xrf,Du:2021fmf,Liu:2020hcv}) and $\Xi_c\bar D\*$ systems (where they are not). Moreover, the analogy relies on the assumption that $P_c(4440)$ and $P_c(4457)$ are both $\Sigma_c\bar D^*$ molecules, and we recently argued that this assumption is not consistent with experimental constraints \cite{Burns:2021jlu}. (Scenarios with different interpretations for $P_c(4457)$ \cite{Burns:2019iih,Burns:2022uiv} do not have the same problem.)
An additional awkward feature of the molecular scenario for $P_{\psi s}^\Lambda(4338)$, which has hardly been discussed in the literature, is that it is not bound with respect to the $\Xi_c\bar D$ thresholds, but somewhat above:
\begin{align}
\Xi_c^0\bar D^0&=4335.28\pm 0.33~\textrm{ MeV},\\
\Xi_c^+D^-&=4337.37\pm 0.28~\textrm{ MeV}.
\end{align}
The situation here is similar to $P_c(4457)$, which is widely interpreted as a $\Sigma_c\bar D^*$ bound state, despite having a mass which is consistent with the threshold not only for $\Sigma_c\bar D^*$, but also $\Lambda_c(2595)\bar D$. Signals at (rather than below) threshold are more amenable to non-resonant interpretations, and we recently demonstrated that $P_c(4457)$ can be explained as a cusp or an enhancement due to the logarithmic triangle singularity \cite{Burns:2022uiv}. In this paper we explore related possibilities for $P_{\psi s}^\Lambda(4338)$.
\section{Model}
We assume, as in our previous work \cite{Burns:2020epm,Burns:2022uiv}, that the distribution can be described by triangle diagrams which couple to the final state through interactions which respect heavy-quark symmetry, and that the dominant diagrams are those with color-favored weak vertices. Hence we consider $\bar B\to D_s^{(*)-}\bar D\*$ transitions, noting that such branching fractions range from approximately 1-3\%. We can then form the triangle diagram shown in Fig.~\ref{fig:prod} (left), involving virtual $\Lambda_c$ exchange, followed by rescattering into the final $J/\psi\,\Lambda$. Generically, the distribution associated with this diagram peaks around the $\Lambda_c^+ D_s^-$ threshold (where it has a cusp). We notice that the $J/\psi \Lambda$ distribution~\cite{ChenParticleZoo} has exactly this shape, and we regard this as strong support for this proposed production mechanism. We also notice that, similar to the other exotic hadron systems~\cite{Burns:2021jlu,Burns:2022uiv}, the tree-level diagrams for the $ J/\psi \, \Lambda\, \bar{p}$ final state are color-suppressed; hence it is natural to assume that the color-favored triangle diagram is a dominant contribution.
\begin{figure*}
\centering
\includegraphics[width=0.45\textwidth]{T1.pdf}
\includegraphics[width=.45\textwidth]{T2.pdf}
\caption{Diagram $T_1$ (left) has a color-favored weak transition, whereas $T_2$ (right) is color-suppressed but enhanced at the $\Xi_c\bar D$ threshold due to the triangle singularity. The solid circles indicate non-perturbative final-state interactions, as described in the text.}
\label{fig:prod}
\end{figure*}
The proximity of $P_{\psi s}^\Lambda(4338)$ to $\Xi_c\bar D$ threshold also suggests a role for diagrams in which $\Xi_c\bar{D}$ rescatters to $ J/\psi\, \Lambda$, and here we notice an intriguing possibility. Such a novel intermediate state can be realized via an electroweak decay, such as $B^- \to \bar\Lambda_c \Xi_c$ or $B^- \to \bar\Sigma_c \Xi_c$, as shown in Fig.~\ref{fig:prod}~(right). Although the electroweak vertex is color-suppressed, we notice that the first of these modes has been observed in experiment \cite{Belle:2018kzz}, with quite significant branching fraction $(9.51\pm 2.1\pm 0.88)\times 10^{-4}$. Even accounting for some suppression at the electroweak vertex, such diagrams can make a significant contribution to the amplitude near the $\Xi\bar D$ threshold, if the mass of the $\bar \Lambda_c$ or $\bar \Sigma_c$ leads to a logarithmic singularity in the triangle diagram. The same idea has been applied in different contexts to explain other exotic hadron phenomena, as reviewed in ref.~\cite{Guo:2019twa}.
We will concentrate on the $\bar\Sigma_c$ (rather than $\bar \Lambda_c$) diagram, since only this is capable of producing the specific charge combination $\Xi_c^+D^-$ which, given the mass of $P_{\psi s}^\Lambda(4338)$, appears to be relevant. Solving the non-relativistic version of the Landau equations \cite{Guo:2019twa}, we find that the requisite $\bar\Sigma_c$ mass is 2810~MeV. Amazingly, there is a three-star resonance $\Sigma_c(2800)$ whose mass is almost exactly right. The skeptical reader may suspect
that there is a large number of possible $\Sigma_c$ states, and
that we are simply choosing one with the right mass in
order to make our proposed mechanism work. On the contrary -- aside from the familiar ground states $\Sigma_c(2455)$ and $\Sigma_c(2520)$, the only $\Sigma_c$ baryon in the Particle Data Group tables
is $\Sigma_c(2800)$~\cite{ParticleDataGroup:2020ssz}.
Hence we will attempt to fit the $J/\psi\, \Lambda$ distribution, adopting the following model for the $B^-\to J/\psi \, \Lambda\, \bar{p}$ amplitude:
\begin{equation}
\mathcal{A} = b + g_1 T_1 + g_2 \frac{1}{\sqrt{6}} \left[ 2 T_2^{(--)} - T_2^{(-)}\right],
\label{eq:A}
\end{equation}
where $b$ (the background, a complex constant) and $g_{1,2}$ (the production couplings) are fit to data, and $T_1$ and $T_2$ are the sub-amplitudes corresponding to the diagrams in the left and right panels of Fig.~\ref{fig:prod}, respectively, computed as described below.
Because the $P_{\psi s}^\Lambda(4338)$ peak is much closer to the threshold for $\Xi_c^+D^-$ rather than $\Xi_c^0\bar D^0$, we do not expect isospin symmetry to be respected in this system. This was discussed in refs~\cite{Yan:2022wuz,Meng:2022wgl}, and is similar to related observations in, for example, $X(3872)$ and the $P_c$ states~\cite{Swanson:2003tb,Close:2003sg,Burns:2015dwa,Guo:2019fdo,Guo:2019kdc}. In $T_2$ we therefore consider the $\Xi_c^+D^-$ and $\Xi_c^0\bar D^0$ diagrams separately, and refer to these as $T_2^{(--)}$ and $T_2^{(-)}$, corresponding to $\bar\Sigma_c^{--}$ and $\bar\Sigma_c^-$ states in the triangle diagrams. The remaining factors in equation~\eqref{eq:A} arise from an isospin decomposition of the amplitude. Notice that the weighted combination of $T_2^{(--)}$ and $T_2^{(-)}$ corresponds to $\Xi_c\bar D$ in a linear combination of isospin 0 and 1. Hence the production mechanism itself does not respect isospin, even before considering the mass difference between $\Xi_c^+D^-$ and $\Xi_c^0\bar D^0$. This is a different (and more significant) source of isospin mixing than that due to the $\Xi_c\bar D$ masses; similar effects could be present in the $X(2900)$ system~\cite{Burns:2020xne}.
Our calculation of the amplitudes $T_1$ and $T_2$ follows the method outlined in detail in our previous paper~\cite{Burns:2022uiv}, so here we just mention some key points. The triangle diagrams are computed in the nonrelativistic limit, incorporating form factors for the strong decay vertex and the final state interactions. The form factor scale is fixed at 800~MeV, as in the previous study. Because modeling the electroweak vertex is difficult, and because it does not vary much over the rather narrow phase space available, we have chosen to employ a constant electroweak vertex, whose strength is ultimately absorbed into production couplings $g_1$ and $g_2$ which, as discussed below, are fit to data.
In diagram $T_1$, the functional dependence of the amplitude is insensitive to the choice of the ``$\bar D\*$'' mass. Hence instead of separately considering contributions from $\bar D$ and $\bar D^*$, we compute a single diagram, using the physical $\bar D^*$ mass.
In diagram $T_2$, we notice that, in order to generate a prominent enhancement at $\Xi_c\bar D$ threshold, we need the $\bar\Sigma_c\to\bar D\bar p$ vertex to be S-wave. The corresponding $\Sigma_c$ state therefore has $J^P$ quantum numbers $1/2^-$. The quantum numbers of $\Sigma_c(2800)$ have not been measured, but the mass is consistent with expectations for the 1P multiplet, which includes states with quantum numbers $1/2^-$, $3/2^-$ and $5/2^-$ \cite{Copley:1979wj,Chen:2007xf,Cheng:2006dk,Cheng:2015naa,Wang:2021bmz}. We therefore assume that there is a $1/2^-$ state around 2800~MeV, though we are not necessarily assuming it is $\Sigma_c(2800)$ itself. For the purposes of the calculation, we fix the ``$\bar\Sigma_c$'' mass to 2801~MeV (the measured mass of $\Sigma_c(2800)^{++}$~\cite{ParticleDataGroup:2020ssz}), but we notice that the fit quality is not highly sensitive to this choice. (Although the logarithmic singularity strictly arises within a specific and narrow window of $\bar\Sigma_c$ masses, in practice we find very strong enhancements in the triangle diagram over a range of $\bar\Sigma_c$ masses around 2800~MeV.)
Most particles in the triangle diagrams are very narrow, so we ignore their widths and instead introduce a small imaginary part $\epsilon$ in the energy denominators. The exception is $\bar\Sigma_c$, whose width is included explicitly. We consider two cases: $\Gamma=70$~MeV (from the measured width of the $\Sigma_c(2800)$), and $\Gamma=15$~MeV (a model prediction for the width of the $1/2^-$ state in the 1P multiplet~\cite{Wang:2021bmz}).
The triangle diagrams couple to the $J/\psi\, \Lambda$ final state via non-perturbative final-state interactions, represented by the solid circles in Fig.~\ref{fig:prod}. Following ref.~\cite{Burns:2022uiv}, we assume a separable form for the interaction potential, and so obtain the non-perturbative T-matrix by solving the Bethe-Heitler equation, $T= V + VGT$, using algebraic methods. The final-state interactions are responsible for couplings among all the relevant channels so far discussed ($\Lambda_c^+D_s^-$, $\Xi_c^+D^-$, $\Xi_c^0 \bar D^0$, $\Lambda J/\psi$), as well as others which are related to these by heavy-quark symmetry and which also couple to $1/2^-$ in S-wave. In particular, we include $\Lambda \eta_c$, $\Sigma J/\psi$ and $\Sigma\eta_c$ as possible final states of interest, noting that the $\Sigma$ modes are relevant because of the explicit isospin mixing in the model. We do not include other channels such as $\Lambda_c D_s^*$, $\Xi_c\bar D^*$ and $\Xi_c^{(\prime,*)}\bar D\*$, whose thresholds are beyond the kinematic boundary for $J/\psi \Lambda$ in $B^-\to J/\psi \, \Lambda\, \bar{p}$.
We choose to model the final-state interactions as contact terms constrained by heavy-quark symmetry, following an approach which has been widely applied to $\Sigma_c\*\bar D\*$ systems~\cite{Liu:2018zzu,Liu:2019tjn,Sakai:2019qph,Valderrama:2019chc,Liu:2019zvb,Du:2019pij,Peng:2020xrf,Du:2021fmf}, and more recently $\Xi_c^{(\prime,*)}\bar D\*$ systems \cite{Peng:2019wys,Yan:2021nio,Peng:2020hql,Liu:2020hcv,Yan:2022wuz}. We previously tabulated the relevant contact terms for S-wave interactions among isospin 1/2 channels $\Lambda_c\bar D\*$, $\Sigma_c\*\bar D\*$, $NJ/\psi$ and $N\eta_c$~\cite{Burns:2022uiv}. Most of the contact terms we need for the present case can be extracted from those by assuming, as in refs~\cite{Peng:2020hql,Yan:2021nio,Yan:2022wuz}, that the interactions are invariant under rotations in $SU_3$ flavor space. Hence the matrix elements for the $\|SU_3\textrm{~flavor},SU_2\textrm{~isospin}\>$ states $\|\mathbf 8,\mathbf 1\>$ and $\|\mathbf 8,\mathbf 3\>$, formed out of $\Lambda_c^+D_s^-$, $\Xi_c^+D^-$, $\Xi_c^0 \bar D^0$ using $SU_3$ isoscalar factors~\cite{Kaeding:1995vq}, are identical to those of $\Lambda_c\bar D$, which is the $\|\mathbf 8, \mathbf 2\>$ state with the same spin structure. In a similar way, matrix elements involving $\LambdaJ/\psi$ and $\SigmaJ/\psi$ ($\Lambda\eta_c$ and $\Sigma\eta_c$) are the same as the corresponding terms in our previous paper involving $NJ/\psi$ ($N\eta_c$). In summary, we have
\begin{align}
\<\mathbf 8,\mathbf1\|V\|\mathbf 8,\mathbf 1\>&=\<\mathbf 8,\mathbf 3\|V\|\mathbf 8,\mathbf 3\>=A,\\
\<\mathbf 8,\mathbf1\|V\|\Lambda J/\psi\>&=\<\mathbf 8,\mathbf 3\|V\|\Sigma J/\psi\>=\frac{\sqrt 3}{2}D,\\
\<\mathbf 8,\mathbf1\|V\|\Lambda \eta_c\>&=\<\mathbf 8,\mathbf 3\|V\|\Sigma \eta_c\>=\frac{1}{2}D,
\end{align}
where $A$ and $D$ are contact terms which are (somewhat) constrained by our previous analysis of $\Lambda_b\to \jp\, p K^-$ decays \cite{Burns:2022uiv} (see below).
The $\Lambda_c^+D_s^-$, $\Xi_c^+D^-$, $\Xi_c^0 \bar D^0$ basis states also combine into a flavor singlet, which implies an additional contact term which is independent of the other contact terms \cite{Yan:2021nio}. We call this $A'$, by analogy with $A$ above:
\begin{align}
\<\mathbf 1,\mathbf 1\|V\|\mathbf 1,\mathbf 1\>=A'.
\end{align}
Although the potentials are assumed to respect $SU_3$ and $SU_2$ symmetries, the transition amplitudes will not, because of mass differences among the constituents with different flavors. Hence we formulate the potential in the particle basis, and show the result in Table~\ref{tab:V}, where we have introduced $\Delta = (A'-A)/3$. Note that, as in our previous work, we are assuming that potentials coupling two ``closed-charm'' states (such as $\Lambda J/\psi\to\Lambda J/\psi$) are zero.
\begin{table}
\begin{tabularx}{\columnwidth}{X|XXX|XXXX}
\hline
& $\Lambda_c^+ D_s^-$ & $\Xi_c^+ D^-$ & $\Xi_c^0 \bar{D}^0$ & $\Lambda J/\psi$ & $\Lambda \eta_c$ & $\Sigma J/\psi$ & $\Sigma \eta_c$ \\
\hline
$\Lambda_c^+ D_s^-$ & $A+\Delta$ & $\Delta$ & $-\Delta$ & $\frac{D}{\sqrt{2}}$ & $\frac{D}{\sqrt{6}}$ & 0 & 0 \\
$\Xi_c^+ \bar{D}^-$ & & $A+\Delta$ & $-\Delta$ & $-\frac{D}{2 \sqrt{2}}$ & $-\frac{D}{2 \sqrt{6}}$ & $\frac{\sqrt{3} D}{2 \sqrt{2}}$ & $\frac{D}{2 \sqrt{2}}$ \\
$\Xi_c^0 \bar{D}^0$ & & & $A+\Delta$ & $\frac{D}{2 \sqrt{2}}$ & $\frac{D}{2 \sqrt{6}}$ & $\frac{\sqrt{3} D}{2 \sqrt{2}}$ & $\frac{D}{2 \sqrt{2}}$ \\
\hline
$\Lambda J/\psi$ & & & & 0 & 0 & 0 & 0 \\
$\Lambda \eta_c$ & & & & & 0 & 0 & 0 \\
$\Sigma J/\psi$ & & & & & & 0 & 0 \\
$\Sigma \eta_c$ & & & & & & & 0 \\
\hline
\end{tabularx}
\caption{Contact terms in the $1/2^-$ channel.}
\label{tab:V}
\end{table}
With our model for the potentials, analysis of $\Lambda_b\to \jp\, p K^-$ decays \cite{Burns:2022uiv} indicates that $D$ is constrained very roughly to be a number of order 1 GeV$^{-2}$, and we will adopt that value in this work. The contact term $A$ is not well-constrained by $\Lambda_b\to \jp\, p K^-$ decays, although we know that it cannot be large and negative, as it would imply $\Lambda_c\bar D\*$ bound states, which are apparently not seen in the data. Hence in this work we consider $A\ge 0$.
\section{Results}
We first attempt a fit with only diagram $T_1$ (fixing $g_2=0$), corresponding to the conventional molecular scenario. To understand the possibilities, we note that in the isospin basis, the diagonal $\Xi_c\bar D$ potentials are
\begin{align}
V(\Xi_c\bar D,I=0)&=A+2\Delta,\label{eq:1}\\
V(\Xi_c\bar D,I=1)&=A.\label{eq:2}
\end{align}
Comparing to the $\Lambda_c^+D_s^-$ potential in Table~\ref{tab:V}, it suggests there may be some region of parameter space (with $\Delta<0$) in which $\Xi_c\bar D$ ($I=0$) binds, but not $\Lambda_c^+D_s^-$. But in practice we find the opposite, namely a prominent signal at (or below) $\Lambda_c^+D_s^-$ threshold, rather than $\Xi_c\bar D$. One reason is that $\Delta$ not only contributes to the diagonal potential, but also the off-diagonal coupling between the $\Lambda_c^+D_s^-$ and $\Xi_c\bar D$ channels,
generating an effective attraction in the lower channel ($\Lambda_c^+D_s^-$), and also broadening any peak associated with $\Xi_c\bar D$ (increasing its decay width). Ref~\cite{Yan:2022wuz} also observed that binding in $\Xi_c\bar D$ inevitably also implies binding in $\Lambda_c^+D_s^-$. We also remark that, in our model, the $\Lambda_c^+D_s^-$ state is always more prominent in the amplitude, since it is produced directly in the triangle diagrams -- in contrast to $\Xi_c\bar D$, which arises in diagram $T_1$ only through final-state interactions (via $\Delta$). All of this is problematic for the molecular scenario.
Hence we proceed with the full model, including both diagrams $T_1$ and $T_2$. The results are much better: in Fig.~\ref{fig:fit1} we give two illustrative examples of successful fits where, in both cases, the bulk of the distribution is captured by diagram $T_1$, while the $P_{\psi s}^\Lambda(4338)$ peak is from diagram $T_2$, due to the anticipated triangle singularity. The shape of the triangle peak is sensitive to the chosen width of the virtual $\bar\Sigma_c$, which in turn has implications for the contact terms $A$ and $\Delta$. The two parameter sets illustrated in Fig.~\ref{fig:fit1} are summarized in Table~\ref{tab:par}.
\begin{table}
\centering
\begin{tabularx}{\columnwidth}{XXXl}
\hline
& $\Gamma(\bar\Sigma_c)$ / MeV&$A$ / GeV$^{-2}$ &$\Delta$ / GeV$^{-2}$\\
\hline
Set A & 70&6&$-7$\\
Set B & 15& 0&$-1$\\
\hline
\end{tabularx}
\caption{Parameter sets A (green) and B (blue) in Fig~\ref{fig:fit1}.}
\label{tab:par}
\end{table}
The effect of a broader $\bar\Sigma_c$ is to broaden the triangle peak in diagram $T_2$. Hence in Set~A, in order to obtain a sufficiently sharp peak, we adjust $A$ and $\Delta$ to effectively add some attraction to the $\Xi_c\bar D$ channel, but in such a way that we do not have a bound state in $\Xi_c\bar D$ or $\Lambda_c^+D_s^-$. With the narrower $\bar\Sigma_c$ of Set~B, the triangle peak is already sharp enough that no extra attraction is really needed, although with our chosen $\Delta = -1$~GeV$^{-2}$ there is a small attraction, as well as coupling between the $\Xi_c\bar D$ and $\Lambda_c^+D_s^-$ channels. Note that the values of $A$ and $\Delta$ are not very well constrained by the fits.
To better understand our results, let us consider one of the fits (Set A) in some more detail. The fit parameters are
\begin{align}
b& = 0.93 + 6.63 i \ \textrm{GeV}^{-2},\\
g_1& = 5766, \\
g_2& = -316.1.
\end{align}
In Fig.~\ref{fig:fit1} we show separately (with thin dotted lines) the contributions from diagrams $T_1$ and $T_2$, and the background. Notice that $T_1$ nicely captures the overall shape of the distribution, including the peak around the $\Lambda_c^+D_s^-$ threshold which, as mentioned previously, is a natural consequence of the assumed dominance of the color-favored triangle diagram. While it is amusing that the peak corresponds to the single data point near 140 candidates, we regard this as a fluke. It is also reassuring that the background is relatively small compared to $T_1$, and this is as expected, since tree-level production of the final state is color-suppressed.
We are also satisfied that the fit has $|g_2|<<|g_1|$, since it is consistent with expectations that the electroweak vertex in $T_2$ is suppressed compared to $T_1$ by at least the number of colors, with further suppression expected due to the orbital excitation in $\bar\Sigma_c$. As has been anticipated, despite a small coupling $g_2$, diagram $T_2$ still makes a prominent contribution to the fit because of the enhancement due to the triangle singularity.
The chi-squared value for the fit with Set A parameters is 1.9. This can be improved substantially by using a more complicated background model; however, our purpose is not to obtain a very precise fit, but to establish the physical mechanism that explains the data.
\begin{figure}[ht]
\centering
\includegraphics[width=\textwidth]{fit3-final.pdf}
\caption{The $J/\psi \Lambda$ invariant mass spectrum for parameter sets A (green) and B (blue), compared to the experimental data from ref.~\cite{ChenParticleZoo}. The thin dotted lines show the separate contributions of $T_1$, background, and $T_2$ (top to bottom), for parameter set A.
}
\label{fig:fit1}
\end{figure}
From the matrix elements in Table~\ref{tab:V}, we can make some general observations about the prospects of observing $P_{\psi s}^\Lambda(4338)$ in other final states. (Detailed predictions for distributions in various final states are difficult, without a model for the backgrounds in each case.)
For example, in $B^-\to \eta_c\, \Lambda\, \bar{p}$ decays, we expect a $P_{\psi s}^\Lambda(4338)$ signal in the $\eta_c\, \Lambda$ distribution, but suppressed in comparison to the $J/\psi\,\Lambda$ signal by a factor of approximately 3 (in rate).
As noted previously, we expect $P_{\psi s}^\Lambda(4338)$ to exhibit isospin mixing, as the $\Xi_c\bar D$ pair in diagram $T_2$ is a linear combination of isospin 0 and 1. If we ignore the (small) additional contribution to mixing due to the charged/neutral mass difference, and assume that $P_{\psi s}^\Lambda(4338)$ is solely due to the triangle singularity (in the sense that there are no non-perturbative final state interactions in $\Xi_c\bar D$), we find that in diagram $T_2$, the signals in isospin 1 modes ($J/\psi\, \Sigma^0$ and $\eta_c \Sigma^0$) are suppressed compared to the corresponding isospin 0 modes ($J/\psi\, \Lambda$ and $\eta_c \Lambda$), but only by a factor of 3 in rate (before accounting for phase space differences). Isospin mixing in this system is therefore a very significant effect. Moreover, the $P_{\psi s}^\Lambda(4338)$ peak will feature more prominently in the $J/\psi\, \Sigma^0$ and $\eta_c \Sigma^0$ distributions since (at leading order) they receive no contribution from the diagram $T_1$, which dominates the $J/\psi\Lambda$ spectrum.
But the factor of three only applies in the perturbative limit, and indeed we find that non-perturbative final state interactions can change the outcome considerably. For example, in parameter set A (Table~\ref{tab:par}), the values of $A$ and $\Delta$ imply attraction in isospin 0, but repulsion in isospin 1 -- see equations~\eqref{eq:1} and \eqref{eq:2}. So by introducing final state interactions to enhance the triangle peak in isospin 0, we effectively suppress the peak in isospin 1. Indeed we have verified that with our parameter set A, the suppression of the isospin 1 peak is very much stronger than the factor of 3 which applies in the perturbative limit.
Because the magnitude of isospin mixing is correlated with the parameters $A$ and $\Delta$, future experimental measurement of isospin 1 modes could be used to constrain these parameters, which cannot be determined with current data.
\section{Conclusion}
We have demonstrated that a simple model based on the assumed dominance of color-favored weak transitions, and approximate heavy-quark and $SU_3$ flavor symmetries, can describe the $B^-\to J/\psi \, \Lambda\, \bar{p}$ data, including the prominent $P_{\psi s}^\Lambda(4338)$ peak. Notably, we are not assuming a molecular nature for $P_{\psi s}^\Lambda(4338)$ -- on the contrary, we showed that the molecular scenario does not work. Our conclusions (and methods) are very similar to those of our recent analysis of $\Lambda_b\to\jp\, p K^-$ data \cite{Burns:2021jlu,Burns:2022uiv}, in which we argued that the $\Sigma_c\bar D^*$ molecular interpretation of $P_c(4457)$ is problematic, but that several viable alternatives give an excellent fit to data. Our results underline the importance of exploring alternatives to the prevailing molecular interpretation of states which are located at thresholds.
\acknowledgments
Swanson's research was supported by the U.S. Department of Energy under contract DE-SC0019232.
|
\section{Introduction}
\label{introduction}
Consider a weakly separable model with a binary endogenous variable:%
\begin{eqnarray}
Y &=&g(v_1(X,D),v_2(X,D),...v_J(X,D),\varepsilon ) \label{outcomeeq} \\
D &=&1\left \{ \theta (Z)-U>0\right \} \label{treateq}
\end{eqnarray}%
where $( v_1(X,D),v_2(X,D),...v_J(X,D) ) \equiv v(X,D) $ is a $J$-vector of
unknown linear or nonlinear indices in the outcome equation (1.1) and $D$ is
a binary endogenous variable defined by equation (1.2). Here $%
X\in \mathbb{R}^{d_{x}}$ and $Z\in \mathbb{R}^{d_{z}}$ are vectors of
observable exogenous variables, which may have overlapping elements. This paper is about the identification and estimation of the
average treatment effect (ATE).\footnote{%
The ATE is one of several parameters of interest in the program evaluation
literature. In their seminal work, \citeasnoun{imbensangrist} focus on a
different parameter of interest, the Local Average Treatment Effect (LATE)
which is the ATE for the subset of the population referred to as compliers.
However, as pointed out in \citeasnoun{HV05}, \citeasnoun{heckvyt-handbook}, %
\citeasnoun{CHV10}, and \citeasnoun{santos18}, there are settings where the
LATE by itself is not always the focal point of policy studies.} Similar to %
conditions in \citeasnoun{vytlacilyildiz}, we require that there is some continuous element in $Z$
excluded from $X$ so that $\theta(Z)$ varies continuously conditional on $X$, and that we can vary $X$ after conditioning on $\theta(Z)$.%
\footnote{%
This does not necessarily require a second exclusion restriction that there be an
element in $X$ not in $Z$. As explained in \citeasnoun{vytlacilyildiz}, what
is required is that the element in $Z$ not in $X$ be continuously
distributed.}
It is worth noting that the method we propose also applies directly in a more general setup where the error in the outcome equation is specific to $D$, i.e., with $\varepsilon$ replaced by $\varepsilon_D$, provided the marginal distribution of $\varepsilon_d$ is the same for $d\in\{0,1\}$.
The literature on program evaluation abounds in examples in which self-selected, endogenous treatment depends on exogenous instruments that have no direct impact on potential outcomes.
For instance, \citeasnoun{vytlacilyildiz} provided an example where $D$ indicates an individual's enrollment in job training at date 0, and $Y$ is employment status at date 1. In this case, $Z$ may contain variables summarizing local labor market conditions at date 0. They noted these variables affect the opportunity cost of training, and consequently the training decision at date 0, but do not directly affect the individual's employment at date 1.
In the system of equations above, $U$ is the unobservable
random variable normalized to follow the standard uniform distribution, and
the error term $\varepsilon$ in the outcome equation is allowed to be a
random vector with a known dimension. We assume $\left( X,Z\right) $ are independent of $%
(\varepsilon ,U)$. Note that we allow $v(X,D)$ to be a vector of multiple
indices, whereas existing methods, such as \citeasnoun{vytlacilyildiz}, can only be
applied when it is a single index.
\citeasnoun{HV05} and \citeasnoun{carneiroLee2009} maintained that $(\varepsilon,U)$ is independent of $Z$ conditional on $X$, and used local instrumental variables to estimate marginal treatment effect (MTE) for realized values of $U$ over the support of $\theta(Z)$ given $X$. In their case, integrating out MTE to ATE would require the support of $\theta(Z)$ given $X$ to be the complete unit interval, i.e. $[0,1]$.
In comparison, we propose a method that relaxes such a support condition by exploiting more structure in the potential outcome equation and its full distribution.
Moreover, while \citeasnoun{carneiroLee2009} use $E[1\{Y\leq y\}|X,D,U=p]$ for \textit{each fixed $y$} to recover the conditional distribution of potential outcomes at $y$, we use the full distribution of the observed outcomes to find the matching covariates. This helps us to recover ATE without the aforementioned full support condition.
As in \citeasnoun{vytlacilyildiz}, our method requires that at least some covariates in the potential outcome equation are exogenous.
Since \citeasnoun{vytlacilyildiz}, important work has considered
identification and estimation of similar models, but under alternative
conditions, notably on the supports of $D$, and $\theta(Z)$ as well
as the dimension of $\varepsilon$. \citeasnoun{neweyimbens} assume $D$ is
continuous and monotone in the error term, $\theta(Z)$ is continuous with
large support, and $\varepsilon$ is scalar. \citeasnoun{kasy14} allows $%
\varepsilon$ to be a vector as we do here but imposes that both $D$ and $%
\theta(Z)$ are continuously distributed and further assumes a monotone
relationship between the two. \citeasnoun{dhault2014} and %
\citeasnoun{torgo2014} assume $D$ is continuous, $\varepsilon$ is scalar,
and assume monotonicity in the error term in each of the two equations,
though they allow for $\theta(Z)$ to be discrete. \citeasnoun{vuongxu}
assume $D$ is discrete and $\theta(Z)$ is continuous, restricts $\varepsilon$
to be scalar and requires $Y$ in (\ref{outcomeeq}) to be monotone in $%
\varepsilon$. \citeasnoun{JunEtal2016} study an extension of %
\citeasnoun{vytlacilyildiz}, and also require the same monotonicity
condition as in the latter. \citeasnoun{fengjun} shows how
to identify nonseparable triangular models where the endogenous variable is
discrete but restricted to have larger support than the instrument variable.%
\footnote{%
These papers focus on point identification. For partial identification of a
model with a binary outcome, see \citeasnoun{shaikhvytlacil11} and %
\citeasnoun{mourifie15}. \citeasnoun{santos18} explore partial
identification of relevant treatment effect parameters in models without
structure imposed on the outcome equation. In such settings, attaining point
identification requires support conditions on the propensity score function
that are stronger than imposed here.}
As in the conventional framework, two potential outcomes $Y_{1}$ and $Y_{0}$
satisfy%
\begin{equation*}
Y_D = g(v(X,D),\varepsilon) \mbox{ for } D=0,1. \text{\ }
\end{equation*}%
We only observe $\left( Y,D,X,Z\right) $, where $Y=DY_{1}+(1-D)Y_{0}$. In this model, we do not impose parametric distribution on the error terms $(\varepsilon,U)$ or a linear index structure. %
Similarly, \citeasnoun{vytlacilyildiz} also do not require such restrictions, but assume that $v(X,D) \in \mathbb{R}$ is a single
index, and
\begin{equation}
E\left[ g(v,\varepsilon )|U=u\right] \mbox{ is strictly increasing in }v \in
\mathbb{R}\mbox{ for all }u.
\end{equation}%
We do not impose any such monotonicity structure. That is because our approach exploits the full information in the distribution of the outcome variable, instead of just its mean.
Indeed, when the outcome distribution function is more informative than the mean, our method is applicable to more general settings;
in particular, not only do we not rely on such a monotonicity assumption, but we also allow for multiple indices.
In Section \ref{Identification} we present the identification argument and discuss its required conditions.
In Sections \ref{Examples} and \ref{Extension} we provide some examples in
which such a monotonicity condition fails, but the average effect of the
binary endogenous variable is still identified. To reiterate, we allow the potential outcomes to be weakly separable in multiple indices, that is, $%
v(X,D)=(v_{1}(X,D),v_{2}(X,D),...,v_{J}(X,D)) \in \mathbb{R}^J$.
We consider the identification and estimation of the average treatment
effect of $D$ on $Y$, $E(Y_{1}|X\in A)$, $E(Y_{0}|X\in A)$, and $%
E(Y_{1}-Y_{0}|X\in A)$, for some set $A$, without the aforementioned
monotonicity. Indeed, for the case with multiple indices $v(X,D) \in \mathbb{%
R}^J$, the monotonicity condition is no longer well defined.
\setcounter{equation}{0}
\section{Identification\label{Identification}}
Our identification strategy is based on the notion
of \emph{matching covariates}.
Let $Supp(V)$ denote the support of a generic random vector $V$.
Consider the identification of $E(Y_{1}|X=x)$ for some $x\in Supp(X)$.
Under the assumption that $%
(\varepsilon ,U)\bot (X,Z)$,%
\begin{eqnarray}
&&\text{ }E(Y_{1}|X=x)=E(Y_{1}|X=x,Z=z) \notag \\
&&\text{ }=E(DY_{1}|X=x,Z=z)+E[(1-D)Y_{1}|X=x,Z=z] \notag \\
&&\text{ }=P(z)E(Y|D=1,X=x,Z=z)+[1-P(z)]E(Y_{1}|D=0,X=x,Z=z) \label{EQ1}
\end{eqnarray}%
where $P(z)\equiv E(D|Z=z) = \theta(z)$ because $U$ is normalized to follow a standard uniform distribution. The only term not directly identifiable
on the right-hand side of (\ref{EQ1}) is:\footnote{
If the support of $P(Z)$ given $x$ covers the full closed interval $[0,1]$, then $E(Y_1|X=x)$ is directly identified as $E(Y|D=1,X=x,Z=z)$ at $z$ s.t. $P(z)=1$. However this means point identification hinges on the event ``$P(Z)=1$ given $x$''.
}%
\begin{equation*}
E(Y_{1}|D=0,X=x,Z=z)=E[g(v(x,1),\varepsilon )|U\geq P(z)]\text{.}
\end{equation*}
The main idea behind our approach would at first seem to be similar to that of \citeasnoun{vytlacilyildiz}%
, which is to use exogenous variations of covariates in the outcome equation to
find some $(\tilde{x},\tilde{z})\in Supp(X,Z)$ such that%
\begin{equation}
P(z)=P(\tilde{z}) \text{ and } v(x,1)=v(\tilde{x},0) \label{EQ2}
\end{equation}%
so that%
\begin{eqnarray} \label{EQ23}
E(Y|D=0,X=\tilde{x},Z=\tilde{z})&=&E(Y_{0}|D=0,X=\tilde{x},Z=\tilde{z}) \\
&=&E[g(v(\tilde{x},0),\varepsilon )|U\geq P(\tilde{z})]=E[g(v(x,1),\varepsilon
)|U\geq P(z)] \notag \\
&=& E[Y_1|D=0,X=x, Z=z]\text{.} \notag
\end{eqnarray}
However, unlike \citeasnoun{vytlacilyildiz}, we utilize the full distribution of $Y$ (rather than its first moment only) while searching for such pairs of $(x,%
\tilde{x})$ in (\ref{EQ2}). This allows us to relax the single-index and monotonicity conditions.
To fix ideas, let there be continuous components in $Z$ that are excluded from $X$.
For any $p$ on the support of $P(Z)$ given $X=x$, and for any $y$, define
\begin{eqnarray}
h_{1}^{\ast }(x,y,p) &=&E[D1\left \{ Y\leq y\right \} |X=x,P(Z)=p] \notag \\
&=&E\left[ 1\left \{ U<P(Z)\right \} 1\left \{ g(v(X,1),\varepsilon )\leq
y\right \} |X=x,P(Z)=p\right] \notag \\
&=&\int_{0}^{p}F_{g|u}(y;v(x,1))du\text{,} \label{EQ3}
\end{eqnarray}%
where
\begin{equation*}
F_{g|u}(y;v(x,d))\equiv E[1\{g(v(x,d),\varepsilon )\leq y\}|U=u]\text{,}
\end{equation*}%
with $v(x,d)$ being a realized index at $X=x$ and the expectation in the
definition of $F_{g|u}$ is with respect to the distribution of $\varepsilon $
given $U=u$. The last equality in (\ref{EQ3}) holds because of independence
between $(\varepsilon ,U)$ and $(X,Z)$.
Assume that for all $y,x$ and $d \in \{ 0,1 \}$, the function $F_{g|u}(y;v(x,d))$ is continuous in $u$ over $[0,1]$. This implies $h^*_1$ is differentiable in $p$.
By construction, $h_{1}^{\ast
}(x,y,p)$ is directly identified from the joint distribution of $(D,Y,X,Z)$
in the data-generating process. Furthermore, for any pair $p_{1}>p_{2}$,
define:%
\begin{equation*}
h_{1}(x,y,p_{1},p_{2})\equiv h_{1}^{\ast }(x,y,p_{1})-h_{1}^{\ast
}(x,y,p_{2})=\int_{p_{2}}^{p_{1}}F_{g|u}(y;v(x,1))du\text{.}
\end{equation*}%
Likewise, define
\begin{eqnarray*}
h_{0}^{\ast }(x,y,p) &=&E((1-D)1\left \{ Y\leq y\right \} |X=x,P(Z)=p) \\
&=&E\left[ 1\left \{ U\geq P(Z)\right \} 1\left \{ g(v(X,0),\varepsilon
)\leq y\right \} |X=x,P(Z)=p\right] \\
&=&\int_{p}^{1}F_{g|u}(y;v(x,0))du.
\end{eqnarray*}%
and let
\begin{equation*}
h_{0}(x,y,p_{1},p_{2})\equiv h_{0}^{\ast }(x,y,p_{2})-h_{0}^{\ast
}(x,y,p_{1})=\int_{p_{2}}^{p_{1}}F_{g|u}(y;v(x,0))du\text{.}
\end{equation*}%
Let $\mathcal{P}_{x}$ denote the support of $P(Z)$ given $X=x$; let $ Int( \mathcal{P}_{x}\cap \mathcal{P}_{\tilde{x}} )$ denote the interior of the intersection of $\mathcal{P}_{x}$ and $\mathcal{P}_{\tilde{x}}$. Similar to \citeasnoun{HV05} and \citeasnoun{carneiroLee2009}, we use continuous, exogenous variation in $Z$ to establish here that for any $ x \neq \tilde{x}$ with $ Int( \mathcal{P}_{x}\cap \mathcal{P}_{\tilde{x}} ) \neq \emptyset$ and any $y$,%
\begin{equation}
h_{1}(x,y,p,p^{\prime })=h_{0}(\tilde{x},y,p,p^{\prime })\text{ for all }%
p>p^{\prime }\text{ on } Int(\mathcal{P}_{x}\cap \mathcal{P}_{\tilde{x}}) \label{EQ4}
\end{equation}%
if and only if%
\begin{equation}
F_{g|p}(y;v(x,1))=F_{g|p}(y;v(\tilde{x},0))\text{ for all }p\in Int(\mathcal{P}_{x}\cap \mathcal{P}_{\tilde{x}})\text{.} \label{EQ5}
\end{equation}
\noindent Sufficiency of (\ref{EQ5}) is immediate from the definition of $h_{1}$ and $h_{0}$.
To see its necessity, note that for all $p>p^{\prime }$ on $Int(\mathcal{P}_{x}\cap \mathcal{P}_{%
\tilde{x}})$,%
\begin{equation*}
\left. \frac{\partial h_{1}(x,y,\tilde{p},p^{\prime })}{\partial \tilde{p}}%
\right \vert _{\tilde{p}=p}=\left. \frac{\partial h_{1}^{\ast }(x,y,\tilde{p}%
)}{\partial \tilde{p}}\right \vert _{\tilde{p}=p}=F_{g|p}(y;v(x,1))
\end{equation*}%
{and}%
\begin{equation*}
\left. \frac{\partial h_{0}(\tilde{x},y,\tilde{p},p^{\prime })}{\partial
\tilde{p}}\right \vert _{\tilde{p}=p}=-\left. \frac{\partial }{\partial
\tilde{p}}h_{0}^{\ast }(\tilde{x},y,\tilde{p})\right \vert _{\tilde{p}%
=p}=F_{g|p}(y;v(\tilde{x},0))\text{.}
\end{equation*}%
Thus (\ref{EQ4}) and (\ref{EQ5}) are equivalent.
Next, we collect identifying assumptions as follows:\medskip
\noindent ASSUMPTION A-1: The distribution of $U$ is absolutely continuous
with respect to Lebesgue measure, and is normalized to standard uniform, and $\theta(z)$ in (\ref{treateq}) is continuous in $z$. \medskip
\noindent ASSUMPTION A-2: The random vectors $(U,\varepsilon )$ and $(X,Z)$ are independent.
There exists at least one continuously distributed component in $Z$ that is excluded from $X$.
\medskip
\noindent ASSUMPTION A-3: Both $g(v(X,1),\varepsilon )$ and $g(v(X,0),\varepsilon )$ have finite first moments conditional on $U=u$ for all $u \in [0,1]$.
\noindent ASSUMPTION A-4: For all realized values of $y,x$ and $d\in\{0,1\}$, $E[1\{g(v(x,d),\epsilon)
\leq y\}|U=u] $ is continuous in $u$ over $[0,1]$. \medskip
\noindent
ASSUMPTION A-5: For any $x\neq \tilde{x}$ such that $Int(\mathcal{P}_{x}\cap \mathcal{P}_{\tilde{x}})$ is nonempty,
$F_{g|p}(y;v(x,1))=F_{g|p}(y;v(\tilde{x},0))$ for all $y$ and $p\in Int(\mathcal{P}_{x}\cap \mathcal{P}_{\tilde{x}})$ if and only if $v(x,1)=v(\tilde{x},0)$.\medskip
Assumptions A-1 and A-3 are common regularity conditions in the literature. Assumption A-2 consists of a standard condition of instrument exogeneity. It allows $P(Z)$ to have continuous, exogenous variation conditional on $X$. The continuity of instrument is also common in the literature on treatment effects. Examples include \citeasnoun{heckmanvytlacil} and \citeasnoun{vytlacilyildiz}.
As noted above, Assumption A-4 ensures the identifiable functions $h^*_1(x,y,p),h^*_0(x,y,p)$ are both differentiable in the last argument $p$.
It is important to note that Assumption A-5 allows the support of $P(Z)$ to be a strict subset of the interval $(0,1)$.
This is because our method does not use an identification-at-infinity argument, which would require $P(Z)$ to have full support [0,1] given $X$ in order to identify $E(Y_0|X)$ directly.
(See Footnote 4 above.)
We also note that Assumption A-5 relaxes two limitations of Assumption 4 in \citeasnoun{vytlacilyildiz}. Specifically, to identify pairs $(x,\tilde{x})$ with $v(x,1)=v(\tilde{x},0) $, \citeasnoun{vytlacilyildiz} rely on two assumptions that $v(X,D) \in \mathbb{R}$ is a single index and that $E\left[ g(v(x,1),\varepsilon )|U=u\right] =E\left[ g(v(\tilde{x},0),\varepsilon )|U=u\right] $ if and only if $v(x,1)=v(\tilde{x},0) $.
The latter holds under a maintained assumption that $E\left[
g(v(x,d),\varepsilon )|U=p\right] $ is a strictly monotonic function of $%
v(x,d)$.
In comparison, we construct an identification strategy without the single index and monotonicity restrictions by matching conditional distributions $F_{g|p}(\cdot ;v(x,1))$ and $\ F_{g|p}(\cdot ;v(\tilde{x},0))$.
The role of Assumption A-5 in our method can be illustrated by drawing an analogy with a standard identifying condition in nonlinear regression: $ Y = f(X,\theta ^*) + \varepsilon$. In this case, identification of $\theta^*$ requires: ``$f(x,\theta^* ) = f(x,\theta)$ for all $x$ if and only if $\theta^* =\theta$ ''.
To see how this is related to Assumption A-5 in our setting, let $ \theta_1, \theta_0$ be shorthand for $v(x,1), v(\tilde{x},0)$ respectively, and let $m(y,p,\theta) \equiv F_{g|p}(y,\theta )$.
Then our method requires: ``$m(y,p,\theta _{1})=m(y,p,\theta _{0})$ for all $(y,p)$ if and only if $\theta _{1}=\theta _{0}$ ''.
It is worth emphasizing that Assumption A-5 only presents our identification condition in the weakest form, for the sake of generality.
Later in Section \ref{Examples}, we exploit the structures embedded in specific examples to show how Assumption A-5 can be satisfied under intuitive, mild primitive conditions.
Next, we specify conditions under which such pairs of covariates exist on the support. Define $\mathcal S \equiv \{(x,\tilde{x}):v(x,1)=v(\tilde{x},0)\} $ and $ \mathcal T \equiv \{(x,\tilde{x}): \exists z, \tilde{z} \text{ with } (x,z),(\tilde{x},\tilde{z})\in Supp(X,Z) \text{ and } P(z)=P(\tilde{z}) \in Int(\mathcal{P}_x\cap \mathcal{P}_{\tilde{x}})\}$.
Let $ \mathcal X^1 \equiv \{x:\exists \tilde{x} \text{ with }(x,\tilde{x})\in \mathcal S \cap \mathcal T\}$ and $ \mathcal X^0 \equiv \{x:\exists \tilde{x} \text{ with }(\tilde{x},x)\in \mathcal S \cap \mathcal T\}$.
\noindent ASSUMPTION A-6: $\Pr (X\in \mathcal X^1)>0$ \ and $\Pr (X\in \mathcal X^0)>0$.\medskip
This condition is similar to Assumption A-4 in \citeasnoun{vytlacilyildiz}. By definition, for each $x\in\mathcal X^1$, we can find $\tilde{x}\in Supp(X)$ such that there exists $(z,\tilde{z}) $ with $ (x,z),(\tilde{x},\tilde{z}) \in Supp(X,Z)$, $v(x,1)=v(\tilde{x},0)$, and $P(z)=P(\tilde{z})$.
This requires variation in $ x $ while holding $P(Z)$ fixed.
As noted earlier (Footnote 2), this does \textit{not} necessarily require a ``second exclusion restriction'' that there is an element in $X$ that is excluded from $Z$.
In general, there exist multiple such values of $\tilde{x}$ that can be matched with this $x$.
Hence for each $x\in\mathcal X^1$, we define the set of such matched values as
\begin{equation*}
\lambda_0(x) \equiv \{\tilde{x}:
h_{1}(x,y,p,p^{\prime })=h_{0}(\tilde{x},y,p,p^{\prime }) \text{ }\forall p>p^{\prime }\text{ on } Int(\mathcal{P}_{x}\cap \mathcal{P}_{\tilde{x}})\},
\end{equation*}
and define $ \mathcal{P}_0^*(x) \equiv \bigcup_{\tilde{x} \in \lambda_0(x)}(\mathcal{P}_{\tilde{x}}\cap \mathcal{P}_x)$.
Note that by definition of $\mathcal{X}^1$, the set $\mathcal{P}_x \cap \mathcal{P}_{\tilde{x}}$ must be non-empty for $x \in \mathcal{X}^1$ and $ \tilde{x} \in \lambda_0(x)$ under our maintained assumptions.
Likewise, by symmetry, for each $x\in\mathcal X^0$, we define
\begin{equation*}
\lambda_1(x) \equiv \{\tilde{x}:
h_{0}(x,y,p,p^{\prime })=h_{1}(\tilde{x},y,p,p^{\prime }) \text{ }\forall p>p^{\prime }\text{ on } Int(\mathcal{P}_{x}\cap \mathcal{P}_{\tilde{x}})\},
\end{equation*}
and let $ \mathcal{P}_1^*(x) \equiv \bigcup_{\tilde{x} \in \lambda_1(x)}(\mathcal{P}_{\tilde{x}}\cap \mathcal{P}_x)$.
The theorem below shows how to use such matched values to identify the conditional mean of potential outcomes.
\begin{theorem}
\label{theorem2.1} Suppose Assumptions (A-1)-(A-6) hold. For each $x\in\mathcal X^1$, \begin{equation} \label{id_EY1}
E(Y_1|X=x) = E(DY|X=x,P(Z)=p) + E[(1-D)Y|X\in\lambda_0(x),P(Z)=p]
\end{equation}
for any $p\in \mathcal{P}_0^*(x)$.
For each $x\in\mathcal X^0$,
\begin{equation} \label{id_EY0}
E(Y_0|X=x) = E(DY|X\in\lambda_1(x),P(Z)=p) + E[(1-D)Y|X=x,P(Z)=p]
\end{equation}
for any $p\in \mathcal{P}_1^*(x)$.
\end{theorem}
\begin{proof} As established in the text, (\ref{EQ4}) and (\ref{EQ5}) are equivalent under Assumptions (A-1)-(A-4).
Then by Assumption (A-5), we know that for each $x\in\mathcal X^1$ and $\tilde{x}\in \lambda_0(x) $, there exists $z,\tilde{z}$ such that $(x,z),(\tilde{x},\tilde{z})\in Supp(X,Z)$, $v(x,1)=v(\tilde{x},0)$ and $P(z)=P(\tilde{z})=p\in \mathcal{P}^*_0(x)$.
It follows from (\ref{EQ1}) and (\ref{EQ23}) that
\begin{equation}\label{EQ1Repeat}
E(Y_{1}|X=x) = E(DY|X=x,Z=z)+E[(1-D)Y|X=\tilde{x},Z=\tilde{z}].
\end{equation}%
By (A-2), the right-hand side of (\ref{EQ1Repeat}) equals the right-hand side of (\ref{id_EY1}).
A symmetric argument proves (\ref{id_EY0}).
\end{proof}
It is worth mentioning that the conditions for this theorem do not directly restrict the support of potential outcomes.
As we show in the next section, our method applies in important applications where the potential outcome is either discrete (e.g., determined by multinomial choices), or multi-dimensional with both discrete and continuous components (e.g., determined in a Roy model).
\setcounter{equation}{0}
\section{Examples}
\label{Examples}
In this section, we present several examples in which the potential outcomes depend on multi-dimensional indices with an endogenous treatment.
More importantly, we show how the specific structure embedded in each application naturally leads to transparent, primitive conditions that imply Assumption A-5, thus corroborating the wide applicability of this general approach.
In the first and third examples, the monotonicity condition in \citeasnoun{vytlacilyildiz} does not hold; in the second example, the identification requires a generalization of the monotonicity condition into an invertibility condition in higher dimensions
\noindent \textbf{Example 1. (Heteroskaedastic shocks in Uncensored or Censored outcomes)} First, consider a triangular system where a continuous \textit{uncensored} outcome is determined by double indices $v(X,D)\equiv (v_{1}(X,D),v_{2}(X,D))$:\footnote{
\citeasnoun{abrevaya2019estimation} adopted a different approach to identify ATE in the same model with uncensored, continuous outcome and multiplicative, heteroskaedastic shocks above, using a binary instrument $Z$.
They showed the conditional mean of the observed outcome $Y$, when scaled by the conditional covariance of $Y1\{D=d\}$ and $Z$, is a weighted sum of the conditional means of potential outcomes: $v_1(x,0)$ and $v_1(x,1)$.
Thus, using the means of the scaled $Y$ conditional on $Z = 0,1$ respectively, one can construct a linear system that identifies $v_1(x,0)$ and $v_1(x,1)$, provided the proper rank condition holds.
Their method leverages this particular specification of multiplicative endogeneity, but does not generalize to the case of censored outcomes or in the other examples we consider later.
}
\begin{equation*}
Y=g(v(X,D),\varepsilon )=v_{1}(X,D)+v_{2}(X,D)\varepsilon \text{ for }D\in
\{0,1\} \text{.}
\end{equation*}
The selection equation determining the actual treatment is the same as (1.2).
In this case the concept of monotonicity in $v\in \mathbb{R}^{2}$ is not well-defined, so the procedure proposed in \citeasnoun{vytlacilyildiz} is not suitable here.\footnote{%
For this particular design, the approach proposed in \citeasnoun{vuongxu}
should be valid. But it will not be for a slightly modified model, such as $%
Y=v_1(X,D)+(e_2+v_2(X,D)*e_1)$, whereas ours will be.} Nevertheless, we can
apply the method in Section \ref{Identification}\ to identify the average
treatment effect by using the \textit{distribution} of outcomes to find pairs
of $x$ and $\tilde{x}$ such that $v(x,1)=v(\tilde{x},0)$.
To see how Assumption A-5 holds, assume the range of $v_2(\cdot)$ is positive.
Note that%
\begin{eqnarray*}
F_{g|u}(y;v(x,d)) &=&\Pr\left[ v_{1}(x,d)+v_{2}(x,d)\varepsilon \leq y|U=u%
\right] \\
&=&F_{\varepsilon |u}\left( \frac{y-v_{1}(x,d)}{v_{2}(x,d)}\right)
\end{eqnarray*}%
for $d=0,1$. Suppose the distribution of $\varepsilon $ conditional on $u$ is increasing over $\mathbb{R}$.
Then for all $y$ and $x \in \mathcal{X}^1$ and matched values $\tilde{x} \in \lambda_0(x)$,
\begin{equation*}
F_{g|u}(y;v(x,1))=F_{g|u}(y;v(\tilde{x},0))
\text{ if and only if }%
\frac{y-v_{1}(x,1)}{v_{2}(x,1)}=\frac{y-v_{1}(\tilde{x},0)}{v_{2}(\tilde{x}%
,0)}\text{.}
\end{equation*}%
Differentiating with respect to $y$ yields $v_{2}(x,1)=v_{2}(\tilde{x},0)$, %
which implies $v_{1}(x,1)=v_{1}(\tilde{x},0)$.%
Next, consider a similar model with continuous \textit{censored} outcomes, where
\begin{equation*}
Y=g(v(X,D),\varepsilon )=max\{v_{1}(X,D)+v_{2}(X,D)\varepsilon,0\} \text{ for } D\in
\{0,1\}.
\end{equation*}
In this case, our identification argument for the uncensored case above applies almost immediately.
The only adjustment needed is to confine the values of $y$ used for constructing matched pairs $(x,\tilde{x})$ over the \textit{uncensored} segment, i.e. $y > 0$.
In contrast, other methods for identifying ATE in the uncensored model which rely strictly on the linear structure with multiplicative heterogeneity, such as \citeasnoun{abrevaya2019estimation}, no longer applies.
\noindent \textbf{Example 2. (Multinomial potential outcome)} Consider a
triangular system where the outcome is multinomial. The multinomial response
model has a long and rich history in both applied and theoretical
econometrics. Recent examples in the semiparametric literature include %
\citeasnoun{lee95}, \citeasnoun{pakesporter2014}, \citeasnoun{ahnetal2018}, \citeasnoun{shietal2018}, %
and \citeasnoun{KOT2019}. None of those papers allow for dummy endogenous variables or potential outcomes.
In this example, let the observed outcome be
\begin{equation*}
Y=g(v(X,D),\varepsilon )=\arg \max_{j=0,1,...,J}y_{j,D}^{\ast },
\end{equation*}%
where
\begin{equation*}
y_{j,D}^{\ast }=v_{j}(X,D)+\varepsilon _{j} \mbox{ for } j =
1,2,...,J\text{; }y_{0,D}^{\ast }=0\text{.}
\end{equation*}%
In this case, the index $v\equiv (v_{j})_{j\leq J}$ and the errors $%
\varepsilon \equiv (\varepsilon _{j})_{j\leq J}$ are both $J$-dimensional.
The selection equation that determines $D$ is the same as (1.2). In this
case, we can replace $1\{Y\leq y\}$ by $1\{Y=y\}$ in the definition of $%
h_{1},h_{0},h_{1}^{\ast },h_{0}^{\ast }$ and $F_{g|u}(\cdot ;v)$. Then for $%
d=0,1$ and $j\leq J$,\
\begin{eqnarray*}
F_{g|u}(j;v(x,d)) &\equiv &E[1\{g(v(x,d),\varepsilon )=j\}|U=u] \\
&=&\Pr \left \{ v_{j}(x,d)+\varepsilon _{j}\geq v_{j^{\prime
}}(x,d)+\varepsilon _{j^{\prime }}\text{ }\forall j^{\prime }\leq J\mid
U=u\right \} \text{.}
\end{eqnarray*}%
By \citeasnoun{ruud2000} and \citeasnoun{ahnetal2018}, the mapping from $%
v\in \mathbb{R}^{J}$ to $(F_{g|u}(j;v):j\leq J)\in \mathbb{R}^{J}$ is smooth
and invertible provided that $\varepsilon \in \mathbb{R}^{J}$ has
non-negative density everywhere. This implies Assumption A-5.\bigskip
\noindent \textbf{Example 3}. \textbf{(Potential outcome from a Roy model)}
Consider a treatment effect model with an endogenous binary treatment $D$
and with the potential outcome determined by a latent Roy model. The Roy
model has also been studied extensively from both applied and theoretical
perspectives. See, for example, the literature survey in %
\citeasnoun{heckvyt-handbook} and the seminal paper in %
\citeasnoun{heckmanhonoreb}.
Here the observed outcome consists of two pieces:\ \ a continuous measure $%
Y=DY_{1}+(1-D)Y_{0}$ and a discrete indicator $W=DW_{1}+(1-D)W_{0}$ for $%
d=0,1$. These potential outcomes are given by
\begin{equation*}
Y_{d}=\max_{j\in \{a,b\}}y_{j,d}^{\ast }\text{ and }W_{d}=\arg \max_{j\in
\{a,b\}}y_{j,d}^{\ast }
\end{equation*}%
where $a$ and $b$ index potential outcomes realized in different sectors,
with
\begin{equation*}
y_{j,d}^{\ast }=v_{j}(X,d)+\varepsilon _{j} \mbox{ for } j\in \{a,b\}\text{.}
\end{equation*}%
The binary endogenous treatment $D$ is determined as in
equation (1.2). For example, $D\in \{1,0\}$ indicates whether an individual
participates in a professional training program, $W_{d}\in \{a,b\}$
indicates the potential sector in which the individual is employed, $%
y^*_{j,d}$ is the potential wage from sector $j$ under treatment $D=d$, and $Y_{d}\in \mathbb{R}$ is the potential wage if the treatment status is $D=d$.\footnote{
Note this differs from the classical approach that formulates treatment effects using Roy models (e.g. \citeasnoun{heckmanUrzuaVytlacil2006}) in that the potential outcome $Y_d$ itself is determined by a latent Roy model.}
As before, we maintain that $(X,Z)\bot (\varepsilon ,U)$.
The parameter of interest is
\begin{equation*}
\Pr \{Y_{1}\leq y,W_{1}=a|X\}.
\end{equation*}%
By the independence condition that $(X,Z)\bot (\varepsilon ,U)$ and an application of the law of total probability, this conditional probability can be expressed in terms of directly identifiable quantities and the following counterfactual quantity
\begin{eqnarray}
&&\Pr \{Y_{1}\leq y,W_{1}=a\mid X=x,Z=z,D=0\} \notag \\
&=&\Pr \left \{ v_{b}(x,1)+\varepsilon _{b}<v_{a}(x,1)+\varepsilon _{a}\leq
y\mid U\geq P(z)\right \} \text{.} \label{CF}
\end{eqnarray}%
Again, we seek to identify this counterfactual quantity by finding $\tilde{x}$ such that there exists $\tilde{z} $ with $ (\tilde{x},\tilde{z}) \in Supp(X,Z)$, $P(z) = P(\tilde{z}) $, and
\begin{equation}\label{fullEQ}
v_{a}(x,1)=v_{a}(\tilde{x},0)\text{ and }v_{b}(x,1)=v_{b}(\tilde{x},0)\text{.}
\end{equation}%
This would allow us to recover the counterfactual conditional probability in (\ref{CF}) as
\begin{equation*}
\Pr \{Y_{0}\leq y,W_{0}=a\mid X=\tilde{x},Z=\tilde{z},D=0\} \text{.}
\end{equation*}
To find such a pair of $(x,\tilde{x})$, define $h_{d,W}(x,p,p^{\prime
}),h_{d,W}^{\ast }(x,p)$ by replacing $1\{Y\leq y\}$ with $1\{W=a\}$ in the
definition of $h_{d},h_{d}^{\ast }$ in Section \ref{Identification}.
Similarly, define $h_{d,Y}(x,y,p,p^{\prime }),h_{d,Y}^{\ast }(x,y,p)$ by
replacing $1\{Y\leq y\}$ with $1\{Y\leq y,W=a\}$ in the definition of $%
h_{d},h_{d}^{\ast }$ in Section \ref{Identification}. Then
\begin{eqnarray*}
h_{d,W}(x,p_{1},p_{2}) &=&\int_{p_{2}}^{p_{1}}\Pr \{v_{b}(x,d)+\varepsilon
_{b}<v_{a}(x,d)+\varepsilon _{a}|U=u\}du\text{;} \\
h_{d,Y}(x,y,p_{1},p_{2}) &=&\int_{p_{2}}^{p_{1}}\Pr \{v_{b}(x,d)+\varepsilon
_{b}<v_{a}(x,d)+\varepsilon _{a}\leq y|U=u\}du\text{;}
\end{eqnarray*}%
and $h_{d,W}(x,p_{1},p_{2})$ and $h_{d,Y}(x,y,p_{1},p_{2})$ are both
identified over their respective domains by construction.
Assume $(\varepsilon _{a},\varepsilon _{b})$ is continuously distributed with positive density over $\mathbb{R}^{2}$ conditional on all $u$, and the distribution of $ (\varepsilon_0,\varepsilon_1) $ given $U=u$ is continuous in $u $ over $[0,1]$.
Then the statement
\begin{eqnarray*}
\text{\textquotedblleft }h_{1,W}(x,p,p^{\prime })=h_{0,W}(\tilde{x}%
,p,p^{\prime })\text{, }h_{1,Y}(x,y,p,p^{\prime })=h_{0,Y}(%
\tilde{x},y,p,p^{\prime })\text{ }
\text{for all }y\text{ and }p > p^{\prime }\text{ on }\mathcal{P}_{x}\cap \mathcal{P}_{\tilde{x}%
}\text{\textquotedblright }
\end{eqnarray*}%
holds true if and only if (\ref{fullEQ}) holds.
To see this, first note that matching $%
h_{1,W}(x,p,p^{\prime })=h_{0,W}(\tilde{x},p,p^{\prime })$ requires
\begin{equation}
v_{a}(x,1)-v_{b}(x,1)=v_{a}(\tilde{x},0)-v_{b}(\tilde{x},0)\text{,}
\label{CF1}
\end{equation}
while matching $h_{1,Y}(x,y,p,p^{\prime })=h_{0,Y}(\tilde{x},y,p,p^{\prime
}) $ at the same time requires
\begin{equation}
v_{a}(x,1)=v_{a}(\tilde{x},0)\text{.} \label{CF2}
\end{equation}%
Thus requiring (\ref{CF1}) and (\ref{CF2}) to hold jointly is equivalent to (\ref{fullEQ}).
It is worth mentioning that the identification strategy above also applies in a more general setup where the error term in the potential outcome is specific to the treatment, i.e., with $\varepsilon_j$ replaced by $\varepsilon_{j,d}$ in $y^*_{j,d}$.
In such cases, the argument above remains valid under a ``rank similarity'' condition (that the marginal distribution of $\varepsilon_{j,d}$ is the same for $d\in\{0,1\}$).
The rank similarity condition has been used for identifying treatment effects in instrumental quantile regression models such as \citeasnoun{chernozhukovhansenjoe}.
\setcounter{equation}{0}
\section{Extension}
\label{Extension}
The identification strategy we use requires finding matched pairs for $x$ in $\mathcal{X}^1$ and $\mathcal{X}^0$.
In some cases, with the outcome being continuous, we can construct similar arguments for identifying a counterfactual quantity in a treatment effect model by matching different elements on the support of continuous outcomes.
To the best of our knowledge, this approach has not been explored in the literature on the effects of endogenous treatments.
The following example illustrates this point.
\noindent \textbf{Example 4.} \textbf{(Potential outcome with random
coefficients)} Random coefficient models are prominent in both the
theoretical and applied econometrics literature. They permit a flexible way to allow for conditional heteroscedasticity and unobserved heterogeneity.
For a survey and recent developments, see \citeasnoun{hsiaochapter}, \citeasnoun{hoderlein2010analyzing}, \citeasnoun{arellano2012identifying}, and \citeasnoun{masten2018random}.
We consider a treatment effect model where the potential outcome is determined through random coefficients:
\begin{equation*}
Y=DY_{1}+(1-D)Y_{0}\text{ where }Y_{d}=(\alpha _{d}+X^{\prime }\beta _{d})%
\text{ for }d=0,1
\end{equation*}%
and the binary endogenous treatment $D$ is determined as in equation (1.2). The \textit{random} intercepts $\alpha _{d}\in \mathbb{R}$
and the \textit{random} vectors of coefficients $\beta _{d}$ are given by
\begin{equation*}
\alpha _{d}=\bar{\alpha}_{d}(X)+\eta _{d}\text{ and }\beta _{d}=\bar{\beta}%
_{d}(X)+\varepsilon _{d}
\end{equation*}%
where for any $x \in Supp(X)$ and $d \in \{0,1\}$, $(\bar{\alpha}_{d}(x),\bar{\beta}%
_{d}(x))\in \mathbb{R}^{K+1}$ is a vector of constant parameters while $\eta
_{d}\in \mathbb{R}$ and $\varepsilon _{d}\in \mathbb{R}^{K}$ are
unobservable noises.
As before, suppose some elements of $Z$ in the treatment equation are excluded from $X$. We allow the vector of unobservable terms $(\varepsilon_{1},\varepsilon _{0},\eta _{0},\eta _{1},U)$ to be arbitrarily correlated, and assume:%
\begin{equation}
\left( X,Z\right) \text{ }\bot \text{ }(\varepsilon _{1},\varepsilon _{0},\eta
_{0},\eta _{1},U)\text{.} \label{indep}
\end{equation}%
Aslo, normalize the marginal distribution of $U$ to standard uniform, so that $\theta (Z)$ is directly identified as $P(Z)\equiv E(D|Z)$.
Our goal is to identify the distribution of potential outcomes $%
Y_{d}$ given $X=x$ for $d=0,1$.
From this result, we can recover other parameters of interest such as average treatment effects, quantile treatment effects, etc.
As a preliminary step, we start by pinpointing a counterfactual item that is crucial for this identification question.
Let $G_{P|x}$ denote the distribution of $P\equiv P(Z)$ given $X=x$ (recall that its support is denoted as $\mathcal{P}_x$).
This conditional distribution is directly identifiable from the data-generating process.
By construction,%
\begin{equation*}
\Pr \{Y_{1}\leq y|X=x\}=\int \Pr \{Y_{1}\leq y|X=x,P=p\}dG_{P|x}(p)\text{,}
\end{equation*}%
where%
\begin{eqnarray} \label{dist_ltp}
&&\Pr \{Y_{1}\leq y|X=x,P=p\} \notag \\
&=&E\left[ D1\{Y_{1}\leq y\}|X=x,P=p\right] +E\left[ (1-D)1\{Y_{1}\leq
y\}|X=x,P=p\right] \text{.}
\end{eqnarray}%
The first term on the right-hand side of (\ref{dist_ltp}) is identified as%
\begin{equation*}
E[D1\{Y\leq y\}|X=x,P=p]\text{.}
\end{equation*}%
The second term on the right-hand side of (\ref{dist_ltp}), denoted by $\phi_0(x,y,p)$ , is counterfactual and can be written as%
\begin{eqnarray*}
\phi _{0}(x,y,p)&\equiv& E[1\{U\geq P\}1\{\alpha _{1}+X^{\prime }\beta
_{1}\leq y\}|X=x,P=p] \\
&=&E[1\{U\geq p\}1\{\bar{\alpha}_{1}(x)+\eta _{1}+x^{\prime }(\bar{\beta}%
_{1}(x)+\varepsilon _{1})\leq y\}] \\
&=&\int_{p}^{1}\Pr \{\eta _{1}+x^{\prime }\varepsilon _{1}\leq y-\bar{\alpha}%
_{1}(x)-x^{\prime }\bar{\beta}_{1}(x)|U=u\}du\text{.}
\end{eqnarray*}%
Hence identification of the conditional distribution of $Y_1$ amounts to identification of $\phi_0(\cdot) $.
As noted at the beginning of this section, we will identify the counterfactual $\phi_0(\cdot)$ by finding matched elements on the support of the observed outcome $Y$.
This takes two steps.
First, we show that if for a given pair $(x,y)$ one can find $t(x,y)$ such that
\begin{equation}\label{defn_txy}
y-\bar{\alpha}_{1}(x)-x^{\prime }\bar{\beta}_{1}(x)=t(x,y)-\bar{\alpha}%
_{0}(x)-x^{\prime }\bar{\beta}_{0}(x)\text{,}
\end{equation}%
then one can use $t(x,y)$ to identify $\phi_0(x,y,p)$ for any $p\in \mathcal{P}_x$.
Specifically, for any $p$ on the support of $P$ given $X=x$, define%
\begin{eqnarray*}
h_{1}^{\ast }(x,y,p)&\equiv& E\left[ D1\left\{ Y\leq y\right\} |X=x,P=p%
\right] \\
&=&E\left[ 1\left\{ U<P\right\} 1\left\{ \alpha _{1}+X^{\prime }\beta
_{1}\leq y\right\} |X=x,P=p\right] \\
&=&E\left[ 1\left\{ U<p\right\} 1\left\{
\alpha _{1}+x^{\prime }\beta _{1}\leq y\right\} \right] \\
&=&\int_{0}^{p}\Pr \{\eta _{1}+x^{\prime }\varepsilon _{1}\leq y-\bar{\alpha}%
_{1}(x)-x^{\prime }\bar{\beta}_{1}(x)|U=u\}du,
\end{eqnarray*}%
where the second equality uses (\ref{indep}). Likewise, under (\ref{indep})
we have:
\begin{eqnarray*}
h_{0}^{\ast }(x,y,p) &\equiv& E\left[ (1-D)1\left\{ Y\leq y\right\} |X=x,P=p%
\right] \\
&=&\int_{p}^{1}\Pr \{\eta _{0}+x^{\prime }\varepsilon _{0}\leq y-\bar{\alpha}%
_{0}(x)-x^{\prime }\bar{\beta}_{0}(x)|U=u\}du\text{.}
\end{eqnarray*}%
Assume\footnote{%
Such a distributional equality condition has been used to motivate the \emph{rank similarity}
condition imposed frequently in the econometrics literature -- see, for example, %
\citeasnoun{chernozhukovhansen}, \citeasnoun{vytlacilyildiz}, \citeasnoun{chenstackhan}, \citeasnoun{franlef}, \citeasnoun{dongshen}.}
\begin{equation}
F_{\left( \eta _{1},\varepsilon _{1}\right) |U=u}=F_{\left( \eta _{0},\varepsilon
_{0}\right) |U=u}\text{ for all }u\in \lbrack 0,1]\text{.} \label{eqCF}
\end{equation}%
Under (\ref{eqCF}), we have%
\begin{equation}
\phi _{0}(x,y,p)=\int_{p}^{1}\Pr \{ \eta _{0}+x^{\prime }\varepsilon _{0}\leq y-%
\bar{\alpha}_{1}(x)-x^{\prime }\bar{\beta}_{1}(x)|U=u\}du\text{.}
\label{swapEps}
\end{equation}%
Then by definition of $t(x,y)$ in (\ref{defn_txy}),%
\begin{eqnarray*}
h_{0}^{\ast }(x,t(x,y),p) &\equiv &\int_{p}^{1}\Pr \{ \eta _{0}+x^{\prime
}\varepsilon _{0}\leq t(x,y)-\bar{\alpha}_{0}(x)-x^{\prime }\bar{\beta}%
_{0}(x)|U=u\}du \\
&=&\int_{p}^{1}\Pr \{ \eta _{0}+x^{\prime }\varepsilon _{0}\leq y-\bar{\alpha}%
_{1}(x)-x^{\prime }\bar{\beta}_{1}(x)|U=u\}du \\
&=&\phi _{0}(x,y,p).
\end{eqnarray*}%
Thus the counterfactual $\phi _{0}(x,y,p)$ would be identified as $h_{0}^{\ast }(x,t(x,y),p)$.
The second step is to show that for each pair $(x,y)$ we can indeed uniquely recover $t(x,y) $ using quantities that are identifiable in the data-generating process.
To do so, we define two auxiliary functions as follows: for $p_{1}>p_{2}$ on the support of $P$ given $X=x$, let
\begin{eqnarray*}
h_{1}(x,y,p_{1},p_{2}) &\equiv &h_{1}^{\ast }(x,y,p_{1})-h_{1}^{\ast
}(x,y,p_{2}) \\
&=&\int_{p_{2}}^{p_{1}}\Pr \{ \eta _{1}+x^{\prime }\varepsilon _{1}<y-\bar{%
\alpha}_{1}(x)-x^{\prime }\bar{\beta}_{1}(x)|U=u\}du\text{;}
\end{eqnarray*}%
and%
\begin{eqnarray*}
h_{0}(x,y,p_{1},p_{2}) &\equiv &h_{0}^{\ast }(x,y,p_{2})-h_{0}^{\ast
}(x,y,p_{1}) \\
&=&\int_{p_{2}}^{p_{1}}\Pr \{ \eta _{0}+x^{\prime }\varepsilon _{0}<y-\bar{%
\alpha}_{0}(x)-x^{\prime }\bar{\beta}_{0}(x)|U=u\}du\text{.}
\end{eqnarray*}%
Suppose $\eta _{d}+x^{\prime }\varepsilon _{d}$ is continuously distributed
over $\mathbb{R}$ for all values of $x$ conditional on all $u \in [0,1]$.
Then for any fixed pair $(x,y)$ and $p_{1}>p_{2}$,%
\begin{equation*}
h_{1}(x,y,p_{1},p_{2})=h_{0}(x,t(x,y),p_{1},p_{2})
\end{equation*}%
if and only if%
\begin{equation*}
t(x,y)=y-\bar{\alpha}_{1}(x)-x^{\prime }\bar{\beta}_{1}(x)+\bar{\alpha}%
_{0}(x)+x^{\prime }\bar{\beta}_{0}(x)\text{.}
\end{equation*}%
\setcounter{equation}{0}
\section{Estimation}
\label{estimation}
In this section, we outline estimation procedures from a random sample of the observed variables that are motivated by our identification results.
We first describe an estimation procedure for the parameter $E[Y_1]$ in the first three examples.
Recall $\mathcal{P}_{x}$ denotes the support of $P(Z) \equiv P $ given $X=x$.
Let $f_P(.|x)$ denote the density of $P(Z)$ given $X=x$, and define
\begin{equation*}
\mathcal{P}_{x}^{c} \equiv \left \{ p\text{: }f_{P}(p|x)>c\right \} \mbox{ for a known } c>0.
\end{equation*}%
For simplicity, assume
\begin{equation*}
1-c_{0}>P(Z)>c_{0} \mbox{ for a known } c_0 > 0 \mbox{ almost surely}.
\end{equation*}%
Define a measure of distance between $h_{1}(x_{1},\cdot )$ and $%
h_{0}(x_{0},\cdot )$ as follows:
\begin{eqnarray*}
&&\left \Vert h_{1}(x_{1},\cdot )-h_{0}(x_{0},\cdot )\right \Vert \\
&=&\left \{ \int \int \int \left( \int_{p_{2}}^{p_{1}} (F_{g|u}(y;v(x_1,1))
- F_{g|u}(y;v(x_0,0)) du \right)^{2} I \left( p_{1},p_{2}\in
\mathcal{P}_{x}^{c}\right) w(y)dydp_{1}dp_{2}\right \}^{1/2}
\end{eqnarray*}
where $w(y)$ is a chosen weight function.
Consider the case when $h_{0}(x,y,p_{1},p_{2})$, $h_{1}(x,y,p_{1},p_{2})$ and $P(z)$ are known. For any given $x_{i}$, let $\tilde x_{i}$ be such that%
\begin{equation*}
\left \Vert h_{0}(\tilde x_{i},\cdot ) - h_{1}(x_{i},\cdot ) \right \Vert =0,
\end{equation*}
which, under Assumption A-5 in Section \ref{Identification}, is equivalent to
\begin{equation*}
v(\tilde x_{i},0) = v(x_{i},1).
\end{equation*}%
Let $P_i$ be shorthand for $P(Z_i)$ and define
\begin{equation*}
Y_{i}^{*} \equiv E[Y|D=0, \| h_0( X,\cdot) - h_1(X_i,\cdot) \|=0,P=P_i].
\end{equation*}%
This conditional expectation equals $E[ Y | D=0, v( X , 0 ) = v( X_i , 1), P=P_i] $, which in turn equals $E( Y_1 | D=0, X=X_i, P=P_i) $.
The parameter of interest $\Delta \equiv E[Y_{1}]$ can be written as $ \Delta = E[D_iY_i+(1-D_i)Y^*_i]$.
Therefore, we estimate $\Delta$ by its sample analog, after replacing $Y^*_i$ with its Nadaraya-Watson estimates.
That is, we estimate $\Delta$ by
\begin{equation*}
\hat{\Delta}=\frac{1}{n}\sum_{i=1}^{n}\left( D_{i}Y_{i}+(1-D_{i})\hat{Y}%
_{i}\right),
\end{equation*}%
where $\hat{Y}_i$ is a kernel regression estimator of $Y^*_i$, using knowledge of $h_0,h_1$ and $P(Z)$.
Likewise, for estimating the conditional mean $E(Y_1|X \in A)$ where $A$ is a generic subset of the support of $X$, we use a weighted version%
\begin{equation*}
\hat{\Delta}_{A}=\frac{\frac{1}{n}\sum_{i=1}^{n}1\left \{ X_{i}\in A\right
\} \left( D_{i}Y_{i}+(1-D_{i})\hat{Y}_{i}\right) }{\frac{1}{n}%
\sum_{i=1}^{n}1\left \{ X_{i}\in A\right \} }.
\end{equation*}%
Limiting distribution theory for each of these estimators follows from
identical arguments in \citeasnoun{vytlacilyildiz}. Here we formally state
the theorem for the first estimator:
\begin{theorem}
\label{theorem1} Suppose Assumptions (A-1) to (A-6) hold, and $Y_1$ has positive and finite second moments. Then
\begin{equation*}
\sqrt{n}(\hat \Delta-\Delta)\overset{d}{\rightarrow} \mathbb{N}(0,\Sigma),
\end{equation*}
where
\begin{equation*}
\Sigma=Var(E[Y_1|X,P,D])+E[PVar(Y_1|X,P,D=1)].
\end{equation*}
\end{theorem}
Next, we describe an estimation procedure for the distributional treatment effect in Example 4, where potential outcomes depend on random coefficients.
In this case, the parameter of interest is, for a given value $y \in \mathbb{R}$,
\begin{equation*}
\Delta_2(y)=\Pr \{Y_{1}\leq y\} \text{.}
\end{equation*}%
For fixed values of $y$ and $p_{1}>p_{2}$, we propose to estimate $t(x,y)$ as
\begin{equation*}
\hat{t}(x,y,p_{1},p_{2})=\arg
\min_{t}(h_{1}(x,y,p_{1},p_{2})-h_{0}(x,t,p_{1},p_{2}))^{2},
\end{equation*}%
and then average over values of $p_{1},p_{2}$:
\begin{equation*}
\hat{\tau}(x,y)=\frac{1}{n(n-1)}\sum_{i\neq j}I[P_{i}>P_{j}]\hat{t}%
(x,y,P_{i},P_{j}).
\end{equation*}%
An infeasible estimator for $\Delta_2(y)$, which assumes $%
t(x,y)$ is known, would be%
\begin{equation*}
\hat{\Delta}_2(y)=\frac{1}{n}\sum_{i=1}^{n} \left( D_{i}1\{Y_{i}\leq
y\}+(1-D_{i})1\{Y_{i}\leq t(X_{i},y)\} \right) \text{.}
\end{equation*}%
In practice, for feasible estimation, one needs to replace $t(x,y)$ with its estimator $\hat{\tau}(x,y)$.
We conclude this section with some discussion about the computational aspects.
The computational costs for finding the suitable pairs $(x,\tilde{x})$ are modest in comparison with typical semi- or nonparametric methods in the literature.
To see this, consider the case with $J=2$, where $v=(v_{1},v_{2})\in \mathbb{R}^{2}$ consists of two index functions $v_{j}(x,d):\mathbb{R}^{K}\times \{0,1\}\rightarrow R$ for $j=1,2$.
The actual dimensions that matter in implementation are: (i) $K$ in the estimation of $h_{1}$ and $h_{0}$ in the first stage, and (ii) the dimension of the indexes to be matched in the second stage: $J=2$.
Therefore, the dimensionality that causes difficulty in estimation is $\max \left\{J,K\right\} $, not $J\times K$.
In our view, this is no more difficult than implementing estimators that are based on matching or pairwise comparison, such as Blundell and Powell (2003).
\setcounter{equation}{0}
\section{Simulation Study}
\label{simulations}
This section presents simulation evidence for the performance of the
estimator in Section \ref{estimation}, for both the Average Treatment Effect and the Distributional Treatment Effect.
\subsection{Designs with continuous outcomes}
We report results for both our estimator and that in \citeasnoun{vytlacilyildiz}, for several designs where the potential outcomes are real-valued and continuous.
These include designs where the monotonicity condition fails, and designs where the disturbance terms in the outcome equation are multi-dimensional.
Throughout all designs, we model the treatment or dummy endogenous variable
as
\begin{equation*}
D=I[Z-U>0],
\end{equation*}
where $Z,U$ are independent standard normal. We experiment with the
following designs for the outcome.
\begin{description}
\item[(Design 1)] $Y=X+0.5\cdot D +\varepsilon$, where $X$ is standard normal while $(\varepsilon, U)$ are bivariate normal, with mean 0, variance 1, and correlation $\rho_v \in \{0, 0.25, 0.5\}$.
\item[(Design 2)] $ Y=X+0.5\cdot D+(X+D)\cdot \varepsilon $, where $X$ is standard normal while $(\varepsilon, U)$ are bivariate normal, with mean 0, variance 1, and correlation $\rho_v \in \{0, 0.25, 0.5\}$.
\item[(Design 3)] $ Y=(X+0.5 \cdot D+\varepsilon)^2$, where $X$ is standard normal while $(\varepsilon, U)$ are bivariate normal, each with mean 0, variance 1, and correlation $\rho_v \in \{0, 0.25, 0.5\}$
\end{description}
We note that the monotonicity condition holds in Design 1 but fails in the other two designs. For each of these designs, we report results for estimating $E[Y_1]$, i.e., the mean potential outcome under treatment $D=1$.
The two estimators reported in the simulation study are our estimator proposed in Section \ref{estimation} and the one proposed in \citeasnoun{vytlacilyildiz}.
The summary statistics, scaled by the true parameter value, Mean Bias, Median Bias, Root Mean Squared Error, (RMSE), and Median Absolute Deviation (MAD) are evaluated for sample sizes of $n = 100, 200, 400$ for 401 replications.
Results for each of these designs are reported in Tables 1 to 3 respectively. In implementing our estimator, we assume the propensity score function is known, and conduct next stage estimation using a nonparametric kernel estimator with normal kernel function, and a bandwidth of $n^{-1/5}$. This rate reflects ``undersmoothing" as there are two regressors, the propensity score and the regressor $X$. For the estimator in \citeasnoun{vytlacilyildiz}, which involves the derivative of conditional expectation functions as well, we also report results for an infeasible version of their estimator, assuming such functions, as well as the propensity scores, are known.
To implement the second stage of our estimator, in calculating the distance $\|h_1(x_i,\cdot)-h_0(x_0,\cdot)\|$ we used an evenly spaced grid of values for $y$, and selected $n/50$ grid points, with $n $ denoting the sample size.
The results indicate the desirable properties of our estimator, generally agreeing with Theorem \ref{theorem1}. In all designs, our estimator has small values for bias and RMSE, with the value of RMSE decreasing as the sample size grows.
In contrast, the procedure based on \citeasnoun{vytlacilyildiz} only performs well in Design 1, with the sizes of bias and RMSE comparable to those using our method.
As in our estimator, these values decrease as the sample size grows, which is expected, as the monotonicity condition they require is satisfied in this design.
In this case, their approach has smaller standard errors largely due to the relatively simpler structure of the infeasible version of the estimator, but their biases persist even when the sample size increases.
For Designs 2 and 3, where the monotonicity condition is violated, the estimator proposed in \citeasnoun{vytlacilyildiz} does not perform well. Table 2 shows that in Design 2 both the bias and RMSE of their estimator are generally decreasing slowly with the sample size.
Results for their estimator are better in Design 3 in Table 3, but the bias hardly converges with the sample size, and is much larger compared to our estimator.
\begin{center}
\scalefont{0.75}{
Table 1
\begin{tabular}{||l|ccc|ccc||}
\hline
\ & \ & CKT & & & VY & \\ \hline\hline
$\rho _{v}$ & $0$ & $1/4$ & $1/2$ & $0$ & $1/4$ & $1/2$ \\ \hline
n=100 & \ & \ & \ & \ & \ & \\
MEAN BIAS & -0.0170 & 0.0229 & -0.0435 & -0.1302 & -0.1676 & -0.2018 \\
MEDIAN BIAS & -0.0137 & 0.0124 & -0.0653 & -0.1318 & -0.1678 & -0.2087 \\
RMSE & 0.4936 & 0.4800 & 0.4945 & 0.3308 & 0.3337 & 0.3546 \\
MAD & 0.3289 & 0.3328 & 0.3156 & 0.2200 & 0.2271 & 0.2546 \\
n=200 & \ & \ & \ & \ & \ & \\
MEAN BIAS & 0.0032 & -0.0024 & -0.0069 & -0.0864 & -0.1299 & -0.1766 \\
MEDIAN BIAS & -0.0102 & -0.0141 & -0.0314 & -0.0934 & -0.1277 & -0.1679 \\
RMSE & 0.3355 & 0.3367 & 0.3521 & 0.2293 & 0.2457 & 0.2711 \\
MAD & 0.2240 & 0.2228 & 0.2517 & 0.1594 & 0.1676 & 0.1865 \\
n=400 & \ & \ & \ & \ & \ & \\
MEAN BIAS & -0.0187 & 0.0101 & -0.0055 & -0.0584 & -0.11134 & -0.1593 \\
MEDIAN BIAS & -0.0261 & 0.0128 & -0.0065 & -0.0592 & -0.1162 & -0.1572 \\
RMSE & 0.2496 & 0.2489 & 0.2578 & 0.2049 & 0.1867 & 0.2167 \\
MAD & 0.1523 & 0.1732 & 0.1659 & 0.1197 & 0.1345 & 0.1605 \\ \hline\hline
\end{tabular}
\bigskip \bigskip
Table 2
\begin{tabular}{||l|ccc|ccc||}
\hline
\ & \ & CKT & & & VY & \\ \hline\hline
$\rho _{v}$ & $0$ & $1/4$ & $1/2$ & $0$ & $1/4$ & $1/2$ \\ \hline
n=100 & \ & \ & \ & \ & \ & \\
MEAN BIAS & 0.0109 & 0.0397 & -0.0671 & -0.1509 & -0.2875 & -0.4207 \\
MEDIAN BIAS & 0.0151 & 0.0227 & -0.0939 & -0.1590 & -0.2918 & -0.4262 \\
RMSE & 0.5089 & 0.2737 & 0.4853 & 0.3524 & 0.4199 & 0.5289 \\
MAD & 0.3395 & 0.2447 & 0.3105 & 0.2419 & 0.30898 & 0.4310 \\
n=200 & \ & \ & \ & \ & \ & \\
MEAN BIAS & 0.0322 & 0.0143 & -0.0311 & -0.1273 & -0.2559 & -0.3875 \\
MEDIAN BIAS & 0.0159 & 0.0054 & -0.0543 & -0.1310 & -0.2553 & -0.3884 \\
RMSE & 0.3487 & 0.3444 & 0.3455 & 0.2622 & 0.3407 & 0.4475 \\
MAD & 0.2317 & 0.2297 & 0.2552 & 0.1782 & 0.2624 & 0.3884 \\
n=400 & \ & \ & \ & \ & \ & \\
MEAN BIAS & 0.0088 & 0.0269 & -0.0294 & -0.0962 & -0.2247 & -0.3708 \\
MEDIAN BIAS & 0.0007 & 0.0244 & -0.0309 & -0.0982 & -0.2255 & -0.3769 \\
RMSE & 0.2578 & 0.2557 & 0.2549 & 0.1920 & 0.2764 & 0.4037 \\
MAD & 0.1649 & 0.1733 & 0.1606 & 0.1354 & 0.2283 & 0.3769 \\ \hline\hline
\end{tabular}
\pagebreak
Table 3
\begin{tabular}{||l|ccc|ccc||}
\hline
\ & \ & CKT & & & VY & \\ \hline\hline
$\rho _{v}$ & $0$ & $1/4$ & $1/2$ & $0$ & $1/4$ & $1/2$ \\ \hline
n=100 & \ & \ & \ & \ & \ & \\
MEAN BIAS & -0.0097 & -0.0070 & 0.0019 & -0.0691 & -0.0898 & -0.1066 \\
MEDIAN BIAS & -0.0233 & -0.0101 & -0.0240 & -0.0799 & -0.0925 & -0.1178 \\
RMSE & 0.1893 & 0.2085 & 0.2126 & 0.1546 & 0.1630 & 0.1701 \\
MAD & 0.1398 & 0.1342 & 0.1374 & 0.1125 & 0.1178 & 0.1315 \\
n=200 & \ & \ & \ & \ & \ & \\
MEAN BIAS & -0.0108 & -0.0069 & -0.0068 & -0.0609 & -0.0765 & -0.0968 \\
MEDIAN BIAS & -0.0148 & -0.0033 & -0.0099 & -0.0674 & -0.0769 & -0.1017 \\
RMSE & 0.1372 & 0.1434 & 0.1424 & 0.1163 & 0.1262 & 0.1369 \\
MAD & 0.0949 & 0.0989 & 0.0953 & 0.0855 & 0.0887 & 0.1078 \\
n=400 & \ & \ & \ & \ & \ & \\
MEAN BIAS & -0.0073 & -0.0014 & -0.0026 & -0.0583 & -0.0725 & -0.0889 \\
MEDIAN BIAS & -0.0149 & -0.0023 & -0.0029 & -0.0610 & -0.0751 & -0.0887 \\
RMSE & 0.1084 & 0.0994 & 0.0989 & 0.0924 & 0.1007 & 0.1131 \\
MAD & 0.0697 & 0.0685 & 0.0654 & 0.0689 & 0.0788 & 0.0901 \\ \hline\hline
\end{tabular}
}
\end{center}
We also report estimator performance in samples simulated from a model where potential outcomes are determined by random coefficients and dummy endogenous variables.
It is important to note that for this design, the estimator in \citeasnoun{vytlacilyildiz} does not apply.
This is because different values of $x$ lead to different distributions of the composite error $\eta_d + x^{\prime }\epsilon_d$.
Our contribution in Section \ref{Extension} is to propose a new approach based on matching different values of the observed outcome $y$, rather than the exogenous covariates $x$.
Based on the counterfactual framework discussed in Section \ref{Extension}, here the treatment variable $D$ is modeled as the same way as in the first three designs, with the regressor $X$ being standard normal.
For both $Y_0,Y_1$, the intercepts were modeled as constants (0 and 1, respectively) and the additive error terms were each standard normal.
For the random slopes, the means were 1 and 2 respectively, and the additive error terms were also standard normal, independent of all other disturbance terms and each other.
Here we use the procedure in Section \ref{Extension} to estimate the parameter $\Delta_2=P(Y_1<y)$, where in the simulation, we set $y=1$.
Results for this design with random coefficients are reported in Table 4.
The same four summary statistics are reported for sample sizes $n \in \{100,200,400\}$, based on 401 replications.
The estimator proposed in Section \ref{estimation} performs well.
The bias and RMSE are much smaller for a bigger sample with $n=400$ than for smaller samples with $n = 100$ and $n=200$, indicating convergence at the parametric rate.
\pagebreak
\begin{center}
{\scalefont{0.75}
Table 4
\begin{tabular}{||l|ccc||}
\hline
\ & CKT & & \\ \hline\hline
$\rho _{v}$ & $0$ & $1/4$ & $1/2$ \\ \hline
n=100 & \ & \ & \\
MEAN BIAS & 0.0109 & -0.0086 & 0.0038 \\
MEDIAN BIAS & 0.0000 & -0.0064 & 0.0126 \\
RMSE & 0.1011 & 0.0979 & 0.0955 \\
MAD & 0.0600 & 0.0648 & 0.0652 \\
n=200 & \ & \ & \ \\
MEAN BIAS & -0.0050 & -0.0150 & 0.0095 \\
MEDIAN BIAS & -0.0100 & -0.0161 & 0.0029 \\
RMSE & 0.0669 & 0.0669 & 0.0665 \\
MAD & 0.0400 & 0.0454 & 0.0457 \\
n=400 & \ & \ & \ \\
MEAN BIAS & 0.0012 & -0.0132 & 0.0074 \\
MEDIAN BIAS & 0.0049 & -0.0162 & 0.0077 \\
RMSE & 0.0501 & 0.0494 & 0.0495 \\
MAD & 0.0349 & 0.0325 & 0.0360 \\ \hline\hline
\end{tabular}
}
\end{center}
\subsection{Designs with discrete potential outcomes}
We also report the performance of our estimator in a sample drawn from
Example 2, where the observed outcomes are determined in a multinomial
choice model:%
\[
Y_{i}(D_{i})=\arg \max_{j\in \{0,1,2\}}Y_{i,j}^{\ast }(D_{i})\text{,}
\]%
where the potential outcomes are
$Y_{i,j}^{\ast
}(d)\equiv v_{j}(X_{i,j},d)+\varepsilon _{i,j}$ with%
\[
v_{j}(X_{i,j},d)=\alpha _{j}(d)+X_{i,j}\beta _{j}(d)
\]
for $j=1,2$, and $v_0(X_{i,j},d) =0$ by way of normalization.
The exogenous covariates $X_{i,j}$ are drawn independently from standard normal, and the intercepts and slope coefficients are:%
\begin{eqnarray*}
\alpha _{1}(0)=0\text{; }\alpha _{2}(0)=1\text{; }\alpha _{1}(1)=1\text{;
}\alpha _{2}(1)=2\text{; }
\beta _{1}(0)=0.8\text{; }\beta _{2}(0)=1\text{; }\beta _{1}(1)=1%
\text{; }\beta _{2}(1)=2\text{.}
\end{eqnarray*}%
For each individual $i$, the binary treatment $D_{i}$ is determined as before:
\[
D_{i}=1\{U_{i} < Z_{i}\}\text{,}
\]%
where the instrument $Z_{i}$ is independently drawn from a standard uniform distribution.
The marginal distribution of the selection error $U_{i}$ is standard uniform.
Conditional on $U_{i}=u$, the outcome errors $\varepsilon_{i,j}$ are independent across $j=0,1,2$ and are distributed as type-1 extreme value with unit variance and means $(0,\delta u,2\delta u)$ respectively, where $\delta$ is a parameter to be specified in the data-generating process.
We adopt this specification as it allows for substantial dependence between $U_{i}$ and $\varepsilon _{i}\equiv (\varepsilon_{i,j})_{j=0,1,2}$.
Our simulation study uses sample sizes $n\in \{250,500,1000,2000\}$.
For each sample size $n$, we generate $S=400$ independent samples from the data-generating process above.
Throughout this section, we focus on estimating a conditional distribution of \textit{potential} outcomes $\Pr\{Y_d=j|X \in \omega\}$ for $d\in \{0,1\}$ and $ j \in \{1,2\}$, where $\omega \equiv\{x:x_{j}\in \lbrack -1,1]$ for $j=1,2\}$ is a subset of the support of covariates.
As a benchmark, we first implement the \textit{infeasible} version of the estimator in Section \ref{estimation}, where knowledge of $h_{1}^{\ast }(\cdot )$ and $h_{0}^{\ast }(\cdot )$ are used for finding $(x,\tilde{x})$ with $v(x,1)=v(\tilde{x},0)$ for estimating $\Pr\{Y_1=j|X \in \omega\}$ (or $v(x,0)=v(\tilde{x},1)$ for estimating $\Pr\{Y_0=j|X \in \omega\}$).
In this case, $F_{g|u}$ has a known close form, and we use numerical integration via mid-point approximation to calculate $h_{1}^{\ast }$ and $h_{0}^{\ast }$.
Table 5 shows the mean bias and mean squared error (M.S.E.) for the infeasible estimator calculated from the $S$ simulated samples.
We report these measures for $Pr\{Y_d = j | X \in \omega \}$ for $d=0,1$ and $j=1,2$.
In the last two columns of the table, we report the M.S.E. for the \textit{full} vector summarizing the probability mass function of $Y_d$, i.e., $[\Pr\{Y_d=1|X \in \omega \}, \Pr\{Y_d=2|X \in \omega \}]$.
The M.S.E. in Table 5 diminishes at a root-n rate that is proportional to the sample size.
While in most cases the mean bias decreases with the sample size, the root-n rate of convergence appears to be substantially driven by the diminishing variance of the estimator.
The performance does not vary substantively across different designs with the parameter values that affect the strength of correlation between the structural error $\varepsilon _{i,j}$ and the selection error $U_{i}$, i.e., the parameter values $\delta=(1/4,1/3,1/2)$.
We then construct a feasible version of the estimator by using $h_{1}^{\ast }$ and $h_{0}^{\ast }$ with their corresponding kernel estimates.
Specifically, we use bivariate Gaussian kernels with bandwidths $1.06\hat{\sigma}^{-1/7}$, where $\hat{\sigma}$ denotes the sample standard deviation of components in $(X_{i},P_{i})$.
Table 6 reports the mean bias and M.S.E. of this feasible estimator in the same data-generating process as in Table 5.
The estimation errors in Table 6 are overall larger than those reported in Table 5 but demonstrate similar patterns of convergence in most cases, even though the convergence appears to be slower for the MSE of $F_{Y_{0}|X\in \omega}$ and $F_{Y_{1}|X\in \omega}$ when $\delta $ is large.
The difference in the magnitude of estimation errors across Table 5
and Table 6 is attributable to the estimation error in $h_{1}^{\ast }$ and $h_{2}^{\ast }$. All in all, we conclude our estimator has decent
finite-sample performance in these designs.
\pagebreak
{\scalefont{0.75}
\begin{center}
Table 5. Performance of Infeasible Estimator
\begin{tabular}{cc|cc|cc|cc|cc|c|c}
\hline
& & \multicolumn{2}{|c|}{$\Pr\{Y_0 = 1|\omega\}$} & \multicolumn{2}{|c|}{$\Pr\{Y_0 = 2|\omega\}$}
& \multicolumn{2}{|c|}{$\Pr\{Y_1 = 1|\omega\}$} & \multicolumn{2}{|c|}{$\Pr\{Y_1 = 2|\omega\}$} &
$F[Y_0|\omega]$ & $F[Y_1|\omega]$ \\ \hline\hline
$\delta $ & $n$ & Bias & MSE & Bias & MSE & Bias & MSE & Bias & MSE & MSE &
MSE \\ \hline
\multicolumn{1}{r}{1/4} & \multicolumn{1}{r|}{250} & \multicolumn{1}{|r}{
0.01305} & \multicolumn{1}{r|}{0.00225} & \multicolumn{1}{|r}{-0.00120} &
\multicolumn{1}{r|}{0.00328} & \multicolumn{1}{|r}{-0.00753} &
\multicolumn{1}{r|}{0.00254} & \multicolumn{1}{|r}{0.00276} &
\multicolumn{1}{r|}{0.00275} & \multicolumn{1}{|r|}{0.00554} &
\multicolumn{1}{|r}{0.00254} \\
\multicolumn{1}{r}{1/4} & \multicolumn{1}{r|}{500} & \multicolumn{1}{|r}{
0.00970} & \multicolumn{1}{r|}{0.00113} & \multicolumn{1}{|r}{-0.00215} &
\multicolumn{1}{r|}{0.00175} & \multicolumn{1}{|r}{-0.00693} &
\multicolumn{1}{r|}{0.00121} & \multicolumn{1}{|r}{0.00227} &
\multicolumn{1}{r|}{0.00144} & \multicolumn{1}{|r|}{0.00288} &
\multicolumn{1}{|r}{0.00121} \\
\multicolumn{1}{r}{1/4} & \multicolumn{1}{r|}{1000} & \multicolumn{1}{|r}{
0.00910} & \multicolumn{1}{r|}{0.00077} & \multicolumn{1}{|r}{-0.00269} &
\multicolumn{1}{r|}{0.00095} & \multicolumn{1}{|r}{-0.00389} &
\multicolumn{1}{r|}{0.00078} & \multicolumn{1}{|r}{0.00146} &
\multicolumn{1}{r|}{0.00080} & \multicolumn{1}{|r|}{0.00172} &
\multicolumn{1}{|r}{0.00078} \\
\multicolumn{1}{r}{1/4} & \multicolumn{1}{r|}{2000} & \multicolumn{1}{|r}{
0.00871} & \multicolumn{1}{r|}{0.00041} & \multicolumn{1}{|r}{-0.00447} &
\multicolumn{1}{r|}{0.00051} & \multicolumn{1}{|r}{-0.00376} &
\multicolumn{1}{r|}{0.00038} & \multicolumn{1}{|r}{0.00110} &
\multicolumn{1}{r|}{0.00043} & \multicolumn{1}{|r|}{0.00092} &
\multicolumn{1}{|r}{0.00038} \\ \hline
\multicolumn{1}{r}{1/3} & \multicolumn{1}{r|}{250} & \multicolumn{1}{|r}{
0.01825} & \multicolumn{1}{r|}{0.00269} & \multicolumn{1}{|r}{-0.00955} &
\multicolumn{1}{r|}{0.00339} & \multicolumn{1}{|r}{-0.00350} &
\multicolumn{1}{r|}{0.00248} & \multicolumn{1}{|r}{-0.00262} &
\multicolumn{1}{r|}{0.00276} & \multicolumn{1}{|r|}{0.00608} &
\multicolumn{1}{|r}{0.00275} \\
\multicolumn{1}{r}{1/3} & \multicolumn{1}{r|}{500} & \multicolumn{1}{|r}{
0.00696} & \multicolumn{1}{r|}{0.00126} & \multicolumn{1}{|r}{-0.00144} &
\multicolumn{1}{r|}{0.00166} & \multicolumn{1}{|r}{-0.00656} &
\multicolumn{1}{r|}{0.00123} & \multicolumn{1}{|r}{0.00330} &
\multicolumn{1}{r|}{0.00130} & \multicolumn{1}{|r|}{0.00292} &
\multicolumn{1}{|r}{0.00144} \\
\multicolumn{1}{r}{1/3} & \multicolumn{1}{r|}{1000} & \multicolumn{1}{|r}{
0.01170} & \multicolumn{1}{r|}{0.00079} & \multicolumn{1}{|r}{-0.00842} &
\multicolumn{1}{r|}{0.00097} & \multicolumn{1}{|r}{-0.00315} &
\multicolumn{1}{r|}{0.00070} & \multicolumn{1}{|r}{0.00007} &
\multicolumn{1}{r|}{0.00081} & \multicolumn{1}{|r|}{0.00175} &
\multicolumn{1}{|r}{0.00080} \\
\multicolumn{1}{r}{1/3} & \multicolumn{1}{r|}{2000} & \multicolumn{1}{|r}{
0.00918} & \multicolumn{1}{r|}{0.00045} & \multicolumn{1}{|r}{-0.00625} &
\multicolumn{1}{r|}{0.00060} & \multicolumn{1}{|r}{-0.00486} &
\multicolumn{1}{r|}{0.00043} & \multicolumn{1}{|r}{0.00273} &
\multicolumn{1}{r|}{0.00044} & \multicolumn{1}{|r|}{0.00105} &
\multicolumn{1}{|r}{0.00043} \\ \hline
\multicolumn{1}{r}{1/2} & \multicolumn{1}{r|}{250} & \multicolumn{1}{|r}{
0.01322} & \multicolumn{1}{r|}{0.00235} & \multicolumn{1}{|r}{-0.00770} &
\multicolumn{1}{r|}{0.00323} & \multicolumn{1}{|r}{-0.00359} &
\multicolumn{1}{r|}{0.00253} & \multicolumn{1}{|r}{-0.00158} &
\multicolumn{1}{r|}{0.00295} & \multicolumn{1}{|r|}{0.00559} &
\multicolumn{1}{|r}{0.00248} \\
\multicolumn{1}{r}{1/2} & \multicolumn{1}{r|}{500} & \multicolumn{1}{|r}{
0.01426} & \multicolumn{1}{r|}{0.00130} & \multicolumn{1}{|r}{-0.01140} &
\multicolumn{1}{r|}{0.00196} & \multicolumn{1}{|r}{-0.00231} &
\multicolumn{1}{r|}{0.00120} & \multicolumn{1}{|r}{0.00081} &
\multicolumn{1}{r|}{0.00132} & \multicolumn{1}{|r|}{0.00326} &
\multicolumn{1}{|r}{0.00123} \\
\multicolumn{1}{r}{1/2} & \multicolumn{1}{r|}{1000} & \multicolumn{1}{|r}{
0.01142} & \multicolumn{1}{r|}{0.00075} & \multicolumn{1}{|r}{-0.00782} &
\multicolumn{1}{r|}{0.00094} & \multicolumn{1}{|r}{-0.00512} &
\multicolumn{1}{r|}{0.00064} & \multicolumn{1}{|r}{0.00378} &
\multicolumn{1}{r|}{0.00071} & \multicolumn{1}{|r|}{0.00169} &
\multicolumn{1}{|r}{0.00070} \\
\multicolumn{1}{r}{1/2} & \multicolumn{1}{r|}{2000} & \multicolumn{1}{|r}{
0.01007} & \multicolumn{1}{r|}{0.00048} & \multicolumn{1}{|r}{-0.00772} &
\multicolumn{1}{r|}{0.00061} & \multicolumn{1}{|r}{-0.00438} &
\multicolumn{1}{r|}{0.00039} & \multicolumn{1}{|r}{0.00353} &
\multicolumn{1}{r|}{0.00042} & \multicolumn{1}{|r|}{0.00109} &
\multicolumn{1}{|r}{0.00043} \\ \hline
\end{tabular}
\bigskip
Table 6. Performance of Feasible Estimator
\begin{tabular}{cc|cc|cc|cc|cc|c|c}
\hline
& & \multicolumn{2}{|c|}{$\Pr\{Y_0 = 1|\omega\}$} & \multicolumn{2}{|c|}{$\Pr\{Y_0 = 2|\omega\}$}
& \multicolumn{2}{|c|}{$\Pr\{Y_1 = 1|\omega\}$} & \multicolumn{2}{|c|}{$\Pr\{Y_1 = 2|\omega\}$} &
$F[Y_0|\omega]$ & $F[Y_1|\omega]$ \\ \hline\hline
$\delta $ & $n$ & Bias & MSE & Bias & MSE & Bias & MSE & Bias & MSE & MSE &
MSE \\ \hline
\multicolumn{1}{r}{1/4} & \multicolumn{1}{r|}{250} & 0.00522 & 0.00360 &
0.05045 & 0.00785 & -0.04019 & 0.00480 & -0.00071 & 0.00401 & 0.01145 &
0.00533 \\
\multicolumn{1}{r}{1/4} & \multicolumn{1}{r|}{500} & 0.00931 & 0.00237 &
0.04097 & 0.00489 & -0.03585 & 0.00303 & -0.00602 & 0.00237 & 0.00726 &
0.00400 \\
\multicolumn{1}{r}{1/4} & \multicolumn{1}{r|}{1000} & 0.01099 & 0.00131 &
0.03946 & 0.00332 & -0.03488 & 0.00215 & -0.00750 & 0.00130 & 0.00464 &
0.00252 \\
\multicolumn{1}{r}{1/4} & \multicolumn{1}{r|}{2000} & 0.01193 & 0.00093 &
0.04014 & 0.00277 & -0.03345 & 0.00167 & -0.00784 & 0.00071 & 0.00370 &
0.00171 \\ \hline
\multicolumn{1}{r}{1/3} & \multicolumn{1}{r|}{250} & 0.00340 & 0.00395 &
0.04905 & 0.00792 & -0.03382 & 0.00415 & -0.00671 & 0.00415 & 0.01187 &
0.00725 \\
\multicolumn{1}{r}{1/3} & \multicolumn{1}{r|}{500} & 0.00409 & 0.00226 &
0.04718 & 0.00525 & -0.03287 & 0.00278 & -0.00559 & 0.00228 & 0.00751 &
0.00567 \\
\multicolumn{1}{r}{1/3} & \multicolumn{1}{r|}{1000} & 0.00856 & 0.00128 &
0.03914 & 0.00322 & -0.03292 & 0.00191 & -0.00630 & 0.00125 & 0.00450 &
0.00486 \\
\multicolumn{1}{r}{1/3} & \multicolumn{1}{r|}{2000} & 0.01029 & 0.00086 &
0.03921 & 0.00261 & -0.02912 & 0.00137 & -0.00819 & 0.00077 & 0.00346 &
0.00468 \\ \hline
\multicolumn{1}{r}{1/2} & \multicolumn{1}{r|}{250} & 0.00267 & 0.00369 &
0.04291 & 0.00680 & -0.03508 & 0.00399 & 0.00350 & 0.00367 & 0.01050 &
0.00643 \\
\multicolumn{1}{r}{1/2} & \multicolumn{1}{r|}{500} & 0.00563 & 0.00217 &
0.04076 & 0.00450 & -0.03205 & 0.00266 & 0.00067 & 0.00215 & 0.00667 &
0.00417 \\
\multicolumn{1}{r}{1/2} & \multicolumn{1}{r|}{1000} & 0.00426 & 0.00123 &
0.04284 & 0.00357 & -0.02865 & 0.00169 & -0.00184 & 0.00109 & 0.00481 &
0.00238 \\
\multicolumn{1}{r}{1/2} & \multicolumn{1}{r|}{2000} & 0.01084 & 0.00089 &
0.03538 & 0.00230 & -0.02820 & 0.00128 & -0.00173 & 0.00058 & 0.00319 &
0.00207 \\ \hline
\end{tabular}
\end{center}
}
\section{Concluding Remarks}
In this paper, we consider identification and estimation of
weakly separable models with endogenous binary treatment. Existing approaches are based on a monotonicity condition, which is violated in models with multiple unobserved idiosyncratic shocks.
Such models arise in many important empirical settings, including cases where potential outcomes are determined by Roy models, multinomial choice models, or random coefficients with dummy endogenous variables.
We establish new identification results for these models which are constructive and conducive to estimation procedures.
A simulation study indicates adequate finite sample performance of our method.
This paper leaves several open questions for future research.
For example, it may not be feasible to locate pairs of $(x,\tilde{x})$ that satisfy the matching criterion perfectly due to limited, say, discrete, support of $X$ or $Z$.
In this case, it remains an open question how or whether the partial identification approach proposed in \citeasnoun{shaikhvytlacil11} can be applied in the current setting where multiple indices in potential outcomes defy any notion of monotonicity.
Besides, our method requires the selection of the number and location of cutoff points, so a data-driven method for selecting these would be useful. Furthermore, the relative efficiency of our proposed estimation approach needs to be explored, perhaps by deriving efficiency bounds for these new classes of models.
\pagebreak
{\normalsize
\bibliographystyle{econometrica}
|
\section{Introduction}
The first nuclear excited state in $^{229}$Th ($Z=90$), $\approx8$\,eV above the ground state, has the lowest excited-state energy among all intrinsic nuclear excitations observed so far \cite{Seiferle2019energy,Sikorsky2020Energy,von2020ThEn}.
It also has a long radiative lifetime, on the order of $10^3 - 10^4$\,s \cite{Minkov_Palffy_PRL_2017}.
This unique nuclear isomeric state (denoted as $^{229m}$Th below) opens opportunities for building a high-precision nuclear clock, studying fundamental physics, as well as for other interdisciplinary research such as investigating the interplay between the nuclear and electronic degrees of freedom of the atom. It has stimulated many related theoretical and experimental studies; see, for example, Refs.\,\cite{Beeks2021Th,Peik2021} and references therein.
However, the isomeric-state energy has so far only been determined with a relative uncertainty of $2\%$ \cite{Seiferle2019energy,Sikorsky2020Energy},
and the radiative decay was only very recently observed at ISOLDE/CERN \cite{ISOLDE-IS658_2022}.
Along the road of pursuing better understanding of $^{229m}$Th, there were attempts of direct resonant excitation of the low-lying nuclear transition \cite{Jeet2015DirectExc,Yamaguchi2015DirectExc,Stellmer2018OpticalExc,Wense2017DirectExc,von2020frequency-comb}. For instance, excitation of $^{229}$Th ions doped in a crystal host using vacuum ultraviolet (VUV) photons was experimentally attempted \cite{Jeet2015DirectExc}, but population of the isomeric state was not observed. It was also proposed to utilize the internal conversion (IC) in neutral $^{229}$Th atoms for isomer detection \cite{Wense2017DirectExc}, which could shorten the time required to search for the excitation and improve the signal-to-noise ratio (SNR) compared to measuring radiative decay of the isomeric state.
There have also been proposals for investigating excitation or decay of $^{229m}$Th using the electron-bridge process in $^{229}$Th ions in low-charge states (see, for example, Refs.\,\cite{Porsev2010EB,Bilous2018EB}) or in highly charged $^{229}$Th ions ($^{229}$Th$^{35+}$ in Ref.\,\cite{Bilous2020EB_35+}).
$^{229m}$Th could also be produced through excitation of the second nuclear exited state at 29\,keV.
X-ray pumping of the isomeric state via the 29\,keV level was demonstrated in Ref.\,\cite{Masuda2019ThXray}.
So far, most studies involving $^{229m}$Th have dealt with nonrelativistic atoms or ions in low charge states, posing high demand on the light source required for excitation.
For instance, in the direct resonant excitation case a tunable VUV light source with a rather high spectral density must be applied, which could involve multiple steps of frequency doubling of light from a sufficiently tunable and powerful master laser, or high-harmonic generation of frequency-comb light \cite{Wense2017DirectExc,von2020frequency-comb,von2020LasExcTheory}.
An alternative approach to the study of the low-lying isomeric state could be to use relativistic highly charged ions (HCI) stored in a high-energy storage ring \cite{Krasny:2015ffb,Budker2021GF_nucl}.
With the Lorentz boost of photon energies in the ion frame, high-power lasers with wavelengths in the visible range or longer combined with a Fabry–P\'erot cavity can be used for excitation, offering high isomer-production rates.
We propose several approaches to excitation of the isomeric state in relativistic highly charged thorium ions in Sec.\,\ref{Sec:iso_exc}, sorted by the required photon energy in the ion frame.
We discuss direct excitation of the isomeric state using bare nuclei or partially stripped ions (PSI) in Sec.\,\ref{Subsec:reso_exc}, excitation via an intermediate electronic excited state in Li-like thorium ions in Sec.\,\ref{Subsec:exc_271eV},
as well as excitation via the second nuclear excited state in Sec.\,\ref{Subsec:exc_29keV}.
Notably, in H- or Li-like thorium ions,
the tightly bound unpaired electron produces a strong magnetic field at the nucleus leading to hyperfine mixing of the nuclear ground and isomeric states. This is known as the nuclear hyperfine mixing (NHM) effect; see Refs.\,\cite{Karpeshin1998NHM,Tkalya2016_NHM,Shabaev2021NHM}.
It is predicted that the NHM effect reduces the isomeric-state lifetime
in H- or Li-like thorium ions
by orders of magnitude compared to bare nuclei \cite{Shabaev2021NHM}, corresponding to significantly higher excitation rates in the direct resonant excitation scenario. This
could allow for implementing a $\pi-$pulse to transfer population from the nuclear ground state to the isomeric state (see Sec.\,\ref{subsubsec:res_H-like}).
The NHM effect also offers novel opportunities of nuclear isomer excitation using electronic transitions.
We discuss isomer excitation in Li-like $^{229}$Th ions using Stimulated Raman Adiabatic Passage (STIRAP) via an intermediate electronic excited state as well as using a single laser simultaneously exciting the electronic and nuclear states (see Sec.\,\ref{Subsec:exc_271eV}).
The candidate storage rings for implementing these excitation schemes
can be, for example, those at GSI/FAIR
\cite{Litvinov-2013,Shevelko2018IonLifetimeHESR,Sanchez2020HESR,Spiller2020SIS100},
HIAF~\cite{Yang-2013},
or the proposed Gamma Factory (GF) at CERN \cite{Krasny:2015ffb,Krasny2019PoP,jaeckel2020quest,Budker2020_AdP_GF,Budker2021GF_nucl}.
The energy of relativistic highly charged ions in storage rings can be flexibly adjusted, complementary to tuning the laser photon energy, in order to meet resonance conditions for excitation or detection of isomers (see Sec.\,\ref{Sec:search_res}).
Possible ways to achieve stored intense thorium beams in a high-energy storage ring and ion-bunch parameters for various operational and planned facilities are discussed in Sec.\,\ref{Sec:ion_source}.
Methods for probing the produced isomers
and determining the energy of the isomeric state
are presented
in Sec.\,\ref{Sec:Detect}, based on either nuclear transitions (Sec.\,\ref{Subsec:Rad_decay} and \ref{Subsec:Fur_exc}) or electronic transitions (Sec.\,\ref{Subsec:elec_tran}).
According to our estimate, the energy of the isomeric state can be measured with a precision of better than $10^{-4}$ (Sec.\,\ref{Subsec:Rad_decay}) or even down to
below $10^{-6}$ (Sec.\,\ref{Subsec:Fur_exc} and \ref{Subsec:elec_tran}), which would be orders of magnitude improvement of the present value. This could pave the way for further improvement in determining the isomeric-state energy, complementary to the other studies mentioned above, thus bringing a precise nuclear clock and fundamental-physics experiments with this system closer to reality.
Some schemes proposed here could also be realized using x-ray pulses combined with moderately accelerated $^{229}$Th ions or ions at rest (see, for instance, Ref.\,\cite{Kirschbaum2022}).
Besides, the approaches investigated for $^{229}$Th could also be adapted to studying low-energy nuclear states in other nuclei such as $^{229}$Pa, which is briefly discussed in Sec.\,\ref{Sec:other_low_state}.
\section{Preparation of stored and cooled intense beams of Thorium-229}
\label{Sec:ion_source}
There are several technological prerequisites for the studies proposed to become possible.
One of the major challenges is to obtain HCI stored in accelerators. An efficient method for producing beams of heavy HCI is to start with low-charge ions from an ion source, and then accelerate them to high energies before focusing onto a stripper target to remove more of the bound electrons; see, for example, Refs.\,\cite{Atanasov-2015, Steck2020HeavyIonRing,Kroger2021PSI_source}.
The relatively long half-life of about 7880\,y makes it possible in principle to prepare a $^{229}$Th beam directly from an ion source, like it is done for instance for beams of $^7$Be\,\cite{Limata-2008}.
In such case, $^{229}$Th nuclei need to be produced via irradiation of uranium or thorium targets and chemically purified from inevitable contaminants.
Still, the electrodes will be highly active and the overall process is technologically challenging.
Alternatively, a projectile-fragmentation nuclear reaction can be employed for in-flight production of $^{229}$Th beams.
Here, either $^{238}$U or $^{232}$Th primary beams can be accelerated and fragmented in a target, thereby creating $^{229}$Th ions of interest. The kinetic energy of the beam together with the target material and its thickness can be chosen such that the Th ions emerge from
the target in the desired atomic charge state\,\cite{Litvinov-Bosch-2011}.
Assuming, for example, a $^{238}$U ($^{232}$Th) beam at 600\,A\,MeV with a realistic intensity of 10$^{10}$ particles/s impinging on a 1\,g/cm$^2$ $^9$Be production target, we obtain after the target $3\cdot10^5$ ($3\cdot10^6$) $^{229}$Th ions per second \cite{LISE}. These rates can easily be scaled if higher primary beam intensities are available.
Accumulation schemes would be needed to achieve the required $^{229}$Th intensities.
Another issue is related to the transport of laser beams into the ultra-high vacuum environment of the ring.
Various solutions have been realized at several facilities, see, for example, Ref.\,\cite{Nortershauser-2015}. A high-photon-flux extreme ultraviolet (XUV) source based on high harmonic generation was recently commissioned at the CRYRING\,\cite{Hilbert-2019, Hilbert-2020}. For this purpose viewports are placed at dedicated locations to enable the spacial overlap of the stored beam with the injected laser photons.
Since the photon beam cannot be bent, a special design of bending dipole magnets is required, which might be a challenging task for already existing machines and/or if superconducting magnets are considered\,\cite{Winters2015LasCoolingSIS100}.
Presently, stored and cooled beams of highly-charged stable or radioactive ions are routinely available at GSI \cite{Franzke-2008}. The experimental storage ring ESR \cite{Franzke-1987} offers beams at energies of about 4-400\,A\,MeV for a variety of precision studies\,\cite{Litvinov-2013,Bosch-2013}. Relevant here are the possibilities to collide the stored beams with photon\,\cite{Ullmann-2017}, electron\,\cite{Brandau-2009}, or atomic\,\cite{Glorius-2019} targets.
The production and in-ring purification of stored beams of secondary $^{234}$Pa and $^{237}$U ions was demonstrated in Refs.\,\cite{Brandau-2009,Brandau-2010}. Following these experiments, a pure beam of few-electron $^{229}$Th ions was prepared in the ESR in 2022\,\cite{Brandau-PC}. Also an efficient accumulation scheme is established at the ESR\,\cite{Steck2020HeavyIonRing}. At lower energies, the CRYRING\,\cite{Lestinsky-2016} facility can be employed. We also note a proposal to study $^{229m}$Th in the experimental cooler-storage ring CSRe at IMP in China\,\cite{Ma-2015}.
Importantly, the ESR, CRYRING and CSRe are specially designed for laser-spectroscopy studies.
As of today, no facility exists to provide stored and cooled beams of $^{229}$Th at much higher energies. Beams with energies of up to 5\,A\,GeV will be available at the FAIR facility in Darmstadt. The design parameters are to provide accelerated $^{238}$U$^{28+}$ beams with an intensity of 10$^{13}$ particles/s \cite{Durante-2019}. Technically, acceleration of $^{232}$Th beams is also possible. The secondary $^{229}$Th ions produced from projectile fragmentation will be separated from contaminants via the high-acceptance in-flight fragment separator Super-FRS\,\cite{Geissel-2003s} and injected into the collector ring CR\,\cite{cr}.
The CR is designed for fast stochastic cooling of secondary beams.
Afterwards, the beam will be transferred to and accumulated in the high-energy storage ring HESR\,\cite{hesr} for precision experiments.
An additional storage ring, RESR,
between the CR and HESR, is foreseen at a later stage of FAIR to achieve a more efficient beam accumulation.
Two $0^\circ$ laser-ion intersection regions are being prepared in the HESR\,\cite{Sanchez2020HESR, Stohlker-2014,Stohlker-2015,Stohlker-2015b}. A similar scheme will be possible to realize at the HIAF facility under construction in China\,\cite{Yang-2013}.
At CERN, the in-flight production of $^{229}$Th beam can be envisioned between the Proton Synchrotron (PS) and the Super Proton Synchrotron (SPS). Also the isotope separation on-line (ISOL) production method can be employed as utilized at the ISOLDE facility\,\cite{Catherall2017_ISOLDE, Borge_2017}. Here, a post-accelerated to 10\,A\,MeV $^{229}$Th beam will need to be transported from the ISOLDE hall to the CERN accelerator chain. The advantage of this approach would be the much larger $^{229}$Th production yields owing to several hundred times thicker production targets. The proposal for the Gamma Factory\,\cite{Budker2021GF_nucl} also contains a dedicated storage ring for storing beams of radioactive nuclei from ISOLDE. In such configuration, the beams of $^{229}$Th ions will be stored at about 10\,A\,MeV energy\,\cite{Grieser-2012} and irradiated with high-energy photons from the GF.
Parameters of some synchrotron facilities and ion storage rings are listed in Table\,\ref{tab:ion-beam_facilities} sorted by the relativistic Lorentz factor $\gamma=1/(1-\beta^2)^{1/2}$ ($\beta=v/c$ with $v$ and $c$ being the speed of ions and light, respectively). Technical solutions for production and cooling of highly-charged $^{229}$Th beams and/or coupling of laser beams are in place for the ESR and are planned for HESR and BRING. This would also need to be done for the other machines.
It is not our aim in the present work to investigate possible upgrades of these facilities, but to show that the challenges can be met at the present level of technology.
The parameters in Table\,\ref{tab:ion-beam_facilities} assume a stored beam of $^{229}$Th$^{90+}$ ions, though such beams cannot presently be provided at some of the listed facilities.
\begin{table*}[htpb]
\centering
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}} llccc clc}
\hline
\hline
Facility &Lab. &$\gamma_\textrm{max}$ &$\Delta\gamma/\gamma$ &$N_\textrm{ion}$/bunch &1-$\sigma$ radius &Circumference, $f_b$ &$N_b$\\% &lifetime\\
\hline \\[-0.2cm]
LHC\,\cite{Budker2021GF_nucl} &CERN &2950 &$\approx10^{-4}$--($10^{-6}$) &$10^8$ &16\,$\mu$m &26.7\,km, 11.2\,kHz &592--1232\\% &40\,h\\
SPS\,\cite{Krasny2019PoP}\,\footnote{The SPS parameters are from Ref.\,\cite{Krasny2019PoP} for Li-like Pb ions, whose beam parameters should be similar to those for Li-like Th ions.} &CERN &109 &$\approx10^{-4}$--($10^{-6}$) &$10^8$ &0.5\,mm &6911\,m, 43.4\,kHz &$\leq$36\\% &100\,s\\
RHIC\,\cite{Workman2022_Particle_Data,RHIC-2003} &BNL &100 &$7.5\cdot10^{-4}$ &$4.5\cdot10^{10}$ &100\,$\mu$m &3834\,m, 78.2\,kHz &106\\% &100\,s\\
SIS100\,\cite{Winters2015LasCoolingSIS100,Spiller2020SIS100} &GSI &100 &10$^{-3}$--(10$^{-7}$) &10$^{13}$ &
&1084\,m, 280\,kHz &\\
Main Ring\,\cite{JPARC} & J-PARC & 13/22 & & & &1567.5\,m, 191\,kHz &\\
PS
&CERN & 11.4 & & & &628\,m, 478\,kHz &\\
HESR\,\cite{Kovalenko2015HESR,Shevelko2018IonLifetimeHESR,Sanchez2020HESR} &GSI &6
$\approx10^{-5}$ &$10^8$ &2.1\,mm &575\,m, 522\,kHz &1, coasting \\% &$1000\,$s\\
BRING1\,\cite{Yang-2013} & HIAF & 4.4 & $10^{-3}$ & $2\cdot10^{11}$ & -- & 600\,m, 490\,kHz & -- \\
SRING\,\cite{Yang-2013} & HIAF & 2.1 $(\beta=0.879)$ & $\approx10^{-5}$ & 10$^8$ & 1 mm & 273\,m, 966\,kHz & 1-2, coasting \\
ESR\,\cite{Franzke-1987} &GSI &1.5 $(\beta=0.745)$ &$\approx10^{-5}$ &$10^8$ &1\,mm &108\,m, 2.07\,MHz &1-2, coasting\\
\hline
\hline
\end{tabular*}
\caption{Relativistic HCI beams expected at different accelerators.
Here, $\gamma_\textrm{max}$ is the maximal relativistic Lorentz factor of
highly charged $^{229}$Th$^{q+}$ ions ($q\approx90$ in this work).
The Lorentz boost of laser photon energies is $(1+\beta)\cdot\gamma$ in the ion frame.
$\Delta\gamma/\gamma$ is roughly the FWHM relative range of $\gamma$, due to the ion energy spread.
The lower bound of $\Delta\gamma/\gamma$ may be reached by laser cooling, which might require reducing the intensity of the ion bunch to alleviate intrabeam scattering \cite{Eidam2018LaserCooling}.
The number of particles $N_\textrm{ion}$ per bunch relates to expected maximum particle intensities that can be stored in the machine irrespective to the actual production method of $^{229}$Th beams.
The ``coasting" refers to unbunched ion beams, i.e., the ion beam filling the whole storage ring almost uniformly.
}
\label{tab:ion-beam_facilities}
\end{table*}
\section{Excitation of the thorium isomer}
\label{Sec:iso_exc}
\begin{figure}[!htpb]\centering
\includegraphics[width=\linewidth]{229Th.png}
\caption{Low-lying nuclear levels of $^{229}$Th. These levels belong to two rotational bands (indicated by the horizontal offset) labelled by Nilsson quantum numbers, spin, parity, energy and (radiative) half-life $T_{1/2}^{(\textrm{rad})}$. $T_{1/2}^{\textrm{rad}}$ of the isomeric state is derived from Ref.\,\cite{Minkov_Palffy_PRL_2017}.
$T_{1/2}^{\textrm{rad}}$ of the second excited state is obtained using experimental results in Refs.\,\cite{Masuda2019ThXray,Sikorsky2020Energy} (see the main text). 9.3\% and 90.7\% are the radiative branching ratios.
The nuclear magnetic dipole moment $\mu$ and electric quadrupole moment $Q$ are presented for the ground and isomeric states \cite{Campbell2011Th,Safronova2013MD_EQ,Thielking2018laser}.
}
\label{fig:229Th}
\end{figure}
We begin with introducing the low-lying nuclear energy levels of $^{229}$Th related to this work (see Fig.\,\ref{fig:229Th}).
The isomeric state has an energy of $E_{\textrm{iso}}=8.19(12)$\,eV \cite{Peik2021}, an average of two experimental values from Refs.\,\cite{Seiferle2019energy,Sikorsky2020Energy}.
The radiative half-life of the isomeric state in bare nuclei is $T_{1/2}^{\textrm{rad}}\approx5.2\times10^3\,$s, derived theoretically in Ref.\,\cite{Minkov_Palffy_PRL_2017} using a reduced transition probability $B(M1)\approx0.0076\,$W.u. (Weisskopf units).
The second nuclear excited state, with an energy of $E_{\textrm{2nd}}=29,189.93(7)$ \cite{Masuda2019ThXray} and a radiative half-life of about $25\,$ns, can likewise only decay via radiative transitions, while the internal conversion channel is energetically forbidden in highly charged thorium ions discussed in this work.
The cross-band transition from the second excited state to the ground state has a radiative width of $\Gamma_{\gamma}^{\textrm{cr}}=1.70(40)\,$neV, determined experimentally in Ref.\,\cite{Masuda2019ThXray}. The branching ratio of this transition is BR$^{\textrm{cr}}_\gamma=9.3(6)\%$ \cite{Sikorsky2020Energy}. We can further derive the branching ratio, BR$^{\textrm{in}}_\gamma=90.7(6)\%$, and radiative width, $\Gamma_{\gamma}^{\textrm{in}}=16.6(41)\,$neV, for the in-band transition, from the second excited state to the isomeric state (see also Ref.\,\cite{Kirschbaum2022}).
The nuclear magnetic dipole moment $\mu$ and electric quadrupole moment $Q$ of the ground and isomeric states shown in Fig.\,\ref{fig:229Th} are from Refs.\,\cite{Campbell2011Th,Safronova2013MD_EQ,Thielking2018laser}.
\begin{figure*}[!htpb]\centering
\includegraphics[width=0.95\textwidth]{229Th_exc.pdf}
\caption{
Laser excitation of $^{229m}$Th isomers in the storage ring. The laser pulse nearly counter propagates with the relativistic highly charged thorium ions. The intermediate state of the excited ion depends on the specific excitation scheme. The intermediate state shown here corresponds to exciting the 279\,eV transition in Li-like $^{229}$Th ions, see Sec.\,\ref{subsubsec:exc_279eV}. The purple arrow denotes the photon emitted from the $^2P_{1/2}$ to $^2S_{1/2}$ electronic decay, after which the ion remains in the nuclear isomeric state. Produced $^{229m}$Th will circulate in the storage ring. The orange arrow denotes the radiative decay of the isomeric state.
}
\label{fig:229Th_exc}
\end{figure*}
Excitation of $^{229m}$Th could be pursued via using relativistic $^{229}$Th ions in high-energy storage rings and lasers with wavelengths within the visible range or longer;
see Fig.\,\ref{fig:229Th_exc}.
In the following, we assume dealing with ultrarelativistic ($\beta\approx1$)
and highly charged ions (HCI), from Li-like $^{229}$Th ions to bare $^{229}$Th nuclei.
We note that, heavy-ion storage rings have successfully been employed for studying deexcitation of nuclear excited states in highly charged ions \cite{Litvinov-2003,Reed-2010,Sun-2010,Zeng-2017}.
Activities of ground-state $^{229}$Th nuclei would be negligible due to the relativistic time dilation together with the rather small number of $^{229}$Th ions stored in the ring.
\subsection{Resonant excitation of the isomer}
\label{Subsec:reso_exc}
\subsubsection{Excitation in bare Th nuclei}
\label{subsubsec:res_bare}
Assuming nearly head-on collisions between laser photons and the bunches of relativistic ions, due to the relativistic Doppler effect, the photon energy in the ion frame becomes \cite{Budker2021GF_nucl}
\begin{equation}
\label{Eq:energy_relation}
\hbar\omega'=(1+\beta\cos{\theta_l})\gamma \hbar\omega\approx (1+\beta)\gamma\hbar\omega\approx 2\gamma \hbar\omega\,,
\end{equation}
where $\theta_l\approx0$ is the incident angle of laser photons and $\hbar\omega$ is the laser-photon energy ($\hbar$ is the reduced Planck constant).
The last relation in Eq.\,\eqref{Eq:energy_relation} holds for $\beta\approx1$, i.e., $\gamma^2\gg1$.
We use unprimed quantities ($\theta_l$ and $\hbar\omega$ here) for the lab frame, and primed quantities for the ion frame, unless otherwise specified.
Resonant excitation requires $\hbar\omega'=E_\textrm{iso}$.
The width of the photon absorption spectrum in our case is dominated by the Doppler width, $\Gamma_D\approx\hbar\omega'\Delta\gamma/\gamma$, due to the ion-energy spread characterized by $\Delta\gamma/\gamma$. Since $\Delta\gamma/\gamma$ is typically smaller than the relative uncertainty of the isomeric-state energy, we need to scan across the possible energy range of the isomeric state to search for the resonance. This can be done by tuning $\gamma^{\textrm{ave}}$ (the central value of $\gamma$) of the ion bunch and/or the laser-photon energy, which is discussed in Sec.\,\ref{Sec:search_res}. We assume below $2\gamma^{\textrm{ave}}\cdot\hbar\omega=E_\textrm{iso}$.
We calculate the population transfer using the density matrix approach in the ion frame, solving the master equation \cite{Buervenich2006}
\begin{equation} \label{Eq:DM_TLS}
\frac{d\rho}{dt} =
\frac{1}{i\hbar}[H_{\textrm{eff}},\rho]
- \frac{\gamma_{\textrm{SE}}}{2}
\begin{pmatrix}
-2\rho_{ee} & \rho_{ge}\\
\rho_{eg} & 2\rho_{ee}
\end{pmatrix},
\end{equation}
where the interaction-picture Hamiltonian $H_{\textrm{eff}}$ in the rotating wave approximation is given by
\begin{equation}
H_{\textrm{eff}} = -\hbar\begin{pmatrix}
0 & \frac{\Omega_R(t)}{2}\\
\frac{\Omega_R(t)}{2} & \Delta
\end{pmatrix},
\end{equation}
with $\Omega_R(t)$ being the time-dependent Rabi frequency, and $\Delta=2\gamma\omega-E_{\textrm{iso}}/\hbar=(\gamma/\gamma^{\textrm{ave}}-1)E_{\textrm{iso}}/\hbar$ being the detuning for taking account of the ion energy spread.
$\gamma_{\textrm{SE}}=\Gamma_{\textrm{rad}}/\hbar$ is the spontaneous decay rate of the excited state [$\Gamma_\textrm{rad}= \hbar/(T^\textrm{rad}_{1/2}/\ln{2})\approx8.8\times10^{-20}\,$eV]. We assumed employing fully coherent laser pulses, which should be practical for lasers used here with wavelengths within the visible range or longer.
In order to resonantly excite a large fraction of the ions,
we can tune the laser spectral linewidth in the ion frame to
be close to the Doppler width due to ion-energy spread; see Ref.\,\cite{Bieron2022OpticalPump}. This can be achieved by adjusting the duration of the laser pulse. The pulse is assumed to have a Gaussian temporal profile in intensity
$I'(t)=I_{0}'\exp{[-(4\ln{2}) t^2/t_\textrm{p}'^2]}$, where
$I_{0}'$ is the peak intensity
and $t_\textrm{p}'$ is the full width at half maximum (FWHM in intensity) pulse duration in the ion frame. The corresponding spectral width of the laser light
in the ion frame is $\Gamma_\textrm{p}'\approx (4\ln{2})\hbar/t_\textrm{p}'$.
$\Gamma_\textrm{p}'\approx\Gamma_D$ leads to
\begin{equation} \label{Eq:pulse_dur}
t_\textrm{p}'\approx4\ln{(2)}\frac{\gamma}{\Delta\gamma}\frac{1}{\omega'}\,.
\end{equation}
Alternatively, we can introduce incident-angle divergence of the laser beam to match the Doppler width as frequently done in experiments with atomic beams; see, for example, Ref.\,\cite{Nguyen2000DivLas} and references therein.
However, this might make the approximation of nearly head-on collisions between the laser beam and ion bunch invalid.
Therefore, in the following we consider head-on collisions and tuning the pulse duration of Gaussian shaped laser pulses according to Eq.\,\eqref{Eq:pulse_dur} to match the laser-photon spectrum with the photon absorption spectrum considering the ion energy spread.
Equation\,\eqref{Eq:DM_TLS} can be simplified into
\begin{equation} \label{Eq:DM_TLS_v2}
\frac{d\rho}{dt} =
\frac{1}{i\hbar}[H_{\textrm{eff}},\rho]\,
\end{equation}
since the excited state here has a low radiative decay rate, i.e., $\gamma_{\textrm{SE}}\cdot t_\textrm{p}'\ll1$.
The time-dependent Rabi frequency is
\begin{equation}
\label{Eq:Rabi_red}
\Omega_{\textrm{R}}(t)=\frac{\mu B'(t)}{\hbar}\approx\left(\frac{6\pi c^2 \Gamma_\textrm{rad} I'}{\hbar^2\omega'^3}\right)^{1/2}\,,
\end{equation}
where $\mu$ is the magnetic dipole ($M1$) transition moment, related to the radiative width of the isomeric state, and $B'=(2\mu_0 I'/c)^{1/2}$ is the magnetic-field amplitude of the laser light with $\mu_0$ being the vacuum permeability.
The population-transfer fraction induced by one laser pulse is dependent on the detuning, denoted as $\rho_{ee}(\Delta)$. The average population-transfer fraction can be estimated as
\begin{equation} \label{Eq:ave_popu_tran}
\rho_0 = \int \frac{1}{\sqrt{2\pi}\sigma}\exp{\left(-\frac{\Delta^2}{2\sigma^2}\right)} \rho_{ee}(\Delta) d\Delta
\end{equation}
by averaging over the whole ion ensemble.
Here, we assume that the detuning due to the ion energy spread has a Gaussian distribution with the FWHM being $\Gamma_D/\hbar$, so that $2(2\ln{2})^{1/2}\sigma =\Gamma_D/\hbar$. We consider adjusting the laser beam waist $w_l$ to match the 1-$\sigma$ radius $w_b$ of the ion bunch, leading to $S_l\approx S_b$, where $S_l=\pi w_l^2$ and $S_b=\pi w_b^2$ are the transverse cross sections of the laser beam and ion bunch, respectively.
The number of ions excited with a single laser pulse is about $\rho_0\cdot N_{\textrm{ion}}$ with $N_\textrm{ion}$ being the number of ions per bunch.
Implementing a $\pi$ pulse for ions in the resonant group ($\Delta\approx0$) requires a peak laser-pulse intensity of
\begin{equation} \label{Eq:Intensity_pi}
I_{0\pi}'= \frac{\ln{2}}{3}\frac{\hbar^2\omega'^3}{t_\textrm{p}'^2c^2\Gamma_\textrm{rad}}\approx \frac{1}{48\ln{(2)}}\frac{\hbar^2\omega'^5}{c^2\Gamma_{\textrm{rad}}}\left(\frac{\Delta\gamma}{\gamma}\right)^2
\end{equation}
determined from $\int\Omega_{\textrm{R}}(t)dt=\pi$.
This intensity would be high when having
a narrow radiative width and a large relative ion-energy spread.
It could be more practical to transfer population into the isomeric state through repeatedly pumping the ion bunch than using single $\pi$ pulses.
For this,
we need the same repetition frequency and proper synchronization of laser pulses and ion bunches.
Thus, the repetition rate of the laser pulses is adjusted to be $f_l=N_b\cdot f_b$, with $N_b$ and $f_b$ being the number of ion bunches and the circulating frequency of a single ion bunch in the storage ring, respectively.
The population-transfer fraction in each ion bunch, denoted as $\rho_{\textrm{ex}}$, depends on the number of laser pulses having collided with individual ion bunches, denoted as $n_{\textrm{col}}$.
The population evolution under repeated pumping ($\rho_{\textrm{ex}}$ versus $n_{\textrm{col}}$) can be obtained by repeatedly calculating Eqs.\,\eqref{Eq:DM_TLS_v2} and \eqref{Eq:ave_popu_tran} in a loop [Eq.\,\eqref{Eq:ave_popu_tran} resets initial
conditions of population for Eq.\,\eqref{Eq:DM_TLS_v2}].
\begin{figure}[!htpb]\centering
\includegraphics[width=0.95\linewidth]{Th_excitation.png}
\caption{Production of $^{229m}$Th isomers through (a) resonant excitation or (b) laser excitation via an intermediate nuclear state (the 29\,keV level).
The second scenario requires a large Lorentz factor of the ion bunch, which could be achieved at the Large Hadron Collider (LHC) among facilities listed in Table\,\ref{tab:ion-beam_facilities}.
}
\label{fig:Th_excitation}
\end{figure}
To give a specific example, we consider exciting relativistic $^{229}$Th ions at the SPS; see also Fig.\,\ref{fig:Th_excitation}\,(a) and Table\,\ref{tab:ion-beam_facilities}.
There can be up to $N_b=36$ ion bunches circulating in the storage ring with $S_b\approx\pi(0.5\textrm{mm})^2$ and $N_{\textrm{ion}}\approx10^8$ per bunch.
A pulsed CO$_2$ laser ($\hbar\omega=0.117\,$eV) having an energy of $E_{\textrm{p}}=100\,\mu$J per pulse and operating at a repetition rate of $f_l=1.56\,$MHz synchronized to the circulating ion bunches can be used to excite the nuclear isomeric state in $^{229}$Th ions.
The required Lorentz factor of the ion bunch is $\gamma\approx35$ ($\Delta\gamma/\gamma\approx10^{-4}$ at the SPS).
We assume using a Fabry–P\'erot cavity offering a factor of $\approx10^5$ enhancement of the laser power inside the cavity, leading to $E_{\textrm{p}}=10\,$J. The peak intensity of the laser pulse can be derived from $E_{\textrm{p}}\approx I_0\cdot S_b\cdot t_p $.
The average power is
$E_\textrm{p}\cdot f_l = 1.56\times10^7\,$W.
These laser-pulse parameters are for the lab frame.
We have
\begin{equation} \label{Eq:para_frame}
\begin{split}
E_{\textrm{p}}'&=2\gamma E_{\textrm{p}}\, , \\
t_\textrm{p}'&=t_\textrm{p}/2\gamma\, ,\\
I'& =(2\gamma)^2 I\, ,
\end{split}
\end{equation}
when switching between the ion frame and the lab frame.
We obtain $t_\textrm{p}'\approx2.2\,$ps and $I_0'\approx4.0\times10^{20}\,$W/m$^2$. The $I_0'$ is orders of magnitude below the intensity required for a $\pi$-pulse $I_{0\pi}'\approx7.9\times10^{23}\,$W/m$^2$ [see Eq.\,\eqref{Eq:Intensity_pi}].
The numerical results show
$\rho_{ee}(\Delta=0)\approx1.25\times10^{-3}$ and $\rho_0\approx8.9\times10^{-4}$ [$\rho_0\approx0.71\cdot\rho_{ee}(\Delta=0)$]; see also Table\,\ref{tab:repeat_pump}.
The population transfer is presented in Fig.\,\ref{fig:Rep_pump_SPS_HESR}\,(a).
The maximal population-transfer fraction is $\rho_{\textrm{max}}\approx0.50$. $\rho_{\textrm{ex}}=0.99\rho_{\textrm{max}}$ is reached after about $n_{\textrm{col}}\approx2600$ collisions, corresponding to an excitation time of $t_{\textrm{ex}}=n_{\textrm{col}}/f_b\approx60\,$ms.
We note that with $\rho_0\ll\rho_{\textrm{max}}$, the increase of $\rho_{\textrm{ex}}$ can be well described by a simple exponential behavior (see Fig.\,\ref{fig:Rep_pump_SPS_HESR})
\begin{equation} \label{Eq:exp_behavior}
\rho_{\textrm{ex}} = \rho_{\textrm{max}}\left[1-\exp{\left(-\frac{\rho_0}{\rho_{\textrm{max}}}n_{\textrm{col}}\right)}\right]\,.
\end{equation}
Here, the excitation process, using repeated pumping, is effectively incoherent since the optical phase for consecutive interactions of an ion with the laser light is random.
\begin{figure}[!htpb]\centering
\includegraphics[width=0.95\linewidth]{Rep_pump_SPS_HESR.png}
\caption{
Repeated pumping at (a) SPS and (b) HESR. The blue solid curve is derived from repeatedly calculating the population transfer using Eqs.\,\eqref{Eq:DM_TLS_v2} and \eqref{Eq:ave_popu_tran}. The thick orange dashed curve is the prediction using Eq.\,\eqref{Eq:exp_behavior}, which fits well with the blue curve.
The characteristic time for reaching population saturation is about $60\,$ms in both cases, see the text.
}
\label{fig:Rep_pump_SPS_HESR}
\end{figure}
We consider another similar example of resonant excitation of $^{229m}$Th isomers, this time at HESR.
Electron cooling of ion bunches with energies of up to about 3.5\,A\,GeV and variable stochastic cooling up to the highest energies will be available in the HESR \cite{Stohlker-2015b}.
As a consequence, a smaller ion-beam energy spread (down to $\Delta\gamma/\gamma\lesssim10^{-5}$) can be achieved.
Assuming using cooled relativistic ions with $\gamma\approx6$ ($\Delta\gamma/\gamma\approx10^{-5}$), the energy of laser photons should be $\hbar\omega\approx8.19/12=0.683\,$eV.
Consider that there is one ion bunch containing around $10^8$ $^{229}$Th ions interacting with laser pulses with $f_l=522$\,kHz (synchronized to the ion bunch) and again, $E_{\textrm{p}}=10\,$J (inside a Fabry–P\'erot cavity).
Following a procedure similar to the one discussed above for SPS, we obtain, for HESR,
$t_\textrm{p}'\approx 22\,$ps, $I_0'\approx3.9\times10^{17}\,$W/m$^2$ ($I_0'\ll I_{0\pi}'\approx 7.9\times10^{21}\,$W/m$^2$),
$\rho_{ee}(\Delta=0)\approx1.22\times10^{-4}$, $\rho_0\approx8.6\times10^{-5}$, and $\rho_{\textrm{max}}\approx 0.50$; see also Table\,\ref{tab:repeat_pump}.
We have, again, $\rho_0\approx0.71\cdot\rho_{ee}(\Delta=0)$, and the increase of $\rho_{\textrm{ex}}$ can be well described by Eq.\,\eqref{Eq:exp_behavior}; see Fig.\,\ref{fig:Rep_pump_SPS_HESR}\,(b).
$\rho_{\textrm{ex}}=0.99\rho_{\textrm{max}}$ is reached after about $n_{\textrm{col}}\approx 2.9\times10^4$ collisions, corresponding to an excitation time of $t_{\textrm{ex}}=n_{\textrm{col}}/f_b\approx56\,$ms.
Direct resonant excitation could also be pursued at the LHC, where the large $\gamma$ could allow excitation using the magnetic field produced in an undulator (e.g., a Halbach array made of permanent magnets) or a microwave source \cite{Budker2021GF_nucl}.
\begin{table}[htpb]
\centering
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}} lcccc cc}
\hline
\hline
Facility &$\gamma$ &$t_\textrm{p}'\,$(ps) &$I_0'\,$(W/m$^2$) &$\rho_{ee}(\Delta=0)$ &$\rho_0$ &$t_\textrm{ex}\,$(ms) \\
\hline \\[-0.2cm]
SPS &35 &2.2 &$4.0\times10^{20}$ &$1.25\times10^{-3}$ &$8.9\times10^{-4}$ &60\\
HESR &6 &22 &$3.9\times10^{17}$ &$1.22\times10^{-4}$ &$8.6\times10^{-5}$ &56\\
\hline
\hline
\end{tabular*}
\caption{
Repeated pumping of $^{229m}$Th isomers at the SPS and HESR. $\gamma$, $t_\textrm{p}'$ and $I_0'$ are some of the considered parameters of the ion bunch and laser beam. $\rho_{ee}(\Delta=0)$ and $\rho_0$ are the population-transfer fraction of the ion group resonant with the laser pulse and the transfer fraction averaged over all ions, respectively, in the first round.
$t_\textrm{ex}$ is the time needed to achieve $99\%$ of the maximal population transfer, $\rho_{\textrm{max}}\approx0.5$ in both cases (corresponding to $5\times10^7$ ions per bunch).
}
\label{tab:repeat_pump}
\end{table}
\subsubsection{Excitation in H- or Li- like Th ions}
\label{subsubsec:res_H-like}
In H- or Li-like Th ions, the lifetime of the isomeric state could be reduced by orders of magnitude compared to bare nuclei because of the nuclear hyperfine mixing (NHM) effect; see, for example, Refs.\,\cite{Karpeshin1998NHM,Tkalya2016_NHM,Shabaev2021NHM}.
Direct resonant excitation of $^{229m}$Th isomers using H- or Li-like Th ions was proposed at GSI; see, for example, Ref.\,\cite{Brandau2020_NHM_Th}.
Due to the increased radiative width, significantly higher excitation rates of isomers can be obtained.
Guided by Eq.\,\eqref{Eq:Intensity_pi}, the peak laser pulse intensity required for implementing a $\pi-$pulse would be reduced by orders of magnitude due to the increase in the radiative width.
We calculate below the transition dipole moment considering the NHM effect and the laser-pulse intensity needed for implementing a $\pi-$pulse to transfer population from the nuclear ground state to the isomeric state.
We show that, considering the NHM effect, Eq.\,\eqref{Eq:Intensity_pi} still works as a proper estimate, and the required intensity could be experimentally achievable.
The electronic ground state of H-like $^{229}$Th ions, $^2S_{1/2}$, has two hyperfine components, $F=2,3$ for nuclei in the ground state ($F=1,2$ for nuclei the isomeric state).
$F$ is the total angular momentum quantum number.
As a result of the NHM effect, the two $F=2$ levels with
different nuclear states ($3/2^+$ and $5/2^+$) are mixed and can be represented as \cite{Shabaev2021NHM}
\begin{align} \label{Eq:NHM}
&\overline{\ket{\frac{5}{2}^+,F, M_{F}}}=\sqrt{1-b^2}\ket{\frac{5}{2}^+,F, M_{F}}-b\ket{\frac{3}{2}^+,F, M_{F}}\nonumber\\
&\overline{\ket{\frac{3}{2}^+,F, M_{F}}}=\sqrt{1-b^2}\ket{\frac{3}{2}^+,F, M_{F}}+b\ket{\frac{5}{2}^+,F, M_{F}}\,,
\end{align}
where $b$ is the NHM coefficient.
We use below $b=-2.9\times10^{-2}$ and $-4.9\times10^{-3}$ for H- and Li-like thorium ions, respectively, according to the theoretical calculations in Ref.\,\cite{Shabaev2021NHM}.
We note that only states with the same $F$ and the same magnetic quantum number $M_F$ are mixed.
Theoretical results of the hyperfine structure (HFS) energies and radiative transition rates between the HFS levels are presented in Ref.\,\cite{Shabaev2021NHM} and also shown in Fig.\,\ref{fig:H-like_Th}.
\begin{figure}[!htpb]\centering
\includegraphics[width=\linewidth]{H-like_Th_NHM.png}
\caption{
Energy levels of the nuclear ground and isomeric states and related radiative transition rates ($\Gamma_{\textrm{rad}}/\hbar$) in bare thorium nuclei (left) and H-like thorium ions (right). The values are from Ref.\,\cite{Shabaev2021NHM}, using $B(M1)=0.008$\,W.u. as the reduced transition probability between the nuclear isomeric and ground states.
Due to the NHM effect in H-like thorium, the two $F=2$ hyperfine energy levels are mixed, leading to significantly higher radiative transition rates from the isomeric state to the ground state compared to bare nuclei. Similar effect also exists in Li-like thorium ions; see the text.
}
\label{fig:H-like_Th}
\end{figure}
The higher HFS level of H-like $^{229g}$Th (``g" denotes the nuclear ground state)
has a lifetime of about 183\,ms in the ion frame.
Therefore, after $^{229g}$Th ions are produced and circulate in the storage ring for a sufficiently long time, all the ions are in the lower HFS level, $\overline{\ket{5/2^+,F=2}}$.
Interaction of the ion beam with the magnetic fields of the ring lattice and collisions with rest gas atoms and electrons of the cooler may alter the population of the hyperfine states. However, no re-population of the upper HFS was so far observed in sensitive HCI experiments at the ESR\,\cite{Seelig-1998, Litvinov-2007,Winckler-2009, Atanasov-2012, Kienle-2013, Ozturk-2019}. Therefore, in the following we neglect the machine-induced (de)excitation of hyperfine states.
We consider exciting the transition from $\ket{g}=\overline{\ket{5/2^+,F=2}}$ to $\ket{e}=\overline{\ket{3/2^+,F=2}}$.
The radiative width of the corresponding decay channel $\ket{e}\rightarrow\ket{g}$
is $\Gamma_\textrm{rad}\approx1.4\times10^{-14}\,$eV, and
the transition energy is $\hbar\omega'=8.31(13)\,$eV \cite{Shabaev2021NHM}.
Again, as discussed above, we adjust the laser pulse duration according to Eq.\,\eqref{Eq:pulse_dur} and search for the resonance, so that we have $\Gamma_\textrm{p}'\approx\Gamma_D$ and $2\gamma^{\textrm{ave}}\cdot\hbar\omega\approx8.31\,$eV.
The transition magnetic dipole moment between sublevels $\overline{\ket{5/2^+,F=2,M_F}}$ and $\overline{\ket{3/2^+,F=2,M_F'}}$ is
\begin{equation} \label{Eq:dipole_NHM_res}
\begin{split}
\mu&=\left| \overline{\bra{\frac{3}{2}^+,F, M_{F}'}}M_{L\lambda} \overline{\ket{\frac{5}{2}^+,F, M_{F}}} \right|\\
&\approx \left| \overline{\bra{\frac{3}{2}^+,F, M_{F}'}}M_{L\lambda}^{(e)} \overline{\ket{\frac{5}{2}^+,F, M_{F}}} \right|
\\
&= \left|\braket{FM_FL\lambda|FM_F'}\right| \frac{|b|\sqrt{1-b^2}}{\sqrt{2F+1}}
\\
&\Big|
\bra{\frac{5}{2}^+,F}|M^{(e)}_{L}|\ket{\frac{5}{2}^+,F}-\bra{\frac{3}{2}^+,F}|M^{(e)}_{L}|\ket{\frac{3}{2}^+,F}\Big|\,,
\end{split}
\end{equation}
where
$\braket{FM_FL\lambda|FM_F'}$ is the Clebsch–Gordan coefficient,
$\braket{a||M_L||b}$ is the reduced matrix element between states $\ket{a}$ and $\ket{b}$,
$M_{L\lambda}=M_{L\lambda}^{(e)}+M_{L\lambda}^{(n)}$ is the magnetic multipole operator composed of the electronic part $M_{L\lambda}^{(e)}$ and the nuclear part $M_{L\lambda}^{(n)}$ with $L$ and $\lambda$ representing the radiative transition multiplarity and photon polarization. Here, we have $L=1$ for the $M1$ transition. $\mu$ is dominated by the electronic part as a result of the NHM,
which can be seen from the fact that
$|b|\gg \mu_N/\mu_B\approx5\times10^{-4}$
($\mu_N$ and $\mu_B$ are the nuclear and Bohr magneton, respectively).
For simplicity, we define the reduced magnetic dipole moment
\begin{equation}
\begin{split} \label{Eq:red_dipole}
\mu_{\textrm{r}}&=\frac{\mu}{|\braket{FM_FL\lambda|FM_F'}|} =\frac{|b|\sqrt{1-b^2}}{\sqrt{2F+1}}\cdot
\\
&\Big|
\bra{\frac{5}{2}^+,F}|M^{(e)}_{L}|\ket{\frac{5}{2}^+,F}-\bra{\frac{3}{2}^+,F}|M^{(e)}_{L}|\ket{\frac{3}{2}^+,F}\Big|\,,
\end{split}
\end{equation}
which is independent of the magnetic quantum numbers ($M_{F}, M_{F}',\lambda$).
$\mu_{\textrm{r}}$ can be derived from the radiative rate of the $\ket{e}\rightarrow\ket{g}$ transition \cite{Shabaev2021NHM}:
\begin{equation} \label{Eq:rad_rate}
\begin{split}
\frac{\Gamma_{\textrm{rad}}}{\hbar}&= \frac{\mu_0}{3\pi\hbar}\left(\frac{\omega'}{c}\right)^3\frac{1}{2F+1}\\
&\sum_{M_{F}',M_{F},\lambda}
|\overline{\bra{\frac{5}{2}^+FM_{F}}}
M_{L\lambda}
\overline{\ket{\frac{3}{2}^+FM_{F}'}}|^2\\
&\approx \frac{\mu_0}{3\pi\hbar}\left(\frac{\omega'}{c}\right)^3\frac{1}{2F+1}\\
&\sum_{M_{F}',M_{F},\lambda}
|\braket{FM_FL\lambda|FM_F'}|^2\cdot \mu_{\textrm{r}}^2\\
&=\frac{\mu_0}{3\pi\hbar}\left(\frac{\omega'}{c}\right)^3\mu_{\textrm{r}}^2\,.
\end{split}
\end{equation}
We get the reduced Rabi frequency
\begin{equation}
\label{Eq:Rabi_red_NHM}
\Omega_{\textrm{r}}=\frac{\mu_r B'}{\hbar}\approx\left(\frac{6\pi c^2 \Gamma_\textrm{rad} I'}{\hbar^2\omega'^3}\right)^{1/2}\,,
\end{equation}
having the same form as Eq.\,\eqref{Eq:Rabi_red}.
Consequently,
the peak laser-pulse intensity
required in the ion frame for a (nominal) $\pi-$pulse, $\int\Omega_{\textrm{r}}(t)dt=\pi$, can be deduced using Eq.\,\eqref{Eq:Intensity_pi}.
Such intensity would produce $\rho_{ee}(\Delta=0)\approx1$. Considering the ion-energy spread, the average excitation fraction is $\rho_0\approx0.7\cdot\rho_{ee}(\Delta=0)\approx0.7$; see Eq.\,\eqref{Eq:ave_popu_tran}.
We return to the example of exciting $^{229m}$Th at the SPS using relativistic ion bunches with $\gamma\approx35$ ($\Delta\gamma/\gamma\approx10^{-4}$) but for now dealing with H-like $^{229}$Th ions. The required laser-beam parameters (in the lab frame) for implementing the $\pi-$pulse are
$I_{0}=1.1\times10^{15}$\,W/m$^2$ and
$E_\textrm{p}= 140\,$mJ
($t_\textrm{p}\approx150\,$ps).
For the HESR, we derive
$I_{0}=3.7\times10^{14}$\,W/m$^2$ and $E_\textrm{p}\approx 1.4\,$J ($t_{\textrm{p}}\approx260\,$ps).
The pulse energies required here are smaller than $E_\textrm{p}=10$\,J assumed in Sec.\,\ref{subsubsec:res_bare}, thus easier to achieve experimentally. Also, when exciting isomers using $\pi-$pulses, we do not need high repetition rate of the laser pulses.
For Li-like $^{229}$Th ions,
the NHM coefficient $b$ is smaller compared to H-like $^{229}$Th ions, leading to lower radiative decay rates of the HFS levels \cite{Shabaev2021NHM}.
The higher HFS level of Li-like $^{229g}$Th in the electronic ground state $1s^22s_{1/2}$ ($^2S_{1/2}$)
has a long lifetime of about 27\,s in the ion frame.
Therefore, instead of considering all ions decaying into the lower HFS level, we can assume that both HFS levels ($\overline{\ket{5/2^+,F=2}}$ and $\ket{5/2^+,F=3}$) are populated and the relative population is proportional to the degeneracy factor $g_{F}=2F+1$.
In this case, besides exciting transitions from $\overline{\ket{5/2^+,F=2}}$, we also need to consider transitions from
$\ket{5/2^+,F=3}$ to $\overline{\ket{3/2^+,F=2}}$.
We note that
Eq.\,\eqref{Eq:Rabi_red_NHM}
also holds for those transitions.
The radiative transition rates ($\Gamma_{\textrm{rad}}/\hbar$) of $\overline{\ket{3/2^+,F'=2}}\rightarrow\overline{\ket{5/2^+,F=2}}$, $\overline{\ket{3/2^+,F'=2}}\rightarrow\ket{5/2^+,F=3}$
and $\ket{3/2^+,F'=1}\rightarrow\overline{\ket{5/2^+,F=2}}$
are about 0.72, 0.39, and 0.45\,s$^{-1}$ \cite{Shabaev2021NHM},
respectively.
The required laser intensity for a (nominal) $\pi-$pulse corresponding to those transitions can be obtained using
Eq.\,\eqref{Eq:Intensity_pi}.
One can also produce polarized Li-like $^{229g}$Th ions first via optical pumping using the electronic transition between $1s^22s_{1/2}$ ($^2S_{1/2}$) and $1s^22p_{1/2}$ ($^2P_{1/2}$); see Sec.\,\ref{Subsec:elec_tran},
from which fully polarized ions, population only in the $\ket{5/2^+,F=3,M_F=3}$ state, could be obtained. Afterwards, a $\pi-$pulse could be applied to transfer the population into the isomeric state $\overline{\ket{3/2^+,F=2,M_F=2}}$. The peak laser-pulse intensity required for $\int\Omega_{\textrm{R}}(t)dt=\pi$ ($\Omega_{\textrm{R}}\approx\Omega_{\textrm{r}}|\braket{331-1|22}|$) is $I_{0}\approx8.5\times10^{16}$W/m$^2$, corresponding to $E_{\textrm{p}}\approx11\,$J, at the SPS.
\subsection{Excitation via the electronic excited state in Li-like $^{229}$Th}
\label{Subsec:exc_271eV}
\subsubsection{
Forbidden $E1$ transitions opened by the NHM effect
}
We present in Fig.\,\ref{Fig:Li-like_Th} the low-lying energy levels in Li-like ${}^{229}$Th ions and transitions between hyperfine components that change both the nuclear and electronic states. These transitions are all electric-dipole ($E1$) allowed as a result of the NHM effect, i.e., mixing of the $F=2$ levels in the electronic $1s^22s_{1/2}({}^2S_{1/2})$ and/or $1s^22p_{1/2}({}^2P_{1/2})$ states.
\begin{figure}[!htpb]
\centering
\includegraphics[width=\linewidth]{Li-like_Th.png}
\caption{
The low-lying energy levels in Li-like ${}^{229}$Th ions and transitions between hyperfine components that change both the nuclear and electronic states.
The transitions denoted by the solid arrows, although occurring between different nuclear states, are all electric-dipole transitions, due to the NHM effect. Considering that the mixing only occurs between states with the same total angular momentum quantum number $F$, the transitions represented by the blue (orange) arrows are allowed due to mixing of the two $F=2$ levels in the electronic ground (excited) state, ${}^2S_{1/2}$ (${}^2P_{1/2}$).
The transitions denoted by green arrows include contributions from mixing in both ${}^2S_{1/2}$ and ${}^2P_{1/2}$ manifolds.
}
\label{Fig:Li-like_Th}
\end{figure}
First, consider the transitions from the nuclear ground state to the isomeric state, denoted by the arrows pointing from the left top to right bottom in Fig.\,\ref{Fig:Li-like_Th}. We denote the higher and lower levels as
\begin{equation}
\begin{split}
\ket{e} &=\overline{\ket{{}^2P_{1/2},I_g,F_e}}\\
&\approx \ket{{}^2P_{1/2},I_g,F_e} - \delta_{F_e,2}\cdot b_e \ket{{}^2P_{1/2},I_e,F_e}\,,\\
\ket{g} &=\overline{\ket{{}^2S_{1/2},I_e,F_g}}\\
&\approx \ket{{}^2S_{1/2},I_e,F_g} + \delta_{F_g,2}\cdot b_g \ket{{}^2S_{1/2},I_g,F_g}\,,
\end{split}
\end{equation}
where $\delta_{F_e,2}\cdot b_e$ and $\delta_{F_g,2}\cdot b_g$ (both $\ll1$) are the NHM factors of the HFS levels $F_e$ in the electronic excited ${}^2P_{1/2}$ state and $F_g$ in the electronic ground state ${}^2S_{1/2}$, respectively.
Here, $\delta_{F_{g(e)},2}$ is the Kronecker delta symbol meaning that the NHM occurs between the two $F_{g(e)}=2$ levels.
$b_g\approx-4.9\times10^{-3}$ is derived in Ref.\,\cite{Shabaev2021NHM}. However, to the best of our knowledge, $b_e$ has not been calculated yet.
In Ref.\,\cite{Shabaev2021NHM}, the authors also calculated the NHM effect for boron-like ${}^{229}$Th ions in the electronic ground state, $1s^22s^22p_{1/2}$ (${}^2P_{1/2}$). The NHM factor in this case is $b_g(\textrm{B}) \approx -1.7\times10^{-3}$.
$b_e$ of Li-like ${}^{229}$Th ions should satisfy $|b_g(\textrm{B})|< |b_e(\textrm{Li})|<|b_g(\textrm{Li})|$, thus we roughly estimate it as $b_e \approx -3\times10^{-3}$.
The reduced $E1$ transition dipole between the states is
\begin{equation} \label{Eq:red_dip_NHM}
\begin{split}
d_{e\rightarrow g}
&= \bra{g}|E_L^{(e)}|\ket{e} \\
&\approx \delta_{F_g,2}\cdot b_g \bra{{}^2S_{1/2},I_g,F_g}|E_L^{(e)}|\ket{{}^2P_{1/2},I_g,F_e}\\
&\quad -\delta_{F_e,2}\cdot b_e \bra{{}^2S_{1/2},I_e,F_g}|E_L^{(e)}|\ket{{}^2P_{1/2},I_e,F_e}\,,
\end{split}
\end{equation}
where $\bra{{}^2S_{1/2},I,F_g}|E_L^{(e)}|\ket{{}^2P_{1/2},I,F_e}$ can be further related to the reduced matrix element between the fine-structure levels according to \cite{Auzinsh2010LightAtom}
\begin{equation}
\begin{split}
\bra{\xi_g,J_g,I,F_g}|&E_L^{(e)}|\ket{\xi_e,J_e,I,F_e} = (-1)^{J_g+I+F_e+1}\cdot\\
&(g_{F_e}g_{F_g})^{1/2}\begin{Bmatrix}
J_g & F_g &I\\
F_e & J_e &1
\end{Bmatrix}\bra{\xi_gJ_g}|E_L^{(e)}|\ket{\xi_eJ_e}\,.
\end{split}
\end{equation}
Here, $J_{g(e)}=1/2$ is the total electronic angular momentum of the ${}^2S_{1/2}({}^2P_{1/2})$ fine structure level, $\xi_{g(e)}$ incorporates all other electronic quantum numbers, and $g_{F_g(F_e)}$ is the degeneracy factor of the $F_g(F_e)$ HFS level.
Therefore, Eq.\,\eqref{Eq:red_dip_NHM} can be rewritten as
\begin{equation} \label{Eq:red_dip_NHM_New}
\begin{split}
\frac{d_{e\rightarrow g}}{d_{\xi_eJ_e\rightarrow\xi_gJ_g}}
&\approx \delta_{F_g,2}\cdot (5g_{F_e})^{1/2}b_g \begin{Bmatrix}
1/2 & 2 &5/2\\
F_e & 1/2 &1
\end{Bmatrix}\\
&\,\,+\delta_{F_e,2}\cdot (5g_{F_g})^{1/2}b_e \begin{Bmatrix}
1/2 & F_g &3/2\\
2 & 1/2 &1
\end{Bmatrix}\,,
\end{split}
\end{equation}
where we have used the notation $d_{\xi_eJ_e\rightarrow\xi_gJ_g} = \bra{\xi_gJ_g}|E_L^{(e)}|\ket{\xi_eJ_e}$ and neglected the overall phase factor $(-1)^{F_e}$.
The radiative width of the transition from state $\ket{e}$ to $\ket{g}$, denoted as $\Gamma_{eg}$, can be derived from
\begin{equation} \label{Eq:rad_width}
\frac{\Gamma_{eg}}{\Gamma_{\xi_eJ_e\rightarrow\xi_gJ_g}} = \frac{2J_e+1}{2F_e+1}\left(\frac{\omega_{eg}}{\omega_{\xi_eJ_e\rightarrow\xi_gJ_g}}\right)^3\left| \frac{d_{e\rightarrow g}}{d_{\xi_eJ_e\rightarrow\xi_gJ_g}}\right|^2\,.
\end{equation}
For $E1$ transitions between states $a$ and $b$, we have \cite{Budker2004AP,Auzinsh2010LightAtom}
\begin{equation} \label{Eq:E1_dipole}
\Gamma_{a\rightarrow b}
= \frac{\omega_{ab}^3}{3\pi\epsilon_0 c^3}\cdot \frac{1}{g_a}\cdot|d_{a\rightarrow b}|^2\,,
\end{equation}
where $g_a$ is the degeneracy factor of state $a$.
The radiative width of the $^2P_{1/2}\rightarrow{}^2S_{1/2}$ transition is
$\Gamma_{\xi_eJ_e\rightarrow\xi_gJ_g}=\hbar/T_{\xi_eJ_e}=1.12\times10^{-5}\,$eV, with $T_{\xi_eJ_e}=5.90\times10^{-11}\,$s being the radiative lifetime of the $^2P_{1/2}$ state \cite{Theodosiou1991Li-Ion}.
The transitions from the nuclear excited state to ground state, denoted by the arrows pointing from the right top to left bottom in Fig.\,\ref{Fig:Li-like_Th}, can be treated similarly.
The estimated radiative widths are presented in Table\,\ref{tab:rad_width_NHM}.
These widths are larger than that of the purely nuclear radiative transition (without considering the NHM effect) by about nine orders of magnitude.
\begin{table}[htpb]
\centering
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}} l cc }
\hline
\hline
Transition & $E_{if}$\,(eV) & $\Gamma_{if}\,$(eV) \\
${}^2P_{1/2},I_i^P, F_i\rightarrow {}^2S_{1/2},I_f^P, F_f$ & &\\
\hline \\[-0.2cm]
$5/2^+,3\rightarrow 3/2^+,2$ &263 &$1.4\times10^{-10}$ \\%&3.4 \\
$5/2^+,2\rightarrow 3/2^+,1$ &263 &$4.6\times10^{-11}$ \\%&9.9 \\
$5/2^+,2\rightarrow 3/2^+,2$ &263 & $2.0\times10^{-10}$ \\% & -- \\
$3/2^+,1\rightarrow 5/2^+,2$ &279 &$2.4\times10^{-10}$ \\%&1.9 \\
$3/2^+,2\rightarrow 5/2^+,3$ &279 &$8.5\times10^{-11}$ \\%&5.3 \\
$3/2^+,2\rightarrow 5/2^+,2$ &279 &$2.9\times10^{-10}$ \\%& -- \\
\hline
\hline
\end{tabular*}
\caption{
Energies $E_{if}$ and radiative widths $\Gamma_{if}$ estimated for transitions between HFS levels in Li-like ${}^{229}$Th ions. These transitions change both the nuclear and electronic states, see also Fig.\,\ref{Fig:Li-like_Th}.
For comparison, we note that the radiative widths of the pure electronic transition, $^2P_{1/2}\rightarrow{}^2S_{1/2}$, and the pure nuclear transition, $I_e^P=3/2^+\rightarrow I_g^P=5/2^+$ (without NHM), are about
$1.12\times10^{-5}\,$eV and $8.8\times10^{-20}\,$eV, respectively.
}
\label{tab:rad_width_NHM}
\end{table}
\subsubsection{
Isomer excitation via Stimulated Raman Adiabatic Passage
}
\label{subsubsec:STIRAP_Li-like}
In this Section we discuss the possibility of exciting $^{229m}$Th isomers via the intermediate electronic excited state $1s^22p_{1/2}({}^2P_{1/2})$ in Li-like thorium ions, relying on the NHM effect. This is essentially addressing a three-level system in a so-called $\Lambda$ configuration with two laser fields that transfer the population between the nuclear ground state and isomeric state. In the following we discuss coherent population transfer via Stimulated Raman Adiabatic Passage (STIRAP).
In STIRAP \cite{Bergmann1998_STIRAP, bergmann_perspective_2015}, two coherent laser fields, comprising the pump and the Stokes pulses, couple to a three-level $\Lambda$ system as shown in Fig.\,\ref{Fig:STIRAP_Li-likeTh}(a).
STIRAP is characterized by its robustness against fluctuations of experimental parameters and is independent of spontaneous decay rates and the branching ratio of the upper state.
During the STIRAP process a so-called dark state \cite{Bergmann1998_STIRAP}
\begin{equation} \label{Eq:STIRAP_dark_state}
\ket{D}=\frac{\Omega_\textrm{S} \left(t \right)}{\sqrt{\Omega_\textrm{S} \left(t \right)^2 +\Omega_\textrm{P} \left(t \right)^2}}\ket{1}- \frac{\Omega_\textrm{P} \left(t \right)}{\sqrt{\Omega_\textrm{S} \left(t \right)^2 +\Omega_\textrm{P} \left(t \right)^2}}\ket{3}
\end{equation}
is formed, with
$\Omega_\textrm{S/P}$ being the time dependent Rabi frequency of the Stokes/pump pulse.
By adjusting laser parameters such as the delay time $\Delta \tau$ and the pulse energy, the entire population can be transferred from $\ket{1}$ to $\ket{3}$ via the time-dependent dark state without occupying the intermediate state $\ket{2}$.
To give a concrete example, we consider our system initially in the $\ket{1}=\ket{{}^2S_{1/2},5/2^+,F=3,M_F=3}$ state. This corresponds to fully polarized thorium ions, which can be produced via optical pumping, see also Sec.\,\ref{Subsec:elec_tran}. Our three-level $\Lambda$ scheme is shown in Fig.\,\ref{Fig:STIRAP_Li-likeTh}. A pump laser couples the initial state $\ket{1}$ and the intermediate state $\ket{2}=\ket{{}^2P_{1/2},5/2^+,F=3,M_F=2}$, while the Stokes laser couples the state
$\ket{2}$ and the final state $\ket{3}=\overline{\ket{{}^2S_{1/2},3/2^+,F=2,M_F=1}}$, which, due to the NHM, is a superposition state of $\ket{{}^2S_{1/2},3/2^+,F=2,M_F=1}$ and $\ket{{}^2S_{1/2},5/2^+,F=2,M_F=1}$; see Eq.\,\eqref{Eq:NHM}.
\begin{figure}[!htpb]
\centering
\includegraphics[width=\linewidth]{STIRAP_Li-likeTh.png}
\caption{
(a) A generic STIRAP scheme. The pump and Stokes and pulses interact with a three-level $\Lambda$ system. Under STIRAP conditions, the population from state 1 can be fully transferred to state 3, while the population of state 2 remains nominally zero throughout the process.
(b) Excitation of $^{229m}$Th via the electronic state $1s^22p_{1/2}$(${}^2P_{1/2}$) in Li-like thorium. Red solid lines denote energy levels of thorium ions in the nuclear isomeric state $I^P=3/2^+$. Other energy levels are for thorium ions in the nuclear ground state $I^P=5/2^+$. Here, as an example, we consider a pump laser driving the transition between $\ket{5/2^+,{}^2S_{1/2},F=3}$ and $\ket{5/2^+,{}^2P_{1/2},F=3}$, and a Stokes laser driving the transition between $\overline{\ket{3/2^+,{}^2S_{1/2},F=2}}$ and $\ket{5/2^+,{}^2P_{1/2},F=3}$. The latter transition is allowed due to the NHM effect; see the text.
}
\label{Fig:STIRAP_Li-likeTh}
\end{figure}
The energies and radiative widths of transitions between HFS levels of ${}^2S_{1/2}$ and ${}^2P_{1/2}$ are presented in Tables\,\ref{tab:hfs_tran_gnd} and \ref{tab:hfs_tran_iso} in Sec.\,\ref{Subsec:elec_tran}, for Li-like $^{229g}$Th and $^{229m}$Th, respectively. These transitions do not involve change of the nuclear states. Estimates about the transitions changing both the nuclear and electronic state are presented in Table\,\ref{tab:rad_width_NHM}.
The reduced transition dipole moment of these $E1$ transitions is related to the radiative width, see Eq.\,\eqref{Eq:E1_dipole}.
The Rabi frequency in the ion frame corresponding to the pump laser is given by
\begin{equation} \label{Eq:E1_Rabi_pump}
\begin{split}
\Omega_\textrm{P}= \frac{d_\textrm{P}E'}{\hbar}
&=\frac{|\braket{F_gM_{F_g}L\lambda|F_eM_{F_e}}|}{g_{F_e}^{1/2}}\frac{d_rE'}{\hbar}\\
&=|\braket{F_gM_{F_g}L\lambda|F_eM_{F_e}}|\left(
\frac{6\pi c^2I_\textrm{P}'\Gamma_{F_e\rightarrow F_g}}{\hbar^2\omega_\textrm{P}'^3}
\right)^{1/2}\,,
\end{split}
\end{equation}
where $d_\textrm{P}$ is the transition dipole moment amplitude, $\braket{F_gM_{F_g}L\lambda|F_eM_{F_e}}=\braket{331-1|32}$, $\hbar\omega_\textrm{P}'\approx271\,$eV is the transition energy,
$\Gamma_{F_e\rightarrow F_g}=5.0\times10^{-6}\,$eV
is the radiative width of the transition $\ket{{}^2P_{1/2}, F_e}\rightarrow \ket{{}^2S_{1/2}, F_g}$
in Li-like $^{229g}$Th ions, $I_\textrm{P}'(t)=I_{\textrm{P}0}'\exp{[-(4\ln{2}) (t-\tau_\textrm{P}')^2/t_\textrm{p}'^2]}$ is the intensity of the pump-laser pulse, which is assumed to have a Gaussian shape, with $\tau_\textrm{P}'$ denoting the peak position, and $E'=[2I_\textrm{P}'(t)/\epsilon_0 c]^{1/2}$ is the time-dependent amplitude of the electric field in the ion frame with $\epsilon_0$ being the vacuum permittivity.
The reduced transition dipole moment of the transition driven by the Stokes laser, considering the NHM, can be estimated as $d_r$ of the transition between $\ket{2}$ and $\ket{{}^2S_{1/2},5/2^+,F=2,M_F=1}$ multiplied by the NHM coefficient; see Eq.\,\eqref{Eq:red_dip_NHM}.
Consequently, the Rabi frequency corresponding to the Stokes laser can be derived as
\begin{equation} \label{Eq:E1_Rabi_Stokes}
\begin{split}
\Omega_\textrm{S}= |b_g|\cdot|\braket{F_gM_{F_g}L\lambda|F_eM_{F_e}}|\left(
\frac{6\pi c^2I_\textrm{S}'\Gamma_{F_e\rightarrow F_g}}{\hbar^2\omega_\textrm{P}'^3}
\right)^{1/2}\,,
\end{split}
\end{equation}
where $\braket{F_gM_{F_g}L\lambda|F_eM_{F_e}}=\braket{2111|32}$, $\Gamma_{F_e\rightarrow F_g}=6.2\times10^{-6}\,$eV for Li-like $^{229g}$Th ions,
$I_\textrm{S}'(t)=I_{\textrm{S}0}'\exp{[-(4\ln{2}) (t-\tau_\textrm{S}')^2/t_\textrm{p}'^2]}$ with $\tau_\textrm{S}'=\tau_\textrm{P}'-\Delta\tau'$, where \mbox{$\Delta\tau'\geq0$} is the delay of the pump pulse relative to the Stokes pulse.
The excitation scheme in Fig.\,\ref{Fig:STIRAP_Li-likeTh} could be implemented at the SPS using relativistic Li-like thorium ions with $\gamma=100$ ($\Delta\gamma/\gamma=10^{-4}$). Other parameters of the ion bunches remain those listed in Table\,\ref{tab:ion-beam_facilities}.
Once more we consider $S_l=S_b$ and $\Gamma_{\textrm{P}}'=\Gamma_D$ (see Sec.\,\ref{subsubsec:res_bare}), which leads to the laser pulse duration $t_\textrm{p}'=67\,$fs.
The durations of the two pulses with similar photon energies are approximately the same.
The laser photon energies are about $\hbar\omega_{\textrm{P}}=1.36\,$eV and $\hbar\omega_{\textrm{S}}=1.31\,$eV.
The population transfer in a Raman scheme can be accomplished by implementing two consecutive $\pi$-pulses (a pump pulse followed by a Stokes pulse in contrast to the STIRAP case) such that
$\int\Omega_\textrm{P}(t)dt=\pi$ and $\int\Omega_\textrm{S}(t)dt=\pi$, requiring $I_{\textrm{P}0}'=2.2\times10^{18}\,$W/m$^2$ and $I_{\textrm{S}0}'=2.8\times10^{22}\,$W/m$^2$.
The latter corresponds to a pulse energy of $E_\textrm{p}=7.8\,$J.
The lifetime of the intermediate state $\ket{2}$ is $\tau_2\approx5.9\times10^{-11}\,$s, much longer than $t_\textrm{p}'$.
STIRAP is generally a more robust approach to population transfer and does not require driving a precise $\pi$-pulse.
To show the population transfer, we solve the master equation in a simplified case neglecting damping as Eq.\,\eqref{Eq:DM_TLS_v2}. The relaxation due to the spontaneous radiative decay of the intermediate state is neglected during the STIRAP process because the lifetime of the intermediate state $\tau_{2}$ is much longer than the pulse duration and delay between the two pulses ($t_\textrm{p}'/\tau_{2}\approx10^{-3}\ll1$). We are now dealing with a three-level system, with $H_{\textrm{eff}}$ being \cite{Bergmann1998_STIRAP}
\begin{equation}\label{Eq:DM_lambda}
H_{\textrm{eff}} = -\hbar\begin{pmatrix}
0 &\frac{\Omega_\textrm{P}(t)}{2} &0\\
\frac{\Omega_\textrm{P}(t)}{2} & \Delta_\textrm{P} &\frac{\Omega_\textrm{S}(t)}{2}\\
0 &\frac{\Omega_\textrm{S}(t)}{2} &\delta
\end{pmatrix}\,,
\end{equation}
where $\Delta_\textrm{P(S)}=2\gamma\cdot\omega_\textrm{P(S)}-E_{12(32)}$ is the single photon detuning, $\delta = \Delta_\textrm{P} - \Delta_\textrm{S} $ is the so-called two-photon detuning, $E_{12(32)}$ is the energy difference between the states $\ket{1}(\ket{3})$ and $\ket{2}$.
We again assume searching for the resonance, thus $2\gamma^{\textrm{ave}}\cdot\hbar\omega_\textrm{P(S)}\approx E_{12(32)}$, and consider zero two-photon detuning, i.e., $\delta\approx0$.
Through adjusting the laser-pulse intensities and the delay between the two pulses, we identify the lowest $I_{\textrm{S}0}'$ needed for achieving nearly 100\% population transfer as $I_{\textrm{S}0}'\approx6\times10^{22}\,$W/m$^2$ with the delay $\Delta\tau'=0.07\,t_{\textrm{p}}'$; see Fig.\,\ref{Fig:STIRAP_Intensity_Delay}. The evolution of populations in different states under these optimal parameters is shown in Fig.\,\ref{Fig:STIRAP_Optimal}\,(a), where the transient population in the intermediate state means that the transfer process is not ideally adiabatic, which should be limited by the laser-pulse intensity.
For instance, if $I_{\textrm{S}0}'$ is increased by an order of magnitude together with a suitable delay between the two pulses, the maximal transient population of the intermediate state would be smaller than five percent, see Fig.\,\ref{Fig:STIRAP_Optimal}\,(b).
An empirical criterion based on the adiabaticity condition \cite{bergmann_perspective_2015,Vitanov_2017} gives the order of magnitude required for the peak Rabi frequency amplitude,
$\Omega_0^2 \geq 200\ln{2}/t_{\textrm{p}}'^2$,
for coherent pulses, leading to $\widetilde{I_\textrm{S0}}'\approx9\times10^{23}\,$W/m$^2$ and
$\widetilde{I_\textrm{P0}}'\approx7\times10^{19}\,$W/m$^2$.
\begin{figure}[!htpb]
\centering
\includegraphics[width=\linewidth]{STIRAP_Intensity_Delay.png}
\caption{Population transfer into the isomeric state $\ket{3}$ via the STIRAP set-up under different laser-pulse intensities and pulse delays $\Delta\tau'$ (given in the unit of $t_{\textrm{p}}'$ on the horizontal axis).
The laser intensities are given for $I_{\textrm{S}0}'$. $I_{\textrm{P}0}'$ used here is related to $I_{\textrm{S}0}'$ by letting the peak value of $\Omega_\textrm{P}$ equal to that of $\Omega_\textrm{S}$; see the text.
}
\label{Fig:STIRAP_Intensity_Delay}
\end{figure}
\begin{figure}[!htpb]
\centering
\includegraphics[width=\linewidth]{STIRAP_Optimal.png}
\caption{Population evolution under (a) $I_{\textrm{S}0}'=6\times10^{22}\,$W/m$^2$ with delay $\Delta\tau'=0.07\,t_{\textrm{p}}'$ (the minimum $I_{\textrm{S}0}'$ for 100\% population transfer) (b) $I_{\textrm{S}0}'=6\times10^{23}\,$W/m$^2$ with $\Delta\tau'=0.76\,t_{\textrm{p}}'$. $\rho_{ii}$ denotes the relative population of the state $\ket{i}$. The peak of the Stokes pulse intensity corresponds to $t'/t_\textrm{p}'=3$ in both cases.
}
\label{Fig:STIRAP_Optimal}
\end{figure}
We note that in Fig.\,\ref{Fig:STIRAP_Intensity_Delay}, $I_{\textrm{P}0}'$ is related to $I_{\textrm{S}0}'$ by letting the peak value of $\Omega_\textrm{P}$ equal to that of $\Omega_\textrm{S}$. The minimum $I_{\textrm{S}0}'\approx6\times10^{22}\,$W/m$^2$ corresponds to $I_{\textrm{P}0}'\approx4.8\times10^{18}\,$W/m$^2$.
A larger $I_{\textrm{P}0}'$, for instance, increased by a factor of five, would actually reduce the population transfer, and would require a larger $I_{\textrm{S}0}'$ instead of relieving the requirement on $I_{\textrm{S}0}'$ in order to achieve again the 100\% population transfer. Reducing $I_{\textrm{P}0}'$ has a similar effect.
The population transfer displayed in Figs.\,\ref{Fig:STIRAP_Li-likeTh} and \ref{Fig:STIRAP_Optimal} is for ions corresponding to $\Delta_\textrm{P}\approx\Delta_\textrm{S}\approx0$.
The ion-energy spread can be further considered using Eq.\,\eqref{Eq:ave_popu_tran}, resulting in a relatively small correction, since we have adjusted the pulse duration to match the laser-photon spectrum and the photon-absorption spectrum considering the ion-energy spread.
For instance, the 100\% population transfer for the scenario shown in
Fig.\,\ref{Fig:STIRAP_Optimal}\,(a) corresponds to an average population transfer of $\rho_0=85\%$ when considering the ion energy spread.
The laser-beam parameters and corresponding excitation rates are presented in Table\,\ref{tab:elec_STIRAP}.
\begin{table*}[htpb]
\centering
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}} lcc ccccc}
\hline
\hline
$\gamma$ &$\hbar\omega_\textrm{S}\,$(eV) &$\hbar\omega_\textrm{P}\,$(eV) &$I_\textrm{S0}\,$(W/m$^2$) &$I_\textrm{P0}\,$(W/m$^2$) &$t_\textrm{p}\,$(ps) &$\rho_0$ &rate\,(s$^{-1}$) \\
\hline \\[-0.2cm]
100 &1.31 &1.36 &$1.5\times10^{18}$ &$1.2\times10^{14}$ &13 &0.85 &$1.3\times10^{14}$\\
\hline
\hline
\end{tabular*}
\caption{
Isomer excitation via STIRAP at the SPS using an intermediate electronic excited state in Li-like $^{229}$Th ions. The laser-beam parameters, including the photon energy $\hbar\omega_\textrm{P/S}$, peak intensity $I_{\textrm{S/P0}}$ and pulse duration $t_\textrm{p}$, are given in the lab frame,
related to the ion-frame parameters according to Eqs.\,\eqref{Eq:para_frame}. $\rho_0$ is the isomer excitation fraction considering the ion-energy spread.
The excitation scheme and the population transfer are displayed in Figs.\,\ref{Fig:STIRAP_Li-likeTh} and \ref{Fig:STIRAP_Optimal}\,(a), respectively. Note that for real STIRAP with almost no transient population of the intermediate state, one order of magnitude higher intensity would be needed; see Fig.\,\ref{Fig:STIRAP_Optimal}\,(b).
The average isomer excitation rate in the last column is estimated as $\rho_0 N_{\textrm{ion}} N_b f_b$.}
\label{tab:elec_STIRAP}
\end{table*}
Due to the NHM, one can also excite the transitions with energies of about 279\,eV, exciting both the nuclear and electronic states (see Fig.\,\ref{Fig:Li-like_Th}).
Therefore, in the STIRAP scheme shown in Fig.\,\ref{Fig:STIRAP_Li-likeTh}, we can also use the state $\overline{\ket{{}^2P_{1/2},3/2^+,F=2}}$ as the intermediate state.
If one does not start with fully polarized ions, population in $\overline{\ket{{}^2S_{1/2},5/2^+,F=2}}$ level can be transferred to the isomeric state, $\overline{\ket{{}^2S_{1/2},3/2^+,F=2}}$, via STIRAP with any of the four HFS levels of ${}^2P_{1/2}$
being the intermediate state.
\subsubsection{
Isomer excitation via the 279 eV level using a single laser
}
\label{subsubsec:exc_279eV}
Besides STIRAP, one can also excite the nuclear isomeric state using only a single laser to drive the 279\,eV transition (see Fig.\,\ref{Fig:Li-like_Th}), which involves simultaneous excitation of both the nuclear and electronic states, enabled by the NHM effect.
The required laser-pulse intensity for implementing a $\pi-$pulse (between two HFS levels) can be derived using Eq.\,\eqref{Eq:Intensity_pi} and the radiative widths listed in Table\,\ref{tab:rad_width_NHM}.
After driving the transition, thorium ions will first quickly decay ($T_{\xi_eJ_e}=5.9\times10^{-11}\,$s in the ion frame) from the electronic excited state to the electronic ground state but remain in the nuclear isomeric state which has a much longer lifetime (about 1s in the ion frame).
Detecting the spontaneously emitted photons from the 271\,eV electronic transition after exciting the 279\,eV transition would unambiguously demonstrate isomer excitation. This would also serve as an experimental demonstration of the expected NHM effect in $^{229}$Th ions.
We note that, after excitation, the branching ratio of direct decay into the state with both the electron and nucleus in the ground state is negligible, on the order of $10^{-5}$; see Table\,\ref{tab:rad_width_NHM}.
While an excitation scheme using a single laser beam is, in principle, simpler to implement, STIRAP may still be beneficial for certain experiments. Indeed, the method is tolerant to single-photon detuning and provides a robust and complete transfer to the isomeric state. Additionally, since STIRAP is a coherent process, all ions will experience the same photon recoil during the excitation process, while spontaneous-emission step leads to a random ``kick.'' The latter may become important when deep laser cooling is implemented in stored highly relativistic HCI \cite{Winters2015LasCoolingSIS100,Eidam2018LaserCooling}.
\subsection{Excitation via the 29\,keV level}
\label{Subsec:exc_29keV}
While direct resonant excitation of the isomeric state can, in principle, be implemented at many accelerators,
laser ($\sim$eV) excitation via the 29\,keV level requires a relatively large Lorentz factor, which can be achieved at a smaller number of accelerators, e.g., the LHC. However, in contrast to the large uncertainty in the energy of the isomeric state, the energy of the second nuclear excited state has been measured precisely as
$E_{\textrm{2nd}}=29,189.93(7)$\,eV \cite{Masuda2019ThXray} with a relative uncertainty about $2.4\times10^{-6}$, smaller than $\Delta\gamma/\gamma$ typically achievable at storage rings.
Therefore, the resonance can be found more easily compared to the direct excitation scenario.
The precise knowledge of the energy of the 29\,keV level would also allow one to determine the central value of $\gamma$, $\gamma^\textrm{ave}$, with a similar relative precision via detection of secondary photons from radiative decay of this state; see Sec.\,\ref{Subsec:Fur_exc}, thereby removing one of the major systematic uncertainties in storage-ring based laser spectroscopy experiments\,\cite{Ullmann-2017,Lochmann-2014}.
Some other key properties of this state are shown in Fig.\,\ref{fig:229Th}.
\subsubsection{Incoherent population transfer}
\label{subsubsec:inco_exc_29keV}
X-ray pumping of the isomer state via the 29\,keV level has been demonstrated in Ref.\,\cite{Masuda2019ThXray}, where $^{229}$Th nuclei in a solid target were excited using 29\,keV synchrotron radiation.
Here, we discuss laser excitation of the isomeric state using relativistic $^{229}$Th ions.
$^{229}$Th ions with
$\gamma\approx2950$
($\Delta\gamma/\gamma\approx10^{-4}$) can be obtained
at the LHC; see Table\,\ref{tab:ion-beam_facilities}.
Thereby a laser with photon energy
$\hbar\omega_{12}\approx4.95\,$eV
can be used for resonant excitation of this transition with nearly head-on collisions between laser photons and the ion bunch.
Suppose that we employ a pulsed laser with an energy of 10\,$\mu$J per pulse and a repetition rate of $f_l=N_b f_b\approx13.8\,$MHz ($N_b=1232$) synchronized to the ion bunch, and a cavity further enhancing the laser power by a factor of 10$^5$ (leading to $E_{\textrm{p}}=1\,$J).
We consider again $S_l=S_b$ and $\Gamma_{\textrm{P}}'=\Gamma_D$ (see Sec.\,\ref{subsubsec:res_bare}),
which leads to a pulse duration $t_\textrm{p}\approx3.7\,$ps and peak intensity $I_0\approx3.2\times10^{20}\,$W/m$^2$.
When exciting the $M1$ transition from the nuclear ground state to the 29\,keV state, the system can be approximated as a two-level system since the radiative decay of the excited state is negligible due to its long lifetime compared to the pulse duration.
The population evolution induced by individual laser pulses can be derived by solving Eq.\,\eqref{Eq:DM_TLS_v2}.
The Rabi frequency of the transition can be derived from Eq.\,\eqref{Eq:Rabi_red} with $\Gamma_\textrm{rad}$ replaced by the partial width $\Gamma_{\gamma}^\textrm{cr}\approx 1.70\,$neV in this case.
After the ion bunch is excited by a laser pulse, it circulates in the storage ring before interacting with the next laser pulse. One needs to consider the spontaneous radiative decay of the 29\,keV level during this period, which could end up in both the nuclear ground and isomeric states with the relative radiative branching ratios being about 9\% and 91\%, respectively.
Only radiative branching ratios are considered here because IC channels are energetically forbidden in highly charged ${}^{229}$Th ions with no more than six bound electrons (only $2p_{3/2}$ or higher orbitals have ionization energies smaller than 29\,keV \cite{NIST_ASD}).
\begin{table}[htpb]
\centering
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}} lcccc c}
\hline
\hline
$\gamma$ &$t_\textrm{p}\,$(ps) &$I_0\,$(W/m$^2$) &$\rho_{ee}(\Delta=0)$ &$\rho_0$ &$t_\textrm{ex}\,$(s) \\
\hline \\[-0.2cm]
2950 &3.7 &$3.2\times10^{20}$ &$6.0\times10^{-4}$ &$4.2\times10^{-4}$ &0.56\\
\hline
\hline
\end{tabular*}
\caption{
Repeated pumping of $^{229m}$Th isomers via the 29\,keV level at the LHC. $\gamma$, $t_\textrm{p}$ and $I_0$ are some of the considered parameters of the ion bunch and laser beam. $\rho_{ee}(\Delta=0)$ and $\rho_0$ are the population-transfer fraction of the ion group resonant with the laser pulse and the transfer fraction averaged over all ions, respectively, in the first round.
$t_\textrm{ex}$ is the time needed to achieve $99\%$ of the maximal population transfer, $\rho_{\textrm{max}}\approx1$, corresponding to $\approx10^8$ ions per bunch. If one considers implementing a $\pi-$pulse, the required intensity is $I_{0\pi}\approx2100 I_0$.
}
\label{tab:repeat_pump_lhc}
\end{table}
\begin{figure}[!htpb]\centering
\includegraphics[width=0.95\linewidth]{Rep_pump_LHC.png}
\caption{
Repeated pumping at the LHC. The blue solid curve is derived from the theoretical calculation described in the text.
The thick orange dashed curve is the prediction using Eq.\,\eqref{Eq:exp_inc_lhc}, which fits well with the blue curve.
Nearly full population transfer is reached after $n_\textrm{col}\approx6300$ collisions ($t_\textrm{ex}=n_\textrm{col}/f_b\approx0.56\,$s).
}
\label{fig:Rep_pump_LHC}
\end{figure}
The parameters and results of repeated isomer excitation via the 29\,keV level at LHC are presented in Table\,\ref{tab:repeat_pump_lhc} and Fig.\,\ref{fig:Rep_pump_LHC}.
We note that with $\rho_0\ll\rho_{\textrm{max}}$, the increase of $\rho_{\textrm{ex}}$ can be approximated as
\begin{equation} \label{Eq:exp_inc_lhc}
\rho_{\textrm{ex}} = \rho_{\textrm{max}}\cdot\left[1-\exp{\left(-\frac{\rho_0/f_\textrm{dec}}{\rho_{\textrm{max}}}n_{\textrm{col}}\right)}\right]\,,
\end{equation}
where
\begin{equation}
f_\textrm{dec} = 1-\exp{\left(-\frac{1/f_b}{\gamma\cdot T_{1/2}^{\textrm{rad}}/\ln{2}}\, \right)} = 0.57
\end{equation}
is the decay fraction of ions in 29\,keV level during one circulation in the storage ring and $\gamma\cdot T_{1/2}^{\textrm{rad}}/\ln{2}$ is the radiative lifetime of the 29\,keV level in the lab frame. Here, we have $\rho_\textrm{max}\approx1$, meaning that almost all ions can be finally transferred to the isomeric state.
\subsubsection{STIRAP}
\label{subsubsec:STIRAP_29keV}
We consider now exciting $^{229m}$Th isomers via STIRAP using the three-level $\Lambda$ system comprising the nuclear ground and first two excited states. Population from the ground state $|1\rangle$ can be coherently transferred to the isomeric state $|3\rangle$ via STIRAP with almost no population of the 29\,keV state denoted as $|2\rangle$.
This scenario at the GF was proposed in Ref.\,\cite{Budker2021GF_nucl} and investigated theoretically in more detail in Ref.\,\cite{Kirschbaum2022}. The corresponding nuclear transition energies are $E_{12}=E_\textrm{2nd}\approx29.190\,$keV and $E_{32}=E_\textrm{2nd}-E_{\textrm{iso}}\approx29.182\,$keV.
Considering $\gamma\approx2950$ at the LHC, the required photon energies for the pump and Stokes pulses are $\hbar\omega_\textrm{P}\approx\hbar\omega_\textrm{S}\approx 4.95\,$eV. Following Sec.\,\ref{subsubsec:res_bare}, we set $S_l=S_b$ and $\Gamma_\textrm{p}'=\Gamma_D$, leading to FWHM duration $t_{\textrm{p}}\approx\SI{3.7}{\pico \second}$ for both Gaussian shaped pulses.
The coherent population-transfer dynamics can be evaluated using the density matrix formalism following a procedure similar to the one discussed in Sec.\,\ref{Subsec:exc_271eV}.
We solve Eq.\,\eqref{Eq:DM_TLS_v2} with the Hamiltonian given in Eq.\,\eqref{Eq:DM_lambda}. Thereby we neglect the radiative decay of the 29\,keV state, since its radiative half-life is much longer than the pulse duration and the delay between the Stokes and pump pulses.
The interaction with the laser fields is governed by the time-dependent Stokes and pump Rabi frequencies given by Eq.\,\eqref{Eq:Rabi_red}.
The partial radiative widths of the 29\,keV level related to the pump and Stokes transition are $\Gamma_{\gamma}^{\textrm{cr}}\approx 1.70\,$neV and $\Gamma_{\gamma}^{\textrm{in}}\approx 16.6\,$neV, respectively (see the beginning of Sec.\,\ref{Sec:iso_exc}).
The empirical adiabaticity criterion \cite{bergmann_perspective_2015,Vitanov_2017} mentioned in Sec.\,\ref{subsubsec:STIRAP_Li-like} provides the peak intensities required for the pump and Stokes pulses for robust STIRAP,
$\widetilde{I_\textrm{P0}'}\approx 7.4\times10^{32}$\,W/m$^2$
and $\widetilde{I_\textrm{S0}'}\approx 7.6\times10^{31}$\,W/m$^2$, corresponding to
$\widetilde{I_\textrm{P0}}\approx 2.1\times10^{25}$\,W/m$^2$ (pulse energy $E_\textrm{p}\approx6.6\times10^4\,$J)
and $\widetilde{I_\textrm{S0}}\approx 2.2\times10^{24}$\,W/m$^2$
in the lab frame. These intensities are orders of magnitude higher than the $I_0$ used for repeated incoherent pumping in Sec.\,\ref{subsubsec:inco_exc_29keV}. However, the high laser-pulse repetition rate required for accumulating excitation via repeated pumping becomes unnecessary here.
In Fig.\,\ref{Fig:STIRAP_lhc}(a) we present our numerical results for the coherent population transfer to the isomeric state after a Stokes and pump pulse sequence as a function of time delay between pulses.
Nearly full population transfer can be reached for a large range of delay times $\Delta\tau$ using the peak intensities given above.
For lower intensities, nearly full population transfer can still be achieved but in a narrower range of $\Delta\tau$. Figure\,\ref{Fig:STIRAP_lhc}\,(a) presents the population transfer to state $\ket{3}$ under different scaled pump and Stokes intensities using scaling factors 0.05, 0.1, 0.3, 1 and 3, and different delays between the two pulses.
Figure\,\ref{Fig:STIRAP_lhc}\,(b) shows an example where full population transfer into state $\ket{3}$ with almost no transient population of state $\ket{2}$ is achieved for peak intensity values $0.6\cdot\widetilde{I_\textrm{P/S,0}'}$ with $\Delta\tau'=0.78\,t_{\textrm{p}}'$.
Similar to the discussion in Sec.\,\ref{subsubsec:STIRAP_Li-like},
the population transfer shown in Fig.\,\ref{Fig:STIRAP_lhc} is calculated for ions which satisfy the condition $\Delta_\textrm{P}\approx\Delta_\textrm{S}\approx0$. Ion-energy spread results in only a relatively small correction.
\begin{figure}[!htpb]
\centering
\includegraphics[width=\linewidth]{STIRAP_LHC.png}
\caption{
(a) Population transfer into the isomeric state $\ket{3}$ via the STIRAP set-up under different laser-pulse intensities and pulse delays.
(b)
Population evolution under $0.6\cdot \widetilde{I_{\textrm{P/S,0}}'}$ and $\Delta\tau'=0.78\,t_{\textrm{p}}'$. The peak intensity of the Stokes pulse occurs at $t'/t_\textrm{p}'=5$.
}
\label{Fig:STIRAP_lhc}
\end{figure}
A more detailed discussion of the available ion bunch and laser beam parameters for this STIRAP scenario is presented in Ref.\,\cite{Kirschbaum2022}. Given that the 29\,keV state $|2\rangle$ has a well-known energy and a favorable branching ratio to the isomeric state, repeated incoherent pumping could be more favorable for near-term experimental implementation. Its simplicity (lower laser intensity and only one laser frequency) should be weighed against more stringent requirements for laser-frequency tuning as opposed to robust STIRAP tolerant to laser detuning.
\subsection{Isomer excitation with x-ray lasers}
Some schemes discussed above (those in Sec\,\ref{Subsec:exc_271eV} and \ref{Subsec:exc_29keV}) could also be realized using x-ray pulses combined with moderately accelerated or even static ${}^{229}$Th ions.
The photon energy needed for the schemes discussed in Sec\,\ref{Subsec:exc_271eV} is about $270\,$eV, well within reach of many x-ray laser facilities.
Excitation via STIRAP can be done by use of a two-color x-ray free electron laser (FEL). These devices are capable of generating intense x-ray pulses by guiding relativistic bunches of electrons through a periodic array of dipole magnets known as an undulator~\cite{mcneilXrayFreeelectronLasers2010}. State-of-the-art FELs allow for the generation of two pulses of different energies with tunable time delays~\cite{pratWidelyTunableTwocolor2022}, which potentially can be used to perform STIRAP operations. Such experiments have been proposed for purely nuclear transitions~\cite{liaoNuclearCoherentPopulation2011, mansourzadeh-ashkaniSuperpositionNuclearStates2021} and for electronic ones~\cite{piconStimulatedRamanAdiabatic2015}.
One can also use just a single x-ray laser to excite the 279\,eV transition in order to transfer population into the isomeric state, following the scheme discussed in Sec.\,\ref{subsubsec:exc_279eV}.
The highly charged ions could be produced and trapped using an electron beam ion trap (EBIT) installed as an end station of an FEL. The EBIT operates by means of a magnetically focused electron beam which ionizes neutral particles which are introduced as, for example, a tenuous atomic beam~\cite{levineElectronBeamIon1988}. Ions remain trapped in the negative space-charge potential of the electron beam where they are sequentially ionized further until the ionization energy is beyond the electron beam energy. Production of HCI up to H-like uranium was demonstrated\,\cite{marrsProductionTrappingHydrogenlike1994}; production of Li-like thorium is possible with transportable devices such as the FLASH-EBIT, which has already been used for one-photon spectroscopy at FELs\,\cite{epp2007rontgen, bernittUnexpectedlyLowOscillator2012}.
We stress, however, that due to the need to produce $^{229}$Th in a nuclear reaction,
the number of ions will inevitably be tiny and also unavoidable contaminants needs to be considered.
Subsequent to the STIRAP or single-laser excitation, the state of the thorium nucleus can be determined in various ways. For example, this can be done by extracting the ions and neutralizing them using an electrode or ultrathin carbon foil to open the internal conversion (IC) channel which results in detectable IC electrons in case the isomer would have been excited~\cite{Bilous2020EB_35+}.
We note that isomer excitation via the 29\,keV level using x-ray lasers is discussed in Ref.\,\cite{Kirschbaum2022}.
\section{Detection of the thorium isomer}
\label{Sec:Detect}
Besides excitation of $^{229m}$Th isomers discussed above, accelerators with relativistic HCI also offer unique opportunities for detection of those isomers and precision measurements of the isomeric-state energy.
We now present three different schemes for isomer detection, relying on (i)
the radiative decay of $^{229m}$Th, (ii) further excitation of the isomeric state to the 29\,keV level followed by radiative decay, and (iii) laser spectroscopy of electronic transitions; see Figs.\,\ref{fig:Th_probe_nucl} and \ref{fig:Th_probe_elec}. In contrast to the first two approaches, the third one requires using PSI and electronic transitions instead of only nuclear transitions.
Besides the demonstration of isomer excitation, we also discuss the precision that can be achieved in each scheme in determining the energy of the isomeric state.
\begin{figure}[!htpb]\centering
\includegraphics[width=\linewidth]{Th_probe_nucl.png}
\caption{Detection of $^{229m}$Th using nuclear transitions. (a) radiative decay of the isomeric state. (b) further excitation of the isomeric state to the 29\,keV level.
The blue wavy arrow denotes incident photons exciting the isomeric state (could be pursued at the LHC). The red ones represent photons emitted from radiative decay of the excited state.
}
\label{fig:Th_probe_nucl}
\end{figure}
\subsection{Radiative decay of the isomeric state}
\label{Subsec:Rad_decay}
Until recently \cite{ISOLDE-IS658_2022}, there was no successful and unambiguous detection of photons from the radiative decay of $^{229m}$Th; see, for example, experiments in Refs.\,\cite{Jeet2015DirectExc,Yamaguchi2015DirectExc,Stellmer2018OpticalExc}. This could be due to unexpected non-radiative decay processes in the crystal/solid-state environment of these experiments \cite{Peik2021}.
In contrast, in HCI
radiative decay should be the predominant decay channel of $^{229m}$Th.
Photons from radiative decay of the isomeric state
will not be emitted at a fixed location in the storage ring but rather at arbitrary locations along the ring perimeter during the circulation of the ion bunch. This is due to the long radiative half-life of $^{229m}$Th, even when there is lifetime reduction due to the NHM effect.
The average photon-detection rate can be estimated as
\begin{equation}
\label{Eq:det_rad}
p_\text{d}=N_{\text{iso}}N_{\text{b}}f_{\text{b}}\left[\exp{\left(-\frac{t_1}{\tau}\right)}-\exp{\left(-\frac{t_2}{\tau}\right)}\right]\approx N_{\text{iso}}N_{\text{b}}f_{\text{b}}\frac{\Delta t}{\tau}\,,
\end{equation}
where $N_{\text{iso}}$ is the number of isomers per bunch, $\tau=\gamma T_{1/2}^\textrm{rad}/\ln{2}$ is the radiative lifetime of the isomeric state in the lab frame, and $\Delta t=t_2-t_1$ is the time duration related to the photon collection during each ion-bunch circulation with $t_1$ and $t_2$ being the starting and ending times, respectively.
We assume that during each circulation, photons emitted during the ions travelling in, say, $1\,$m can be detected, i.e., $c\Delta t=1\,\textrm{m}$. This photon-collection length refers to the lab frame and can be made significantly longer with the help of dedicated mirror systems, see, for example, Ref.\,\cite{Sanchez-2017}.
Here, the time delay $t_1$ ($t_1\ll\tau$) is introduced to help filter out photons from the pulsed excitation laser and to collect photons only from the radiative decay of $^{229m}$Th.
Since the lifetime of $^{229m}$Th is long compared to the laser pulse duration, background associated with the excitation laser could be entirely eliminated.
We note that in the ion frame, photons from spontaneous emission (referred to as secondary photons below) are emitted in a broad angular distribution.
As seen in lab frame, however, due to the relativistic effect, the photons are mainly emitted within a small angle ($\approx1/\gamma$) relative to the ion-propagation direction.
Photons re-emitted along the
ions-propagation direction are boosted in energy by another factor of $\approx2\gamma$, so the maximal energy of the secondary photons is about $2\gamma\cdot\hbar\omega'$; see, for instance, Refs.\,\cite{Budker2020_AdP_GF,Budker2021GF_nucl} for more details.
For an estimate, we assume that almost all ions in a bunch are excited to the isomeric state, i.e., $N_{\textrm{iso}}\approx N_{\textrm{ion}}$ is the number of isomers per ion bunch.
We obtain $p_\textrm{d}\approx3.9\,\textrm{s}^{-1}$, $2.0$\,s$^{-1}$ and 0.21\,s$^{-1}$ at the HESR ($\gamma=6$), SPS ($\gamma=35$) and LHC ($\gamma=2950$), respectively, for bare nuclei ($T_{1/2}^\textrm{rad}\approx5.2\times10^3\,$s).
The low counting rates are mainly due to the large $\tau$, and
can be to some degree improved by using ion bunches with a smaller $\gamma$, leading to shorter lifetime in the lab frame.
$\gamma$ could be reduced to around 2, 10 and 200 at the HESR, SPS and LHC, respectively. Then, we get $p_\textrm{d}\sim10\,$s$^{-1}$ at these accelerators. As mentioned above, higher counting rates could also be obtained through increasing the photon-collection length.
For H-like and Li-like thorium ions with the NHM, $\tau$ could be reduced by a factor of about
$10^5$ and $10^4$, respectively, compared to bare nuclei, leading to increase of the photon detection rate by the same factor.
This can also serve as an experimental evidence of the NHM effect, which has not been measured in atoms or atomic ions as yet \cite{Shabaev2021NHM}.
Detection of secondary photons from the radiative decay of the isomeric state directly demonstrates excitation of $^{229m}$Th,
which means the resonance condition for excitation, $E_\textrm{iso}=2\gamma\hbar\omega$,
is fulfilled (($\hbar\omega$ is the laser photon energy)).
Therefore, the isomeric-state energy can be obtained, with a relative precision of $\lesssim\Delta\gamma/\gamma$. This precision corresponds to the uncertainty in $\gamma$ and depends on how well the laser-photon spectrum
matches the photon-absorption spectrum considering the ion-energy spread.
After improving the match
(see Sec.\,\ref{Sec:search_res}), $E_\textrm{iso}$ could be determined more precisely (with a relative precision similar to that of $\gamma^{\textrm{ave}}$) through measuring $\gamma^{\textrm{ave}}$.
$\gamma^{\textrm{ave}}$ can be determined via measuring the energy of secondary photons in the lab frame, $E_{121}=2\gamma E_\textrm{iso}$.
From $E_{121}=2\gamma E_\textrm{iso}=4\gamma^2\hbar\omega$,
we derive $\gamma=(E_\textrm{121}/4\hbar\omega)^{1/2}$ and $E_\textrm{iso}=2\gamma\hbar\omega=
(E_{121}\hbar\omega)^{1/2}$; see also Refs.\,\cite{Wojtsekhowski2021local,Budker2021GF_nucl}. $\gamma^{\textrm{ave}}$ can be measured with a relative precision of $1/2(p_\textrm{d} t_\textrm{det})^{1/2}$ with $t_\textrm{det}$ being the measurement time. In the case of bare nuclei, because of the low detection rate, an 1-hour measurement produces $\gamma^{\textrm{ave}}$ with a relative precision of about $3\times10^{-3}$, which gives no improvement compared to $\Delta\gamma/\gamma$.
For H-like thorium ions, an 1-hour measurement produces $\gamma^{\textrm{ave}}$ with a relative precision of $\lesssim10^{-5}$.
\subsection{Further excitation of the isomeric state}
\label{Subsec:Fur_exc}
Since the long radiative half-life of the isomeric state reduces detection rates of secondary photons, it could be more efficient to further excite $^{229m}$Th from the isomeric state to the second nuclear excited state at 29\,keV which has a significantly shorter half-life, $T_{1/2}^\textrm{rad}\approx25\,$ns, although a large $\gamma$ is needed for such excitation.
For the isomer-production process discussed in Sec.\,\ref{subsubsec:inco_exc_29keV},
when exciting the transition from the ground state to the 29\,keV state, there are initially (before a significant fraction of the nuclei are optically pumped)
$N_0\approx N_\textrm{ion}\cdot\rho_0/\textrm{BR}_{\gamma}^{\textrm{in}}/f_\textrm{dec}\approx 8.1\times10^4$
excitations into the 29\,keV state induced by a single laser pulse.
The corresponding average photon-detection rate is
$p_\textrm{d} \approx N_0N_bf_b\Delta t/\tau\approx3.5\times10^7\,$s$^{-1}$
[see Eq.\,\eqref{Eq:det_rad}]. We again assumed a 1-m collection length and a slight delay in detection $t_1\ll\tau$ (here $\tau$ is for the 29\,keV state).
After almost all ions are excited to the isomeric state, which can be known from the decrease of the detection rate of photons from the radiative decay of the 29\,keV level, we can decrease $\gamma$ or the laser-photon energy
to change the ion-frame laser-photon energy by about $E_\textrm{iso}$,
and observe increased photon collection rates due to excitation of the isomeric state.
Excitations of the isomeric state and the ground state can be distinguished since the transition widths are both around 29.2\,keV$\times\Delta\gamma/\gamma\approx2.9\,$eV in the ion frame, while the energy difference between the isomeric and ground state is about 8.2\,eV.
We note that this probing method could also be applied to isomers produced at the PS or SPS, since ion bunches at those facilities can be transferred to the LHC and further accelerated for excitation of the 29\,keV transition.
Besides probing isomer production, a precise determination of the isomeric-state energy can be realized in this process via detecting secondary photons from radiative decay of the 29\,keV level, utilizing the recent precise measurement of the energy of the 29\,keV state \cite{Masuda2019ThXray}.
Following a procedure similar to the one in Sec.\,\ref{subsubsec:inco_exc_29keV},
we obtain excitation events from the isomeric state to the 29\,keV state as
$N_0\approx8.0\times10^5$
induced by one laser pulse for $N_\textrm{iso}\approx10^8$ isomers per bunch.
The average photon-collection rate is
$p_\textrm{d}\approx3.5\times10^8\,$s$^{-1}$.
Since nearly 9\% of the ions excited to the 29\,keV level will decay into the ground state, the number of isomers per ion bunch as well as the photon-collection rate will decrease within a second. To keep the high photon-collection rate, a second laser can be employed to excite ions from the nuclear ground state to the isomeric state via the 29\,keV level.
We assume two lasers with photon energy $\hbar\omega_{12}$ and $\hbar\omega_{32}$ in the lab frame are
tuned to be resonant with the transitions from the nuclear ground and isomeric state to the 29\,keV level, respectively.
Here, as in Sec\,\ref{subsubsec:STIRAP_29keV}, states $\ket{1}$, $\ket{2}$ and $\ket{3}$ represent the nuclear ground state, 29\,keV state and isomeric state, respectively.
$E_\textrm{iso}$ can be obtained from
\begin{equation}
\label{Eq:mea_29keV}
E_\textrm{iso}=2\gamma (\hbar\omega_{12}-\hbar\omega_{32})=\frac{E_{\textrm{2nd}}}{\hbar\omega_{12}}(\hbar\omega_{12}-\hbar\omega_{32})\,.
\end{equation}
Here, $2\gamma \hbar\omega_{12}=E_{\textrm{2nd}}$ and $2\gamma \hbar\omega_{32}=E_{\textrm{2nd}}-E_\textrm{iso}$ are the resonance conditions.
After matching the laser-photon spectrum in the ion frame and the photon-absorption spectrum considering the ion-energy spread (see Sec.\,\ref{Sec:search_res}),
$\gamma=E_{\textrm{2nd}}/2\hbar\omega_{12}$ can be determined with a relative precision
similar to that of $E_{\textrm{2nd}}$, about $2.4\times10^{-6}$ as measured in Ref.\,\cite{Masuda2019ThXray}. $E_\textrm{iso}=2\gamma (\hbar\omega_{12}-\hbar\omega_{32})$ could be determined with almost the same precision.
We have assumed that $\hbar\omega_{12(32)}^\textrm{ave}$, i.e., the central values of photon energies of the two lasers, can be known with a better precision.
We note that, this scheme, exploiting the recent measurement of $E_{\textrm{2nd}}$, does not require precise knowledge of energies of secondary photons, thus it might be easier to implement experimentally.
A higher precision of $E_{\textrm{iso}}$ could be derived via measuring energies of secondary photons. Secondary photons from the radiative decay of the 29\,keV level into the nuclear ground state and the isomeric state have maximal energies of
$E_{\textrm{321}}=2\gamma E_{\textrm{2nd}}$ and $E_{\textrm{323}}=2\gamma (E_{\textrm{2nd}}-E_{\textrm{iso}})$, respectively, in the lab frame.
Detecting those high-energy photons could be realized using the photon detector described in Ref.\,\cite{Wojtsekhowski2021local}.
With a measurement time of $t_\textrm{det}=1\,$h,
$E_{323}$ can be determined with a relative uncertainty
$\Delta E_{323}/E_{323}\approx
1/(p_\textrm{d}\cdot t_\textrm{det}\cdot\textrm{BR}_\gamma^{\textrm{in}})^{1/2}\approx9.4\times10^{-7}$.
$\gamma=(E_{323}/4\hbar\omega_{32})^{1/2}$
can be determined with a relative precision of $4.7\times10^{-7}$, allowing for measuring $E_\textrm{iso}$ with almost the same precision.
$E_{\textrm{2nd}}$ can also be measured with a similar precision in this process due to
$E_{\textrm{2nd}}= 2\gamma\hbar\omega_{12}$,
which would be better than the precision obtained in Ref.\,\cite{Masuda2019ThXray}.
\subsection{Nuclear spectroscopy via electronic transitions}
\label{Subsec:elec_tran}
Population transfer into the isomeric state can also be probed via laser spectroscopy of electronic transitions,
see, for example, Ref.\,\cite{Peik2003_spectroscopy}.
The electronic-transition spectra would change mainly
due to (1) the difference in the mean-square charge radii of thorium nuclei in the ground state and the isomeric state; and (2) the change in the hyperfine structure.
We address in the following Li-like $^{229}$Th ions and the electronic transition between the fine structure components $1s^22s_{1/2}$ ($^2S_{1/2}$) and $1s^22p_{1/2}$ ($^2P_{1/2}$); see Fig.\,\ref{fig:Th_probe_elec}.
\begin{figure*}[!htpb]\centering
\includegraphics[width=0.95\textwidth]{Th_probe_elec.png}
\caption{Detection of $^{229m}$Th via laser spectroscopy on the $^2S_{1/2}\rightarrow{}^2P_{1/2}$ electronic transition ($\approx271\,$eV) of Li-like thorium ions in
(a) the nuclear ground state and (b) the isomeric state.
The $^2S_{1/2}$ and $^2P_{1/2}$ energy levels without considering the HFS are denoted by dashed lines on the left.
The HFS levels are shown as solid lines on the right and corresponding energy level splittings are indicated. All energies are given in eV.
See the text for details of the transition energies, $E^{229g}$ and $E^{229m}$, and the hyperfine splittings.
}
\label{fig:Th_probe_elec}
\end{figure*}
\subsubsection{Isomer shift and hyperfine splitting}
\label{subsubsec:En_Li-like}
The energy of this transition in Li-like $^{232}$Th ions (in the nuclear ground state) has been calculated as
$E^{232g} = 270.913(98)\,$eV
\cite{Yerokhin2018energy}.
The isotope shift of a transition can be obtained from
\begin{equation}
\delta E^{A,A'}=2\pi\hbar\Delta \widetilde{K}^{RMS}\frac{M'-M}{M'M}+2\pi\hbar F_{FS} \delta\braket{r^2}^{A,A'}
\end{equation}
in the first-order perturbation-theory approximation \cite{Li2012IsotopeShift}.
Here, $\Delta \widetilde{K}^{RMS}$ is the relativistic mass-shift parameter, $M'$ and $M$ are the nuclear masses, $F_{FS}$ is the field-shift factor, and $\delta\braket{r^2}^{A,A'}=\braket{r^2}^{A}-\braket{r^2}^{A'}$ is the difference in the mean-square charge radii.
For Li-like Th$^{87+}$, we can use $\Delta \widetilde{K}^{RMS}=-7.832\times10^5$\,GHz$\cdot$u (u is the unified atomic mass unit) and $F_{FS}=-1.518\times10^5\,$GHz/fm$^2$ \cite{Li2012IsotopeShift}. The uncertainties of $\Delta \widetilde{K}^{RMS}$ and $F_{FS}$, which are not presented here, have negligible contribution to the error of the isomer shift $\delta E^{229m,229g}$ and the error of transition energies between hyperfine components of $^2S_{1/2}$ and $^2P_{1/2}$ in $^{229m,229g}$Th. The former is dominated by the uncertainty in $\delta\braket{r^2}^{229m,229g}$, while the latter is dominated by the uncertainty in $E^{232g}$.
Considering $\delta\braket{r^2}^{232g,229g}= 0.334(1)\,$fm$^2$ \cite{Angeli2013ChargeRadii} and $\delta \braket{r^2}^{229m,229g} = 0.0105(13)\,$fm$^2$ \cite{Safronova2018ChargeRadii}, the energy of the transition $^2S_{1/2}\rightarrow {}^2P_{1/2}$ can be derived as $E^{229g}= 271.122(98)$\,eV and $E^{229m}=271.116(98)$\,eV for nuclei in the ground and isomeric state, respectively.
The isomeric shift is $\delta E^{229m,229g}=2\pi\hbar F_{FS} \delta\braket{r^2}^{229m,229g}=-0.0066(8)$\,eV.
Now we calculate the hyperfine splitting in Li-like $^{229}$Th ions. The $^2S_{1/2}$ and $^2P_{1/2}$ levels have two hyperfine components, $F=I-1/2,I+1/2$.
Since the electric quadrupole interaction is zero in the Li-like ion system with $J=1/2$, we only consider the magnetic dipole interaction between the electrons and the nucleus.
Therefore, the energy shifts of HFS have the form of $E_\textrm{hf}=A_\textrm{hf}\braket{\vec{I}\cdot\vec{J}}/\hbar^2=A_\textrm{hf}[F(F+1)-I(I+1)-J(J+1)]/2$; see Tables\,\ref{tab:hfs_Energy_gnd} and \ref{tab:hfs_Energy_iso} for $^{229g}$Th and $^{229m}$Th, respectively. The $A_\textrm{hf}$ coefficients listed in those tables are calculated through the multi-configuration Dirac Fock (MCDF) method using the GRASP software package \cite{GRANT1980207}. The energy shifts of the $F=2$ levels due to the NHM are negligible, about $2\times10^{-4}$\,eV in Li-like $^{229}$Th ions \cite{Shabaev2021NHM}.
\begin{table}[htpb]
\centering
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}} lcc}
\hline
\hline
HFS &$A_\textrm{hf}\,(2\pi\hbar\times10^6$\,MHz) &$E_\textrm{hf}$\,($10^{-2}\,$eV)\\
\hline \\[-0.2cm]
$^2S_{1/2}$, $F=2$ &9.66(19) &-6.99(14) \\
$^2S_{1/2}$, $F=3$ &9.66(19) &5.00(10) \\
$^2P_{1/2}$, $F=2$ &3.20(6) &-2.31(4) \\
$^2P_{1/2}$, $F=3$ &3.20(6) &1.65(3) \\
\hline
\hline
\end{tabular*}
\caption{The hyperfine splitting of $^2S_{1/2}$ and $^2P_{1/2}$ electronic levels of Li-like $^{229g}$Th ions.
}
\label{tab:hfs_Energy_gnd}
\end{table}
\begin{table}[htpb]
\centering
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}} lcc}
\hline
\hline
HFS &$A_\textrm{hf}\,(2\pi\hbar\times10^6\,$MHz) &$E_\textrm{hf}$\,($10^{-2}$\,eV)\\
\hline \\[-0.2cm]
$^2S_{1/2}$, $F=1$ &-16.6(19) &8.6(9) \\
$^2S_{1/2}$, $F=2$ &-16.6(19) &-5.1(5) \\
$^2P_{1/2}$, $F=1$ &-5.5(8) &2.8(4) \\
$^2P_{1/2}$, $F=2$ &-5.5(8) &-1.7(2) \\
\hline
\hline
\end{tabular*}
\caption{Similar to Table\,\ref{tab:hfs_Energy_gnd} but for $^{229m}$Th.}
\label{tab:hfs_Energy_iso}
\end{table}
\subsubsection{Photoabsorption spectra with unpolarized and polarized nuclei}
\label{subsubsection:spectra_pol}
The excitation rate of
the transition from a HFS level
$\ket{g}=\ket{^2S_{1/2},I,F_g}$
to
$\ket{e}=\ket{^2P_{1/2},I,F_e}$
is proportional to the
radiative width of the decay channel $\ket{e}\rightarrow\ket{g}$, given by \cite{Auzinsh2010LightAtom}
\begin{equation}
\label{Eq:nat_width}
\Gamma_{F_e\rightarrow F_g}=g_{F_g} \cdot g_{J_e}
\begin{Bmatrix}
J_e & F_e &I\\
F_g& J_g &L
\end{Bmatrix}^2 \Gamma_{\xi_eJ_e\rightarrow \xi_gJ_g},
\end{equation}
where $g_{F_g}=2F_g+1$ and $g_{J_e} = 2J_e+1$ are the degeneracy factors,
$L=1$ is the multiplarity of the $E1$ transition,
$\Gamma_{\xi_eJ_e\rightarrow\xi_gJ_g}=1.12(1)\times10^{-5}\,$eV is the radiative width of the $^2P_{1/2}$ state \cite{Theodosiou1991Li-Ion}.
We list the $g_{F_e}\cdot\Gamma_{F_e\rightarrow F_g}$ values directly relevant to the relative excitation rates, and the transition energy $E_{if}$ in Tables\,\ref{tab:hfs_tran_gnd} and \ref{tab:hfs_tran_iso} for $^{229}$Th ions in the nuclear ground and isomeric state, respectively.
\begin{table}[htpb]
\centering
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}} lcc}
\hline
\hline
Transition &$E_{if}$\,(eV) &$g_{F_e}\cdot\Gamma_{F_e\rightarrow F_g}$\\
$^2S_{1/2}\rightarrow {}^2P_{1/2}$& &($10^{-5}$\,eV)
\\
\hline \\[-0.2cm]
$F_g=3\rightarrow F_e=2$ &271.049(98) &4.34(4) \\
$F_g=3\rightarrow F_e=3$ &271.089(98) &3.47(4) \\
$F_g=2\rightarrow F_e=2$ &271.169(98) &1.24(1) \\
$F_g=2\rightarrow F_e=3$ &271.208(98) &4.34(4) \\
\hline
\hline
\end{tabular*}
\caption{The energies ($E_{if}$) and $g_{F_e}\cdot\Gamma_{F_e\rightarrow F_g}$ of transitions between hyperfine components in Li-like $^{229g}$Th ions.}
\label{tab:hfs_tran_gnd}
\end{table}
\begin{table}[htpb]
\centering
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}} lcc}
\hline
\hline
Transition &$E_{if}$\,(eV) &$g_{F_e}\cdot\Gamma_{F_e\rightarrow F_g}$\\
$^2S_{1/2}\rightarrow {}^2P_{1/2}$& &($10^{-5}$\,eV)
\\
\hline \\[-0.2cm]
$F_g=1\rightarrow F_e=2$ &271.012(98) &2.79(2) \\
$F_g=1\rightarrow F_e=1$ &271.057(98) &0.558(3) \\
$F_g=2\rightarrow F_e=2$ &271.149(98) &2.79(2) \\
$F_g=2\rightarrow F_e=1$ &271.194(98) &2.79(2) \\
\hline
\hline
\end{tabular*}
\caption{Similar to Table\,\ref{tab:hfs_tran_gnd} but for $^{229m}$Th.}
\label{tab:hfs_tran_iso}
\end{table}
With the ion-energy spread (assumed to have a Gaussian line shape) taken into account, the relative excitation rate as a function of photon energies in the ion frame is
\begin{equation}
\label{Eq:Exc_rate}
R_{\textrm{exc}}=
\frac{g_{F_e}}{\sum\limits_{F_g} g_{F_g}}\frac{\Gamma_{F_e\rightarrow F_g}}{\Gamma_{\xi_eJ_e\rightarrow\xi_gJ_g}}\frac{1}{\sqrt{2\pi}\sigma_E}\exp\left\{-\frac{(E- E_{if})^2}{2\sigma_E^2}\right\}\,,
\end{equation}
where $2\sqrt{2\ln{2}}\sigma_E\approx(\Delta\gamma/\gamma)E_{if}$ is the FWHM energy of the transition resulting from the ion-energy spread, and the relative population of hyperfine level $F_g$ is assumed to be
proportional to the level degeneracy.
We note that the ion-frame lifetime of the HFS level $F_g=3$ in Li-like $^{229g}$Th is about 27\,s. The lifetimes of HFS levels $F_g=1$ and $2$ in Li-like $^{229m}$Th are about 2.2\,s and 0.88\,s, respectively, considering the NHM \cite{Shabaev2021NHM}.
The photon-absorption spectra in the ion frame for $^{229g}$Th and $^{229m}$Th are displayed in Fig.\,\ref{fig:Probe_spec_Li-like}\,(a) for $\Delta\gamma/\gamma=10^{-4}$.
Through tuning photon energies and monitoring the variation of detection rates of secondary photons, excitation of the isomeric state can be probed. For instance, with photon energies tuned to $271.01\cdot(1\pm0.5\Delta\gamma/\gamma)$\,eV in the ion frame, there is a significant contrast in excitation rates of ions in the nuclear ground state and the isomeric state. Essentially, only ions in the isomeric state can be excited (from $F_g=1$ to $F_e=2$).
\begin{figure}[!htpb]\centering
\includegraphics[width=1.0\linewidth]{Probe_spec_Li-likeTh.png}
\caption{
Relative excitation rates
of (a) transitions from $\ket{^2S_{1/2},I,F_g}$ to $\ket{^2P_{1/2},I,F_e}$ in unpolarized Li-like $^{229}$Th ions; see Eq.\,\eqref{Eq:Exc_rate}, and (b) transitions from $\ket{^2S_{1/2},I, F_gM_{F_g}}$ to $\ket{{}^2P_{1/2},I, F_eM_{F_e}}$ in fully polarized Li-like $^{229}$Th ions; see Eq.\,\eqref{Eq:Exc_rate_pol}.
The horizontal axis denotes photon energies in the ion frame. Solid lines and dashed lines are for ions with nuclei in the ground and isomeric state, respectively.
}
\label{fig:Probe_spec_Li-like}
\end{figure}
Up to now, we have not discussed the effects of nuclear and electronic polarization.
In fact, it is possible to produce polarized PSI by optical pumping, see, for instance, Refs.\,\cite{Budker2020_AdP_GF,Bieron2022OpticalPump}.
They would be easier to probe using laser photons with suitable polarization.
Ions could be polarized in a single path through the interaction region with efficient optical pumping. If these ions are probed promptly, it may be unnecessary to preserve the polarization for a round trip in the storage ring.
Preserving the polarization of PSI in the ring is under investigation, see, for instance, Ref.\,\cite{Bondarevskaya2011_pol_theory}.
Suppose that we have produced fully polarized ions, where only the sublevel of $^2S_{1/2}$ with the maximum $M_{F_g}$ is populated. This could be realized via optical pumping using broadbrand $\sigma^+$ light to drive transitions between $^2S_{1/2}$ and $^2P_{1/2}$.
Then we can use $\sigma^-$ light for probing.
The transitions that will be excited in $^{229g}$Th are from $F_g,M_{F_g}=3,3$ to $F_e,M_{F_e}=2,2$ and $3,2$. For ions in the isomeric state, they are transitions from $F_g,M_{F_g}=2,2$ to $F_e,M_{F_e}=1,1$ and $2,1$.
Radiative widths of the corresponding decay channels from $F_e,M_{F_e}$ to $F_g,M_{F_g}$ are given by
\begin{align}\label{Eq:nat_width_sublevel}
\Gamma_{F_eM_{F_e}\rightarrow F_gM_{F_g}}=
&g_{F_e}\cdot g_{J_e}\begin{Bmatrix}
J_e & F_e &I\\
F_g & J_g &L
\end{Bmatrix}^2 \Gamma_{\xi_eJ_e\rightarrow \xi_gJ_g}\nonumber\\
& \cdot\left |\braket{F_eM_{F_e}L\lambda|F_gM_{F_g}}\right|^2
\,,
\end{align}
where we have $L=1$ for $E1$ transitions and $\lambda=M_{F_g}-M_{F_e}$.
The relative excitation rate displayed in Fig.\,\ref{fig:Probe_spec_Li-like}\,(b) is
\begin{equation}
\label{Eq:Exc_rate_pol}
R_{\textrm{exc}}= \frac{\Gamma_{F_eM_{F_e}\rightarrow F_gM_{F_g}}}{\Gamma_{\xi_eJ_e\rightarrow\xi_gJ_g}}\frac{1}{\sqrt{2\pi}\sigma_E}\exp\left\{-\frac{(E-E_{if})^2}{2\sigma_E^2}\right\}.
\end{equation}
We can see that the two spectra in the figure are well separated and therefore can be more easily distinguished than those in Fig.\,\ref{fig:Probe_spec_Li-like}\,(a).
Excitation of the 271\,eV transition can be implemented, for example, at the SPS or LHC. Generally, electronic transitions can be excited more efficiently compared to nuclear transitions. Significantly higher photon-detection rate can be obtained, thus affording a higher statistical precision in a given detection time.
The photon-detection rate using the electronic transition between $^2S_{1/2}$ and $^2P_{1/2}$
can reach $p_\textrm{d}\sim N_{\textrm{ion}}N_b f_b$, which is $p_\textrm{d}\sim 10^{14}\,$s$^{-1}$ and $10^{15}\,$s$^{-1}$ at the SPS and LHC, respectively.
Note that the nuclear isomeric state has a radiative lifetime of about $\gamma\cdot1\,$s in the lab frame. During the long lifetime, we can tune the laser to excite transitions between different HFS levels. Exciting different transitions, which could also be achieved using multiple lasers, is needed for continuous detection here because there is no cycling transition between HFS levels of ${}^2S_{1/2}$ and ${}^2P_{1/2}$.
Measuring the energy of the secondary photons, denoted as $E_{sps}$,
could determine $\gamma^\textrm{ave}$ with a high precision
according to
$\gamma=(E_{sps}/4\hbar\omega_{sp})^{1/2}$,
where $\hbar\omega_{sp}$ is the lab-frame photon energy
of the laser exciting the electronic transition.
Depending on how the isomers are produced, the isomeric-state energy could be determined with a precision similar to that of $\gamma^\textrm{ave}$.
Another potential advantage of using PSI is that laser cooling via electronic transitions could be implemented
(see, for example, Ref.\,\cite{Krasny2020LaserCooling}), which would further reduce the ion-energy spread allowing for
better resolution of the transitions in Fig.\,\ref{fig:Probe_spec_Li-like}.
Laser cooling of ultrarelativistic Li-like Pb ions via the $^2S_{1/2}\rightarrow{}^2P_{1/2}$ electronic transition
was proposed for SPS \cite{Krasny2019PoP}.
At lower energies, laser cooling of a bunched Li-like carbon and oxygen ion beam was already demonstrated at ESR and CSRe ion storage rings, see Refs.\,\cite{Wen2014_las_cool,Wang2018} and the review \cite{Steck2020HeavyIonRing}.
\section{Search for the resonance}
\label{Sec:search_res}
In the direct excitation case, we need to scan across the possible energy of the isomeric state, which is currently known as $8.19(12)$\,eV \cite{Peik2021}, to search for the resonance. Besides tuning the energy of laser photons, we can also tune the energy of relativistic HCI, i.e., the relativistic Lorentz factor $\gamma$, and the ion energy spread.
We can start with ions having a relative energy spread of $\Delta\gamma/\gamma\approx10^{-3}$, and tune $\gamma$ to scan across the energy range, $8.19\pm(3\times0.12)$\,eV, which would require about 90 scan steps.
The resonant excitation rate of isomers could be sufficiently high for experiments even for bare nuclei, where there is no NHM and the isomeric state has a narrow radiative width.
For instance, it takes around 0.6\,s to excite a significant fraction of ions in an ion bunch composed of bare nuclei at the SPS for $\Delta\gamma/\gamma=10^{-3}$ (see also the examples in Sec.\,\ref{subsubsec:res_bare}).
Detection of isomer excitation can be done by using the schemes presented in Sec.\,\ref{Sec:Detect}.
After finding the resonance, i.e, finding the $\gamma$ required for $E_{\textrm{iso}}=2\gamma\hbar\omega$, the transition energy is known with a relative precision of about $\Delta\gamma/\gamma\approx10^{-3}$.
At the second stage, we can reduce the relative ion-energy spread from $10^{-3}$ to $10^{-4}$, and, again, tune the relativistic factor $\gamma$ to scan across the narrower energy range.
Through this process, the transition energy would be determined with a precision improved by one order of magnitude.
One can continue such procedure if the ion-energy spread can
be further reduced.
Determining the ion-frame transition energy requires knowing the central value of $\gamma$, $\gamma^{\textrm{ave}}$, with a high precision, which could be done via, for example, measuring secondary photons (see Ref.\,\cite{Budker2020_AdP_GF} and schemes in Sec.\,\ref{Sec:Detect}).
Prior to precision measurements of $\gamma^{\textrm{ave}}$,
we should achieve a better match between the photon-absorption spectrum considering the ion-energy spread and the laser-photon spectrum in the ion frame in terms of their centers, i.e.,
$|E_\textrm{iso}-2\gamma^{\textrm{ave}}\hbar\omega|/E_\textrm{iso}\ll\Delta\gamma/\gamma$.
This could be accomplished by utilizing a laser with a narrow bandwidth and scanning the laser-photon energy $\hbar\omega$, maximizing the isomer-production rate.
Matching the two spectra would allow more precise determination of the isomeric-state energy via measuring $\gamma^{\textrm{ave}}$; see Sec.\,\ref{Sec:Detect}.
As mentioned above, further improvement
could be granted by employing laser cooling
of PSI, for instance, as projected for Li-like lead ions at the SPS \cite{Krasny2019PoP}, which could reduce the ion-energy spread by another one or two orders of magnitude.
The transitions related to an intermediate electronic excited state discussed in Sec.\,\ref{Subsec:exc_271eV}
have a relative energy uncertainty of about $5\times10^{-4}$.
When exciting these transitions, one can start with ions having a relative energy spread of $\Delta\gamma/\gamma\approx10^{-4}$ to search for the resonance.
For transitions discussed in Sec.\,\ref{Subsec:exc_29keV}, the relative uncertainty of the transition energy between the nuclear ground (isomeric) state and the 29\,keV state is about $2.4\times10^{-6}$ ($4.8\times10^{-6}$), which is more precise than the typical ion energy spread, $\Delta\gamma/\gamma\approx10^{-3}-10^{-5}$ in storage rings. Therefore, these transitions could be excited directly without varying $\gamma$ to scan across a large energy range.
In these cases, we also need to match the photon-absorption spectrum and the laser-photon spectrum, in order to drive the transition efficiently and to facilitate further measurement of the isomeric-state energy.
\section{other low-lying nuclear transitions}
\label{Sec:other_low_state}
The excitation and probing schemes of $^{229m}$Th discussed above could also be applied to other low-lying nuclear states with experimental parameters such as the laser-photon energy and ion energy chosen case by case.
$^{229}$Th is unique in featuring a nuclear excited state in the eV range. The next confirmed low-energy nuclear excitation occurs in $^{235}$U at 76\,eV \cite{NNDC}.
Notably, there exists another interesting case, $^{229}$Pa (a nucleus with the same mass number $A$ as $^{229}$Th but with one more proton), where there is a possible low-energy excitation, perhaps, as low as $60\pm50$\,eV \cite{Ahmad2015Pa}.
A survey of other relatively low-lying nuclear transitions can be found in Ref.\,\cite{Budker2021GF_nucl}, where the proposed Gamma Factory can access those transition energies.
The 76\,eV transition in $^{235}$U, having an extremely narrow radiative width ($\Gamma_{\textrm{rad}}/\hbar \approx 9.9\times10^{-25}\,$s$^{-1}$) and $E3$ multipolarity \cite{Budker2021GF_nucl}, discourages laser excitation of the isomeric state.
We discuss below the potentially interesting case of $^{229}$Pa.
$^{229}$Pa is expected to have octupole deformation in its nuclear ground state ($I^P=5/2^+$), which may have a nearly-degenerate first excited state of opposite parity ($I^P=5/2^-$) \cite{Ahmad1982Pa,Ahmad2015Pa}. This would make $^{229}$Pa orders of magnitude more sensitive to hadronic CP violation compared to other nuclei such as $^{199}$Hg \cite{Singh2019T-vio}.
However, the existence of the low-lying nuclear excited state has not been unambiguously proven. A recent analysis showed that the parity doublet could have a splitting of $60\pm50$\,eV, where the large error still makes the existence of the doublet uncertain \cite{Ahmad2015Pa}.
Laser spectroscopy of relativistic highly charged $^{229}$Pa ions could provide alternative approaches to probing the doublet \cite{Singh2021}. The $I^P=5/2^-$ state could be excited by either direct resonance excitation or excitation via an intermediate state.
An $E1$ transition could be allowed between the nuclear ground and first excited states.
For direct laser excitation, we need to scan across a large energy range, $60\pm50\,$eV, to search for the resonance. This can be realized, for example, at SPS by tuning the relativistic factor of the HCI from $\gamma=10$ to 110 together with a laser having a fixed energy of about 0.5\,eV in the lab frame. It could take up to about 2400 scan steps assuming using ions with a relative energy spread of $\Delta\gamma/\gamma=10^{-3}$.
One could probe the existence of the doublet and the energy splitting via measuring the radiative decay of the $I^P=5/2^-$ state.
Furthermore, measurement of the $E3$ transition strength between the doublet would indicate the octupole collectivity of $^{229}$Pa unambiguously.
Besides, one could also excite the transition from the nuclear ground state to a higher excited state at 11.4(2) or 11.6(3)\,keV \cite{NNDC}. Measuring radiative decay of the higher state could determine whether there is a decay channel into a low-lying nuclear state and probe the energy splitting of the ground-state doublet.
The expected small energy splitting in the ground state doublet, if it exists, would enhance the NHM effect in H- or Li-like $^{229}$Pa ions, which might offer more possibilities for laser excitation of the first excited state, similar to those discussed in Sec.\,\ref{subsubsec:res_H-like} and \ref{Subsec:exc_271eV}.
We can also detect the change of the nuclear state via laser spectroscopy on electronic transitions; see, for example, Sec.\,\ref{Subsec:elec_tran}.
We note that neutral $^{229g}$Pa atoms in the nuclear ground state have a half-life of about 1.5(5)\,d\,\cite{ENSDF} due to electron capture and $\alpha$-decay which would pose a severe challenge for producing a beam from an ion source. Hence, $^{229}$Pa ions need to be freshly produced at a radioactive-ion beam facility. Projectile fragmentation of $^{238}$U is well suited for this purpose\,\cite{LISE}. Assuming the same beam/target parameters as in Sec.\,\ref{Sec:ion_source}, we obtain $6\cdot10^6$ $^{229}$Pa ions produced per pulse. Large amounts of $^{229}$Pa can be produced at FRIB\,\cite{Abel2019_PaSource} and RIKEN\,\cite{Kubo-2003} where the worldwide highest uranium primary beam intensities are available. The advantage of using GSI/FAIR, HIAF or CERN facilities would be the higher projectile energy and thus the ability to efficiently produce fully-stripped ions. Furthermore, storage facilities would be available as discussed above for the $^{229}$Th ions. In bare nuclei the 99.52(5)\% electron capture branch\,\cite{ENSDF} is disabled correspondingly increasing the half-life of $^{229g}$Pa$^{91+}$ nuclei. Moreover, if stored at high energy in a storage ring, the half-life is further increased in the lab frame due to the relativistic time dilation.
\section{Conclusion}
We propose multiple approaches towards laser excitation of the nuclear isomeric state in highly charged $^{229}$Th ions at high-energy storage rings. It can be achieved
via direct resonant excitation, excitation via an intermediate electronic excited state, or excitation via an intermediate nuclear excited state.
Due to the Lorentz boost of laser photon energies in the ion frame, lasers with wavelengths in the visible range or longer could be used for excitation.
Searching for the resonance and matching the laser-photon spectrum and the photon-absorption spectrum in the ion frame can be realized through scanning the ion energy, adjusting the ion-energy spread as well as tuning the laser beam.
Among the discussed excitation schemes, the STIRAP excitation via the 29\,keV state may require further development of the laser or storage-ring technologies; schemes utilizing repeated pumping relax the demand on the laser intensity.
Nuclear-hyperfine-mixing assisted excitation schemes are also discussed, which could be most favorable for experimental implementation.
In the direct resonant excitation scenario, the significantly reduced isomeric-state lifetime in H- or Li-like $^{229}$Th ions due to the NHM effect would correspond to much higher excitation rates compared to bare thorium nuclei, thus making it more practical to implement a $\pi$ pulse for isomeric excitation.
Furthermore, the NHM effect allows for excitation of $E1$ transitions that change both the electronic and nuclear states, enabling efficient nuclear isomer excitation via an intermediate electronic excited state in Li-like $^{229}$Th. The transition energies are also easily accessible by x-ray lasers, allowing for isomer excitation in static $^{229}$Th ions using x-ray pulses.
We also present schemes to probe the produced $^{229m}$Th isomers, exploiting the radiative decay of the isomeric state,
radiative decay of the second nuclear excited state,
or laser spectroscopy of electronic transitions in Li-like thorium ions.
Among these detection schemes, the NHM effect helps increase the photon detection rate of the decay of the isomeric state,
the precise knowledge of the second nuclear excited state energy can be exploited to measure the isomeric state energy precisely, and using electronic transitions leads to higher photon detection rates than nuclear transitions.
Our estimates show that the isomeric-state energy could be determined with orders-of-magnitude improvement in precision compared to the current result.
The schemes proposed here for $^{229}$Th could also be adapted to studying low-energy nuclear states in other nuclei including $^{229}$Pa.
\section*{Acknowledgements}
The authors are grateful to Mei Bai, Carsten Brandau, Wick Haxton, Anton Peshkov,
Jaideep Singh, Markus Steck, and Thomas St{\"o}hlker for helpful discussions. This work was supported in part by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) Project ID 390831469: EXC 2118 (PRISMA+ Cluster of Excellence) and the Helmholtz
Excellence Network ExNet02. A.~P\'alffy gratefully acknowledges funding from the DFG in the framework of the Heisenberg Program.
AS acknowledges support by the DFG under Germany's Excellence Strategy EXC-2123 QuantumFrontiers-390837967.
Yury Litvinov acknowledges support by the State of Hesse within the Research Cluster ELEMENTS (Project ID 500/10.006).
PGT and A. P\'alffy acknowledge support via the ‘ThoriumNuclearClock’ project that has received funding from the ERC under the European Union’s Horizon 2020 research and innovation programme (Grant Agreement No. 856415).
|
\section{Introduction}
By the year 2050, roughly 114.6 million people will be living with incurable blindness \cite{bourne_magnitude_2017}.
Although some individuals can be treated with surgery or medication, there are no effective treatments for many people blinded by severe degeneration or damage to the retina, the optic nerve, or cortex. In such cases, an electronic visual prosthesis (\emph{bionic eye}) may be the only option \cite{fernandez_development_2018} (Fig.~\ref{fig:overview}A)
Analogous to cochlear implants, these devices electrically stimulate neurons in the early visual system to elicit neuronal responses that the brain interprets as visual percepts (\emph{phosphenes}).
Current devices generally provide users with an improved ability to localize high-contrast objects and perform basic orientation \& mobility tasks \cite{ayton_update_2020}, but are not yet able to match the acuity of natural vision.
Most current prostheses provide a very limited \ac{FOV}; for example, the artificial vision generated by Argus II~\cite{luo_argusr_2016}, the most widely adopted retinal implant thus far, is restricted to roughly $10 \times 20$ degrees of visual angle. This forces users to scan the environment with strategic head movements while attempting to piece together the information \cite{erickson-davis_what_2021}.
In addition, the limited number of electrodes (60 in Argus II) severely limits the number of independent phosphenes that the device can generate \cite{beyeler_learning_2017}.
Consequently, researchers have suggested ways to simplify the visual scene before it is displayed using image processing and computer vision.
One popular approach is to estimate relative depth in the scene \cite{mccarthy_mobility_2014,han_deep_2021} and then make phosphenes appear brighter the closer they are to the observer, in order to highlight nearby obstacles.
Here we refer to this method of substituting depth for intensity as \emph{DepthOnly} mode (Fig.~\ref{fig:overview}A).
A more recent line of research suggests to extract semantic and structural edges instead \cite{sanchez-garcia_indoor_2019,sanchez-garcia_semantic_2020} (\emph{EdgesOnly} mode), in order to give the user a sense of where important objects are in the scene.
Since these two modes provide seemingly complementary sources of information, it is natural to ask how to best combine them.
A straightforward approach would be to visualize both edge and depth cues at the same time (\emph{EdgesAndDepth}).
However, we hypothesized that users might instead prefer the ability to flexibly switch between the two (\emph{EdgesOrDepth}).
To assess the relative importance of depth and edge information, a side-by-side comparison is needed.
Due to the unique requirements of working with bionic eye recipients (e.g., constant assistance, increased setup time, travel cost), experimentation with different encoding methods remains challenging and expensive.
Instead, embedding models of \ac{SPV} in immersive \ac{VR} allows sighted subjects to act as \emph{virtual patients} by ``seeing'' through the eyes of the patient, taking into account their head and eye movements as they explore an immersive virtual environment \cite{kasowski_immersive_2022,wang_cross-task_2018,thorn_virtual_2020,sanchez-garcia_influence_2020}.
This can speed up the development process by allowing researchers to test theoretical predictions in high-throughput experiments, the best of which can be validated and improved upon in an iterative process with the bionic eye recipient in the loop \cite{kasowski_towards_2021}.
To this end, we make the following contributions:
\begin{enumerate}[topsep=0pt,itemsep=-1ex,partopsep=0pt,parsep=1ex,leftmargin=14pt,label=\roman*.]
\item We embed a psychophysically validated \ac{SPV} model \cite{beyeler_model_2019} in \ac{VR} to allow sighted participants to act as virtual bionic eye patients in an immersive virtual environment.
\item We explore the relative importance of different scene simplification strategies based on depth estimation \cite{han_deep_2021} and semantic edge detection \cite{sanchez-garcia_semantic_2020} as a preprocessing strategy for bionic vision (Fig.~\ref{fig:overview}A).
\item We systematically evaluate the ability of these strategies to support obstacle avoidance and object identification (Fig.~\ref{fig:overview}B) with a user study in immersive \ac{VR}.
\end{enumerate}
In sum, this is the first study to compare the relative importance of depth cues and semantic edges for bionic vision and an essential first towards a bionic eye that uses computer vision to improve a user's scene understanding.
\section{Background}
\label{sec:background}
Retinal implants are currently the only FDA-approved technology to treat blinding degenerative diseases such as \ac{RP} and \ac{ARMD}.
Most current devices acquire visual input via an external camera and perform edge extraction or contrast enhancement via an external \ac{VPU}, before sending the signal through wireless coils to a microstimulator implanted in the eye or the brain (see Fig.~\ref{fig:overview}A).
This device receives the information, decodes it, and stimulates the visual system with electrical current.
Two devices are already approved for commercial use: Argus II (60 electrodes arranged in a $6 \times 10$ grid, Second Sight Medical Products, Inc., \cite{luo_argusr_2016}) and Alpha-IMS (1500 electrodes, Retina Implant AG, \cite{stingl_artificial_2013}).
In addition, PRIMA (378 electrodes, Pixium Vision,
has started clinical trials, with others to follow shortly \cite{ayton_first--human_2014,ferlauto_design_2018}.
A common misconception is that each electrode in the grid can be thought of as a ``pixel'' in an image \cite{dagnelie_real_2007,chen_simulating_2009,lui_transformative_2011,perez-yus_depth_2017,sanchez-garcia_indoor_2019}, and most retinal implants linearly translate the grayscale value of a pixel in each video frame to a current amplitude of the corresponding electrode in the array \cite{luo_argusr_2016}.
This is known as the scoreboard model, which implies that creating a complex visual scene can be accomplished simply by using the right combination of pixels, analogous to the images projected on the light bulb arrays of some sports stadium scoreboards \cite{dobelle_artificial_2000}.
On the contrary, recent work suggests that phosphenes vary in shape and size, and differ substantially across subjects and electrodes \cite{beyeler_model_2019,erickson-davis_what_2021}.
Despite their potential to restore vision to people living with incurable blindness, the number of bionic eye users in the world is still relatively small ($\sim 500$ retinal prostheses implanted to date).
To investigate functional recovery and experiment with different implant designs, researchers have therefore been developing \ac{VR} prototypes that rely on \ac{SPV}.
The classical method relies on sighted subjects wearing a \ac{VR} headset, who are then deprived of natural viewing and only perceive phosphenes displayed in a \ac{HMD}. This viewing mode has been termed \emph{transformative reality} \cite{lui_transformative_2011} (as opposed to \emph{altered reality} typically used to describe low vision simulations \cite{bao_augmented_2019}), which allows sighted users to ``see'' through the eyes of the bionic eye recipient, taking into account their head and/or eye movements as they explore a virtual environment \cite{kasowski_towards_2021}.
However, because most \ac{SPV} studies rely on the scoreboard model \cite{dagnelie_real_2007,chen_simulating_2009,lui_transformative_2011,perez-yus_depth_2017,sanchez-garcia_indoor_2019}, it is unclear how their findings would translate to real bionic eye recipients.
Only a handful of studies have incorporated a great amount of neurophysiological detail into their setup~\cite{josh_psychophysics_2013,vurro_simulation_2014,wang_cross-task_2018,thorn_virtual_2020,kasowski_immersive_2022}, only three of which relied on an established and psychophysically validated model of \ac{SPV} \cite{wang_cross-task_2018,thorn_virtual_2020,kasowski_immersive_2022}.
In addition, being able to move around as one would in real life has shown to significantly increase the level of immersion a user experiences \cite{pasch_immersion_2009}.
However, the level of immersion offered by most \ac{SPV} studies is relatively low, as stimuli are often presented on a screen \cite{ying_recognition_2018, wang_cross-task_2018}.
In contrast, most current prostheses provide a very limited \ac{FOV} (e.g., Argus II: $10 \times 20$ degrees of visual angle), which requires users to scan the environment with strategic head movements while trying to piece together the information \cite{erickson-davis_what_2021}.
Furthermore, Argus II does not take into account the eye movements of the user when updating the visual scene, which can be disorienting for the user.
Ignoring these \ac{HCI} aspects of bionic vision can result in unrealistic predictions of prosthetic performance, sometimes even exceeding theoretical limits for visual acuity (as pointed out by \cite{caspi_assessing_2015}).
\section{Related Work}
\label{sec:related_work}
Most retinal implants are equipped with an external \ac{VPU} that is capable of applying simple image processing techniques to the video feed in real time.
In the near future, these techniques may include deep learning--based algorithms aimed at improving a patient's scene understanding \cite{beyeler_towards_2022}.
Based on this premise, researchers have developed various image optimization strategies, and assessed their performance by having sighted observers (i.e., \emph{virtual patients}) conduct daily visual tasks under \ac{SPV} \cite{boyle_region--interest_2008,dagnelie_real_2007,al-atabany_improved_2010,li_image_2018,mccarthy_mobility_2014,vergnieux_simplification_2017}.
These simulations allow a wide range of computer vision systems to be developed and tested without requiring implanted devices.
Current retinal prostheses are implanted in only one eye, and thus are unable to convey binocular depth cues.
Previous work has therefore explored the possibility of obtaining depth information through additional peripherals, such as an RGB-D sensor, and studied behavioral performance of virtual patients typically navigating an obstacle course under \ac{SPV}.
For example, Ref.~\cite{perez-yus_depth_2017} used depth cues to generate a simplified representation of the ground to indicate the free space within which virtual patients could safely walk around, whereas Ref.~\cite{han_deep_2021} used deep neural networks to estimate per-pixel relative depth and then substituted depth for intensity.
Depth cues were also shown to help avoid nearby obstacles that are notoriously hard to detect with other computer vision algorithms, such as branches hanging from a tree \cite{lieby_substituting_2011}.
Ref.~\cite{mccarthy_mobility_2014} used depth to increase the contrast of object boundaries and showed that this method reduced the number of collisions with ground obstacles.
In addition, retinal prosthesis patients were shown to benefit from distance information provided by a thermal sensor when trying to avoid nearby obstacles and people \cite{sadeghi_thermal_2019}.
Recently, with the development of deep learning in computer vision, semantic segmentation algorithms have become unprecedentedly effective. This can be used as another method to reduce visual clutter in prosthetic vision, be it in outdoor scenes \cite{horne_semantic_2016} or indoor scenes \cite{sanchez-garcia_indoor_2019}.
The latter study combined semantic and structural image segmentation to build a schematic representation of indoor environments, which was then shown to improve object and room identification in a \ac{SPV} task \cite{sanchez-garcia_indoor_2019}.
Semantic segmentation has been applied to simplify both the outdoor scenes \cite{horne_semantic_2016} and the indoor scenes \cite{sanchez-garcia_indoor_2019} for retinal prostheses.
\begin{figure*}[t!]
\centering
\includegraphics[width=\linewidth]{figures/fig-rooms.png}
\Description{Six different room layouts shown from above. Bottom half of each room contains either three cylinders (Rooms A - C) arranged differently, five cylinders (Room D), six cylinder (Rooms E), or eight cylinders (Room F) arranged in hexagonal grid. At the end of the room, there was either one table with three objects on it, or three tables with one object on each.}
\caption{Room layouts. Participants started in the center along the bottom wall.
Participants were instructed to walk towards the other end of the room while avoiding obstacles.
At the end of the room, there was either one table with three objects on it, or three tables with one object on each.
Participants had to identify the medium-sized cube located one on of the tables.}
\label{fig:rooms}
\end{figure*}
However, since the above algorithms were developed in isolation and tested on different behavioral tasks, a side-by-side comparison of their ability to aid scene understanding is still lacking.
A notable exception is Ref.~\cite{han_deep_2021}, which compared a number of scene simplification strategies that rely on depth extraction, semantic segmentation, and visual saliency on a single task.
However, their study was limited to sighted participants viewing \ac{SPV} videos on a screen, and thus did not account for the \ac{FOV} restrictions that are common with current implants.
To address these challenges, we used a neurobiologically inspired computational model of bionic vision \cite{beyeler_model_2019} to generate realistic predictions of \ac{SPV}, and combined it with scene simplification strategies based on depth estimation \cite{han_deep_2021} and semantic edge detection \cite{sanchez-garcia_semantic_2020}.
To allow for a fair comparison between algorithms, we asked virtual patients to avoid obstacles and identify objects in a number of immersive virtual environments.
\section{Methods}
\subsection{Virtual Patients}
To simulate a bionic eye patient, we developed \ac{SPV} simulations in Unity that were streamed in real time to a wireless head-mounted \ac{VR} headset (HTC VIVE Pro Eye with wireless adapter, HTC Corporation).
We followed the procedure outlined in Fig.~\ref{fig:overview}A to simulate different scene simplification strategies and tested them on an obstacle avoidance and object selection task (Fig.~\ref{fig:overview}B).
All our simulations were run on an Intel i9-9900k processor (C\# code) and an NVIDIA RTX 2070 Super GPU with 16GB of DDR4 memory (shader code).
The entire \ac{SPV} workflow was thus as follows:
\begin{enumerate}[topsep=0em,itemsep=-1ex,partopsep=0em,parsep=1ex,leftmargin=6ex,label=\roman*.]
\item Image acquisition: Utilize Unity's virtual camera to acquire the central $60 \times 60$ degrees of visual angle of the display's monocular \ac{FOV} at roughly 90 frames per second and downscale to a target texture of $120 \times 120$ pixels.
\item Scene simplification: Extract semantic edges using the \texttt{QuickOutline} asset and/or depth cues using \texttt{UnityObjectToClipPos} to simulate different scene simplification strategies (explained in Section~\ref{sec:modes}).
\item Electrode activation: Determine electrode activation based on the visual input as well as the placement of the simulated retinal implant. In the current study, a $3 \times 3$ Gaussian blur was applied to the preprocessed image to average the grayscale values around each electrode's location in the visual field. This gray level was then interpreted as a current amplitude delivered to a particular electrode in the array.
\item Phosphene model: Use Unity shaders to convert electrode activation to a visual scene in real time (see Section~\ref{sec:spv}).
\item Phosphene rendering: Render the elicited phosphenes in the \ac{HMD} of the \ac{VR} system.
\end{enumerate}
\subsection{Participants}
We recruited $18$ participants with normal or corrected-to-normal vision from the research participant pool of the Department of Psychological \& Brain Sciences at University of California, Santa Barbara (UCSB) to act as virtual patients in our experiment.
Participants ranged from 18 to 20 years in age ($\mathit{M}=18.8$, $\mathit{SD}=.73$ years), with 6 participants identifying as male and 12 identifying as female.
Of these participants, $4$ had never used \ac{VR} before, $12$ had used VR $1-5$ times before, $1$ had used it $10-20$ times, and $1$ had used it $20+$ times.
Potential participants were excluded if they reported that they were prone to cybersickness.
The study was approved by UCSB's Institutional Review Board.
\subsection{Rooms}
\label{sec:rooms}
Participants were asked to navigate six different virtual rooms that were filled with 3-7 obstacles to avoid and a target object to select among distractor objects (Fig.~\ref{fig:rooms}).
Participants always started along the bottom wall and were instructed to walk towards the other end of the room while avoiding obstacles.
Upon collision with an object, a ``thud'' sound was played through the VIVE headphones.
Once they had passed all obstacles, a ``chime'' sound was played through the VIVE headphones to indicate the end of the obstacle avoidance portion of the task.
Participants then had to navigate to one of three tables and identify a medium-sized cube located on it (Fig.~\ref{fig:rooms-table}).
Each room had either one table with three objects on it, or three tables with one object each.
While the arrangements of objects was pseudo-randomized on each trial (consisting either of a sphere, a cylinder, and a medium cube; or a small cube, a medium cube, and a large cube), participants always had to select the medium cube.
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth]{figures/fig-objects.png}
\Description{Two example views of the table with different objects on it. Top panel shows three different boxes, bottom panel shows a box, a sphere, and a cylinder.}
\caption{Example views of the table with different objects on it. The correct object to identify was always the medium-sized cube (indicated with a white dashed circle) among two distractor objects.
The arrangement of the objects was pseudo-randomized on each trial.
Users had to confirm their selection by pressing a button on the VIVE controllers.}
\label{fig:rooms-table}
\end{figure}
\subsection{Simulated Prosthetic Vision}
\label{sec:spv}
To generate realistic simulations of phosphene appearance, we adapted code from the VR-SPV open-source toolbox \cite{kasowski_immersive_2022}.
This toolbox provides a C\#/shader implementation of the pyschophysically validated axon map model \cite{beyeler_model_2019} that describes phosphenes using two shape parameters, $\rho$ (specifying phosphene size) and $\lambda$ (describing phosphene elongation).
We chose $\rho=300$ and $\lambda=550$ for our simulations, which is roughly in the middle of the range of values reported by Argus II users \cite{beyeler_model_2019}.
To support the real-time simulation of a $20 \times 15$ electrode array implanted in the retina, we pre-calculated an initial mapping of each electrode's effects on the scene before starting the experiment.
\subsection{Modes}
\label{sec:modes}
We considered four different scene simplification modes (Fig.~\ref{fig:modes}):
\begin{enumerate}[topsep=0em,itemsep=-1ex,partopsep=0em,parsep=1ex,leftmargin=6ex,label=\roman*.]
\item \emph{EdgesOnly}: To simulate semantic and structural edge segmentation \cite{sanchez-garcia_semantic_2020}, we drew white outlines around the ground-truth object boundaries of the obstacles, tables, and objects in the room using the \texttt{QuickOutline} asset (Fig.~\ref{fig:modes}, \emph{top}).
Rather than applying an edge extraction algorithm to the camera view (which may be too slow to run in real time), \texttt{QuickOutline} coats each object with a white material, followed by a slightly smaller coat of a black material that makes it look as if only the edges of the object are showing (see Fig.~\ref{fig:modes}A for an example).
In addition, we highlighted the edges of the room to give participants an indication of the room layout.
\item \emph{DepthOnly}: To simulate depth to intensity substitution \cite{han_deep_2021}, we first extracted the ground-truth depth of each object relative to the main camera using \texttt{UnityObjectToClipPos}.
Inverse depth was then linearly mapped to phosphene brightness, so that objects at zero depth appeared white, and objects at distances greater than the furthest wall (\texttt{farClipPlane}) appeared black (Fig.~\ref{fig:modes}, \emph{middle}). In other words, objects appeared brighter the closer they were to the observer.
\item \emph{EdgesAndDepth}: This mode visualized both edge and depth information at the same time (Fig.~\ref{fig:modes}, \emph{bottom}).
\item \emph{EdgesOrDepth}: In this mode, participants randomly started in either \emph{EdgesOnly} or \emph{DepthOnly} mode and were able to toggle between modes via press of a button on the VIVE controller.
\end{enumerate}
In our pilot study, we also included a low-resolution ``straight-through'' \ac{SPV} view of the scene (that did not include any scene simplification) as a baseline.
However, participants were unable to perform either task.
This is consistent with previous work showing that such a mode leads to chance performance in obstacle avoidance tasks (at best) \cite{kasowski_immersive_2022}.
To prevent participant frustration and to save time, we therefore did not include this mode in our final experiment.
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth]{figures/fig-modes.png}
\Description{Example frames for the different modes, showing object outlines for EdgesOnly and shaded gray levels for DepthOnly. The scene includes two cylinders that are closeby and a table located further away in the corner of the room.}
\caption{SPV modes used for scene simplification. \emph{Top:} In \emph{EdgesOnly} mode, only semantic and structural edges are visualized. \emph{Middle:} In \emph{DepthOnly} mode, per-pixel ground-truth depth is inverted and linearly translated to grayscale level. \emph{Bottom:} In \emph{EdgesAndDepth} both edges and depth are visualized. A fourth mode, \emph{EdgesOrDepth}, gave users the ability to toggle between \emph{EdgesOnly} and \emph{DepthOnly} modes by pressing a button on the VIVE controller.}
\label{fig:modes}
\end{figure}
\subsection{Procedure}
\label{sec:procedure}
The order in which the modes were presented and the order in which the rooms were presented was pseudo-randomized across participants.
To reduce cognitive load and context-switching, all rooms for a particular mode were presented in a block.
This allowed participants to get comfortable with one mode at a time.
Participants saw each combination of room and mode only once.
Before the start of each mode, participants were allowed to explore a tutorial room (not part of Rooms A--F) that contained a single obstacle to avoid and a medium-sized cube on a table without distractor objects.
Participants were free to navigate the room for as long as they wanted, until they felt comfortable with the \ac{SPV} mode.
When they indicated that they were ready to start the experiment, they had to grab the medium-sized cube off the table and confirm their selection with a button press.
They were then given $30$ seconds to walk back to the starting position, indicated by a red circle on the floor.
At this point the experiment started with a pseudo-randomly selected room.
A trial was concluded by selecting an object from one of the tables in the room.
We estimated from our pilot study that most participants take between thirty seconds and one minute per trial, but on rare occasions would get lost to the point where they could not complete a trial. We therefore introduced a time limit, which was set at four minutes per trial.
If participants exceeded this limit, the trial ended automatically, \ac{SPV} was turned off, and the participant was given $30$ seconds to walk back to the starting position, before the next trial started.
In \emph{EdgesOrDepth} mode, participants were allowed to toggle between \emph{EdgesOnly} and \emph{DepthOnly} modes as often as they wanted.
Each trial started in the mode that was last active during the last trial (or for Trial 1: whichever was last active in the tutorial room).
This way, if participants preferred one mode over the other, they could just stay on that mode without having to constantly switch back and forth between trials.
\begin{figure*}[t!]
\centering
\includegraphics[width=\textwidth]{figures/fig-performance-overall.png}
\Description{Barplots showing the success rates for obstacle avoidance and object selection. Numbers are explained in the main text. OA success rate was worst for EdgesOnly, with the three other modes performing similarly. OS accuracy was best for DepthOnly and EdgesOrDepth.
Time taken was around 20s for OA and around 40s for OS. }
\caption{Obstacle avoidance (OA) performance, measured by success rate (i.e., the number of trials with zero collisions, Panel A) and time taken (Panel C), and object selection (OS) performance, measured by accuracy (i.e., the fraction of trials where the correct object was selected, Panel B), and time taken (Panel D).
The dashed line in Panel B indicates chance performance (\SI{33}{\percent}).
Vertical bars are the standard error of the mean (SEM).
Statistical significance was determined using paired $t$-tests, corrected for multiple testing using the Holm-Sidak method (*: $p<.05$).}
\label{fig:results-overall}
\end{figure*}
\subsection{Data Collection and Analysis}
During the whole experiment, we recorded participant position (sampled every $0.5$ seconds), time and place of obstacle collisions, and object selection.
Upon collision with an obstacle, we applied a $3$ second timeout window so that prolonged collisions with the same object were only counted once.
We considered the number of collisions (the fewer the better) and time taken (the shorter the better) as the main metrics to quantify obstacle avoidance performance.
The obstacle avoidance part of the task ended as soon as participants entered the top third of the room.
From that time forward, participants were considered to be in the object selection part of the task, even if they accidentally crossed the line back into the obstacle avoidance portion of the room.
We considered the fraction of correctly selected objects (accuracy; the higher the better) and time taken (the shorter the better) as the main metrics to quantify object selection performance.
A trial ended as soon as participants selected an object.
To determine whether performance was significantly different across modes ($C(\mathit{Mode})$) and rooms ($C(\mathit{Room})$), we ran a regression analysis for each metric mentioned above using the Ordinary Least Squares (OLS) model in the \texttt{statsmodel} Python module.
We also factored in participant's gender ($C(\mathit{Gender})$) and VR experience ($C(\mathit{VRExperience})$).
To test for learning effects, we labeled trials depending on whether they were completed in the first or second half of each session ($C(\mathit{FirstHalf})$).
The full model was thus $\mathit{Metric} \sim C(\mathit{Mode}) + C(\mathit{Room}) + C(\mathit{Gender}) + C(\mathit{VRExperience}) + C(\mathit{FirstHalf})$.
$p$-values were calculated using paired $t$-tests of the OLS model, which corrects for multiple testing using the Holm-Sidak method.
\section{Results}
All participants were able to complete the experiment using all four tested modes, taking on average $25 \pm 15$ minutes to complete the 24 different trials.
Trial completion time varied widely across rooms and modes, with a median completion time of 16 seconds for the obstacle course, 12 seconds for the object selection portion, and 31 seconds overall.
90\% of trials were completed in less than 193 seconds, with the obstacle course cleared in under 35 seconds and the object selected in under 155 seconds.
The time limit of 4 minutes was reached on only two trials, which were subsequently removed from further analysis.
\subsection{Obstacle Avoidance}
The number of obstacle avoidance (OA) trials completed without any collisions (``success rate'') was significantly affected by the scene simplification mode ($F(3, 432) = 32.2, p<.001$) and room layout ($F(5, 432) = 2.49, p<.05$), summarized in Fig.~\ref{fig:results-overall}A.
Participants performed the OA portion of the task most successfully using the \emph{EdgesAndDepth} mode, completing on average \SI{89.8}{\percent} of trials without colliding.
However, this was not significantly better than with the \emph{DepthOnly} (\SI{87.9}{\percent}) and \emph{EdgesOrDepth} modes (\SI{87.0}{\percent}; $t$-test $p>0.05$, corrected for multiple comparisons with the Holm-Sidak method).
On the other hand, participants performed worst with the \emph{EdgesOnly} mode, as evidenced by a significantly lower success rate (\SI{64.8}{\percent}) and among the longest times taken (\SI{25}{\second}, Fig.~\ref{fig:results-overall}C).
Participants were significantly faster using \emph{DepthOnly} (\SI{17}{\second}) than they were with \emph{EdgesOrDepth} (\SI{22}{\second}) and \emph{EdgesAndDepth} (\SI{25}{\second}; Fig.~\ref{fig:results-overall}C).
In addition, we found a strong learning effect ($F(1, 432)=84.3$, $p<.001$), which indicated an increased success rate of 28\% from the beginning of a session to the end.
Weaker but significant effects were also found for gender ($F(1, 432) = 18.6, p<.001$) and VR experience ($F(3, 432) = 3.38, p<.05$).
\begin{figure*}[tb]
\centering
\includegraphics[width=\linewidth]{figures/fig-paths.png}
\Description{Example paths taken for all rooms and modes. EdgesOnly led to most collisions. Room F led to most collisions. }
\caption{Birds-eye view of paths taken in the different rooms (columns) using the different scene simplification strategies (rows).
For the sake of clarity, only the paths from every fourth subject are shown.
Color of paths gets more saturated as time moves on.
Collisions are indicated with a black \texttt{+} sign.
Circles indicate the location of the obstacles, and rectangles are the table.
The task switched from obstacle avoidance to object selection as soon as the participant crossed the (to them invisible) horizontal dashed line.}
\label{fig:results-paths}
\end{figure*}
\subsection{Object Selection}
The number of object selection (OS) trials with correct object selections was significantly affected by the scene simplification mode ($F(3, 432) = 6.85, p<.001$) and room layout ($F(5, 432) = 10.2, p<.001$), summarized in Fig.~\ref{fig:results-overall}B.
Participants performed the OS portion of the task most successfully using the \emph{EdgesOrDepth} mode, on average locating and selecting the correct object in \SI{54.6}{\percent} of trials (Fig.~\ref{fig:results-overall}B).
However, this was not significantly better than with the \emph{EdgesOnly} (\SI{46.3}{\percent}) and \emph{DepthOnly} modes (\SI{50.9}{\percent}).
Performance with the \emph{EdgesAndDepth} mode (\SI{40.7}{\percent}) was only slightly better than chance (\SI{33}{\percent}).
OS performance was also affected by gender ($F(1, 432) = 16.2, p<.001$) and VR experience ($F(3, 432) = 5.01, p<.001$).
However, we did not find any learning effects here ($F(1, 432) = .48, p=.488$).
All participants took approximately \SI{40}{\second} to walk towards the correct table and select the correct object, no matter which scene simplification mode was used (Fig.~\ref{fig:results-overall}D).
Since the location of the correct object was pseudo-randomized on each trial, it is possible that most of the time for the OS portion of the task was spent simply walking back and forth between the three tables.
\subsection{Participant Paths}
The paths that participants took are summarized in Fig.~\ref{fig:results-paths}, showing a birds-eye view of the six different rooms.
Here, each line represents the path of a participant, with colors getting more saturated as time progresses, and collisions indicated by black crosses.
The poor OA performance of the \emph{EdgesOnly} mode first reported in Fig.~\ref{fig:results-overall} can also be appreciated here, with participants causing more collisions compared to the other modes across all rooms.
This is especially evident in Room F: Even though participants attempted to avoid the obstacle maze by moving in a slalom fashion, there are only two obstacles that never led to collisions.
It is interesting to note that participants seemed to explore Room A the most, with some participants accidentally walking away from the table in the top-left corner of the room and returning to the OA portion of the task.
On the other hand, most participants walked mostly straight ahead in Room B (except in \emph{EdgesOrDepth} mode) and Room E, thus avoiding most obstacles, although most participants still collided with the obstacle that was closest to the starting position.
Room F was by far the most challenging, leading to a large number of collisions in all four modes.
\subsection{Relative Importance of Depth vs. Edges}
One potential indicator of the relative importance of depth and edge cues was \emph{EdgesOrDepth} mode: here participants had the ability to switch between \emph{EdgesOnly} and \emph{DepthOnly} information at the push of a button as often as they saw fit.
We recorded both the number of times that participants toggled modes as well as the amount of time spent in each mode. The results are shown in Fig.~\ref{fig:results-switching}.
To our surprise, participants spent most of their time in \emph{DepthOnly} mode (Fig.~\ref{fig:results-switching}, \emph{top}), both during the OA as well as the OS portion of the task.
In addition, most participants did not often switch between modes ((Fig.~\ref{fig:results-switching}, \emph{bottom}).
The experiment was set up so that each trial started on the mode that was last active (either during the preceding trial or in the tutorial room), so that participants who preferred one mode over the other would not have to constantly switch back and forth.
During the OA portion of the task, participants tended to toggle modes either never (perhaps because the trial started in their preferred mode), once (perhaps because they trial started in their nonpreferred mode), or twice (perhaps to briefly view the other mode, but then promptly switch back to the preferred one).
The only exception was Room A, where participants tended to switch modes much more often. Since the orders of the rooms was pseudo-randomized, it is not entirely clear why Room A would prompt participants to switch modes more often than other rooms.
\subsection{User Preferences}
Upon completion of the experiment, we asked participants which mode they preferred for the OA portion of the task, the OS portion of the task, and overall (Fig.~\ref{fig:results-preference}).
Consistent with Fig.~\ref{fig:results-overall}A, participants preferred \emph{DepthOnly} and \emph{EdgesOrDepth} to avoid obstacles; $\chi^2(1, N=18) = 8.22, p<.05$.
However, \emph{EdgesAndDepth} also resulted in strong OA performance, yet was only preferred by two participants.
Most participants preferred the ability to switch between depth and edge cues by means of the \emph{EdgesOrDepth} mode.
When it came to selecting objects, user preferences were mixed (Fig.~\ref{fig:results-preference}, \emph{center}) and not significantly difference from a uniform random distribution: $\chi^2(1, N=18)=.667, p=.88$.
This is consistent with the somewhat lackluster performance of participants in the object selection portion of the task (Fig.~\ref{fig:results-overall}B), suggesting that none of the tested modes may be ideal for object recognition.
Lastly, overall preference mirrored the OA result, with most participants preferring \emph{DepthOnly} and \emph{EdgesOrDepth} mode (Fig.~\ref{fig:results-preference}, \emph{right}).
However, this result was not significantly different from a random uniform distribution, although it approached significance ($\chi^2(1, N=18)=7.33, p=.06$).
Notably, none of the subjects preferred \emph{EdgesOnly} mode overall, and only four subjects preferred \emph{EdgesAndDepth} mode.
The majority of participants was split between being comfortable using \emph{DepthOnly} mode and preferring the flexibility to switch between modes using \emph{EdgesOrDepth} mode (even though most time was spent looking at depth cues).
Overall these results further corroborate the relative importance of depth cues for both obstacle avoidance and object identification.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{figures/fig-switches.png}
\Description{Barplots describing user behavior during EdgesOrDepth mode. Top: barplots indicating that most time was spent on DepthOnly mode, in all rooms, during both obstacle avoidance and object selection. Bottom: Barplots indicating how often users toggled between EdgesOnly and DepthOnly mode across the six rooms. In obstacle avoidance, users switched between 1 and 2 times in all rooms. In object selection, people switched most in Room A.}
\caption{\emph{EdgesOrDepth}: Fraction of time spent on each mode (\emph{top}) and number of mode toggles (\emph{bottom}) for each room, for the obstacle avoidance portion of the task (\emph{left}) and the object selection portion of the task (\emph{right}).}
\label{fig:results-switching}
\end{figure}
\section{Discussion}
\subsection{Relative Importance of Depth Cues and Semantic Edges for Bionic Vision}
Here we asked sighted participants to act as virtual bionic eye recipients by navigating virtual indoor environments using simulated prosthetic vision.
To the best of our knowledge, this is the first study that directly compares the relative importance of depth cues and semantic edges for bionic vision.
\begin{figure*}[!bt]
\centering
\includegraphics[width=\linewidth]{figures/fig-preference.png}
\Description{Barplots of user preferences for the four different modes in three panels.
In the leftmost panel, users indicate their preferred mode to avoid obstacles: 8 users preferred EdgesOrDepth, 7 users preferred DepthOnly, 2 users preferred EdgesAndDepth, and 1 user preferred EdgesOnly.
In the middle panel, users indicate their preferred mode to select objects: 6 users preferred EdgesOnly mode, and the other modes received 4 votes each.
In the rightmost panel, users indicate their preferred mode overall: 7 users preferred DepthOnly, 7 users preferred EdgesOrDepth, 4 users preferred EdgesAndDepth, and no users preferred EdgesOnly.}
\caption{Reported user preferences for the different SPV modes.}
\label{fig:results-preference}
\end{figure*}
Boxplots of user response times in milliseconds for the six experimental conditions. Mean values are provided in Table 1, column 4. Interquartile ranges are typically 2-3 seconds, with the Gesture+Voice condition having a broader range of 5 seconds and the smallest being the Voice condition at 2 seconds. Voice and Face have the lowest mean values and low standard deviations. Gesture+Voice has the highest. Face+Voice has few outliers.
We found that participants were significantly better at avoiding obstacles using depth-based methods as opposed to relying on edge information alone (Fig.~\ref{fig:results-overall}A).
These results are consistent with previous studies noting the importance of highlighting nearby obstacles for orientation \& mobility tasks \cite{mccarthy_mobility_2014,lieby_substituting_2011}.
However, the exact way in which depth cues were presented seemed to carry little importance, as evidenced by similar performance of the \emph{DepthOnly}, \emph{EdgesAndDepth}, and \emph{EdgesOrDepth} modes.
Nevertheless, it is interesting that participants completed the obstacle avoidance portion of the task the fastest using \emph{DepthOnly} mode (Fig.~\ref{fig:results-overall}C), even though \emph{EdgesAndDepth} provided the same amount of depth cues at all times.
Theoretically, participants should have performed (asymptotically) the same with the \emph{EdgesOrDepth} mode, as it provides at least as much information as \emph{DepthOnly} mode.
Its purpose was to provide all the functionality of the individual modes (i.e., \emph{DepthOnly} and \emph{EdgesOnly}) while giving the user the power and flexibility to choose whichever mode they deemed best for a particular situation.
Instead, this freedom may have added unnecessary overhead (in terms of time taken to complete the task) and opened up the possibility that users may perform the task in a suboptimal mode.
This is an important \ac{HCI} insight to consider for future designs of bionic eyes that rely on computer vision for scene simplification \cite{beyeler_towards_2022}.
\subsection{Depth and Semantic Edges May Not Be Sufficient for Object Recognition}
Depth cues were also important for object selection (Fig.~\ref{fig:results-overall}B), with participants performing best using the \emph{DepthOnly} and \emph{EdgesOrDepth} modes.
However, performance was more lackluster as compared to the obstacle avoidance portion of the task.
Participants performed near chance levels using \emph{EdgesAndDepth}, and were only slightly better with the other three modes.
Only \emph{DepthOnly} and \emph{EdgesOrDepth} were significantly better than \emph{EdgesAndDepth}, and time taken to complete this subtask was similar across modes.
Overall these results suggest that presenting both edge and depth cues at once may hinder participants' ability to segregate important objects from the background.
Yet none of the tested modes seem to provide sufficient information for participants to excel at this portion of the task.
\subsection{Users Prefer Depth Information Over Edges}
The relative importance of depth information was further corroborated by Fig.~\ref{fig:results-switching}, which indicates that when participants were given a choice between edges and depth, they largely preferred to navigate the room by relying on depth information alone.
In obstacle avoidance and overall, most participants preferred \emph{DepthOnly} and \emph{EdgesOrDepth} over the other two modes (Fig.~\ref{fig:results-preference}), but opinions were split:
Approximately half of participants preferred the flexibility of switching between edges and depth, whereas the other half was content with \emph{DepthOnly} mode.
The lackluster object selection performance is also evidenced by the user preference chart, which was not significantly different from random.
Overall these results suggest that while depth information seems to be the most beneficial for obstacle avoidance and object selection, there may not be one scene simplification made that is preferred by the majority of bionic eye recipients.
\subsection{Limitations and Future Work}
Although this study effectively uses immersive \ac{VR} to address previously unanswered questions about \ac{SPV}, there are several limitations that should be addressed in future work as outlined below.
First, it is important to point out that the study was performed on sighted undergraduate students readily available at UCSB.
Their age, navigational affordances, and experience with low vision may therefore be drastically different from real bionic eye recipients, who not only tend to be older (and prolific cane users), but also receive extensive vision rehabilitation training.
Second, we acknowledge that the navigating with simulated prosthetic vision was a challenging task even for experienced users of \ac{VR}.
As a result, there were isolated cases of participants becoming reckless and charging straight ahead through the obstacle course or simply selecting the first object they encountered.
On such occasion, behavior in the virtual environment would thus be noticeably different from a real-world obstacle course, where collisions have physical consequences.
A next step in this line of research would thus be to provide the different scene simplification modes in a real-world environment.
Depth information could be collected with an RGB-D camera and the different \ac{SPV} modes could be rendered on an augmented-reality \ac{HMD}.
While an immersive virtual environment has clear advantages (e.g., safety, access to ground-truth depth information, perfect control over the layout and illumination of the visual environment), behavioral performance of virtual patients in a real-world obstacle course might be closer to the application that this work would eventually feed into.
The ultimate test, of course, would be to demonstrate that these scene simplification modes can support everyday tasks of real bionic eye users.
Interestingly, we found vast individual differences in task performance, which were not unlike those reported in the literature \cite{he_improved_2020}.
Subjects who did well with one mode tended to do well across all modes (data not shown), suggesting that some people were inherently better at adapting to prosthetic vision than others.
Future work should therefore zero in on the possible causes of these individual differences and compare them to real bionic eye users.
Studying these differences could identify training protocols that can enhance the ability of all device users.
\section{Conclusion}
In sum, this is the first study to compare the relative importance of depth cues and semantic edges for bionic vision and constitutes an essential first towards a bionic eye that uses computer vision to improve a user's scene understanding.
\begin{acks}
This work was supported by the National Institutes of Health (NIH R00 EY-029329 to MB).
\end{acks}
\bibliographystyle{ACM-Reference-Format}
|
\section{\bf Introduction}\label{intro}
\noindent{\it The world is so full of a number of things,}
\noindent{\it I'm sure we should all be as happy as kings.}
{\qquad\qquad\qquad\qquad Robert Louis Stevenson}\medskip
\noindent In the empiricist philosophy of this essay\footnote{This is a much revised version of the unpublished sentiments expressed in the draft paper, Peebles (2021).} experiments and observations have two key functions: the puzzles they reveal inspire theories, and the theories are judged by the degree of empirical success of their predictions. The relative importance of the two functions can be quite different in different cases. Phenomena played a small though certainly significant role in forming Einstein's intuition about a philosophically satisfactory theory of gravity, the general theory of relativity. It has been said that general relativity was ``there, to be recognized,'' essentially by pure thought. But in the empiricist philosophy Einstein's great accomplishment is to be particularly celebrated because the theory Einstein found passes many tests that make the empirical case that general relativity is a remarkably useful approximation to the reality whose nature we seek to discover. Phenomenology played a much greater role when Maxwell gave up seeking a mechanical model for the ether and took the intuitive step of completing the field equations that pass such broad empirical tests. Quantum physics, and the standard model for particle physics, grew in an intermediate manner, a combination of brilliant ideas and helpful phenomenology. The same is true of our present physical cosmology, which grew out of suggestive phenomenology and intuitive leaps, some of which failed while others prove to be remarkably successful. The result, the $\Lambda$CDM theory, agrees with a broad range of well-checked predictions. But the thought that there is an even better classical cosmology to be found is encouraged by intuition, as in the feeling that the stark simplicity of the dark sector of the theory surely is inadequate, and from curious issues in the phenomenology, which is the topic of this paper.
To reduce the chance of misunderstanding I emphasize that the empirical case that the $\Lambda$CDM theory is a good approximation to reality remains compelling. But I argue in this paper that we have empirical evidence that there is a still better theory to be found.
The goal of an improved cosmology, and the far greater great goal of a full reconciliation of the quantum and relativity principles, might be approached by strokes of insight, following Einstein's example, or by incremental advances, following Maxwell: searches for improvements inspired by anomalies. By this I mean empirical evidence that seems likely to disagree with what is expected from standard and accepted theories and ideas. An anomaly might prove to be only apparent, resolved by improved empirical evidence or a better appreciation of the predictions of the theory we already have. Past experience suggests that others will prove to be real, and will be a valuable stimulus to the exploration of new or previously neglected ideas about physical theory. This includes cosmology, and maybe even unification of the basic principles of physics. But a final theory, if there is such a thing, cannot be empirically established, only checked to the extent that world economies can afford.
Another consideration follows from the phenomenon of multiples in scientific discovery: apparently independent appreciations of the same idea (as reviewed by Merton 1961). Examples are common; examples in the development of the present standard cosmology are reviewed in Peebles (2020a; 2022a). Some are results of advances in technology or theory that suggest ideas that are in the literature, and could be appreciated by more than one person. Other examples are results of communication by many means, including nonverbal hints to directions of thinking, which can cause growing community recognition of ideas that are ``there, to be recognized,'' in previously neglected phenomena. This invites the hopeful thought that there are still other phenomena that might be recognized to be useful hints to the improvement of physics, if examined more closely. The thought is particularly relevant for cosmology because the universe has room for an immense number things. I argue that some seem to me to call for closer attention.
Research programs tend to pursue systematic explorations of scientific issues that the community agrees are important. An example in cosmology is the issue of cores and cusps in the distributions of dark and baryonic matter in the centers of galaxies. The issue is important, a guide to the establishment of constraints on the nature of dark matter and the theory of galaxy formation. Research in progress is examining the theory and observations. This is good science: close attention to a specific question might reveal something interesting. But it is good science also to cast about for more obscure issues that seem curious but have not yet been examined as carefully as might be useful. My purpose in this essay is to offer thoughts about anomalies in cosmology that are not so widely discussed, and might aid advances in this subject if given more attention.
Section~\ref{sec:anomalies} is a commentary on more familiar anomalies in physical science that seem relevant to cosmology. The subjects of the next three sections are anomalies, real or apparent, that are less well advertised for the most part. Section~\ref{sec:distributions} reviews open issues in the large-scale distributions of radio galaxies, AGNs and clusters of galaxies. The curious properties of the Local Void are considered in Section~\ref{sec:localvoid}, and apparent anomalies in the properties of galaxies are reviewed in Section~\ref{sec:galaxies}. Section~\ref{SummaryRemarks} presents an overview of what I consider to be the main points of this essay.
We must consider first the reliability of the basis by which phenomena are judged to be anomalous within accepted thinking. The empirical case that the relativistic hot big bang $\Lambda$CDM cosmology is a useful approximation to reality is outlined in Section~\ref{sec:EmpiricalBasis}. This theory is not exact, but it does not seem likely to be far off. Thus my assessments of issues and anomalies in cosmology take it that the tests have persuasively established that the $\Lambda$CDM theory is a good approximation to reality that likely requires refinement.
The anthropic principle is another consideration by which theories are motivated and their anomalies evaluated and maybe resolved. The line of thinking has been particularly influential in research in physical cosmology, and so deserves mention here. What is more, we must expect that social constructions, aided by the anthropic principle, will be forced on the scientific community eventually, if the research community survives that long, as theory outstrips the ability to test predictions. But I argue in Section~\ref{sec:Anthropic} that the anthropic approach is of doubtful use for the present purpose because we have immense room between the extremes of the unapproachably large and unapproachably small scales for continued empirical exploration.
My citations to the literature are limited to a few introductory papers and recent review articles, which I hope aids readability while indicating the present state of thinking about the science and the evidence. I do not comment on issues in cosmology where I have no thoughts to add to what already is in the literature.\footnote{\label{fn:S0s}Choices of interesting issues to explore in what we aim to be the advance of objective science must be subjective. The standard cosmology predicts a small but important primeval abundance of the lithium isotope $^7$Li. The prediction is clear, but I am not comfortable assessing the astrophysics of production and depletion of lithium. Planar alignments of satellites of galaxies might be significant, but I worry about our powerful ability to see patterns even in noise. Thinking about the future of the universe, maybe a big crunch or big rip, offers great adventures of the mind, but it is only empirically interesting if there is something to observe, maybe remnants of the last phase of a cyclic universe. I am particularly uneasy about my lack of attention to the S0 galaxies that are common in clusters, and present but not common among nearby galaxies. An S0 looks somewhat like a spiral galaxy whose arms have been erased leaving a disk of stars usually with a relatively large stellar halo. S0s may have readily interpretable things to teach us about cosmic structure formation, but I have not grasped them.} The declarative sentences in this essay are my opinions. For other reviews of independent selections of issues and anomalies in cosmology I refer to Abdalla, Abell{\'a}n, Aboubrahim, et al. (2022) and Perivolaropoulos and Skara (2022). A different philosophy informs the assessment of the empirical situation in cosmology by Subir Sarkar (2022).
\subsection{Empirical Basis for the Standard Cosmology}\label{sec:EmpiricalBasis}
To help clarify discussions of tests and anomalies we need a definition of the standard $\Lambda$CDM cosmology. The version used in this paper is discussed in Section~\ref{sec:definition}. Section~\ref{sec:tests} outlines the tests that establish this cosmology as a useful approximation to reality.
\subsubsection{Definition}\label{sec:definition}
Einstein's cosmological principle, or assumption, is that the universe is close homogeneous and isotropic in the large-scale average. To be more explicit about the role of this assumption in the standard $\Lambda$CDM theory used in this paper I offer the following definition. The theory applies the standard physics of matter, radiation, and Einstein's general theory of relativity with its cosmological constant to a cosmologically flat universe that is a spatially stationary, isotropic, random process with a close to scale-invariant power law power spectrum of Gaussian and adiabatic departures from homogeneity. This trimmed-down theory has eight free parameters (the density parameters in ordinary matter, dark matter, the CMB, and neutrinos with negligible rest masses; with Hubble's constant, the primeval Gaussian process amplitude and power law index, and the optical depth for scattering of the CMB by intergalactic plasma).
The mathematician and statistician Jerzy Neyman (1962) pointed out that the rational statement of Einstein's cosmological principle is that the universe is assumed to be a realization of a ``stationary stochastic process,'' or as it is put here a stationary random process. The process is what a theory with given parameters is supposed to predict. A measurement to test the theory and constrain the parameters uses a sample, which is considered a realization of the process. A fair sample is a realization that has statistical properties that are usefully close to the predictions of the process. Analyses of what is arguably a fair sample test whether the theoretical process is an adequate approximation to reality.
Neyman remarked that the empirical evaluation of a deterministic classical theory is necessarily indeterministic, because no finite sample can determine parameters to arbitrarily tight precision. To my knowledge Neyman was the first to apply this thinking to our strictly limited sample of the universe, our sample of the stationary random process of the cosmological principle. We can add another reason for the indeterministic nature of physical science: our theories are incomplete. A notable example is the inconsistency of the quantum and relativity principles (Sec.~\ref{sec:QM&GR}).
To make progress we must add to the definition of the $\Lambda$CDM theory the less specific provision that the realization of the random process that is our observable universe is close enough to a fair sample to allow reliable computations of testable predictions. Aspects of this situation are discussed in Section~\ref{sec:theacausaluiverse}. A related provision is that the primeval power law power spectrum of the primeval mass distribution must be truncated to avoid a divergence of spacetime curvature fluctuations. The measured scalar spectral index, $n_s\simeq 0.96$, is associated with primeval spacetime curvature fluctuations that scale with size $r$ as $\delta\phi\propto r^{(1-n_s)/2}\sim r^{0.02}$. The power spectrum has to be truncated at some large scale, but it will be assumed that the wavenumber at the truncation doesn't much matter. Variants of this standard picture, such as a Tilted Universe (Turner 1991), in which we might not have a fair sample, are not explicitly considered here.
Other authors use different definitions of the cosmological principle. Secrest, von Hausegger, Rameez, et al. (2022) take the principle to be ``that the universe on large scales must appear to be the same to all observers, independent of their location.'' I am indebted to Subir Sarkar for pointing out that Milne (1933) presented essentially the same definition. This is of historical interest because Milne introduced the term, Einstein's cosmological principle, and pointed out that Hubble's law follows from this principle without application of any deeper theory. But consider that fluctuations in the distributions of extragalactic objects on scales we can observe within the present Hubble length were at earlier epochs not observable. Why should there not be density fluctuations beyond the present Hubble length that are observable if at all only through their indirect effects? Neyman's philosophy accommodates this. Still more aspects of the situation are discussed in Section~\ref{sec:theacausaluiverse}.
\subsubsection{Cosmological Tests}~\label{sec:tests}
Surveys of the tests that make the case that the $\Lambda$CDM cosmology is a useful approximation to reality, and that an even better theory will look much like $\Lambda$CDM, are widely available in the literature. My contribution is in Peebles (2020a). I offer here a reminder of key points in my positive assessment of the situation.
The measured statistical patterns in the space distribution of the galaxies and in the angular distribution of the thermal cosmic microwave background radiation, the CMB, agree with what is expected from the remnants of acoustic oscillations of the plasma and radiation that acted as a fluid up to recombination at redshift $z\sim 1000$. The precision of the CMB anisotropy measurements and the tight consistency with the $\Lambda$CDM predictions is deeply impressive. But even more impressive is the consistency with the theory and observation of the pattern in the space distribution of the galaxies that in theory also is a remnant of the acoustic oscillations. This cosmological test of consistency is based on two quite different phenomena, the space distribution of the galaxies observed at redshifts less than unity and the pattern in the angular distribution of the CMB that in theory was formed at redshift $\sim 1000$. The patterns in the two distributions are measured by different methods of observation and data reduction, yet they agree with the same $\Lambda$CDM universe. This consistency is a demonstration that we have a good approximation to reality that is as close to convincing as we can get in natural science.
Other tests of consistency, in examples based on what happened over a broad range of redshifts, add weight to the case for the $\Lambda$CDM theory. The values of the cosmic mean mass densities in baryons and dark matter, the helium mass fraction, and the cosmological constant that are needed to fit the CMB anisotropy measurements that probe the state of the universe at $z\sim 1000$ agree with the baryon density that fits the formation of the isotopes of hydrogen and helium at $z\sim 10^9$, the abundances of helium in the Sun, interstellar plasma, and planetary nebulae, the dynamical mass density derived from relative motions of the galaxies at redshifts less than unity, the mass density and cosmological constant derived from the supernova redshift-magnitude relation at $z\sim 1$, the stellar evolution ages of the oldest known stars, and the angular size distance as a function of redshift at $z\lower.5ex\hbox{$\; \buildrel < \over \sim \;$} 1$.
There are discrepancies, real or apparent. The values of Hubble's constant, $H_{\rm o}$, that are needed to fit the CMB anisotropy measurements differ from the relation between distances and recession speeds of relatively nearby galaxies by about 10\%. This well-discussed Hubble tension, if real, is a 10\% error arising from tracing the expansion of the universe to the present by a factor of a thousand from the epoch of formation of the patterns in the distributions of the baryons and CMB. I count this as an impressive success to be added to the rest of the evidence that the $\Lambda$CDM theory is a useful approximation to reality, though of course not exact. If the anomaly in the two measures of $H_{\rm o}$ is real then surely other anomalies are to be found. Another tension is the evidence that the normalization of the mass fluctuation power spectrum required to fit the CMB anisotropy measurements differs from the normalization required to fit measurements at low redshifts: gravitational lensing, the galaxy two-point correlation function in redshift space, and counts of clusters of galaxies. The normalizations, measures of $\delta M/M$, again differ at high and low redshifts by 10\% (in the compilation by Perivolaropoulos and Skara 2022).
Einstein's general theory of relativity passes tight tests on the scale of the Solar System down to the laboratory, on scales $\lower.5ex\hbox{$\; \buildrel < \over \sim \;$} 10^{13}$~cm. Cosmology applies this theory at distances on the order of the Hubble length, $10^{28}$~cm. As a general policy, would you trust an extrapolation of a theory by fifteen orders of magnitude in length scale? But we have a prior example, the enormous range of scales of the successful applications of quantum physics, from superconductors to detection of the higgs boson. The success of the cosmological tests we have so far gives considerable weight to this great extrapolation of general relativity to its application in the $\Lambda$CDM theory.
The key point from these considerations is that we have a broadly consistent story from a considerable variety of ways to observe the nature of the universe by different methods of observation and analysis, by groups that are operating independently and, it is reasonable to expect, interested in finding anomalies that may prove to be interesting. It would be ridiculous to suppose that the network of tests is wrong or misleading but has converged to apparent consistency by some combination of accidental and unintended errors, let alone conspiracies. We have instead excellent reason to expect that a better theory to be discovered will look a lot like $\Lambda$CDM, because the $\Lambda$CDM universe has been shown to look a lot like what is observed. This is the basis for the conclusion that we have a useful approximation to reality, and the hope that empirical anomalies will offer hints to improvements.
The search for improvements includes exploration of the consequences of modified gravity physics. It would be a curious coincidence if the theory required modification just on reaching the scales of cosmology, however. A better immediate prospect for improvement is the physics of the dark sector of $\Lambda$ and dark matter, which seems artificially simple.
We cannot prove as a theorem that no other physical theory could predict the reasonable degree of consistency of theory and observation of these many different ways to probe the universe; we do not do theorems in natural science. But we can conclude from these tests that we have a compelling case that the $\Lambda$CDM theory is a useful approximation to reality, and we can hope to see improvements of the theory as the observations improve.
\subsection{The Anthropic Principle}\label{sec:Anthropic}
If society continues to be willing and able to support curiosity-driven research in the natural sciences there will come a time when physicists have found a final theory of everything that is internally consistent and agrees with all available tests, but the assessment by tests of predictions will be impossible. That would be impossible in principle, if something like a multiverse is involved, or impossible in practice, because the world economy cannot afford tests of the predictions of this final theory. It will be reasonable and sensible to consider this theory to be a persuasive nonempirical establishment of reality (Dawid 2013). But it will be sensible also not to be quite sure of what must be a social construction (Peebles 2022).
We have a precursor to this dilemma, the anthropic principle. It offers one way to deal with an anomaly: postulate an ensemble of all possible universes and observe that we could flourish only in one suited to our needs expressed so as to account for the anomaly. Reactions to this line of thought differ. Some dismiss it as a ``just so'' story. Steven Weinberg (1989) pointed out that it is one way to account for the quantum vacuum energy density, which looks likely to be quite unacceptable large. This section is meant to explain my feeling that the anthropic principle is not an appropriate guide to considerations of anomalies in physical cosmology.
Robert Henry Dicke (1961) introduced a weak form of this argument. Dicke pointed out that the universe has to have been expanding for at least a few gigayears. The time is needed to allow for the evolution of several generations of stars that produced the heavy elements we need, then the formation and cooling of the solar system, and then the evolution of the species up to observers who take an interest in the expanding universe. This is a consistency condition. Better put, it is the assumption that Nature abhors logical inconsistencies.
An argument based on a more adventurous form of the anthropic principle starts from the evidence that there are enormous numbers of planets around stars in our galaxy. This allows room for many planets capable of hosting beings similar to us. The frequency distribution in cosmic times when these beings flourish on different planets might be expected to peak at about $10^{10}$~yr, because this allows time for natural evolution while avoiding the serious slowing of star formation at much greater cosmic times. This time is about what is observed on our planet; we flourish about when might be expected. An empirical test from a modest sampling of what is on nearby planetary systems might be possible, eventually.
Weinberg (1989) discussed a stronger form that postulates a statistical ensemble of universes, a multiverse. If, for example, universes in the ensemble that have shorter expansion times are more numerous, then the odds are that we live in one of the universes with the minimum expansion time consistent with what is required to allow our existence, which seems about right. Weinberg applied this thinking to the curiously small value of the cosmological constant compared to what is expected from quantum physics. Weinberg postulated that the laws of physics in each universe in the ensemble would be different. We could only flourish in a universe with physics similar to ours on the level we require, but that degree of similarity could allow a broad spread of values of the quantum vacuum energy density, $\Lambda$, provided that that depends on deeper physics that does not affect our well-being. There would be universes in the multiverse that satisfy this condition and the value of $\Lambda$ is not so negative that the universe stops expanding and collapses too soon for the span of time we required, and not so positive and large that the rapid expansion driven by $\Lambda$ would have prevented the gravitational assembly of galaxies. If galaxies in the ensemble that have larger absolute values of $\Lambda$ are more common, as might be expected from the large value expected of the quantum vacuum, then we would expect to find ourselves in a universe with a value of $\Lambda$ that is about as large as is consistent with our existence. This is about what is observed.
Martin Rees (2020) rightly celebrates the concept of the multiverse as the next layer in the sequence of revolutions in our understanding of the nature of the world around us. Ideas about this have passed through many layers: the Ptolemy universe with the earth centered in the crystal spheres that hold the astronomical objects; the Copernican universe with the sun at the center; the Kapteyn universe centered on the Milky Way galaxy; Hubble's realm of the nebulae with no center; and the multiverse of which our universe is but a speck. The layers of discovery go down in scale too. Henri Poincar\'e (1902) remarked that the Mariotte/Boyle law is wonderfully simple and accurate for many gases, but these gases examined in sufficiently fine detail break up into the complex motions of enormous numbers of particles. Poincar\'e asked whether gravity examined in sufficiently fine detail might also depart from the simplicity of Newton's law into complex behavior. Poincar\'e suggested we consider that ``then again [there may be] the simple under the complex, and so on, without our being able to foresee what will be the last term.'' Maybe underlying the particle physicists' concept of a theory of everything are yet more layers of Poincar\'e's successive approximations. And why should the layers of structure on large scales not continue to multiverses and beyond?
Weinberg (1989) cautioned that the anthropic upper bound on the absolute value of $\Lambda$ is well above the bound that could be set by astronomical observations we had then. There are other examples of what seem to be excessive satisfaction of the anthropic condition. We need at least one gravitational potential well similar to that of a galaxy to contain and recycle the debris from a few generations of stars to have produced the chemical elements we require. But did we need the observed enormous number of galaxies? Would the already large number of planetary systems among the $\sim 10^{11}$ stars in the Milky Way have been adequate? If the Milky Way would serve, and given that it is present, does physics require all those other galaxies? In implementations of cosmological inflation the numbers of galaxies, and their sizes and densities, depend on the amplitude of the primeval fluctuations in spacetime curvature associated with the primeval departures from an exactly homogeneous mass distribution. A universe identical to ours except that the fluctuations in spacetime curvature are an order of magnitude smaller would develop far fewer galaxies that are less dense, but would that be a problem for our existence? We have so many galaxies to spare. If in the multiverse the universes with smaller primeval curvature fluctuations were more common then application of the anthropic consideration would lead us to expect far fewer galaxies than observed; we don't need so many. If universes with larger primeval curvature fluctuations were more common we would expect to find ourselves in an accidental island of tranquility among the chaos of violent mergers and relativistic collapses. Again, neither situation is observed.
Weinberg's argument is good science; it explores a possible aspect of reality that accounts for the great difference between the value of the $\Lambda$ of cosmology and the value expected from quantum physics. But there are troubling aspects of this approach. The excess of baryons over antibaryons in the Local Group could be attributed to the anthropic principle: we live in a universe drawn from the multiverse that has the excess baryon density acceptable for our existence. But theories of particle physics and physical conditions in the early universe might predict the excess. Given the choice of seeking this better physics or resorting to the anthropic principle, I expect most would choose the former. It can be awkward to base arguments on the Panglossian principle that we inhabit the best of all possible universes suited to our existence.
\section{Anomalies in Physical Science}\label{sec:anomalies}
Some anomalies that have long resisted interpretation are so familiar that they tend to pass without mention. An example is Wigner's (1960) ``Unreasonable Effectiveness of Mathematics in the Natural Sciences.''
\subsection{Physics is Successful} \label{PhysicsisSuccessful}
Eugene Paul Wigner (1960) wrote about the ``two miracles of the existence of laws of nature and of the human mind's capacity to divine them.'' Natural scientists are conditioned to accept these two miracles, or phenomena, as self-evident; they are essential for the discoveries of well-tested science that makes possible the vast range of technology we all experience. But we should pause on occasion to consider that Wigner's two miracles are assumptions. Experience supports them, but of course never offers a proof.
The starting miracle to add to these two is that the world around us exists, and has existed in some form or another for a very long time, with the properties one would expect of physical reality. In natural science we must take this as given. We assume then Wigner's two miracles and notice that they satisfy two consistency conditions. First, if macroscopic physics were not reproducible and lawful it would be difficult to imagine the natural evolution of the species: what is the use of adapting to physical properties of matter that can change on time scales less than cosmic? Second, if it were not possible to discover useful approximations to the laws of nature then I suppose we would not be marveling about it. These thoughts could be taken to suggest that the assumption of lawful behavior, which is such a good approximation, might be found to fail as we probe ever deeply into the nature of the world, at levels where erratic behavior need not have had a deleterious effect on living matter. It calls to mind quantum physics.
\subsection{The Relativistic and Quantum Principles are Inconsistent}\label{sec:QM&GR}
In the empiricist philosophy the use of quantum physics to account for the spectra of galaxies at redshift $z=10$ and the formation of the light isotopes at $z\sim1000$, applied in the strongly curved spacetime of relativity, need not be problematic even though the quantum and relativity principles are not consistent. Maybe this is a signal of an intrinsic inconsistency; maybe physics is not exactly lawful. The improbably large estimate of the quantum vacuum mass density discussed in Section~\ref{sec:Lambda} is a real problem. More abstract, but a worry, is whether the quantum physics of observables operating on state vectors is a useful approximation when extrapolated to describe the whole universe. Would a quantum universe in a pure state really decohere into the classical world of general relativity? Would a viable theory of our universe taken to be a mixed state be any different from our present theory? A reasoned assessment of such issues is beyond my ability and the scope of this essay.
\subsection{The Symmetry of Matter and Antimatter is Broken}\label{sec:baryonnumber}
A familiar anomaly that is essential to our existence is the pronounced local excess of matter over antimatter. The sensitive tests for anti-helium by the Alpha Magnetic Spectrometer indicate the Milky Way contains little antimatter (Poulin, Salati, Cholis, Kamionkowski, and Silk 2019; Aguilar, Ali Cavasonza, Ambrosi, et al., 2021). We can add the evidence from the absence of detection of gamma ray annihilation radiation from dwarf galaxies that have plunged into the Milky Way. And since satellites of the Milky Way and our neighbor M~31 surely have intermingled before falling into one or the other of the large galaxies without producing detectable gamma rays it seems clear that the Local Group is made of baryons, with a tiny fraction of antibaryons.
Should we revisit the question of whether some galaxies are made of anti-baryons? A sharp division of regions of matter and antimatter could invite an unacceptably large surface mass density in domain walls, but an imaginative physicist might find a way around that. The literature on possible extensions of the standard model for particle physics and the physical conditions that would account for baryogenesis continues to grow. I am not competent to review the state of the art.
\subsection{Should Local Physics be Evolving?}\label{sec:EvolutionPhysics}
We have been given leave by string theory to imagine that the dimensionless parameters of physics are evolving, because in the varieties of string theory it is difficult to see what fixes their values. Thus Uzan (2003, Sec.~VI.B) concludes that ``as yet no complete and satisfactory mechanism for the stabilization of the extra dimension and dilaton is known," and with it stabilization of the dimensionless parameters of fundamental physics. This complements an older thought based on a measure of the strength of the gravitational interaction,
\begin{equation}
{\cal G} = {Gm^2\over \hbar c}\sim 10^{-38}, \label{gravitystrength}
\end{equation}
where $m$ is the mass of a nucleon. This number is remarkably small, one might say anomalously so. An anthropic explanation is that if ${\cal G}$ were much larger it would make stellar evolution times too short to allow for evolution of the species by natural selection, even on a suitably small planet near a suitably low mass long-lived star.
Dirac (1937) argued that it is difficult to imagine how the tiny value of ${\cal G}$ could follow from a fundamental theory that might be expected to produce numbers such as $\pi$ and $e$, and integers of modest size and their fractional powers. Dirac suggested that a hint might be drawn from the comparison of ${\cal G}$ to the ratio of the atomic time $e^2/mc^3$ to the Hubble expansion time, then estimated to be $t \sim 2\times 10^9$~yr, giving the dimensionless number
\begin{equation}
{\cal T} = {e^2\over t~mc^3} \sim 10^{-38}. \label{atomictime}
\end{equation}
It is curious that the two exceedingly small numbers, ${\cal G}$ and ${\cal T}$, are similar. The value of ${\cal T}$ is decreasing, assuming the universe is evolving and local physics is not. Maybe ${\cal G}$ is small, and comparable to ${\cal T}$, because ${\cal G}$ is evolving to its natural value, zero, along with ${\cal T}$.
The measure of the electromagnetic interaction,
\begin{equation}
\alpha = {e^2\over\hbar c}\sim {1\over 137}, \label{eq:alpha}
\end{equation}
is not such a small number. But if ${\cal G}$ is not constant then $\alpha$ surely need not be fixed either.
Einstein's cosmological constant, $\Lambda$, represents the effective vacuum energy density,
\begin{equation}
\rho_\Lambda = {3 H_{\rm o}^2(1 - \Omega_m)\over 8 \pi G}
\sim 10^{-29}$~g~cm$^{-3}.
\end{equation}
The ratio of this quantity to the Planck mass density, $\rho_{\rm Planck}=c^5/\hbar G^2$, is
\begin{equation}
{\cal L} = {\rho_{\Lambda}\over\rho_{\rm Planck}} \sim 10^{-123}.
\end{equation}
The empirical evidence is that $\cal L$ is not zero (as reviewed in Secs.~\ref{sec:tests} and \ref{sec:Lambda}), but rather this dauntingly small number. Maybe it calls for application of the anthropic principle, as Weinberg (1989) argued. Or, following Dirac's thinking, maybe this number is so small because local physics has been evolving for a long time, and with it the value of $\cal L$.
Dicke was fascinated by the thought that the laws of physics might change as the universe evolves. My impression is that this was at least in part because checking on the idea led to fascinating explorations of the many possible sources of empirical evidence for or against evolution, drawn from the laboratory, geology, astrophysics, and on to cosmology, always with close attention to the empirical content (e.g. Dicke 1964). My PhD dissertation under his direction was on the theory and empirical constraints on the evolution of the strength $\alpha$ of the electromagnetic interaction. Dicke played an important role in setting up the Lunar Laser Ranging experiment that has yielded tight tests of physics, including the evolution of ${\cal G}$. The experiment has established that at the present epoch the strength of the gravitational interaction is not evolving faster than about one percent of the rate of expansion of the universe (Williams, Turyshev, and Boggs 2004). Dicke's program led Bharat Ratra and me to explore the thought that the value of the cosmological constant $\Lambda$ is so small compared to what is expected from quantum physics because $\Lambda$ is not a constant: it has been slowly evolving to its only reasonable value, zero. Ratra and Peebles (1988) present this thought and a model in which it happens.
The search for evidence of evolution of the effective value of the vacuum energy density $\Lambda$ under its new name, dark energy, is now widely discussed and well supported; an example is the Dark Energy Survey, DES (DES Collaboration 2022). It is equally good science to test whether the strength of the electromagnetic interaction, $\alpha = e^2/\hbar c$, might evolve. Advances in technology to test this are discussed in recent papers by Murphy, Molaro, Leite, et al.(2022) and Webb, Lee, and Milakovi{\'c} (2022). So why isn't there a well-supported Fine-Structure Survey, FSS?
The challenge posed by Dirac (1937) and Dicke (1964) remains: discover what accounts for the measure of the strength ${\cal G}$ of the gravitational interaction in equation~(\ref{gravitystrength}) that is so far from what might be expected to be predicted by a theory of everything. A challenge of the same sort is to account for the great difference between the values of the effective vacuum energy density $\Lambda$ of cosmology and the Planck mass density suggested by quantum physics (Sec.\ref{sec:Lambda}). Maybe these pronounced anomalies will be explained by subtle aspects of a deeper theory that predicts the tiny values of ${\cal G}$ and $\cal L$. Maybe we must resort to the anthropic philosophy. Or maybe both ${\cal G}$ and $\cal L$ have been decreasing, ${\cal G}$ following a course of evolution that happens to have escaped the tight constraint from the Lunar Laser Ranging experiment.
\subsection{The Standard Cosmology is Singular}
The $\Lambda$CDM theory defined in Section~\ref{sec:definition} predicts that the expansion of the universe traces back to a singular state of arbitrarily large density. It is usually supposed that Nature abhors singularities, and that this one necessarily points to incompleteness to be remedied by a better theory. The most widely discussed possible extension of $\Lambda$CDM is cosmological inflation (Guth 1981). It removes the singularity, but Borde, Guth, and Vilenkin (2003) argue that even eternal inflation cannot have been eternal back to the arbitrarily remote past. This is not an argument against inflation, eternal or otherwise, only that the term, eternal, cannot be the whole picture. We make progress by successive approximations.
When the concept of cosmological inflation first gained general attention an implication was taken to be that space sections are flat with close to Gaussian adiabatic departures from homogeneity and a power law power spectrum slightly tilted from scale-invariance. This was before each of these conditions were observationally established, and it no longer matters whether inflation really predicts all this. In the empiricist philosophy of this essay this history makes inflation particularly interesting, though of course the empirical case is not yet persuasive because tests of predictions are still scant.
\subsection{The Standard Cosmology is Acausal}\label{sec:theacausaluiverse}
Hubble's (1936) ``Realm of the Nebulae,'' the observed galaxies, does not have a noticeable edge. This is consistent with the $\Lambda$CDM theory defined in Section~\ref{sec:definition}, but it is an anomaly because in this theory distant galaxies observed in well-separated parts of the sky have not been in causal contact no matter how far back in time the expansion is traced, to the singularity. So how did the galaxies ``know'' how to resemble each other?
The resolution offered by the cosmological inflation picture is that there was a time in the early universe when a near exponential rate of expansion produced a far larger horizon, resulting in causal connection across all we can see. The successful empirical tests of cosmology outlined in Section~(\ref{sec:tests}), which depend on this acausality, are in turn evidence of the effect of inflation in some form. At the time of writing, however, we have only a modest empirical basis for assessments of specific theories of this aspect of what happened in the remote past.
In the standard theory the gravitational growth of primeval departures from an exactly homogeneous mass distribution, which is discussed in Section~\ref{sec:localgravity}, is acausal, as follows. Let the universe be cosmologically flat, ignore the mass in radiation and the pressure of matter, and assign time-orthogonal coordinates that eliminate the decaying mode of the departure from a homogeneous mass distribution. Then in linear perturbation theory the motion of the matter relative to the general expansion of the universe is
\begin{equation}
v^\alpha(\vec x,t) = {Ha(t)f(\Omega)\over 4\pi}
{\partial\over\partial x^\alpha} \int d^3x'{\delta(\vec x',t)\over |\vec x' - \vec x|}.
\label{eq:peculiaracceleration}
\end{equation}
Here $H$ is Hubble's constant, $a(t)$ is the expansion parameter, $f\simeq \Omega^{0.6}$ where $\Omega$ is the density parameter, and $\delta(\vec x,t)$ is the fractional departure from a homogeneous mass distribution. We see that in the gravity physics of the standard cosmology a mass concentration $\delta(\vec x',t)$ that is in principle too far away to be observed can produce a flow $v^\alpha(\vec x,t)$ that in principle can be observed. Grishchuk and Zel'dovich (1978) may have been the first to recognize this.
The situation illustrated in equation~(\ref{eq:peculiaracceleration}) is acausal, but we have learned to live with it by placing it in the initial condition that the universe is a stationary random process. This means that in the early universe the cosmic structure we observe would have been waves in the mass distribution with tiny amplitude and spread across far broader lengths than the particle horizon. Put another way, in standard $\Lambda$CDM the statistical homogeneity of the universe and its primeval power spectrum of departures from homogeneity have acausal origins.
Anther aspect of the acausality of our present cosmology arises in the theory and observation of the angular distribution of the thermal cosmic background radiation temperature $T(\theta, \phi)$ as a function of angular position across the sky. It is conveniently expressed as the spherical harmonic expansion
\begin{equation}
T(\theta, \phi) = \sum a_\ell^mY_\ell^m(\theta, \phi). \label{eq:harmonicexp}
\end{equation}
In the standard theory the real and imaginary parts of the expansion coefficient $a_\ell^m$, if measured in many different realizations of the universe, would have gaussian distributions around zero mean. On small scales, large degree $\ell$, there are in effect many observations of realizations of the random process across the sky, so the measurements of the absolute values $|a_\ell^m|^2$, averaged over $m$, have little scatter in the prediction and measurement. This allows the impressively close checks of the $\Lambda$CDM theory over a large range of values of $\ell$. The situation is different at small $\ell$, large angular scale, because the $|a_\ell^m|^2$ are averaged over only a few $m$. The result is an uncertain measurement to be compared to an uncertain prediction that depends on what is in parts of the universe we cannot observe in principle.
\subsection{Why Dark Matter?}\label{DM}
The starting idea for nonbaryonic dark matter was that it might be a new family of neutrinos with the standard V-A coupling to their leptons and neutrino rest mass $\sim 3$~Gev. The mass was chosen so the remnant abundance from the thermal production and annihilation of these hypothetical neutrinos in the hot early universe would be interesting for cosmology. It was introduced, independently as far as I know, in five papers: Hut (1977); Lee and Weinberg (1977); Sato and Kobayashi (1977); Dicus, Kolb, and Teplitz (1977); and Vysotskij, Dolgov, and Zel'dovich (1977). This candidate for dark matter has become known as weakly interacting massive particles, or WIMPs. Discussions of possible detection of WIMPs began with Steigman, Sarazin, Quintana, and Faulkner (1978), who pointed out that interactions of a sea of massive neutrinos with themselves and the baryons might allow accumulation of these neutrinos in stars and planets, with observable effects, as in what became known as dark stars. Cold dark matter was more formally added to the relativistic cosmological model by Peebles (1982), for the purpose of showing how the smooth distribution of the CMB can be reconciled with the clumpy distribution of the galaxies in the picture of gravitational formation of cosmic structure.
The first laboratory attempts to detect interactions of WIMPs with ordinary matter began in the 1980s. Schumann (2019) reviews the considerable advances in sensitivity since then in a considerable variety of experimental designs, mass scales, and effective isolation from cosmic rays and other local noise. There are other candidates for dark matter, with active research aimed at their detection. Examples are the lightest supersymmetric partner, axions, fuzzy dark matter, supersymmetric dark matter, and black holes. The searches for signs of these objects and more, in the laboratory and astronomical observations, have been energetically pursued for four decades, and at the time of writing they have not yielded a generally accepted detection of any form of dark matter (apart from the small contributions by the neutrinos in the three known lepton families). Several aspects of this situation are to be considered.
First, the absence of detection despite many years of great effort does not falsify the assumption of nonbaryonic dark matter in the $\Lambda$CDM theory. This dark matter is an essential postulate, but the theory places no requirement on its place in an extended standard model for particle physics. More broadly put, we must bear in mind that we have no guarantee that we can discover how to fit well-established phenomena into satisfactory theories. The success so far in turning anomalies into physical theories that pass challenging tests has been so productive that the failure to detect dark matter despite considerable effort might be considered anomalous on historical grounds. But to repeat: the failure of detection would conflict with experience in physics, but it would not conflict with the $\Lambda$CDM theory.
Second, we now have in physics two types of matter, baryonic and dark. Properties of the former are known in great detail; properties of the latter are only roughly constrained. Ideas about how baryonic matter formed look promising but have not yet converged on a generally accepted theory (Sec.\ref{sec:baryonnumber}). Ideas about how dark matter formed must be fragmented because they depend on how dark matter fits the rest of particle physics, which is not known. Since dark matter does not seem to be necessary for our existence the Panglossian philosophy of the anthropic principle discussed in Section~\ref{sec:Anthropic} would have it that the creation of dark matter must necessarily have accompanied the creation of the baryonic matter we certainly need. It could account for the otherwise curious coincidence of the comparable amounts of baryonic and dark matter, as in asymmetric dark matter (e.g., Zurek 2014).
Third, the elegance and predictive power of the rich physics of known matter inspires the thought, or hope, that the dark sector surely is more interesting than a gas of freely moving particles with initially small velocity dispersion along with a constant mass density that has a distinctly odd value. It makes sense to explore the idea that the physics of the dark sector resembles elements of our established particle physics. Influential examples include the Sommerfeld/Coulomb enhancement of scattering by a Yukawa potential, in the influential paper by Arkani-Hamed, Finkbeiner, Slatyer, and Weiner (2009); and the scalar field equation~(2) for Fuzzy Dark Matter in the influential discussion by Hui, Ostriker, Tremaine, and Witten (2017). But there are other possibilities. Maybe the dark matter is in whole or part black holes that were present before the earliest stars. The thought has a long history (e.g., Zel'dovich and Novikov, 1967; Carr and Hawking, 1974) and continues to look interesting (e.g., Carr, K{\"u}hnel, and Sandstad, 2016; Cappelluti, Hasinger, Natarajan, 2022). Or maybe dark matter is something new.
Do galaxies always contain dark matter? Since most of the dark matter is in the outskirts of a galaxy, possible exceptions would be galaxies that have been tidally stripped, and the dwarfs that might have formed by dissipative settling from tidal streams. Apart from such effects, dark halos are expected to be universal in the standard $\Lambda$CDM cosmology. There are possibly interesting challenges to this prediction. The large S0 galaxy NGC~3115 is in the field, which is unusual because most S0s at this luminosity are in clusters. If it has a dark matter halo with mass typical of its stellar mass then the halo of NGC~3115 must be much less dense than usual, the halo core radius much broader (Cappellari, Romanowsky, Brodie, et al. 2015). The low surface brightness satellites NGC1052-DF2 and NGC1052-DF4 of the elliptical NGC~1052 also look like exceptions. Keim, van Dokkum, Danieli, et al. (2022) argue that ``the dark matter halo masses of these galaxies cannot be much greater than their stellar masses.'' It is too soon to declare a challenge to the $\Lambda$CDM theory from the evidence of galaxies with little dark matter, but the development of the empirical evidence will be worth following.
What use is dark matter anyway? We could live on a planet in a solar system in a universe that is identical to ours except that the matter is all baryonic in standard forms, with no dark matter. The larger baryon density would result in a lower residual ionization and a larger molecular hydrogen abundance at decoupling, and the onset of galaxy formation would be delayed by the coupling of all matter to the CMB up to redshift $z\sim 1000$. I am not aware of analyses of the effects on the formation of stars and planets in young galaxies, but there are so many stars in our universe that I expect this alternative universe without dark matter would have ample homes for observers. We could live in a universe similar to ours except that the baryon mass fraction is much smaller, though not zero. Gravity would gather dark matter in halos similar to those of our galaxies, but with far fewer baryons. The dissipation of energy by these baryons would be slowed by the smaller baryon density, though aided by the larger residual ionization allowed by the lower baryon density. If this universe continued expanding into the sufficiently remote future then the baryons in a massive dark halo such as the one around the Milky Way would eventually lose enough energy to become dense enough to collapse to stars and planets and observers. These observers would see far fewer galaxies forming stars, but it is difficult to see how that would adversely affect their well-being. In short, our presence does not seem require an anthropic explanation of the dark matter. Maybe its presence is purely accidental. Maybe it is an anomaly to be resolved.
\subsection{Why Dark Energy?}\label{sec:Lambda}
We have evidence of detection of Einstein's Cosmological Constant, $\Lambda$, from the BAO signature in the CMB angular distribution; the consistent BAO signature in the galaxy spatial distribution; the supernova redshift-magnitude relation; the comparison of stellar evolution ages and the cosmic expansion time; and the dynamical measurements of the cosmic mean mass density. If these pieces of evidence were seriously wrong the consistency of the $\Lambda$CDM cosmology with these very different ways to probe the universe would be far more improbable than most of us would be willing to consider. The community conclusion is instead that we have a compelling empirical case for the presence of something that acts as Einstein's $\Lambda$. This is an argument of reasonableness, of course, not a theorem.
In the standard cosmology we flourish not long after the cosmological constant $\Lambda$ and the mean mass density in matter made equal contributions to the expansion rate. This curious, one might say unlikely, coincidence used to be considered a good argument against the presence of the $\Lambda$ term and for the scale-invariant Einstein-de~Sitter model in which the universe is expanding at escape speed whenever we happen to measure it. The argument has been falsified; we must learn to live with $\Lambda$. The anthropic principle accounts for the coincidence, at least broadly, by the argument that we are in a universe in the multiverse in which the absolute value of $\Lambda$ is about as large as is consistent with our existence (Weinberg 1989). Must we leave it at that? The issue is pressing because it proves to be difficult to see another way out of the expectation that the quantum vacuum mass density is quite unacceptably large.
It is worth reviewing the case for reality of the quantum zero-point energy. Consistency of the theory and measurements of binding energies of molecules requires taking account of zero-point energies. The well-tested consistency of energy and active and passive gravitational masses requires that this real zero-point energy of matter gravitates. The same quantum and gravity physics applies to the electromagnetic field. But the sum of the electromagnetic zero-point energies over laboratory wavelengths amounts to a gravitating mass density far greater than allowed in a relativistic cosmology. Jordan and Pauli (1928) recognized the problem and proposed that the zero-point energy of the electromagnetic field is not real. This was despite the empirical evidence they had of the reality of the zero-point energy of matter fields. How could they, and we, justify distinguishing between zero-point energies based on the same physical theory?
There also are the positive and negative zero-point energies of all the other fields of particle physics, with prescriptions for ultraviolet truncation, to be added to the contributions to the stress-energy tensor by field binding energies. This looks challenging to get right, but the sum surely is vastly different from an acceptable mass density in the relativistic cosmology.
Standard physics is independent of the velocity of the observer. If this is true of the quantum vacuum energy then in general relativity its stress-energy tensor must be of the form $g_{\mu\nu}\Lambda_{\rm qm}$, where $\Lambda_{\rm qm}$ is a constant. This is the form of Einstein's cosmological constant. It would be an elegant result except for the woeful difference between $\Lambda_{\rm qm}$ and the empirical $\Lambda$.
What are we to make of this? Maybe a symmetry principle to be discovered forces the value of the quantum vacuum energy density to vanish, the only natural and reasonable value. Then the cosmological $\Lambda$ would have to be a new parameter of nature. One could instead turn to the anthropic argument discussed in Section~\ref{sec:Anthropic}. Or maybe the value of $\Lambda$ is decreasing from the large value postulated in the inflation picture, and is approaching its natural value, zero, but slowly enough that we flourish while $\Lambda$ still is slightly different from zero. This is discussed further in Section~\ref{sec:EvolutionPhysics}. These thoughts are awkward. Maybe nature agrees and has something better for us to find.
\subsection{Why Magnetic Fields; Why not Cosmic Strings?}
The existence of cosmic magnetic fields is clear; it ranks as an anomaly because of the difficulty of accounting for its presence (e.g. Wielebinski and Beck 2005). Cosmic strings and other topological defects are anomalous because they are natural extensions of standard particle physics, yet they are not part of the standard $\Lambda$CDM theory. As as been said of particle theory, ``what is not forbidden is required.'' This hint from a proven theory is not to be lightly disregarded.
The magnetic field threaded through the Milky Way is made visible by the tendency of interstellar dust particles to be aligned with long axes either perpendicular or parallel to the local magnetic field. The dust absorbs starlight and reradiates the energy at longer wavelengths, preferentially with the electric field of the radiation parallel to the alignment of the dust. The effect is observed in the polarization of starlight that passes through dust clouds and is partially absorbed, and it is observed in the polarization of the radiation reemitted by the dust at longer wavelengths. I recommend looking at the wonderful map of the magnetic field threading the Milky Way from observations of this polarized radiation obtained by the ESA Planck Satellite.\footnote{Click on \url{https://www.esa.int/ESA_Multimedia/Missions/Planck/(result_type)/images} and scroll down.} At greater distances and on larger scales the evidence of magnetic fields is less direct and an active line of research.
The formation of cosmic magnetic fields might be understood within standard physics applied to what is known of the astrophysics (e.g., Daly and Loeb 1990; Kulsrud and Zweibel 2008; Durrer and Neronov 2013; Garaldi, Pakmor, and Springel 2021); or maybe we need new physics (e.g., Ratra 1992; Widrow, Ryu, Schleicher, et al. 2012). Or maybe magnetic fields grew out of fossils from the universe before the big bang, whatever that means.
The thought that cosmic strings and other field defects are natural extensions of the standard model for particle physics was persuasive enough to motivate considerable research on how cosmic strings might be observable and might play a role in the formation of cosmic structure (e.g. Kibble 1980; Vilenkin and Shellard 2000). Although cosmic strings are absent in the present standard cosmology it remains important to look for their effects in cosmic structure, the angular distribution of the CMB, and the gravitational waves produced by cosmic strings. Vachaspati (2021) reviews the present state of this art. Ostriker, Thompson, and Witten (1986) introduced the fascinating idea of magnetized superconducting cosmic strings; maybe they hold the secret to where cosmic magnetic fields came from. And maybe cosmic strings have something to do with the curiosities in the large-scale distributions of AGNs and rich clusters of galaxies that are discussed in Section~\ref{sec:distributions}.
\section{Large-Scale Distributions of Radio Galaxies, Quasars, and Clusters of Galaxies}\label{sec:distributions}
Analytic estimates and numerical simulations of cosmic structure formation in the $\Lambda$CDM cosmology show the growth of mass concentrations that are good approximations to observed rich clusters of galaxies. There are discrepancies in the cosmological parameters that best fit cluster counts and best fit the other constraints, but they are small (e.g., Perivolaropoulos and Skara 2022, Table~2). I would count this as a success for the standard cosmology if there were not the curious distributions of clusters and powerful radio galaxies at redshifts $z\lower.5ex\hbox{$\; \buildrel < \over \sim \;$} 0.02$, and the distributions of radio galaxies and quasars at distances approaching the Hubble length. A standard interpretation is that these unusual objects form where the ambient mass density is unusually large. The evidence that there is more to it is reviewed in Section~\ref{sec:LSC}, on the situation in the region around us some 170~Mpc across, and in Section~\ref{CosPrin}, on scales closer to the Hubble length. A summary assessment is presented in Section~\ref{sec:remarks}.
\subsection{The Extended Local Supercluster}\label{sec:LSC}
G\'erard de Vaucouleurs' (1953) Local Supercluster is observed as concentrations of relatively nearby galaxies near the great circle across the sky that defines the plane of the Local Supercluster. It includes the concentrations of galaxies in our Local Group and in and around the Virgo Cluster of galaxies at about 18 Mpc distance. The pronounced presence of galaxies near this plane extends to perhaps 30 Mpc. To be considered here is the distributions of objects beyond that distance, in the region at redshifts between $z=0.01$ to 0.02, or distances about 45 to 85~Mpc at Hubble's constant
\begin{equation}
H_{\rm o} = 70\hbox{ km s}^{-1}\hbox{ Mpc}^{-1}.
\end{equation}
The lower bound on distance removes our special situation close to the plane of the Local Supercluster. The upper bound defines a region about $170$~Mpc across, with the central one eighth of the volume removed. The galaxies are distributed in clumps that look fairly close to uniformly scattered across this region. But the great clusters of galaxies, and the galaxies that are powerful radio sources, tend to be near the extension of the plane of the Local Supercluster. Tully (1986) and Tully, Scaramella, Vettolani, and Zamorani (1992) pointed out this effect for clusters, and Shaver and Pierre (1989) found the same effect for radio galaxies. Shaver (1991) remarked on the key point, the distinct difference from the far weaker concentration of the general population of galaxies to this plane. This interesting difference is not widely advertised, but it it well established and illustrated in Figure~\ref{fig:LSCf}.
\begin{figure}
\begin{center}
\includegraphics[angle=0,width=3.25in]{LSC.pdf}
\caption{Angular distributions of the sine of the supergalactic latitude of objects at redshifts $z=0.01$ to $0.02$ and galactic latitudes $|b|>10^\circ$. The red histograms are the mean for an isotropic distribution. Panel (a) shows the distributions of the 30 clusters of galaxies in this region plotted in black and the 32 most powerful radio galaxies in blue and shifted slightly to the right. Panel (b) shows the distributions of the 29 galaxies that are most luminous at $60\mu$ plotted in black, and the next 32 most luminous in blue. The counts in the vertical axes are scaled to a common effective area of 30 objects under each histogram. (Figs. adapted from Peebles 2022b.)}\label{fig:LSCf}
\end{center}
\end{figure}
The supergalactic latitude SGB of an object is the angular distance from the great circle defined by the plane of the Local Supercluster. The two panels in Figure~\ref{fig:LSCf} show distributions of the counts of angular positions of objects in equal intervals in sin~SGB, which are equal intervals of solid angle. The data are truncated at galactic latitude $|b|=10^\circ$ to take account of obscuration and confusion near the plane of the Milky Way Galaxy. This reduces the solid angles of the samples, largely at high supergalactic latitudes, SGB close to $\pm 90^\circ$. The effect is seen in the red histograms that show the mean of a random isotropic distribution of points at $|b|>10^\circ$. These red histograms are nearly flat, but suppressed at high supergalactic latitudes by the absence of objects at low galactic latitudes.
The black histogram in Panel (a) in Figure~\ref{fig:LSCf} is the distribution in sin~SGB of the 30 clusters of galaxies at $0.01 < z < 0.02$ detected as X-ray sources (from the NASA HEASARC compilation of clusters detected by the X-ray luminosity of the hot intracluster plasma). The clusters are close to the plane of the Local Supercluster as indicated by the peak at low angle SGB. This is to be compared to the red curve expected of an isotropic distribution truncated at low galactic latitude. At lower redshifts the Virgo and Ursa Major clusters, and the rich groups that contain the radio galaxy Centaurus~A and the giant elliptical galaxy IC 3370, are at supergalactic latitudes $-2.4^\circ$, $2.8^\circ$, $-5.2^\circ$ and $-15.1^\circ$, respectively. Again, they are close to the plane.
The blue histogram shifted slightly to the right in Panel (a) is the distribution of the 32 most luminous galaxies at radio wavelengths $\sim 1.1$~GHz (compiled by van Velzen, Falcke, Schellart, Nierstenh{\"o}fer, and Kampert 2012; the data were downloaded from the VizieR Online Data Catalog). This region contains some $10^4$ galaxies with stellar masses comparable to or larger than the Milky Way, luminosities $L\lower.5ex\hbox{$\; \buildrel > \over \sim \;$} L_\ast$, meaning the 32 most powerful radio sources are exceptional galaxies. They tend to be in clusters, but since clusters and radio sources are found and cataloged in very different ways the consistency of their distributions offers a meaningful test of reproducibility of the concentration to the plane indicated by the peaks at low angle from the plane.
Panel~(b) in Figure~\ref{fig:LSCf} shows the distribution of the galaxies at $0.01<z<0.02$ that are most luminous at $60\mu$ wavelength. This is based on the redshift catalog Saunders, Sutherland, Maddox, et al. (2000) drew from the infrared astronomical satellite sky survey (Neugebauer, Habing, van Duinen, et al., 1984) at wavelengths from 12 to $100\mu$. The data were downloaded from the NASA HEASARC IRASPSCZ catalog, class GALAXY. I refer to these objects as LIRGs, for Luminous Infrared Galaxies, which seems appropriate though they do not necessarily fit the standard definition. The evidence is that these LIRGs are extraordinarily luminous at $60\mu$ because they are passing through phases of rapid formation of stars that generate the dust that absorbs the starlight and reradiates it as infrared radiation (P{\'e}rez-Torres, Mattila, Alonso-Herrero, Aalto, and Efstathiou 2021 and references therein). The black histogram in Panel~(b) in Figure~\ref{fig:LSCf} is the distribution in sin~SGB of the 29 most luminous LIRGs. For a check of reproducibility the blue histogram shifted slightly to the right shows the distribution of the 32 next most luminous. Both show no obvious tendency for these galaxies to be close to the plane of the Local Supercluster, or to avoid it. This is a pronounced difference from the distributions of the most powerful radio galaxies in Panel (a), and an illustration of the different distributions of different kinds of galaxies on scales $\sim 100$~Mpc.
\begin{figure}
\begin{center}
\includegraphics[angle=0,width=3.25in]{allearlylate.pdf}
\caption{The same as Figure~\ref{fig:LSCf} for the distributions of more numerous early- and late- type galaxies in the Huchra et al. (2012) catalog.}\label{fig:commongalaxies}
\end{center}
\end{figure}
Figure~\ref{fig:commongalaxies} shows the distribution relative to the plane of the Local Supercluster of a more numerous sample of galaxies (drawn from the Huchra, Macri, Masters, et al. 2012 catalog based on the Skrutskie, Cutri, Stiening, et al. 2006 identifications of galaxies detected in the Two Micron All Sky Survey, 2MASS). These galaxies are luminous enough to be in the Huchra et al. catalog out to redshift $z\leq 0.02$. Panel (a) shows the distribution of the 2708 early-type galaxies, ellipticals plus S0s (with Huchra morphological types $T\leq 0$), and Panel (b) shows the distribution of the 3276 spiral galaxies (with $1\leq T\leq 9$).\footnote{S0s are mentioned here for completeness, but since S0s are not common among the nearby $L\sim L_\ast$ galaxies they do not figure much in this essay, as noted in footnote~\ref{fn:S0s}. It might be interesting to compare distributions of separate spiral types, but this is not considered here.}
Since the early-type elliptical and S0 galaxies tend to be in clusters it is not surprising that the distribution in Panel~(a) is peaked at low SGB, with the clusters. But the peak is not as pronounced as in Panel~(a) in Figure~\ref{fig:LSCf}, meaning a greater fraction of these less luminous early-type galaxies are well away from the plane compared to the most powerful radio galaxies.
If considered alone the peak in the distribution of the 3276 spiral galaxies in Panel~(b) in Figure~\ref{fig:commongalaxies} might not seem significant; perhaps it only shows the large fluctuations to be expected in the correlated positions of galaxies, in this case a deficit at positive SGB balanced by an excess at SGB close to zero. But the peaks in the other distributions argue for a real tendency of spiral galaxies to be near this special plane. We also see that the tendency of common spirals to be near this plane is weaker than for common ellipticals plus S0s, which in turn is weaker than for powerful radio galaxies.
\begin{figure}
\begin{center}
\includegraphics[angle=0,width=3.25in]{morphologies.pdf}
\caption{The same as Figures~\ref{fig:LSCf} and~\ref{fig:commongalaxies} for distributions of the most luminous elliptical and spiral galaxies (adapted from Peebles 2022b).}\label{fig:morphologies}
\end{center}
\end{figure}
Figure~\ref{fig:morphologies} shows another aspect, a comparison of the distributions in sin~SGB of the elliptical and spiral galaxies with the greatest stellar masses (as indicated by the luminosity at $\sim 2\mu$ wavelength, which is considered a useful indicator of the stellar mass). These data also were drawn from the Huchra, et al. (2012) catalog. The 180 most luminous galaxies in the catalog at $0.01<z<0.02$ and galactic latitudes $|b|>10^\circ$ have absolute magnitudes bounded at apparent magnitude $K_s < 9.5$ at $z=0.02$. Of them, 53 are classified as ellipticals (with Huchra classification $T\leq -5$), 54 are classified spirals ($1\leq T\leq 9$), and the rest are classified S0s and irregulars of various kinds. Figure~\ref{fig:morphologies} shows the distributions in sin~SGB of the ellipticals and spirals.
It is not surprising that the angular distribution of the 53 most luminous ellipticals is similar to that of the most powerful radio galaxies and the clusters of galaxies shown in Figure~\ref{fig:LSCf}, because radio galaxies tend to be in giant ellipticals that tend to be in clusters. But again the data were obtained in different ways, and the samples are different. The pronounced concentration of the most luminous ellipticals to the plane of the extended plane of the Local Supercluster adds to the evidence of this interesting alignment.
The 54 most luminous spirals, with stellar masses that are comparable to the 53 most luminous ellipticals, are not noticeably more or less common at low SGB. Their distribution in the right-hand panel in Figure~\ref{fig:morphologies} resembles that of the most luminous galaxies at $60\mu$ in Panel (b) in Figure~\ref{fig:LSCf}. Within the noise there could be a peak in the distribution similar to that of more common spirals in Panel~(b) in Figure\ref{fig:commongalaxies}.
The arrangements of clusters of galaxies and galaxies of various types relative to the plane of the Local Supercluster are readily studied because we happen to be close to the plane. Other arrangements of objects similar to what is seen within 85~Mpc distance from our special position might include the Pisces-Perseus supercluster (Giovanelli, Haynes, and Chincarini, 1986), the CfA Great Wall (Geller and Huchra 1989), and the Sloan Great Wall (Gott, Juri{\'c}, Schlegel, et al., 2005). They are spread over hundreds of megaparsecs. It would be interesting to know whether radio galaxies and massive ellipticals in the neighborhoods of these more distant configurations are largely confined to a ridge, in the manner of the Local Supercluster, while massive spiral galaxies, and galaxies that are exceptionally luminous in the infrared, are not so particularly concentrated.
I have not discussed the distribution of quasars at $0.01<z<0.02$ because I am not aware of a quasar catalog with measured redshifts that is suitably close to complete across the sky at this relatively short distance. But we might take it that quasars and radio galaxies are related, and that the concentration of radio galaxies to a plane in a region around us $\sim 170$~Mpc across is an analog of a Large Quasar Group of the kind discussed by Clowes, Harris, Raghunathan, et al. (2013).
In the sample at $0.01 < z < 0.02$ the similar distributions of giant ellipticals, radio galaxies, and clusters of galaxies on the one hand, and the different but again similar distributions of giant spirals and luminous infrared galaxies on the other, present us with interesting issues.
\begin{enumerate}[label*=\arabic*.]
\item Why are the clusters of galaxies largely near a plane at $z<0.02$? If clusters formed where primeval mass density fluctuations were particularly large it would require that upward density fluctuations large enough to grow into clusters are confined to a plane in a region some 170~Mpc across. These large upward density fluctuations cannot be at all common at similar distances from us but not near the plane, because clusters are not at all common there. On the face of it this arrangement looks unlikely in the standard $\Lambda$CDM theory, but it could be checked in pure dark matter simulations.
\item Why are primeval conditions capable of producing galaxies with exceptionally large stellar masses far from the plane as well as near it, but only capable of producing clusters near the plane?
\item Why are the most massive elliptical and spiral galaxies, with comparable stellar masses to judge by the $2\mu$ luminosities, so differently distributed? If the clusters formed where the primeval mass density is exceptionally large, close to this preferred plane, it could account for the concentration of giant ellipticals to this plane. But then why are the giant spirals not more noticeably abundant than average near the plane of the Local Supercluster? Maybe the conditions that favor cluster formation were hostile to the formation of spirals near this plane, perhaps more likely to destroy the spiral arms? But if so why do the spirals not show evidence of avoiding the plane? The more common spirals instead show a modest tendency to be near the plane.
\item What accounts for the concentration of powerful radio galaxies to this plane? Radio galaxies seem to require the presence of a massive compact central object, very likely a black hole. Massive black holes are present in many if not all $L\sim L_\ast$ galaxies, spirals as well as ellipticals, including the many that are closer than 85~Mpc and not close to the plane of the Local Supercluster. These more common black holes seem to cause some galaxies to be radio sources, though seldom at the level of power of the radio galaxies that tend to be near the plane of the extended Local Supercluster. What is special about the massive black holes that are associated with the powerful radio sources that tend to be close to this plane?
\item What are we to make of the contrast between the distributions of galaxies that are exceptionally luminous at radio wavelengths and those that are exceptionally luminous at $60\mu$? The former tend to be near the plane, the latter not.
\end{enumerate}
These questions are not widely advertised. A measure of this is the 26 citations to Shaver (1991) in the Astrophysics Data System. One is a self-citation, four are mine, all on Shaver's point, five are on the possible implication of the alignment of radio galaxies for the angular distribution of energetic cosmic rays, and sixteen are on the nature of the space distribution of radio galaxies. Three of these space distribution papers take note of the planar distribution of the relatively nearby radio galaxies and clusters of galaxies. For example, Strauss (1993) remarks on ``the very large planar structures seen in the cluster and radio galaxy distribution by Tully et al. (1992) and Shaver (1991); this discrepancy remains to be explained.'' It certainly is interesting. But I find no discussion of what is to me most interesting, the considerable difference between the space distributions of radio galaxies and clusters of galaxies on the one hand, and the distribution of ordinary large galaxies on the other. The literature on this point might be scant to nonexistent because it is difficult to know what to make of it. That is no excuse, though: we are missing something, which surely is worth investigating.
Figures~\ref{fig:LSCf} to \ref{fig:morphologies} are added illustrations of Shaver's (1991) key point, that different kinds of extragalactic objects can have quite different spatial distributions relative to the plane of the Local Supercluster. Hints to understanding this might be found in the situation on a larger scale that seems analogous, as discussed next.
\subsection{Anomalies on Large Scales}\label{CosPrin}
To be considered here are apparent anomalies in the distributions and motions of objects on scales approaching the Hubble length. This begins in Section~\ref{sec:CMBDipole} with our motion relative to the reference frame set by the near homogeneous sea of thermal cosmic microwave background radiation, the CMB. The standard interpretation of the CMB dipole anisotropy is that it is the effect of our motion through the radiation. This is tested by computing the local peculiar velocity expected from the gravitational acceleration computed from the observed departures from a homogeneous distribution of objects that seem likely to be useful mass tracers. The results discussed in Section~\ref{sec:localgravity} do not disagree with the idea, but they are not very tight. The test considered in Section~\ref{sec:bulkflows} uses estimates of the mean motion relative to the CMB of objects within a given distance from us, computed from the departures of redshifts from the homogeneous Hubble flow. By some measures this mean motion, the bulk flow, is found to approach zero as the distance is increased, about as expected from the standard $\Lambda$CDM theory. But other measures of the bulk flow that look equally reliable are anomalous. A possibly related problem reviewed in Section~\ref{sec:KinematicDipole} is the predicted dipole anisotropy in the angular distributions of objects that are so far away that, in the standard cosmology, the space distribution likely averages out to homogeneity. The Doppler shifts and aberration of the observed angular positions of these objects caused by our motion relative to the CMB are predicted to produce a dipole anisotropy in the angular distributions of these objects, the kinematic dipole. The measured dipoles in the distribution of quasars and in the distributions of radio galaxies cataloged at several radio frequencies are in about the predicted direction, but the dipole amplitudes are too large, an anomaly. The situation from these considerations is reviewed in Section~\ref{sec:remarks}.
\subsubsection{The CMB Dipole Anisotropy}\label{sec:CMBDipole}
Departures from an exactly homogeneous sea of thermal microwave radiation, the CMB, are usefully represented by the spherical harmonic expansions of the CMB temperature and polarization as functions of position across the sky (eq.~\ref{eq:harmonicexp}). The amplitudes $a_\ell^m$ of the spherical harmonic expansion of the temperature at degree $\ell > 1$ are convincingly demonstrated to be remnants of the decoupling of acoustic oscillations in the plasma-radiation fluid. The dipole amplitude, $\ell = 1$, is much larger than predicted from this effect. The standard idea attributes it to our motion relative to the rest frame defined by the CMB, at velocity
\begin{equation}
{\vec v}_{\rm helio} - {\vec v}_{\rm CMB} = 370\hbox{ km s}^{-1}\hbox{ to } l = 264^\circ,\ b=48^\circ,\label{eq:heliocen_wrt_cmb}
\end{equation}
in galactic coordinates and the solar system rest frame (Planck Collaboration et al. 2020a). The adjustment for the solar motion relative to the Milky Way Galaxy and the motion of the Galaxy relative to an estimate of the motion of the center of mass of the Local Group, between the Milky Way and M~31, indicates the Local Group of galaxies is moving relative to the sea of radiation at
\begin{equation}
{\vec v}_{\rm Local Group} - {\vec v}_{\rm CMB}= 620\hbox{ km s}^{-1}\hbox{ to } l = 272^\circ,\ b=30^\circ. \label{eq:LG_wrt_cmb}
\end{equation}
The speed, $\sim 600$~km~s$^{-1}$, is much larger than the relative motions of the galaxies closer than 10~Mpc. A natural interpretation is that we and the nearby galaxies are moving at a near common velocity relative to the CMB, and that this motion is to be associated with the growing departures from an exactly homogeneous mass distribution. A test is discussed next.
\subsubsection{The Peculiar Gravitational Acceleration} \label{sec:localgravity}
In linear perturbation theory and time-orthogonal coordinates chosen to eliminate the decaying mode in a cosmologically flat universe our peculiar motion at coordinate position $\vec r=0$ is predicted to be usefully approximated (in a version of eq.~[\ref{eq:peculiaracceleration}]) as
\begin{equation}
\vec v = {2\beta G \rho_b\over 3 H_{\rm o}\Omega}\int d^3r~\delta_g(\vec r)~{\vec r\over r^3},\label{eq:dynamic_v}
\end{equation}
where
\begin{equation}
\delta_g(\vec r) = {\delta n(\vec r)\over\langle n\rangle},\quad
\delta_\rho={\delta\rho(\vec r)\over\langle\rho\rangle}\simeq{\delta_g\over b},\quad \beta \approx {\Omega^{0.55}\over b}\sim 0.4.\label{eq:dynamicparameters}
\end{equation}
The fractional departure from a homogeneous galaxy distribution, $\delta_g(\vec r)$, might be represented as a sum of Dirac delta functions minus the mean, or the result of smoothing of galaxy counts through a window. The mass density contrast $\delta_\rho$ is written in the simple linear bias model for the relative distributions of galaxies and mass. We have a measure of the bias parameter, $b\sim 1.2$, and the mass density parameter, $\Omega=0.31$, from the fit of theory to observations of the patterns in the distributions of galaxies and the CMB (as discussed in Planck Collaboration et al. 2020b). The evidence from these data is that galaxies are reasonably good mass tracers.
Having adopted general relativity in the standard cosmology we must accept that the mass outside the region we can observe in principle (that is, outside the particle horizon subsequent to inflation, or whatever saved us from the singularity of the standard model) can affect predicted observable peculiar motions (as discussed in Sec.~\ref{sec:theacausaluiverse}). There is evidence that the mass distribution that can be observed is a useful approximation, however. Erdo{\v{g}}du, Huchra, Lahav, et al. (2006) numerically evaluated the integral in equation~(\ref{eq:dynamic_v}) using a version of the Huchra, Macri, Masters, et al. (2012) galaxy redshift catalog. (This valuable catalog was used in the studies of the extended Local Supercluster in Sec.~\ref{sec:LSC}.) Erdo{\v{g}}du et al. found that the predicted velocity of the Local Group relative to the CMB seems to converge at about 100~Mpc distance from the Local Group. The integral computed to this distance indicates the motion of the Local Group is toward $l\sim 265^\circ$, $b\sim 38^\circ$, some $10^\circ$ from the CMB dipole. This is tolerably consistent considering the uncertainties in the use of galaxies as mass tracers. The integral computed to this distance agrees with the observed speed if the combination $\beta$ of bias parameter and mass density parameter in equation~(\ref{eq:dynamicparameters}) is $\beta = 0.40\pm 0.09$. This is consistent with the value derived from the CMB anisotropy spectrum (Planck Collaboration et al. 2020). At greater cutoff distances the computed velocity moves away from the CMB direction, a likely consequence of large effects of small systematic errors in the mass distribution on larger scales.
This test is important but not yet very precise. A better application requires a catalog of positions of useful mass tacers to greater distances with tighter controls on completeness and systematic errors in distances and positions across the sky, a challenging task. And we must live with the unknowable situation outside the Hubble length.
\subsubsection{Cosmic Bulk Flows}\label{sec:bulkflows}
This probe requires measurements of the radial components $v_p$ of peculiar velocities of objects relative to the general expansion of the universe,
\begin{equation}
v_p = cz - H_{\rm o}r. \label{eq:peculiarvelocity}
\end{equation}
To order $v/c$ the radial velocity derived from the measured redshift $z$ of an object is $v = cz$, $H_{\rm o}r$ is the speed of cosmological recession at the physical distance $r$ of the object, and the difference is the radial peculiar velocity $v_p$. In a suitably large sample of objects that are moving relative to us at mean velocity $\vec v_{\rm obs}$ the measured $v_p$ of the objects are expected to vary across the sky as $v_{\rm obs}\cos\alpha$, where $\alpha$ is the angle between the direction to the object and the direction of the mean velocity $\vec v_{\rm obs}$. It is conventional to define the bulk velocity $\vec v_{\rm sample}$ of the sample to be the mean velocity referred to the rest frame in which the CMB has no dipole anisotropy. (This ignores the small intrinsic dipole remnant from the decoupling of baryonic matter and the CMB.) In the standard model $\vec v_{\rm sample}$ is predicted to converge to zero in a catalog that reaches distances large enough to be a fair sample of the statistically homogeneous universe.
Boruah, Hudson, and Lavaux (2020) measured peculiar velocities of galaxies closer than $\sim 60$~Mpc from distances to supernovae of type Ia and distances to spiral galaxies based on the Tully-Fisher relation. The Boruah et al. mean of the peculiar velocities relative to the CMB is
\begin{equation}
{\vec v}_{\rm galaxies} - {\vec v}_{\rm CMB}= 252 \pm 11\hbox{ km s}^{-1}\hbox{ to } l = 293^\circ,\ b=14^\circ (\pm 5^\circ).\label{eq:Hudson}
\end{equation}
This can be compared to the earlier Ma and Scott (2013) measurement on a similar scale,
\begin{equation}
{\vec v}_{\rm galaxies} - {\vec v}_{\rm CMB}\sim 290\pm 10 \hbox{ km s}^{-1}\hbox{ to } l = 280\pm 8^\circ,\ b=5^\circ \pm 6^\circ.\label{eq:Scott}
\end{equation}
from fundamental plane distances to early type galaxies, Type Ia supernovae, and the Tully-Fisher relation for late-type galaxies. (The speed is my estimate from the four results in Ma and Scott Table 2.) Similar results are reported in quite a few analyses. Watkins, Feldman, and Hudson (2009) found bulk flow direction consistent with equations~(\ref{eq:Hudson} and~(\ref{eq:Scott}) and speed 150~km~s$^{-1}$ larger, but that is just twice the estimated uncertainty. Tighter and consistent results are reported by Turnbull, Hudson, Feldman, et al. (2012), who used Type~Ia supernova distances; Hong, Springob, Staveley-Smith, et al. (2014), who used Tully-Fisher distances to spiral galaxies; and Scrimgeour, Davis, Blake, et al. (2016), who used the fundamental plane relation to get distances to early-type galaxies. This bulk velocity seems to be securely established.
The bulk velocity of the sample of objects closer than about 60~Mpc need not be in the same direction as the peculiar velocity of the Local Group, but one might expect it to be fairly close, as it is in these measurements. The speed of a sample relative to the CMB, the bulk velocity, is expected to be smaller when averaged over larger scales, because the mass distribution is assumed to average to homogeneity in sufficiently large volumes. The speeds found from these analyses are roughly half that of the Local Group, and Figure 10 in Boruah et al. indicates the speed is consistent with the probability distribution in the average over a region of this size computed in linear perturbation theory in the $\Lambda$CDM cosmology.
In a remarkable advance the Planck Collaboration et al. (2014) reported a measurement of the mean motion of clusters of galaxies relative to the CMB based on the kinematic kSZ effect (Sunyaev and Zel'dovich 1980). This is the Doppler shift of the CMB scattered by electrons in the intracluster plasma of a cluster of galaxies that is moving relative to the CMB. The conclusion from detections of this kSZ effect is that the average speed of the clusters relative to the CMB is compatible with zero and the speed is less than about $260\hbox{ km s}^{-1}$ at the 95\% confidence level.This is in a sample of clusters with redshifts ranging around $z\sim 0.2$.
The detection of the kSZ effect in the plasma in a cluster of galaxies offers a direct measure of the effect of the motion of the cluster relative to the CMB. The Planck Collaboration observations agree with the expectation that the motion of the clusters averaged over scales approaching the Hubble length is small, continuing the trend from the speed of the Local Group, $\sim 600$~km~$s^{-1}$, to the mean for the galaxies out to 60~Mpc at half that speed, to the still smaller mean speed expected of the still more extended sample of clusters of galaxies.
The motions of clusters of galaxies based on measurements of distance rather than the kSZ effect are more complicated. Lauer and Postman (1994) used distances to Abell clusters (Abell 1958; Abell, Corwin, Olowin, 1989) based on apparent magnitudes of the brightest cluster members, measured in an aperture of fixed metric size and corrected for the redshift of the spectrum and extinction in the Miky Way. They found that the mean peculiar velocity of the Abell clusters relative to the CMB is
\begin{equation}
{\vec v}_{\rm Abell} - {\vec v}_{\rm CMB} \sim 689\hbox{ km s}^{-1}\hbox{ to } l = 343^\circ ,\ b = 52^\circ.\label{eq:LP}
\end{equation}
The direction might be taken to be roughly similar to that of the galaxies (eq.~[\ref{eq:Hudson}]), but the large speed is anomalous. This result remains unexplained. The Migkas, Pacaud, Schellenberger et al. (2021) measurement of the cluster bulk flow used scaling relations among properties of the plasma in clusters rather than properties of the cluster galaxies. The smallest scatter among their measurements is found in the relations between the distance-independent intracluster plasma temperature and the distance-dependent cluster X-ray luminosity, and between the plasma temperature and the distance-dependent integrated Comptonization parameter. The clusters in their sample are at redshifts $z\sim 0.01$ to 0.55, with 50\%\ within $z=0.05$ to 0.18. They found that in a region of the sky centered around $l\sim 276^\circ$, $b \sim -16^\circ$ their cluster peculiar velocities referred to the CMB rest frame are systematically negative, with a suggestion of positive peculiar motions in the opposite part of the sky. It can be interpreted as the effect of a mean bulk flow of the cluster sample at velocity
\begin{equation}
{\vec v}_{\rm cluster} - {\vec v}_{\rm CMB} \sim 900\hbox{ km s}^{-1}\hbox{ to } l \sim 88^\circ,\ b \sim 16^\circ.\label{eq:Migkasetal}
\end{equation}
Again, the speed is anomalously large. The direction disagrees with the Lauer and Postman result (eq.~\ref{eq:LP}), and it is close to opposite to the motions of the Local Group and the bulk flow of the galaxies within distance $\sim 60$~Mpc (eqs.~\ref{eq:LG_wrt_cmb}, \ref{eq:Hudson}).
The situation is interesting. The Abell cluster catalog was compiled by hand, so there may be sampling inhomogeneity, but that need not seriously affect the mean motion of the observed clusters. The Lauer and Postman (1984), and Migkas, et al. (2021), bulk flows are seriously different, maybe because their distances are based on different cluster scaling relations. It also might have something to do with the odd distributions of clusters and radio galaxies closer than 85~Mpc (Sec.~\ref{sec:LSC}), and the odd larger-scale distributions of radio galaxies and quasars to be considered next.
\subsubsection{The Kinematic Dipole}\label{sec:KinematicDipole}
This cosmological test requires a catalog of objects that are far enough away that we can assume their mean spatial distribution is adequately close to the homogeneity and isotropy of the cosmological principle. It of course requires that the efficiency of detection of objects is adequately close to uniform across the sky. In these conditions our motion relative to the mean of this sample is expected to produce a dipole anisotropy in the counts of objects, the results of the Doppler shifts of apparent magnitudes and the aberration of angular positions.
Ellis and Baldwin (1984) pointed out that this consideration provides us with a cosmological test: the dipole anisotropy of counts of objects across the sky is predicted to be, to order $v/c$,
\begin{equation}
\delta N/N = [2+x(1+\alpha)](v/c)\cos\theta, \label{eq:EllisBaldwin}
\end{equation}
where
\begin{equation}
S\propto \nu^{-\alpha}, \ N(>S) \propto S^{-x}. \label{eq:EBparameters}
\end{equation}
The parameter $\alpha$ is a measure of the typical spectrum of an object, $x$ is a measure of the variation of counts of objects with limiting flux density $S$, and $v$ is the heliocentric velocity relative to the mean of the sample of distant objects.
The physics of the Ellis and Baldwin kinematic effect is well established in other contexts. The Compton-Getting effect is the dipole anisotropy of energetic cosmic rays seen by an observer moving through an isotropic sea of cosmic rays. The same effect accounts for the thermal spectrum of the CMB detected in a given direction by an observer moving relative to the CMB rest frame (Peebles and Wilkinson 1968). Kaiser (1987) remarked that this apparent dipole in the mass distribution produces an apparent contribution to the computation of our cosmic gravitational acceleration. Bahr-Kalus, Bertacca, Verde, and Heavens (2021) review current studies of this ``rocket effect."
The evidence to be reviewed here is that the dipole anisotropy in the distribution of objects at distances comparable to the Hubble length is about in the direction expected from the kinematic effect if the dipole anisotropy in the CMB is due to our motion relative to the rest frame defined by the mean mass distribution, but the dipole amplitude is at least twice the prediction. This anomaly is about as well established as the Hubble Tension, yet the literature on the kinematic effect is much smaller than the 344 papers with the phrase ``Hubble Tension'' in the abstract in the SAO/NASA Astrophysics Data System. (I expect the difference is an inevitable consequence of the way we behave.) To illustrate this difference I offer my attempt at a close to complete literature on the kinematic effect (with apology for overlooked publications).
Baleisis, Lahav, Loan, and Wall (1998) considered the possible detection of the kinematic effect in the dipole anisotropy of radio galaxies. Since these objects are detectible past redshift $z=1$ their distribution probes scales large enough that one might hope the clustering of these objects averages out to the wanted uniformity to be seen broken by the kinematic dipole. Scharf, Jahoda, Treyer, et al. (2000) considered detections of the kinematic effect in the angular distributions of X-ray, AGNs, and clusters of galaxies. These are the earliest empirical studies of the effect I have found. Blake and Wall (2002) had the NRAO VLA Sky Survey of radio sources (NVSS; Condon, Cotton, Greisen, et al., 1998). Their dipole estimate is consistent with the direction and amplitude expected from the dipole anisotropy of the CMB. Later studies of the NVSS catalog confirmed the direction, but found larger than expected dipole amplitudes in this catalog at various flux density cuts (Singal, 2011; Gibelyou and Huterer, 2012; Tiwari, Kothari, Naskar, Nadkarni-Ghosh, and Jain 2015; Secrest, von Hausegger, Rameez, Mohayaee, and Sarkar 2022). Rubart and Schwarz (2013) had a partial check by combining NVSS with the Westerbork Northern Sky Survey (WENSS; de Bruyn, Miley, Rengelink, et al., 2000). Colin, Mohayaee, Rameez and Sarkar (2017) combined NVSS with the Sydney University Molonglo Sky Survey (SUMSS; Mauch, Murphy, Buttery, et al., 2003). Both papers reported the anomaly. Bengaly, Maartens and Santos (2018) reported consistency of separate measurements of the dipole anisotropies derived from the NVSS (USA) and TGSS-ADR1 (India; Intema, Jagannathan, Mooley, and Frail, 2017) catalogs, and Siewert, Schmidt-Rubart, and Schwarz (2021) reported separate measurements of the NVSS, TGSS-ADR1, WENSS (The Netherlands) and SUMSS (Australia) catalogs. Since each of these four catalog might be affected by its own systematic errors the reasonable consistency of the separately analyzed catalogs is a valuable check of reliability. The dipole directions from these four independently obtained samples average to about
\begin{equation}
l\sim 240^\circ,\ b\sim 30^\circ\hbox{ for radio galaxies.}\label{eq:radiodipole}
\end{equation}
This is not far from the direction of the heliocentric CMB dipole (eq.~[\ref{eq:heliocen_wrt_cmb}]), as would be expected from the kinematic effect. The radio source dipole amplitudes from the different samples are considerably different, maybe in part because the amplitude depends on frequency, but all are roughly a factor of 4 times the amplitude expected from our motion relative to the CMB rest frame. Siewart et al. conclude that the dipoles ``exceed the expectations derived from the CMB dipole, which cannot strictly be explained by a kinematic dipole alone.''
Darling (2022) reports a different situation from the analysis of the dipole anisotropy of radio sources in the more recent VLASS (USA; Lacy, Baum, Chandler, et al., 2020) and RACS (Australia; McConnell, Hale, Lenc, et al., 2020) surveys. The dipole interpreted as the kinematic effect indicates heliocentric velocity
\begin{equation}
{\vec v}_{\rm helio} - {\vec v}_{\rm radio} \sim 330\pm 130\hbox{ km s}^{-1}\hbox{ to } l \sim 270\pm 55^\circ,\ b \sim 56\pm 25^\circ.\label{eq:Darling}
\end{equation}
This agrees with the heliocentric velocity, direction and speed, relative to the CMB (eq.~[\ref{eq:heliocen_wrt_cmb}]).
Darling estimates that the ``most permissive'' analysis allows effective velocity 740~km~s$^{-1}$ at three standard deviations. Siewert et al. (2021) estimate that at that speed the kinematic would correspond to dipole amplitude $d\sim 0.01$. But this is well below most of the Siewert et al. radio source dipoles. It is a cautionary example of the difficulty of establishing this important measurement.
We have a check on this situation from the angular distribution of the roughly one million objects selected by their mid-infrared colors to be quasars detected by the Wide-field Infrared Survey Explorer (WISE; Wright, Eisenhardt, Mainzer, et al. 2010). In independently selected catalogs, Secrest, von Hausegger, Rameez, et al. (2021, 2022) and Singal (2021) found dipole anisotropies pointing to (in my estimate of Singal's central value)
\begin{align}
& l\sim 210^\circ,\ b\sim 45^\circ,\ v\sim 1700\hbox{ km s}^{-1},\hbox{ Singal (2021)},\nonumber\\
&l= 238^\circ,\ b=31^\circ, v \sim 750\hbox{ km s}^{-1},\hbox{ Secrest et al. (2022)}. \label{eq:kinematicdipole}
\end{align}
The directions of these two quasar dipoles are reasonably similar to the directions of the radio dipole (eq.~[\ref{eq:radiodipole}]), and perhaps not unreasonably far from the CMB dipole, $l = 264^\circ,\ b=48^\circ$ (eq.~[\ref{eq:heliocen_wrt_cmb}]). But with the authors' estimates of the parameters for spectrum and counts in equation~(\ref{eq:EBparameters}) the dipole amplitudes are at least twice that expected from the kinematic dipole (eq.~\ref{eq:EllisBaldwin}) at the velocity expected from the CMB dipole. Darling's (2022) new radio galaxy dipole amplitude puts the three standard deviation upper limit of 750~km~s$^{-1}$ on the effective velocity from the kinematic amplitude, consistent with Secrest et al. but still well below Singal.
Quasars and radio galaxies are related, but the data on the two were obtained and reduced by quite different methods, and each have been analyzed by two or more independent groups with consistent results that seem to make a consistent case for an anomaly. But the result in equation~(\ref{eq:Darling}) from independent radio data and analysis is an important reminder of the hazards of systematic errors in these measurements. I conclude that the present weight of the evidence from the other measures of the radio dipole and the WISE quasar dipole (eqs.~\ref{eq:radiodipole} and \ref{eq:kinematicdipole}) is that there is an anomalously large dipole common to distant radio galaxies and quasars, but the case is not yet persuasive.
Several other points are to be noted. First, the general conclusion has been that, within the standard cosmology and a reasonable degree of biasing in the positions of quasars and radio galaxies relative the mass, intrinsic inhomogeneity is not a likely explanation for the anomalous dipole anisotropies of these distant objects (e.g., Rubart, Bacon, and Schwarz 2014; Tiwari and Nusser 2016; Colin et al. 2017; Dom{\`e}nech, Mohayaee, Patil, and Sarkar 2022; and Murray 2022, who considered the effect of gravitational lensing). This might be checked by the cross-correlation between the surface mass density indicated by CMB lensing with the positions of distant radio galaxies (Robertson, Alonso, Harnois-D{\'e}raps, et al., 2021), but I have not found a discussion of the test. A simple order-of-magnitude argument is that if quasars and radio galaxies were useful mass tracers on scales approaching the Hubble length then the observed dipole anisotropies $\delta N/N\sim 0.02$ would have produced a bulk flow on the order of 2\% of the speed of light, which is absurd.
Second, we have a separate reason to question whether these objects are useful mass tracers, from the curious distributions of objects that contain AGNs closer than 85~Mpc (Sec.~\ref{sec:LSC}). Following this line of thought we might expect that the angular positions of clusters of galaxies at redshifts close to unity have an anomalous dipole anisotropy, perhaps similar to that of quasars and radio sources, though of course not similar to the mass distribution. Perhaps this can be checked with available data. If ordinary $L\sim L_\ast$ galaxies are useful mass tracers, as evidenced by the cosmological tests, then a really deep galaxy catalog would be expected to have dipole direction and amplitude consistent with the kinematic dipole (eq.~\ref{eq:EllisBaldwin}) indicated by our motion defined by CMB dipole. It may be possible to check this.
Third, one is tempted to ask whether the dipole anisotropy in the distributions of distant radio sources and quasars, or the Migkas et al. (2021) cluster bulk flow, are somehow related to the plane of the Local Supercluster. There is no indication of that in the directions they define, supergalactic latitudes $SGB\sim -50^\circ$ and $SGB\sim +50^\circ$. But we do see common evidence of anomalous distributions of objects that contain AGNs.
Fourth, Migkas (private communication 2022) points out that our heliocentric velocity relative to the cluster rest frame found by Migkas et al. (2021, eq.~[\ref{eq:Migkasetal}]) is
\begin{equation}
{\vec v}_{\rm helio} - {\vec v}_{\rm clusters} \sim 1100\hbox{ km s}^{-1}\hbox{ to } l \sim 280^\circ,\ b \sim 5^\circ.\label{eq:Kostas}
\end{equation}
This is not far from our effective heliocentric velocities relative to the radio sources in equation~(\ref{eq:radiodipole}) and the quasars in equation~(\ref{eq:kinematicdipole}). Maybe it is only a coincidence, but it is not to be ignored in this confusing situation.
\subsection{Summary Remarks} \label{sec:remarks}
The association of the CMB dipole anisotropy with the growing mode of the departure from an exactly homogeneous universe is tested by the check of consistency with the prediction from the peculiar gravitational acceleration. There is room for an anomaly because the integral in equation~(\ref{eq:dynamic_v}) seems to converge at $\sim 100$~Mpc distance while the evidence is that the bulk flow of the galaxies on about the same scale is some 250~km~s$^{-1}$ (eq. [\ref{eq:Hudson}]). But assessing the significance of the discrepancy is difficult because the computation of the integral in equation~(\ref{eq:dynamic_v}) is sensitive to small errors in the large-scale galaxy distribution.
The evidence is that the bulk flow of the galaxies relative to the CMB decreases with increasing sample size about as expected from standard ideas, from about 600~km~$^{-1}$ in the average over distances of a few megaparsecs, to about 250~km~$^{-1}$ averaged out to distances $\sim 60$~Mpc. Detection of the kinetic SZ effect on the intracluster plasma suggests an even smaller bulk flow of clusters of galaxies at distances distributed around $\sim 600$~Mpc. Better checks of the convergence of the galaxy bulk flow and the predicted local peculiar gravitational acceleration from the integral over the mass distribution requires improved measures of the space distributions of galaxies and mass. Perhaps it will come from the Euclid mission ``to capture signatures of the expansion rate of the Universe and the growth of cosmic structures'' (Percival, Balogh, Bond, et al. 2019). Nadolny, Durrer, Kunz, and Padmanabhan (2021) present a worked example of how this might go.
There are interesting anomalies in the space distributions of radio galaxies, quasars, and clusters of galaxies. The evidence reviewed in Section~\ref{sec:LSC} (and shown in Figs. \ref{fig:LSCf} and \ref{fig:morphologies}) is that we are in a region some 170~Mpc across in which the most luminous early-type galaxies, powerful radio galaxies, and clusters of galaxies tend to be near the extended plane of the Local Supercluster. In contrast to this the spirals with stellar masses as great as the most massive ellipticals, to judge by the luminosities at $2\mu$, are not noticeably correlated with this plane. The same is true of the most luminous galaxies at $60\mu$, and the much more abundant $L\sim L_\ast$ galaxies that are useful mass tracers.
A possibly related anomaly is found in measurements of the bulk velocities of clusters of galaxies (Lauer and Postman 1994; Migkas, et al. 2021). Both are based on cluster properties: first-ranked galaxy luminosity or cluster plasma scaling relations. If the anomaly from cluster distance measures persist, and the mean motion of clusters of galaxies relative to the CMB measured by the kSZ effect continues to show a small bulk velocity, then we will be forced to the conclusion that scaling relations of cluster properties are not universal. This is not as extreme at it might at first seem, for recall the anomalous distribution of the nearer clusters, in the region 170~Mpc across.
Yer another possibly related anomaly is the dipole anisotropies in the angular distributions of presumably distant and on average uniformly distributed radio galaxies and quasars. All recent analyses agree that the dipole is about in the direction expected from the kinematic effect of our motion relative to the CMB. Though not all recent analyses agree on the amplitude, the weight of the evidence is that the radio galaxy and quasar dipole amplitudes are anomalously large.
We are led to the thought that the properties of large groups and clusters of galaxies, radio galaxies, and quasars defined by colors, were enabled by something that has had a subdominant correlation with the mass distribution everywhere except in rare situations such as the neighborhood of the extended Local Supercluster. This would be a departure from the Gaussian adiabatic initial conditions of the standard cosmology. Thoughts turn to cosmic strings, or primeval isocurvature fluctuations, or black holes left from some earlier epoch. Or, as I have remarked elsewhere in this essay, something completely different.
Also to be considered is the indication that positions of physically related quasars are spread over greater lengths than could have grown out of the Gaussian near scale-invariant adiabatic initial conditions assumed in the standard $\Lambda$CDM cosmology (e.g., Clowes, et al. 2014). If these Large Quasar Groups are physically real associations then, under the assumptions of the standard theory, they are associations among objects that have never been causally connected. This is a familiar situation, of course. The Gaussian initial conditions of the standard cosmology also are acausal in the standard model (Sec.~\ref{sec:theacausaluiverse}).
Some authors conclude that the anomalously large dipole anisotropy of distant quasars and radio galaxies, and the Large Quasar Groups, if physically real, violate the cosmological principle (e.g., Secrest, von Hausegger, Rameez, et al. 2022). This depends on the definition of this principle, of course. It does not violate the definition explained in Section~(\ref{sec:definition}), but it likely violates the assumption of Gaussian near scale-invariant and adiabatic initial conditions that has served so well for many other cosmological tests.
\begin{figure}
\begin{center}
\includegraphics[angle=0,width=3.7in]{LocalVol.pdf}
\caption{Distribution of galaxies closer than 9~Mpc. The galaxies more luminous than $M_B = -20$ are marked by the large red circles. The smaller red circles show the positions of the less luminous galaxies with absolute magnitudes $-20 < M_B < -17$. The still smaller black circles show the dwarfs with known distances at $M_B > -17$.}\label{fig:LocalVol}
\end{center}
\end{figure}
\section{The Local Void}\label{sec:localvoid}
Figure~\ref{fig:LocalVol} shows the distribution of galaxies closer than $D=9$~Mpc, plotted in the supergalactic coordinates discussed in Section~(\ref{sec:LSC}). (The data are from Karachentsev, Karachentseva, Huchtmeier, Makarov, 2004, updated in the NASA HEASARC Updated Nearby Galaxy Catalog. A distance cutoff at $D=8$~Mpc eliminates the interesting dwarf galaxy at the top of the figure; a cutoff at 10~Mpc adds several dwarfs in the low density area to the upper right that are not very close to the two interesting dwarfs that are well within the Local Void.)
The Local Supercluster is the concentration of galaxies running through the center of the figure. The distributions of radio galaxies and clusters of galaxies relative to the plane of the Local Supercluster, at ten times the distance sampled in Figure~\ref{fig:LocalVol}, are illustrated in Figures~\ref{fig:LSCf} and~\ref{fig:morphologies}
The open red circles in Figure~\ref{fig:LocalVol} show positions of the 14 most luminous galaxies, absolute magnitudes $M_B<-20$. Galaxies that are more luminous than any but the brightest in this region are exceedingly rare. The 65 smaller filled red circles mark positions of galaxies with absolute magnitudes in the range $-20 < M_B < -17$, which spans a factor of 16 in luminosity. The positions of the 718 less luminous of the galaxies with useful distance estimates are marked as the still smaller filled black circles. It is expected that many more dwarfs will be added to this sample.
The upper left region in Figure~\ref{fig:LocalVol} is part of the Local Void, a strikingly empty region. Only two of the known $\sim 800$ galaxies within 9~Mpc occupy about a quarter of the volume in this part of the Local Void. This amounts to a space density of galaxies in the near empty region at about one percent of the mean within the full $R<9$~Mpc volume. A common estimate from numerical simulations of structure formation in the $\Lambda$CDM theory is that in low density regions the mean mass density bottoms out at roughly 10\% of the cosmic mean. A recent example is presented in Cautun, Cai, and Frenk (2016). Peebles (2001) argued that this low density of detected galaxies seems distinctly odd. Tikhonov and Klypin (2009) concluded from their numerical simulations that ``The emptiness of voids [is] yet another overabundance problem for the cold dark matter model.'' But Tinker and Conroy (2009) pointed out that the low density of galaxies in the Local Void is consistent with the predicted mass distribution if the most numerous lowest mass dark matter halos contain very few luminous stars. The Tinker and Conroy application of this idea using the halo occupation distribution model applied to high resolution pure dark matter simulations produces acceptably empty voids. This is progress but it is a prescription, not a prediction. It might be tested by other information.
Let us begin with the two known dwarf galaxies in the nearest part of the Local Void. The dwarf galaxy at the top of the figure, ZOA~J1952+1428, was discovered in a blind survey for HI emission by the Arecibo Zone of Avoidance Survey (McIntyre, Minchin, Momjian, et al. 2011). The other well isolated dwarf galaxy lower down and to the left in Figure~\ref{fig:LocalVol} is KK~246, also known as ESO 461-036.
Karachentsev, Dolphin, Tully, Sharina, et al. (2006) present HST images of KK~246 among other nearby galaxies. The optical image of KK~246 looks similar to other dwarfs (to my untrained eye) that are not so extremely isolated. Kreckel, Peebles, van Gorkom, van de Weygaert, and van der Hulst (2011) present maps of the extended atomic hydrogen envelope around KK~246. The long axis of the stellar distribution is tilted relative to this hydrogen envelope. This is curious because the tilt does not seem likely to be a long-lasting feature in a galaxy that looks so well isolated. Perhaps KK~246 was disturbed by a relatively recent merger with another dwarf galaxy, though that would seem odd given the isolation. Perhaps, as Tinker and Conroy (2009) argued, the local void contains numerous dark matter halos that have too few stars and too little atomic hydrogen to be observable. Maybe KK~246 was disturbed by one of them.
Rizzi, Tully, Shaya, et al. (2017) present an HST image of ZOA~J1952+1428; it too has the appearance of other low mass early-type galaxies. McIntyre et al. (2011) found that the mass of the atomic hydrogen envelope is $M_{\rm HI}=10^{7.0}M_\odot$. The optical luminosity, $L_B=10^{7.5}L_\odot$, suggests the mass in HI is less than the mass in stars. This is unusual; Bradford, Geha, and Blanton (2015) find that isolated low mass galaxies typically have considerably more mass in atomic hydrogen than in stars. One might wonder whether ZOA~J1952+1428 has been disturbed by an event that dissipated much of its HI, maybe supernovae, though that has not affected other dwarfs, or something external, though it appears to be isolated.
Another interesting object in the low density region toward the top of Figure~\ref{fig:LocalVol} is the spiral galaxy NGC~6946. It is marked by the open red circle at the largest positive value of SGZ. The ambient density is low there, but the image of this galaxy (to be seen on the web) looks much like the large spirals in the far more crowed region near the plane of the Local Supercluster that runs across the middle of the figure. A quantitative measure of that is the tight relation between the spiral galaxy circular velocity and baryonic mass (McGaugh 2020), which does not offer much room for sensitivity to environment. The atomic hydrogen surrounding NGC 6946 extends well beyond the starlight (Boomsma, Oosterloo, Fraternali, van der Hulst, and Sancisi, 2008), and it has the customary retinue of dwarf satellite galaxies (Karachentsev, Sharina, and Huchtmeier 2000).
The galaxy NGC~6946 is a counterexample to one of those arguments that seem to make intuitive sense. In the standard cosmology the primeval departures from homogeneity are a random Gaussian process. For simplicity reduce this to two waves, a long wavelength one that represents ambient conditions, and a short wavelength one that represents the seeds of galaxy formation. Suppose a seed that happened to be near a maximum density in the long wavelength component, the ambient density, could produce a large galaxy like NGC~6946. That same seed that happened to be near a minimum of the ambient density would have a smaller total density; it would end up as a dwarf. The picture looks reasonable. It works for the most massive galaxies, which are found in particularly dense regions such as clusters of galaxies. It makes sense put another way: a galaxy might be expected to grow larger where the ambient density is larger and better able to supply matter to the growing galaxy. But this intuition does not account for the presence of NGC~6946 in the low density region above the Local Supercluster in Figure~\ref{fig:LocalVol}. And it does not account for the general evidence of similar space distributions of $L\sim L_\ast$ galaxies and the far more numerous dwarf galaxies (Davis, Huchra, Latham, and Tonry 1982; Zehavi, Zheng, Weinberg, et al. 2011.) This failure of intuition is an anomaly.
Conditions in the Local Void are different from our neighborhood, to judge by the scarcity of galaxies. What might be new and interesting there? Maybe dark matter halos with HI but no stars, or dark matter halos without baryons, or even HI clouds without dark matter?
Arrays of telescopes such as MeerKAT are sensitive to the 21-cm line from atomic hydrogen, and will be surveying the Local Void as part of scans of all the sky at all the radio frequencies accessible to the telescopes. But the Local Void is interesting enough to justify a Grand Project: a far deeper search for 21-cm sources confined to the part of the sky and the range of redshifts of the Local Void. This restricted use of an important facility would limit its production, but consider the compelling scientific interest in this unique opportunity to probe a void as deeply as possible.
The image of the dwarf KK~246 in the Local Void is easy to see on the digitized ESO sky survey plates (when I have been shown where to look), and I suppose there is not likely to be more dwarfs this luminous in the Local Void at this distance and not obscured by dust at low galactic latitude. The fainter Local Void dwarf ZOA~J1952+1428 was discovered as a 21-cm source, but the HST image in Rizzi et al. (2017) certainly looks like an unambiguous detection of the stars. This means an optical to infrared search for more of these faint dwarfs in the Local Void is technically feasible, if given expensive resources.
The Local Void is particularly interesting because it can be examined in particular detail. Why are galaxies of stars so scarce in this void? Why does the spiral NGC~6946 with its retinue of dwarfs show so little indication of having been affected by its isolation? Why do the two dwarfs in the part of the Local Void pictured in Figure~\ref{fig:LocalVol} seem unusual despite their apparent isolation? What else is in this void?
\section{Galaxies}\label{sec:galaxies}
A century of research on the nature of galaxies\footnote{A century ago \"Opik (1922) turned earlier thoughts that the spiral nebulae might be other galaxies of stars into a quantitative demonstration. \"Opik started with the assumption that the Andromeda Nebula M~31 ``consists of stellar matter similar to the matter of our Galaxy,'' with the same ratio of mass to luminosity as the estimate for the Milky Way. That combined with the angular size of M~31, its apparent magnitude, and the measured rotation velocity, from the Doppler shift, yields a useful estimate of the distance and mass of this galaxy. (Showing how this follows is a good exercise for the student.) \"Opik's distance is half the correct value, an impressive advance at the time, and clear evidence that M~31 is a massive galaxy of stars, comparable to the size of the Milky Way.} has yielded a rich phenomenology and the challenge of understanding how or whether the phenomenology agrees with the cosmology. The complex nature of galaxies limits this test, but there are regularities that are useful hints to how the galaxies formed, which in turn offer guidance to whether the properties of galaxies fit what is expected in the standard $\Lambda$CDM cosmology. An example from the late 1990s is the prediction by Neta Bahcall and colleagues that in the Einstein-de Sitter model the masses of rich clusters of galaxies grow more rapidly than observed (Bahcall, Fan, and Cen 1997). This was credible early evidence that the mass density is less than predicted by the Einstein-de Sitter model that was popular then. The evidence remains credible and an example of how galaxies serve to test cosmology. Deciding which galaxy regularities, or curiosities, seem worthy of closer attention must be a matter of taste, of course. I offer the following potentially informative lines of thought.
\subsection{Early and Late Type Family Resemblances} \label{sec:earlyandlate}
Galaxies, like snowflakes, are all different if examined closely enough. You can see this by looking at the images of nearby $L\sim L_\ast$ galaxies to be found on the web. Among them are distinctly odd objects, but they look odd because they do not resemble the great majority of nearby galaxies that are readily classified as either elliptical or spiral, or in common usage early or late.\footnote{A century ago the two morphological classes, spiral and elliptical, were well known, as seen in Wolf's (1908) sketches showing examples that include the elliptical NGC 4494 and the spiral M~101. Jeans, Hubble, and others thought a galaxy might evolve from one type to the other, hence the names early and late. This now seems unlikely, but use of the names remains common.} In broad terms, another way to put it is that, with occasional exceptions, late type $L\sim L_\ast$ galaxies are gas-rich and early types are gas-poor. Stars are forming frequently enough in a gas-rich galaxy that the short-lived massive luminous blue stars tilt the spectrum of the galaxy to the blue; it is said to be in the blue cloud in a scatter plot of galaxy color and luminosity. A gas-poor galaxy has relatively few massive young blue stars; it fits in the red sequence in this color-luminosity plot (e.g., Salim, Rich, Charlot, et al. 2007, Fig. 1). S0 galaxies are a complication, but they are not common nearby, and they do not figure much in this essay (which might be a serious omission, as noted in footnote~\ref{fn:S0s}).
The galaxies in each of the two distinct types have their own family resemblances,\footnote{My use of the term, family resemblance, follows the Wikipedia interpretation of Ludwig Wittgenstein's thinking: family members share resemblances, or features, though no member need have all features. I take the term to be equivalent to family traits.} as in the red sequence and blue cloud. Tully and Fisher (1977) pointed out that the luminosity of a spiral galaxy is correlated with the circular velocity of the stars and gas in the disk. McGaugh (2020) presents the extension to include the mass in atomic hydrogen, which gives a tight correlation between the observed baryonic mass of a spiral and the circulation speed in its disk. This is a family resemblance, a characteristic of spiral galaxies. The analog for the early-type family began as the Faber and Jackson (1976) correlation between the luminosity and velocity dispersion of the stars in an elliptical galaxy. It was sharpened to the fundamental plane relating the elliptical galaxy luminosity, velocity dispersion, and radius (Dressler, Lynden-Bell, Burstein, et al. 1987; Djorgovski and Davis 1987). Bernardi, Nichol, Sheth, Miller, and Brinkmann (2006) show an example of this family trait, or regularity, and demonstrate that the regularity is not sensitive to ambient density.
A notable family trait among ellipticals is the correlation of the spectrum of the galaxy with its stellar velocity dispersion: the greater the velocity dispersion the redder the mean spectrum (Zhu, Blanton, Moustakas 2010). But Zhu et al. show that at given velocity dispersion the mean spectra are very similar for ellipticals in more crowded and less crowded environments (apart from more prominent H-$\alpha$ emission in field ellipticals). If ellipticals grew by dry mergers of star clusters that had a range of values of velocity dispersions, and hence a range of different spectra, then one might have predicted a sensitivity of the assembled elliptical to the present local situation, which might be expected to be correlated with the degree of merging. But the effect on the spectra is difficult to see.
The largest galaxies, with luminosities $L\sim 10L_\ast$ prefer dense regions. But the properties of the early and late-type galaxies with $L\sim L_\ast$, the ones that contribute most of the cosmic mean luminosity density, are insensitive to environment. This is notable evidence.
The early family type prefers denser regions. Early and late types have different life histories. Ellipticals have larger abundances of the alpha-process elements---carbon, oxygen, and so on---that are produced in early generations of massive stars, and lower abundances of the iron group elements that are more slowly produced in explosions of type~I supernovae. The abundance pattern in the spiral family brings to mind slower build-up of the elements, which agrees with the different distributions of stellar ages in the two families.
\begin{figure}
\begin{center}
\includegraphics[angle=0,width=3.5in]{BtoT.pdf}
\caption{\small Frequency distributions of ratios of bulge to total luminosities of nearby $L\sim L_\ast$ galaxies, from Peebles (2021).}\label{fig:BtoT}
\end{center}
\end{figure}
Figure~\ref{fig:BtoT} shows measured values of bulge to total luminosity $B/T$ for 32 galaxies that are within 10 Mpc distance and have luminosities $L_K > 10^{10}$ (from Kormendy, Drory, Bender, and Cornell 2010; and Fisher and Drory 2011). The sample is small, but beautiful images of the galaxies mentioned in the next paragraph are to be seen and admired on the web.
The three ellipticals --- Centaurus~A, Maffei~1, and M~105 --- are in the modest peak at the right-hand side of the figure. The stars in these ellipticals are supported by near isotropic motions; we may say these stars are a hot component. The stars in the disk of a spiral galaxy are supported by rotation with a relatively small scatter around the mean: a cool component. The Sombrero Galaxy NGC~4594 at the center of the figure has $B/T=0.5$. It looks like a large spiral centered on an elliptical of similar size. Other disk galaxies further to the left in the figure, including the spirals M~31 and M~81, have a classical bulge, a hot component that rises above the disk. In these galaxies the bulge is more compact than in the Sombrero Galaxy. The pure disk spirals near the peak at the left-hand side of Figure~\ref{fig:BtoT} do not have an appreciable classical bulge, and their images look strikingly flat. Examples are M~101, NGC~253, and the edge-on galaxy NGC~4945. These pure disk spirals might have a pseudobulge, an unusually large surface brightness in the disk near the center. Authorities warn that observations of more distant galaxies at poorer spatial resolution may mistake a pseudobulge for a classical bulge; deciding which it is can be difficult. Some pure disk galaxies have a bar of stars that runs across the center of the galaxy; NGC~1300 is a pronounced example (at greater distance than the other galaxies mentioned here).
There are exceptions to the two families. I mentioned the Sombrero Galaxy. The S0 galaxies have near featureless disk-like distributions of stars with bulges,
giving the impression of spirals that have lost the spiral arms but kept remnants of the disk stars and the dust. Examples of nearby S0s are NGG~404, about 3~Mpc away, and NGC~2784, at about 10~Mpc. The S0 NGC~1460, at about 20 Mpc distance, is an elegant example of a barred galaxy without the spiral arms. These are rare exceptions to the population of galaxies outside clusters of galaxies; common in clusters. There are irregular galaxies; NGC~4490 looks like it is merging or falling apart, and the NASA/IPAC Extragalactic Database seems to be uncertain about the classification of the Circinus Galaxy. These exceptions are real but not common among nearby large galaxies. If the galaxies closer than 10~Mpc are a fair sample of the situation outside clusters of galaxies then they present us with clear and persuasive evidence that galaxies exhibit a distinct bimodality in their family traits.
It is said that elliptical galaxies formed by dry mergers, spirals by wet. Maybe an example is the situation in the group that contains the radio galaxy Centaurus~A. The two largest members are the elliptical NGC~5128, which is the radio source, and the spiral M~83. Figure 1 of Karachentsev, Sharina, Dolphin, et al. (2002) shows that most of the smaller galaxies around the late type M~83 also are late types, and most of the smaller galaxies around the early type Centaurus~A are early types. This agrees with the thought that early type galaxies grew by mergers of dry subhalos while late types grew by wet mergers. It is a description, of course, not an explanation.
People have been wondering about the origin of the early-late bimodality, and more broadly the Hubble sequence of galaxies, for the last century. Modern numerical simulations based on the $\Lambda$CDM cosmology capture aspects of the early and late morphologies (e.g., Vogelsberger, Genel, Springel, et al. 2014). Nelson, Pillepich, Springel, et al. (2018) show in their Figures~1 and~3 distributions of model galaxy color and stellar mass from their simulations of the formation of the central galaxies in dark matter halos. The distributions are quite similar to the Kauffmann, Heckman, White, et al. (2003) results from their analyses of the SDSS observations, which is encouraging. And their models at stellar masses $\sim 10^{10}M_\odot$ show bimodal morphologies. The empirical situation is richer, of course. There are comparable numbers of the most massive galaxies with elliptical and spiral morphologies in the Huchra et al. 2012 catalog (as in Fig.~\ref{fig:morphologies} in Sec.~\ref{sec:LSC}, and in Ogle, Lanz, Appleton, Helou, and Mazzarella 2019), and there are clear examples of spirals and ellipticals at luminosities $L\lower.5ex\hbox{$\; \buildrel < \over \sim \;$} L_\ast$. Understanding the distinct nature of galaxy bimodality remains an interesting challenge.
\subsection{What is the Separatrix for Bistable Galaxy Formation?}\label{separatrix}
Current thinking, which is well motivated by the success of the standard cosmology, is that galaxies grew by gravity out of tiny primeval departures from an exactly homogeneous mass distribution, a stationary random Gaussian adiabatic process. The baryonic and dark matter gathered into mass concentrations, or halos, which grew more massive by merging with other halos and accretion of diffuse matter. Bayons settled, stars formed, and a protogalaxy grew into one or the other of the distinct galaxy families discussed in Section~\ref{sec:earlyandlate}.
Sidney van den Bergh's (1976) thinking about galaxy morphologies a half-century ago was that
\begin{quotation}\noindent
\noindent Canonical views on galaxy evolution suggest that the present morphology of galaxies is predestined by the genetic heritage provided by initial mass and angular momentum. The results discussed above suggest that the evolution of galaxies is also substantially affected by environmental factors.
\end{quotation}
Both thoughts remain empirically well supported. We can add that, if galaxies grew by gravity out of small primeval Gaussian departures from homogeneity, then galaxy formation had to have been a bistable process. The point can be made a little more explicit by recalling the idea of bistable evolution and its separatrix in classical mechanics.
Suppose the state of a system is completely described by the values of $N$ components of particle positions and their $N$ canonical momenta. Let these $2N$ parameters be the coordinates in a $2N$ dimension phase space. The initial condition of the system is represented by its position in this space at a chosen starting time. Imagine an ensemble of initial conditions spread across phase space at this starting time. The equation of motion determines the evolution of the system, its path through phase space, from each initial position. In a bistable situation paths in phase space from the distribution of initial conditions arrive at one or the other of two (or more) basins of attraction. The separatrix is the boundary that separates initial positions in phase space that end up in one of the basins of attraction from the initial positions that end up in the other(s). The orbits of stars may have separatrices (e.g., Yavetz, Johnston, Pearson, Price-Whelan, and Weinberg 2021). The evolution of protogalaxies from their initial conditions is much more complicated; we must take account of dissipation, for example, and consider the complexities of stellar formation and its effects on the evolution of the galaxy. But the example from classical mechanics illustrates the concept of evolution of protogalaxies from initial conditions without manifest bimodality to a bimodal final state. It is what seems to have happened.
So in this way of thinking, what is the separatrix in galaxy formation that determines the evolution of a protogalaxy to a spiral or elliptical morphology? It cannot simply be the mass. There are spirals among the most luminous of galaxies, at $L \sim 10L_\ast$. (The distributions of the most massive spiral and elliptical galaxies relative to the plane Local Supercluster are shown in Fig.~\ref{fig:morphologies}.) At least some of these supermassive late types have the familiar two arms elegantly spiraling out from the center. An early example is UGC 2885 (Rubin, Ford, and Thonnard 1980); Ogle et al. (2019) catalogue others. At lower stellar masses there are more spirals than ellipticals, but both types are observed. Van den Bergh (1976) had good reason to mention mass as part of the genetic heritage, but the story must be more complicated.
The disk of stars in a spiral galaxy is supported largely by rotation, while the stars in an elliptical are supported by a closer to isotropic distribution of orbits. Thus van den Bergh had good reason to consider that the separatrix is related to angular momentum. A dimensionless measure of the angular momentum $L$ of a galaxy is the combination $\Lambda = L~E^{1/2}G^{-1}M^{5/2}$, where $E$ the magnitude of the binding energy and $M$ is the mass (Peebles 1971). In analytic estimates and numerical simulations the distribution of $\Lambda$ is not bimodal (e.g., Efstathiou and Jones 1979). If, despite this, $\Lambda$ is the separatrix, the division between early and late types would require a sharp sensitivity to the values of $\Lambda$ and mass. A more likely picture along this line is that morphology is determined by ``the {\it coherent alignment} of the angular momentum of baryons that accrete over time to form a galaxy'' (Sales, Navarro, Theuns, et al. 2012). The investigation of galaxy morphology and halo spin in numerical simulations by Rodriguez-Gomez, Genel, Fall, et al. (2022) reveals a systematic difference of angular momentum of models classified as spirals and as ellipticals plus S0s. It is not yet bimodality in the spin-stellar mass plane (in their Fig.~1), but perhaps a step in this direction.
Environment matters. The giant $L\sim 10L_\ast$ galaxies in rich clusters likely formed by mergers of cluster members, meaning environment likely is the separatrix between these giants and ordinary $L\sim L_\ast$ galaxies. Maybe another example follows from the larger ratio of early to late types in denser regions. For example, one might imagine that all protogalaxies began evolving toward the spiral morphology, but that violent mergers turned some proto-spirals into proto-early types. It would have happened more frequently in more crowded environments. But recall the separate family resemblances of spirals and ellipticals, which do not seem to be sensitive to environment. And consider the curious separation of early and late types in the Centaurus group (Sec.~\ref{sec:earlyandlate}).
The evidence is that the formation of galaxy morphologies was determined more by nature than nurture, and it is an interesting challenge to identify the character of the separatrix. It might be some combination of mass, angular momentum, and environment, or maybe something completely different. The issue might be resolved by what is learned from numerical simulations of galaxy formation, or maybe by semi-analytic considerations of what appears to be happening. It is a fascinating opportunity for research, provided you bear in mind that people have been trying to solve the puzzle of the early-late bimodality for a long time. It means the resolution must be subtle, but surely it exists.
\subsection{Bulges and Disks of Spiral Galaxies}\label{sec:spirals}
Numerical simulations of galaxy formation produce spiral galaxies that are impressively good approximations to what is observed, but there are three (or more; I invite suggestions) issues that are persistent enough to merit attention. One is that the central concentrations of starlight in simulated galaxies are overly luminous. Another is that the velocity dispersions of the stars moving in the planes of the disks of model spiral galaxies are unrealistically large. And a third is that we do not know the separatrix responsible for the distinct galaxy bimodality. I review these issues in Peebles (2020b); they are outlined and considered further here.
Early attempts to simulate galaxy formation encountered the problem that a cloud of baryonic matter --- gas and plasma --- with the mass and radius typical of an $L\sim L_\ast$ galaxy readily dissipates energy and collapses almost freely. Observed spiral galaxies must have avoided this overcooling problem, and it must be avoided in models for otherwise they would have overly prominent classical bulges or stellar halos. The problem has been tamed by adjustments of the prescriptions for star formation and models for the effects of the stars on the distributions of baryonic and dark matter, but the evidence is that the problem in simulations of galaxy formation persists.
Figure~\ref{fig:BtoT} shows the distribution of the ratio of bulge to total luminosities of the nearby large galaxies. For some galaxies we can add to the hot component in the bulge the hot stellar halo that spreads out to greater distances away from the disk. Estimates of the median value of the luminosity fraction of the stars in the two hot components, bulge plus halo, in spiral galaxies are (from Peebles 2020b)
\begin{equation}
\hbox{simulations: }{B+H\over T}\sim 0.45,\quad \hbox{observations: }{B+H\over T}\sim 0.18. \label{eq:BHoverT}
\end{equation}
The rest of the total luminosity, $T$, is assigned to the disk. The median of the observed fraction is from examinations of ten nearby galaxies by Merritt, van Dokkum, Abraham, and Zhang (2016) and Harmsen, Monachesi, Bell, et al. (2017). The hot fraction in simulations is from reports by Grand, G{\'o}mez, Marinacci, et al. (2017) and Garrison-Kimmel, Hopkins, Wetzel, et al. (2018) of the results of two large research programs. The greater hot fraction in simulations agrees with my visual impression of images of real and model spirals. You are invited to check your impression.
A second anomaly is the large dispersion of model disk stars in the direction of the plane of a simulated spiral galaxy. An illustration uses a simplified model for a spiral galaxy in which the stars move in the plane of a disk with a flat rotation curve, constant circular speed $v_c$. This is a reasonable approximation to many observed and model spiral galaxies. I refer to Peebles (2020b) for the details of the results of computation of stellar orbits. The orbit of a model star is characterized by a parameter, $\epsilon$, that is a measure of the orbital angular momentum, with $\epsilon = 1$ for a star moving in a circular orbit and $\epsilon = -1$ for a star in a circular orbit but moving in the opposite direction from the mean motion of the disk stars.
Numerical solutions give the rms radial velocities at two choices of this circularity parameter:
\begin{align}
\langle (dr/dt)^2\rangle^{1/2} &= 0.32 v_c\hbox{ for } \epsilon = 0.9; \nonumber\\
&= 0.45 v_c \hbox{ for } \epsilon = 0.8. \label{eq:radialveldispn}
\end{align}
I have not found discussions of disc star velocity dispersions in model disk galaxies. My estimate is that in recent suites of numerical simulations (Grand, et al. 2017; Garrison-Kimmel et al. 2018) the most promising of the distributions in $\epsilon$ for a spiral galaxy have at least a quarter of the stars at $\epsilon<0.9$, which means that the radial velocity dispersions are greater than about a third of the circular velocity $v_c$ in a quarter of the stars. A majority of the stars in a promising model have $\epsilon<0.8$, with radial velocity dispersion greater than about half the circular velocity. Observations of the distribution and motions of the stars in our neighborhood of the Milky Way Galaxy (Anguiano, Majewski, Hayes, et al. 2020) indicate the radial velocity dispersion in the thin plus thick disk stars is about $\sigma_r=43$~km~s$^{-1}$ with $v_c\sim 240$~km~s$^{-1}$. The models look much hotter.
The evidence reviewed here is that the model galaxies that emerge from modern simulations of cosmic evolution have unacceptably large populations of stars with large velocity dispersions present in stellar halos, classical bulges, and disks. This looks quite different from the impression of cool populations of stars and gas in the common nearby $L\sim L_\ast$ pure disk galaxies (with the small fraction of the stellar mass in hot stars in the halo). I do not know whether this anomaly has resisted persistent attempts at remediation, or perhaps it has been put aside pending explorations of how to deal with the many other complexities in modeling galaxy formation. But since the art of simulating galaxy formation has a large literature, and the problem with hot star populations in simulated spiral galaxies remains, it ranks as a serious anomaly. Perhaps the situation will be resolved by further advances in numerical simulations based on the $\Lambda$CDM theory. Or again, maybe something is missing.
\subsection{Merger Trees and the Cosmic Web}
The phrases, ``merger tree'' and variants, and ``cosmic web,'' often figure in discussions of how the galaxies formed. Aspects of both are worth considering here.
Merging certainly happens; a clear example of a violent merger is the nearby Antennae Galaxies. Ostriker (1980) introduced considerations of what the remnant of the merger of two $L\sim L_\ast$ galaxies, as in this example, would look like when it had relaxed to a close to steady state. Let us only note that the remnant would have a luminous stellar bulge and halo made of pre-existing stars, which certainly is not seen in the nearby pure disk galaxies. And if the remnant of merging spirals looked like an elliptical it would have an unusual mix of chemical elements. If the local galaxies are close to a fair sample of the situation outside clusters then relaxed remnants of mergers of $L\sim L_\ast$ galaxies are not common, because pure disk galaxies are common, and I expect we would have heard of it if ellipticals with odd chemical abundances were common.
A more modest example of merging is the pure disk edge-on galaxy NGC 5907 with its stellar stream that we may expect eventually will add to the stellar halo of this galaxy (e.g., van Dokkum, Lokhorst, Danieli, et al. 2020). The ample evidence of tails and streams of stars around ellipticals and spirals is suggestive of close passages and mergers (e.g., van Dokkum 2005). And the stellar halo of the Milky Way is growing by tidal disruptions of dwarf galaxies (e.g., Belokurov, Zucker, Evans, et al. 2006)
The concept of galaxy formation as a hierarchical merging process grew out of several considerations. As just noted, galaxies do merge. The distribution of galaxies on scales from $\sim 0.1$~Mpc to $\sim 10$~Mpc is well approximated as a scale-invariant clustering hierarchy. It seems natural that the hierarchy also formed at smaller scales and was erased by merging to form galaxies (Davis, Groth, Peebles, 1977). And hierarchical growth of clustering is seen in numerical simulations of cosmic structure formation. But although the merger tree concept is well motivated by theory and observation it is in the spirit of empiricism to ask whether, absent simulations but given our knowledge of the phenomenology, people would have been led to the hierarchical assembly picture.
Absent the guidance of merger trees people might have settled on the Eggen, Lynden-Bell, and Sandage (1962) account of the formation of the stellar halo and disk of the Milky Way spiral galaxy by a closer to monolithic collapse. The theory would include occasional mergers of galaxies, as observed. It would include formation of substructure in protogalaxies to account for satellites and the spreading of baryons across the disks of the Milky Way and other spiral galaxies, but substructure need not be to the extent of formation of identifiable halos that merge to form identifiable halos in a merger tree. The Eggen et al. picture is more in line with the observation that the distributions of heavy elements in halo stars do not resemble the distributions in stars in dwarf satellites (e.g., Tolstoy, Hill, and Tosi 2009, Fig. 10). Our stellar halo instead would have formed in the closer to monolithic collapse Eggen et al. envisoned, and would have been salted by stars from accreted dwarfs that produced the ``field of streams'' (Belokurov, et al. 2006).
How are we to interpret the description of the formation of early type galaxies by dry mergers and the late type by wet mergers, those rich in diffuse hydrogen? It calls to mind formation of morphology by nature rather than nurture: protogalaxies that have dry or wet natures. That could happen in a merger tree, but more simply in a monolithic collapse.
Cowie, Songaila, Hu, and Cohen (1996) pointed out that lower mass galaxies on average formed the bulk of their stars later. This downsizing effect does not naturally follow from a hierarchical merger tree in which less massive halos formed earlier. The discrepancy need not be serious because galaxy formation is complicated, but it does call to mind a picture similar to Eggen et al. (1962).
Madau and Dickinson (2014) concluded that ``galaxies formed the bulk (75\%) of their stellar mass at $z < 2$.'' The growth of the stellar mass of a pure disk galaxy in the manner described by Madau and Dickinson cannot have been by the merging of subhalos, or galaxies, that contained many stars, because the stars would have ended up in stellar bulges or halos, which are not prominent in these galaxies. The stars in pure disk galaxies had to have formed out of gas or plasma that had settled to the disk, as in cool streams (e.g. Kretschmer, Dekel, and Teyssier, 2022). This has the flavor of the Eggen, Lynden-Bell, and Sandage picture.
The phrase, ``cosmic web,'' also often figures in discussions of galaxy formation. Bond, Kofman, and Pogosyan (1996) introduced the concept as an evocative description of the distribution of dark matter in numerical simulations: large and small concentrations of dark matter are connected by filaments of dark matter in a pattern that calls to mind a web. It also resembles the observed galaxy distribution. Bond et al. pointed out that the filaments in simulations might be observable by detection of atomic hydrogen along the filaments. One might imagine that the filaments also are threaded by a magnetic field, and maybe even cosmic strings.
Detection of HI streams would be particularly interesting because the filaments are expected to be present, connecting collapsing concentrations of dark matter, if the dark matter is adequately well approximated as a continuous fluid with no pressure or viscosity and the initial conditions are continuous. Dark matter consisting of black holes would be expected to form streams if the masses were small enough, and separate dark matter halos if the masses were large enough.
Instrument arrays capable of detecting 21-cm radiation from the hydrogen in primeval filaments of matter connecting dark matter halos may be becoming available (e.g., Tudorache, Jarvis, Heywood, et al., 2022; Greene, Bezanson, Ouchi, et al., 2022). It will be interesting to see the nature of the H{\small I} distribution between mass concentrations, and the constraint that places on the black hole mass in a black hole model of the dark matter.
\subsection{Massive Black Holes}\label{massiveblackholes}
The large luminosities and compact natures of quasars led to the thought that these objects are powered by the energy released by the collapse of matter onto black holes with masses of perhaps a million solar masses (Salpeter 1964; Lynden-Bell 1969). The clear evidence now is that large galaxies contain central compact objects with masses in the range of $10^5$ to $10^{10} M_\odot$. The objects in the centers of the elliptical M~87 and our Milky Way spiral galaxy certainly are compact (Event Horizon Telescope Collaboration et al., 2019; 2022). It makes a good case that these two objects, and the ones in other galaxies, are supermassive black holes of the kind predicted by Einstein's general theory of relativity.
It was natural to suppose that matter would settle to the center of a galaxy and perhaps accumulate to the point of relativistic gravitational collapse. But that picture is complicated by the extreme difference between the density characteristic of an $L\sim L_\ast$ galaxy, perhaps $\sim 10^{-24}$ g~cm$^{-3}$, and the density characteristic of a $10^{9} M_\odot$ black hole, $\sim c^6G^{-3}M^{-2}$, roughly $1$~g~cm$^{-3}$. Feeding the growth of a supermassive black hole by dissipative settling of diffuse baryonic matter certainly is conceivable, but one might instead expect that the settling would result in multiple fragmentation and the formation of star clusters, as in the formation of the first stars (e.g., Abel, Bryan, and Norman, 2002). Statistical relaxation of the star cluster would in time produce core collapse to a central black hole surrounded by a nuclear star cluster. But this is only one of the several current lines of thought that are reviewed by Greene, Strader, and Ho (2020), along with a discussion of how these black holes are detected. Issues that seem particularly relevant are discussed here.
A clue to the formation of these central black holes is the relation between the black hole mass and properties of the galaxy such as the bulge or spheroid mass or stellar velocity dispersion (e.g., Magorrian, Tremaine, Richstone, et al., 1998). We need more than one relation, because some of the pure disk $L\sim L_\ast$ galaxies that are common nearby have massive central black holes with at most modest classical bulges. The familiar example is the Milky Way Galaxy with its central bar, little starlight in a classical bulge, and clear evidence of a central black hole with mass $4\times 10^6 M_\odot$. Another is the galaxy NGC~4945 (Gaspar, D{\'\i}az, Mast, et al. 2022 and references therein). This galaxy is seen nearly edge-on, it looks wonderfully flat, and there is little indication of a concentration of starlight in a classical stellar bulge rising out of the disk. (You can see the image of this galaxy at \url{https://apod.nasa.gov/apod/ap220226.html}.) The evidence is that this galaxy contains an active galactic nucleus operating around a black hole with mass comparable to the one in the center of our Milky Way galaxy.
If evidence accumulates that every $L\sim L_\ast$ galaxy has a central supermassive black hole it will invite the thought that galaxies formed around black holes, whether supermassive or collections of less massive black holes that seeded their formation (e.g., Silk and Rees 1998; Carr and Silk 2018). The thought is encouraged by the observations of quasars at redshifts $z\sim 7$. These quasars presumably depended on the presence of massive black holes, at a cosmic time when the stellar masses of galaxies were much smaller than now.
In the Local Void (discussed in Sec.~\ref{sec:localvoid}) galaxy formation seems to have been suppressed. The thought that galaxies formed around black holes suggests that the Local Void contains black holes, seed or supermassive, that are centered on galaxies that are unusually small for their black hole masses. Maybe closer examinations of the space distributions and redshifts of the stars in the two void dwarfs at the top and to the left in Figure~\ref{fig:LocalVol} could determine whether they contain unusually massive black holes for such small galaxies. There are other void galaxies to examine, and there is a considerable area of sky and range of redshifts in the Local Void for deeper surveys for HI sources, star clusters, and maybe even gravitational lensing by primeval black holes without stars, any of which could be a valuable clue to the origin of massive black holes.
Seth, van den Bosch, Mieske, et al. (2014) present evidence of an ultra-compact dwarf galaxy with an exceptionally massive central compact object, maybe a black hole. It may be near the large elliptical galaxy M~60. Maybe this dwarf is the remnant of a galaxy whose growth around a particularly massive primeval massive black hole was interrupted by tidal stripping. It is easy to invent such scenarios, sometimes difficult to test them, but perhaps an essential part of the search for the explanation of these objects.
Are there normal-looking $L\sim L_\ast$ galaxies that do not have a central massive black hole? The relatively nearby and face-on spiral galaxy M~101 has lanes of dust that spiral in toward the center, ending at a star cluster with mass $\sim 4\times 10^6M_\odot$ (Kormendy et al 2010). It is not yet known whether there is a massive central black hole inside the star cluster. A detection, maybe from the shapes of integrated Doppler-broadened stellar lines, would be interesting, and a seriously tight upper bound on the mass of a central black hole would be even more interesting.
The normal-looking spiral galaxy M~33, the third largest galaxy in the Local Group, does not have a central black hole more massive than about $2\times 10^3M_\odot$ (Gebhardt, Lauer, Kormendy, et al., 2001; Merritt, Ferrarese, and Joseph, 2001). It is difficult to imagine how a merger of two galaxies could have driven both black holes out of the remnant while leaving the pure disk morphology of this galaxy. The more likely interpretation is that the formation of M~33 did not require a primeval black hole. It would mean that the massive black holes in other galaxies need not have served as seeds for galaxy formation, but instead grew together with the galaxies.
In the coevolution picture the accumulation of mass in a growing central black hole would be by dissipative settling and merging at rates that might be expected to differ from galaxy to galaxy. This could be compared to the formation of stellar bars that run across the centers of spiral galaxies: some dominate the shape of the spiral, some are less conspicuous, and others are not noticeable. So it might be with the formation of supermassive black holes. The correlation of mass with host galaxy properties argues against a broad scatter of massive black hole masses at given galaxy mass, but it will be helpful to see the scatter of central black hole masses as a function galaxy properties in larger samples. It might serve as a test of the coevolution picture. And it is to be noted that the coevolution picture does not seem promising for M~33, because there is no evidence of a central massive black hole.
The LIGO detection of gravitational waves from the merging of black holes with masses $\sim 66$ and $85 M_\odot$ (the event GW190521 reported by Abbott, Abbott, Abraham, et al. 2020) was unexpected because the masses are intermediate between the black holes produced by the relativistic collapse of stars and the supermassive black holes in the centers of galaxies. Maybe they are in line with the idea that black holes in this intermediate mass range were seeds for the formation of supermassive black holes. There is a long span of logarithmic time from the end of inflation, or whatever saved us from the spacetime singularity of the standard cosmology, to the formation of the isotopes of hydrogen and helium. During this time cataclysmic events of some sort might have produced massive black holes or their seeds. The favorite thought is directed to the disturbances to the mass distribution as the universe expanded and cooled through first-order cosmic phase transitions. These transitions might have been violent enough to have produced seed black holes, maybe with a broad range of masses set by the variety of cosmic first-order transitions in standard particle physics (e.g., Cappelluti, Hasinger, and Natarajan, 2022). Another thought is that supermassive black holes or their precursors formed during cosmological inflation (e.g., Kallosh and Linde 2022). Yet another is that seed mass black holes formed by collapse of the first generation of gravitationally bound clouds of hydrogen and helium with mass $\sim 10^5M_\odot$ set by the baryon Jeans length (Silk and Rees 1998).
Considerable research on the theory and observations of supermassive black holes in the centers of galaxies has not yet produced a convergence to the theory the origin of these objects. Their presence remains an anomaly.
\subsection{Why the Characteristic Galaxy Luminosity?}\label{sec:Lstar}
The frequency distribution of optical luminosities of galaxies has a characteristic value, $L_\ast$. There are far more galaxies with luminosities less than $L_\ast$, but the $L\sim L_\ast$ galaxies produce most of the cosmic mean optical luminosity density. The largest known galaxies have luminosities $L\sim 10 L_\ast$. That factor of ten is a curiously abrupt cutoff compared to the broad range of luminosities of galaxies that are less luminous than $L_\ast$. What accounts for the value of $L_\ast$ and the cutoff at greater luminosities?
The abrupt cutoff was anticipated by Schechter's (1976) functional form for the galaxy luminosity function with its exponential cutoff. It grew out of a Press and Schechter (1974) argument that ``contains no ad hoc information about an initial spectrum of long-wavelength density perturbations.'' The counterargument in Peebles (1974) agrees with the more recent demonstrations that the formation of cosmic structure is sensitive to the form of the spectrum of primeval departures from homogeneity. The Press and Schechter argument nevertheless produced an analytic form for the luminosity function that remains broadly useful.
In pure dark matter numerical simulations of the growth of cosmic structure in the standard cosmology the halo mass function is not as abruptly truncated at the high mass end as the galaxy luminosity function (e.g., Garrison, Eisenstein, Ferrer, et al., 2018, Fig. 7). That need not be an anomaly; we must bear in mind the complexities of how mass was apportioned to dark matter halos and the baryons to stars. But is the existence of the characteristic luminosity $L_\ast$ and the sharp upper cutoff in galaxy luminosities an accidental result of these complexities, or might both be more readily understandable in an improved cosmology?
\subsection{Why MOND is Successful but Unpopular}\label{MOND}
The rotation speed $v_c$ of the stars and gas in the outer parts of the disk of a spiral galaxy typically is close to independent of radius. The standard interpretation is that the spherically averaged mass density in the outer parts of the galaxy varies with the distance $r$ from the galaxy as $\rho\propto r^{-2}$, which translates to gravitational acceleration $g\propto r^{-1}\propto v_c^2/r$, satisfying the condition that the speed $v_c$ is independent of distance from the galaxy. This is the flat rotation curve observed in many spiral galaxies. The starlight density in the outer parts of a typical spiral galaxy falls off more rapidly then $r^{-2}$. The standard remedy is the postulate that the mass in the outer parts of a galaxy is the nonbaryonic dark matter of the $\Lambda$CDM theory, the dark matter halo.
Milgrom (1983) introduced an influential alternative: modified Newtonian gravity, or MOND. Instead of the hypothetical dark matter Milgrom proposed that the rotation curve is flat in the outer parts of a galaxy because at gravitational acceleration less than a characteristic value, $a_0$, Newton's law is modified to gravitational acceleration $g = \sqrt{GMa_0}/r$. In this limit, and assuming most of the mass is within the radius $r$, then the speed in a circular orbit is $v_c=(GMa_0)^{1/4}$. If the luminosity $L$ of the stars in the galaxy is proportional to the mass $M$ in baryons, then MOND predicts that the value of the circular velocity $v_c$ in the outer flat part of the galaxy rotation curve scales with the luminosity of the galaxy as the universal form $v_c\propto L^{1/4}$. This is close to the empirical Tully and Fisher (1977) relation. It is even closer to the power law relation $v\propto M^{1/4}$ observed when $M$ is the mass in interstellar atomic hydrogen added to the mass in stars (McGaugh 2020).
MOND was proposed after the discovery of the Tully-Fisher relation, but it is reasonable to count the observed $v_c\propto M^{1/4}$ relation as a MOND prediction that passes a tight test. One may ask why this successful prediction receives so little community attention.
Let us note first that in the standard $\Lambda$CDM cosmology the observed $v_c\propto M^{1/4}$ relation does not challenge the theory. It is instead is a property of galaxies, along with the other family traits. The challenge is to explore in this theory how these family resemblances among galaxies grew as the universe expanded. It will be a crisis for $\Lambda$CDM if an explanation for the observed traits cannot be found within the theory. We know far too little about how galaxies formed to hope for a judgement on this point any time soon.
An argument for MOND is the considerable literature on how the properties of galaxies, and groups and clusters of galaxies, can be understood in this picture. It is reviewed by Diaferio and Angus (2016) and Banik and Zhao (2022). If in an alternative world all our present phenomenology of cosmic structure were known but nothing was known about the evidence of a relativistic evolving universe, MOND likely would be a community favorite. But in our world there are two serious reasons for limited interest in MOND.
First, the well-tested $\Lambda$CDM theory with its cold dark dark matter offers a ready and promising framework for development of a theory of the formation of cosmic structure, galaxies and all. It has attracted the attention of active and productive research groups. I have argued for problems with the results, but they are details that have not discouraged the research groups and I am hoping might guide us to adjustments that improve $\Lambda$CDM.
Second, Milgrom's (1983) MOND does not offer a ready framework for development of a viable cosmology. The approach explored by Angus (2009) and Diaferio and Angus (2016) follows thoughts about alternative gravity physics by Bekenstein and Milgrom (1984) and Milgrom (2010). A starting postulate is that there is dark matter, in the form of thermal sterile neutrinos with rest mass 11~eV. The dark matter density parameter is close to the standard model. Gravity physics in this model is enough like general relativity at redshifts $z \lower.5ex\hbox{$\; \buildrel > \over \sim \;$} 1000$, and the warm dark matter is enough like the cold dark matter of the standard cosmology, that the acoustic oscillation pattern imprinted on the CMB is close to the standard prediction and the measurements. This is an important result. At low redshift the neutrino hot dark matter would have largely escaped galaxy potential wells, and gravity physics would have become enough like Milgrom's original MOND that galaxy rotation curves fit measurements without dark matter in halos around galaxies and with the weaker gravity of MOND. This is viable, within the notions of the gravity physics in this picture, though contrived. It is not demonstrated in a theory that allows computations of predictions. Diaferio and Angus (2016) conclude that
\begin{quotation}
It remains to be seen whether [in the adopted gravity physics] gravitational instability in a universe filled with baryonic matter and one species of sterile neutrino with 11 eV mass can form the observed cosmic structure at the correct pace.
\end{quotation}
The neutrino dark matter might be trapped in growing clusters of galaxies, which could be a helpful feature for the application of MOND to clusters. But Diaferio and Angus caution that
\begin{quotation}
the ability to explain the cluster mass discrepancy does not directly imply that MOND combined with 11-eV sterile neutrinos can form clusters in a cosmological context.
\end{quotation}
To my mind it is exceedingly unlikely that an alternative gravity physics and cosmology without cold dark matter, or something that acts much like it, can fit the array of tests passed by the standard general theory of relativity applied to the $\Lambda$CDM cosmology. But good science seeks to replace intuition with worked predictions of adequately specified theories. As Diaferio and Angus conclude, it remains to be seen whether this can be done for a generalization of Milgrom's MOND.
\section{Summary Remarks}\label{SummaryRemarks}
To prevent misunderstandings I repeat the conclusion in Section~\ref{sec:tests}, that the empirical tests give excellent reason to expect that a more advanced physical cosmology will look much like the theoretical $\Lambda$CDM universe, because many well-checked tests show that the $\Lambda$CDM universe looks much our universe. But the great progress in cosmology and the other physical sciences has left anomalies, some of which have been troubling for a long time.
A century ago Pauli and Jordan understood the vast difference between the vacuum mass density allowed by the relativistic theory and the density suggested by quantum physics (Sec.\ref{sec:Lambda}). Applications of quantum physics have grown far broader, giving compelling evidence that this physics is a broadly useful approximation to reality, but I am not aware of a significant advance in resolving the quantum energy density problem, apart from the anthropic argument (in Sec.~\ref{sec:Anthropic}).
A century ago \"Opik (1922) turned earlier thoughts that the spiral nebulae might be other galaxies of stars into a quantitative demonstration. The progress from there to a promising physical basis for analyses of how the galaxies formed was slower than the development of quantum physics, but we have now a well-tested cosmology that might be expected to be an adequate basis for a secure understanding of these objects. The starting fundamental goals for a theory of the galaxies are to understand galaxy stellar masses, the spatial distributions of the stars in galaxies, and the stellar motions. Modern numerical simulations have made encouraging progress to this end, but there are anomalies. The evidence I have seen is that, despite careful attention to the many details required for the numerical simulations, disk star velocity dispersions in the planes of model spiral galaxies are unrealistically large (Sec.~\ref{sec:spirals}). The overcooling problem remains, resulting in unrealistically large classical bulges and stellar halos (Sec.~\ref{sec:spirals}). Many of the nearby $L\sim L_\ast$ are strikingly flat, unlike model spirals.
A century ago Wolf (1908) reviewed evidence of what proves to be the bimodal natures of galaxies. The far richer evidence we have now (Sec.~\ref{sec:earlyandlate}) presents us with an interesting opportunity: identify the separatrix that determines whether a protogalaxy evolves into a spiral or an elliptical (Sec.~\ref{separatrix}). How are we to understand the cool motions of stars in the wonderfully thin galaxies seen nearby, so very different from the motions of stars in elliptical galaxies? The problem has been known for a century, aspects of the situation are the subjects of many papers, but identification of the separatrix remains an open challenge.
Less familiar anomalies tend to be less secure because less thought has been given to the theory and observation, and we can add that it is natural to give less thought to what is contrary to accepted thinking. We have an example in Section~\ref{sec:distributions}, on the large-scale distributions of astronomical objects. In the standard cosmology clusters of galaxies formed where the primeval upward mass density fluctuations were unusually large. Within available accuracy this agrees with numerical simulations of cosmic structure formation. But why would the primeval mass distribution in the $\Lambda$CDM universe be so arranged that upward mass fluctuations capable of evolving into the clusters that at distances less than about 85~Mpc are present only near the extended plane of the Local Supercluster (Fig.~\ref{fig:LSCf} in Sec.~\ref{sec:LSC})? Why are the most massive elliptical galaxies within 85~Mpc close to this plane, while comparably massive spirals are not noticeably correlated with the plane (Fig.~\ref{fig:morphologies})? The powerful radio sources present in some galaxies are thought to be associated with massive central black holes. Many large galaxies contain these black holes. So why are the powerful radio sources within 85~Mpc present in a select few of the large galaxies, those near this plane? Is there something special about these black holes? Tully and Shaver knew aspects of this situation thirty years ago (Sec.~\ref{sec:LSC}). To judge by the sparse citations to Shaver's key point these phenomena have not captured the general attention of the community. But the phenomena surely are real, and interesting, and not likely to have grown out of the Gaussian, adiabatic, and near scale-invariant initial conditions of the present standard cosmology.
On larger scales, the applications of scaling relations to convert cluster luminosities to distances for estimates of peculiar velocities yield indications of unreasonable bulk flows relative to the rest frame defined by the CMB (Sec.~\ref{sec:bulkflows}). This disagrees with observations of the effect of the motion of intracluster plasma on the CMB. This kinetic SZ effect indicates a reasonably small cluster bulk flow. Maybe the situation is confused by subtle systematic errors, though the peculiar cluster velocity measurements have been carefully checked. If the anomalous peculiar velocities are confirmed it means the luminosities of first-ranked cluster members, and the properties of the intracluster plasma, are not well constrained by other cluster properties that are not sensitive to distance, maybe even that cluster properties depend on parameters that are not in the standard cosmology. The thought is speculative, but recall that there are cosmic magnetic fields, and maybe cosmic strings, which might connect clusters and perhaps set hidden parameters.
On scales comparable to the Hubble length the heliocentric dipole anisotropies of quasars and radio galaxies are about in the direction expected from the kinematic effect of our motion relative to the CMB, but the dipole amplitude is unacceptably large. If this were because of a real unexpectedly large dipole anisotropy in the mass distribution on the scale of the Hubble length then the standard cosmology would predict an unacceptably large local peculiar velocity. But recall that at distances $\lower.5ex\hbox{$\; \buildrel < \over \sim \;$} 85$~Mpc the curious distributions of radio galaxies, massive elliptical and spiral galaxies, and clusters of galaxies encourage the thought that the positions of radio galaxies, and so likely quasars, are only weakly related to the mass distribution traced by $L\sim L_\ast$ galaxies. So we have a mystery: what would have caused the anomalous large-scale distributions of massive ellipticals, quasars, radio galaxies and clusters of galaxies? The evidence is that all these objects contain massive black holes. Establishment of the theory of how these black holes formed might help.
The standard $\Lambda$CDM theory passes demanding well-checked tests that reenforce the expectation that some if not all of the curious phenomena discussed here will prove to be results of systematic errors and/or statistical fluctuations. But consider one example, the distributions of the clusters of galaxies and the most luminous galaxies at distances less than about 85~Mpc. The $\Lambda$CDM theory offers a good account of the number density of mass concentrations similar to rich clusters. This adds to the argument for this theory and against the proposed anomalies. But does the distribution of these mass concentrations that grow into the clusters shown in Figure~\ref{fig:LSCf} in Section~\ref{sec:LSC} look reasonable? It certainly looks real. Luminous radio galaxies and the most massive elliptical galaxies also tend to be close to this plane. They are related, but identified and cataloged in different ways. The consistent case for alignment of these object with the extended plane of the Local Supercluster is convincing. The odd thing is that the most massive spirals, and the galaxies that are most luminous at $60\mu$, are not noticeably concentrated to the plane.This curious situation is difficult to reject and maybe suggestive of something interesting to be discovered.
Let us not blame the messengers for the problems with the properties of galaxies reviewed in Section~\ref{sec:galaxies} and the distributions of galaxies discussed in Sections~\ref{sec:distributions} and \ref{sec:localvoid}; research groups are doing the best they can with the theory they have. It has not escaped community attention that the extreme simplicity of the dark sector of the standard $\Lambda$CDM cosmology seems unlikely to be better than a crude approximation to reality, maybe crude enough to be an impediment to progress in understanding cosmic structure.
\section{Acknowledgements}
I am grateful to colleagues for guidance to issues arising. They include Jean Brodie and Elaina Tolstoy for advice about dwarf galaxies; Antonaldo Diaferio and Garry Angus or advice on the generalization of MOND; Simon Driver and Samir Salim for education about galaxy morphologies; Mike Hudson, Tod Lauer, and Kostas Migkas for explanations of cluster bulk flow measurements; Manoj Kaplinghat for discussions of supermassive black holes; Roya Mohayaee and Subir Sarkar for discussions of the kinematic dipole; Dylan Nelson for discussions of simulations of galaxy formation; Patrick Ogle for instruction on supermassive spiral galaxies; and Xavier Prochaska for comments on observations of extragalactic magnetic fields. I am particuarly indebted to Michael Strauss for guidance to the phenomena and Neil Turok for guidance to the theory. Turok encouraged me to write this considerable revision of the draft in Peebles (2021).
|
\section{Introduction }
In quantum information, continuous-time quantum walk is an essential ingredient in transport problems of quantum spin networks. The study of perfect state transfer in quantum spin chains dates back to \cite{Bose2003}. Further research has been established in \cite{Albanese2004,Christandl2004,Christandl2005}. In \cite{ALASTAIR2010}, Kay provided a comprehensive survey on this topic. With entanglement generation in quantum spin chains, the transport problem is said to be fractional revival in \cite{Genest2016}. Investigating fractional revival is both of theoretical and practical significance as it generalizes two quantum transport phenomena, namely, perfect state transfer and periodicity.
Let $\Gamma=(V,E)$ be a connected simple graph with set $V$ of $n$ vertices and set $E$ of edges. Let $A=A(\Gamma)=(a_{u,v})_{u,v\in V}$ be the adjacency matrix of $\Gamma$ defined by
\begin{equation*}
a_{u,v}=\left\{\begin{array}{cc}
1, & \mbox{ if $(u,v)\in E$}, \\
0, & \mbox{ otherwise}.
\end{array}
\right.
\end{equation*}
The matrix $A$ is an $n\times n$ symmetric $\{0,1\}$-matrix and all eigenvalues of $A$ are real numbers. The transfer matrix of $\Gamma$ is defined by the following $n\times n$ unitary matrix
\begin{equation*}
H(t)=\exp(\imath tA)=\sum_{s=0}^{+\infty}\frac{(\imath tA)^s}{s!}=(H_{u,v}(t))_{u,v\in V},\ t\in \mathbb{R}, \imath =\sqrt{-1},
\end{equation*}
where $\mathbb{R}$ is the set of real numbers.
\begin{defn}For two
distinct vertices $x$ and $y$ of $\Gamma$, if
\begin{equation}\label{f-1}
H(t)\mathbf{e}_x=\alpha \mathbf{e}_x+\beta \mathbf{e}_y
\end{equation}
for some complex numbers $\alpha$ and $\beta$ satisfying $\beta\neq 0$ and $|\alpha|^2+|\beta|^2=1$, then the graph $\Gamma$ admits
$(\alpha,\beta)$-fractional revival (or fractional revival, in short, FR) from $x$ to $y$ at time $t$, where $\mathbf{e}_x$ is the standard vector whose $x$-th entry is $1$ and $0$ otherwhere. Factoring a common unimodular
phase factor, $(\alpha,\beta)$-fractional revival is turned into $e^{\imath \theta}(\alpha,\beta)$-fractional revival, where $\alpha$ and $\theta$ are real
scalars and $\beta$ is complex. If $\alpha=0$, then $(0,\beta)$-fractional revival is reduced to perfect state transfer (in short, PST). The graph $\Gamma$ is said to be periodic at some time $t$ if $\beta=0$ in \eqref{f-1}. \end{defn}
The basic problem on this topic is to find graphs having FR. More specifically, for a given connected simple graph $\Gamma=(V,E)$, and $u,v\in V$, define
\begin{eqnarray}\label{f-2}
\nonumber T(u,v)&=&\{t\in \mathbb{R}: t>0, H(t)\mathbf{e}_u=\alpha \mathbf{e}_u+\beta \mathbf{e}_v\}.
\end{eqnarray}
We want to determine $T(u,v)$ for any pair $\{u,v\}$ of vertices, or at least, to determine if $\Gamma$ has FR, that is, $T(u,v)$ is not empty for some pair $\{u,v\}$. PST on various graphs, which is a special case of FR, has been investigated by many researchers. The works in \cite{Basic2009,Cao2020,Cao2020a,Tan_2019,Luo2021,Cheung2011,Coutinho2016,Coutinho2015,Coutinho2015a,Godsil2012,Godsil2012a,Li2021,Zheng2018,Pal2016} and the references therein form a partial list of an already vast literature. With regard to PST, only a few attempts have been made to FR. Genest, Vinet and Zhedanov in \cite{Genest2016} systematically studied quantum spin chains with FR at two sites. Later in \cite{Bernard2018}, a graph exemplified the existence of FR between antipodes. Chan {\it et al.} in \cite{Chan2019} established a general approach to constructing graphs exhibiting FR between cospectral vertices. In addition, they characterized cycles and paths having FR. In a consecutive paper \cite{Chan_2020}, the same authors of \cite{Chan2019} discussed FR on graphs with adjacency matrices from the Bose-Mesner algebra of association schemes. A comprehensive study of FR on graphs between cospectral and strongly cospectral vertices was presented in \cite{Chan2020}. Unlike the work in \cite{Chan2020}, Godsil and Zhang in \cite{Godsil2021} proposed
a class of graphs admitting FR between non-cospectral vertices. Given a fixed element $a$ of an abelian group $G$, Wang, Wang and Liu in \cite{Wang2022} gave a sufficient and necessary condition for Cayley graphs on $G$ having FR between vertices $x$ and $x+a$. The characterization in \cite{Wang2022} is incomplete as the element $a$ is fixed. Up to now, there is no sufficient and necessary condition for Cayley graphs on abelian groups admitting FR between any two distinct vertices.
This paper aims at characterizing FR on Cayley graphs over abelian groups between any two distinct vertices. Our contribution begins by Theorem \ref{lem-2} in the next section with a sufficient and necessary condition for abelian Cayley graphs having FR between any two distinct vertices. Based on such a criteria, we propose two general constructions of abelian Cayley graphs admitting FR in Corollaries \ref{thm-1.61} and \ref{thm-2.81}. As a consequence, two families of abelian Cayley graphs admitting FR are provided in respective Propositions \ref{thm-1.6} and \ref{thm-2.8}. Furthermore, Propositions \ref{thm-1.6} and \ref{thm-2.8} tell us that the pool of feasible graphs having FR is larger than that of graphs admitting PST. In Section \ref{S3}, we propose the definition of plateaued functions over abelian groups. With the help of these functions, a family of abelian Cayley graphs having FR is built in Theorem \ref{thm-3.6}. In Section \ref{S4}, we discuss FR on cublike graphs. A flexible construction of cublike graphs admitting FR is given in Theorem \ref{thm-4.1}. Utilizing bent or semi-bent functions over finite fields of even characteristics, we present the last construction of cublike graphs having FR in Theorem \ref{bentconstruction}. Section \ref{S5} concludes this paper.
\section{A characterization of abelian Cayley graphs having FR}
Let $G$ be an abelian group of order $n$ and let $S$ be a subset of $G$ with $S=-S$ and $G=\langle S\rangle$. Let $\mathbb{C}$, $\mathbb{Z}$ and $\mathbb{Q}$ be, respectively, the set of complex numbers, the set of integers and the set of rational numbers. Let $\Gamma={\rm Cay}(G,S)$ be an abelian Cayley graph whose vertex set is $G$ and two vertices, $u,v$, form an edge of $\Gamma$ if $u-v\in S$. The adjacency matrix of $\Gamma$ is an $n\times n$ matrix $A=(a_{g,h})$ defined by
\begin{equation*}
a_{g,h}=\left\{\begin{array}{cc}
1, & \mbox{ if $g-h\in S$}, \\
0, & \mbox{ otherwise}.
\end{array}
\right.
\end{equation*}
An automorphism of $\Gamma$ is a bijection from $G$ to $G$ preserving the adjacency. The set of automorphism of $\Gamma$ is denoted by ${\rm Aut}(\Gamma)$. For every $z\in G$, define $\sigma_z: G\rightarrow G; x\mapsto x+z$. It is easy to see that $\sigma_z\in {\rm Aut}(\Gamma)$ and $G$ is a subgroup of ${\rm Aut}(\Gamma)$.
Each finite abelian group $G$ can be decomposed as a direct sum of cyclic groups:
\begin{equation*}
G=\mathbb{Z}_{n_1}\oplus \cdots\oplus \mathbb{Z}_{n_r},\ \ n_s\geq 2,
\end{equation*}
where $\mathbb{Z}_m=(\mathbb{Z}/m\mathbb{Z},+)$ is a cyclic group of order $m$. For every $x=(x_1,\cdots,x_r)\in G$ with $x_s\in \mathbb{Z}_{n_s}$, the mapping
\begin{equation*}
\chi_x:G\rightarrow \mathbb{C},\ \chi_x(g)=\prod_{s=1}^r\omega_{n_s}^{x_sg_s}, \ g=(g_1,\cdots,g_r)\in G
\end{equation*}
is a character of $G$, where $\omega_{n_s}=\exp(2\pi i/n_s)$ is a primitive $n_s$-th root of unity in $\mathbb{C}$. The set of characters of $G$, denoted by $\hat{G}$, forms a group under the following operation:
\begin{equation*}
\chi_g\chi_h: G\rightarrow \mathbb{C}; x\mapsto \chi_g(x)\chi_h(x), \forall g,h\in G.
\end{equation*}
Let $\hat{G}$ be the character group of $G$. Recall that for any abelian group $G$, $G$ is isomorphic to its dual group via $G\rightarrow \hat{G}, g\mapsto \chi_g$. It is known that the eigenvalues of $A$ are the real numbers
\begin{equation*}
\lambda_g=\sum_{s\in S}\chi_g(s), \forall g \in {G}.
\end{equation*}
Let $P=\frac{1}{\sqrt{n}}(\chi_g(h))_{g,h\in G}$ and $E_z=p_z^*p_z$ with $p_z$ being the $z$-th column of $P$ and $p_z^*$ being the conjugate transpose of $p_z$. The spectral decomposition of $A$ is given by
\begin{equation}\label{f-3}
A=\sum_{g\in G}\lambda_gE_g.
\end{equation}
Since $\{E_z: z\in G\}$ forms a set of idempotents, namely,
\begin{equation*}
E_gE_h=\left\{\begin{array}{cc}
E_g, & \mbox{ if $g=h$},\\
0, & \mbox{otherwise},
\end{array}
\right.
\end{equation*}
we have the spectral decomposition of $H(t)$ as follows:
\begin{equation}\label{f-4}
H(t)=\sum_{g\in G}e^{\imath \lambda_g t}E_g.
\end{equation}
Thus the $(g,h)$-th entry of $H(t)$ is
\begin{equation}\label{f-5}
H(t)_{g,h}=\frac{1}{n}\sum_{z\in G}e^{\imath t\lambda_z}\chi_z(g-h).
\end{equation}
If FR happens on $\Gamma={\rm Cay}(G,S)$ from a vertex $x$ to a vertex $y$, then by (\ref{f-1}) and noticing that $H(t)$ is a unitary matrix, the $x$-th row of $H(t)$ is determined by
\begin{equation}\label{f-6}
H(t)_{x,u}=\left\{\begin{array}{cc}
\alpha, & \mbox{ if $u=x$}, \\
\beta, & \mbox{ if $u=y$}, \\
0, & \mbox{otherwise}.
\end{array}
\right.
\end{equation}
Since $A$ is an $G$-invariant matrix, namely, $a_{g+z,h+z}=a_{g,h}, \forall g,h,z\in G$, we know that $H(t)$ is also $G$-invariant.
Therefore, we have the following preliminary result.
\begin{lem}\label{lem-1}Let $G$ be an abelian group of order $n$ and let $S$ be a subset of $G$. Let $\Gamma={\rm Cay}(G,S)$ be a Cayley graph with the adjacency matrix $A$. Then $\Gamma$ has $(\alpha,\beta)$-FR from a vertex $x$ to a vertex $y$ at some time $t$ if and only if
\begin{equation}\label{f-9}
H(t)=\alpha I_n+\beta Q,
\end{equation}
where $Q$ is a permutation matrix satisfying $Q^T=Q$ and $Q^2=I_n$, the identity matrix of order $n$. Moreover, $Q$ has no fixed points and $\alpha \overline{\beta}+\overline{\alpha}\beta=0$.
\end{lem}
\begin{proof}Suppose that $ H(t)=\alpha I_n+\beta Q$ for some $t$. By the fact that $Q$ is a permutation matrix with no fixed points, we can assume that there is some $a\neq 0$ such that
\begin{equation*}
Q_{0,a}=1, Q_{0,b}=0, \forall b\neq 0.
\end{equation*}
Then we have
\begin{equation*}
H(t)_{0,0}=\alpha, H(t)_{0,a}=\beta, \mbox{ and } H(t)_{0,b}=0, \forall b\neq 0, a.
\end{equation*}
Therefore the graph $\Gamma$ exhibits $(\alpha,\beta)$-FR from $0$ to $a$.
Conversely, if $\Gamma$ has $(\alpha,\beta)$-FR from $x$ to $y$, then
\begin{equation*}
H(t)\mathbf{e}_x=\alpha \mathbf{e}_x+\beta \mathbf{e}_y,
\end{equation*}
which implies that
\begin{equation*}
H(t)_{x,x}=\alpha, H(t)_{x,y}=\beta \mbox{ and } H(t)_{x,b}=0, \forall b\neq x, y.
\end{equation*}
Since $H(t)$ is $G$-invariant, we have that for every $z\in G$,
\begin{equation*}
H(t)_{x+z,x+z}=\alpha, H(t)_{x+z,y+z}=\beta \mbox{ and } H(t)_{x+z,b+z}=0, \forall b\neq x, y.
\end{equation*}
When $z$ runs through $G$, we get that
$H(t)=\alpha I_n+\beta Q$, where $Q$ is a permutation matrix with no fixed points. The matrix $Q$ is symmetric as $H(t)$ is a symmetric matrix. Since $H(t)$ is a unitary matrix, we obtain that
\begin{equation*}
H(t)\overline{H(t)}=I_n\Leftrightarrow \alpha \overline{\alpha}I_n+\beta \overline{\beta}Q^2+(\alpha \overline{\beta}+\overline{\alpha}\beta)Q=I_n.
\end{equation*}
Therefore, we have
\begin{equation*}
\left\{\begin{array}{l}
Q^2=I_n, \\
\alpha \overline{\beta}+\overline{\alpha}\beta=0.
\end{array}
\right.
\end{equation*}
This completes the proof.
\end{proof}
\textbf{From now on, we only consider the case of $(\alpha,\beta)$-FR with $\alpha\neq 0$. That is, we always assume that $\alpha \beta\neq 0$.}
Recall that $H(t)$ is $G$-invariant, if FR happens on $\Gamma={\rm Cay}(G,S)$ from a vertex $x$ to a vertex $y$, then we have
$H(t)_{x,u}=H(t)_{0,u-x}$, thus we can write
\begin{equation}\label{f-7}
H(t)_{0,u}=\frac{1}{n}\sum_{z\in G}e^{\imath t\lambda_z}{\chi_z(u)}=\left\{\begin{array}{cc}
\alpha, & \mbox{ if $u=0$}, \\
\beta, & \mbox{ if $u=y-x$}, \\
0, & \mbox{otherwise}.
\end{array}
\right.
\end{equation}
Denote $a=x-y$. According to \eqref{f-7} and its Fourier inversion, we get
\begin{equation}\label{f-81}
e^{\imath \lambda_g t}=\alpha+\overline{\chi_a(g)}\beta, \forall g\in G.
\end{equation}
Since $\lambda_g$ is real for every $g\in G$, we derive that
\begin{equation*}
1=e^{\imath \lambda_g t}\overline{e^{\imath \lambda_g t}}=(\alpha+\overline{\chi_a(g)}\beta)(\overline{\alpha}+{\chi_a(g)}\overline{\beta})=1+\overline{\chi_a(g)}\overline{\alpha}\beta+{\chi_a(g)}\alpha\overline{\beta},
\end{equation*}
which is equivalent to
\begin{equation}\label{f-10}
\chi_a(g){\alpha}\overline{\beta}+\overline{\chi_a(g)}\overline{\alpha}{\beta}=0.
\end{equation}
Combing with $\overline{\alpha}\beta+\alpha\overline{\beta}=0$ in Lemma \ref{lem-1}, we have
\begin{equation}\label{f-11}
\chi_a(g)=\overline{\chi_a(g)}, \mbox{ i.e., } \chi_a(g)=\pm 1, \forall g\in G.
\end{equation}
That is, $a$ is of order two in $G$. As a consequence, FR cannot happen on any abelian Cayley graph with the underlying group has an odd order.
Define two subsets of $G$ as follows,
\begin{equation}\label{f-12}
G_0=\{g\in G: \chi_a(g)=1\}, G_1=\{g\in G: \chi_a(g)=-1\}.
\end{equation}
Then $G_0$ is a subgroup of $G$ of order $n/2$ and $G_1$ is a coset of $G_0$.
By (\ref{f-81}), we have
\begin{equation}\label{f-13}
e^{\imath t \lambda_g}=\left\{\begin{array}{cc}
\alpha+\beta, & \mbox{ if $g\in G_0$},\\
\alpha-\beta, & \mbox{ if $g\in G_1$}.
\end{array}
\right.
\end{equation}
Assume the polar decompositions of $\alpha$ and $\beta$ are respective $r_\alpha e^{\imath \theta_\alpha}$ and $r_\beta e^{\imath \theta_\beta}$. According to $|\alpha|^2+|\beta|^2=1$ and $\overline{\alpha}\beta+\alpha\overline{\beta}=0$, we have
\begin{equation}\label{f-14}
\left\{\begin{array}{ll}
r_\alpha^2+r_\beta^2 &=1 \\
\theta_\alpha-\theta_\beta & =\frac{\pi}{2}+k\pi, \mbox{ for some $k\in \mathbb{Z}$}. \end{array}
\right.
\end{equation}
Thus
\begin{equation*}
\alpha\pm\beta=e^{\imath \theta_\beta}((-1)^kr_\alpha \imath\pm r_\beta),
\end{equation*}
Write $r_\beta+(-1)^kr_\alpha \imath =e^{\imath \theta_0}$. Then $\theta_0\in \mathbb{R}$ and
\begin{equation}\label{f-15}
\alpha+\beta=e^{\imath(\theta_0+\theta_\beta)}, \alpha-\beta=e^{\imath({\pi}-\theta_0+\theta_\beta)}.
\end{equation}
Write $t=2\pi T$ and $\theta_0=2\pi u_0, \theta_\beta=2\pi u_\beta$. Due to (\ref{f-13}), it holds that
\begin{equation}\label{f-16}
T\lambda_g-u_0-u_\beta\in \mathbb{Z}, \forall g\in G_0, \mbox{ and } T\lambda_g+u_0-u_\beta\in \frac{1}{2}+\mathbb{Z}, \forall g\in G_1.
\end{equation}
Therefore,
\begin{equation*}
T\left(\sum_{g\in G_0}\lambda_g+\sum_{g\in G_1}\lambda_g\right)-nu_\beta\in \frac{n}{4}+\mathbb{Z}.
\end{equation*}
Since $\sum_{g\in G_0}\lambda_g+\sum_{g\in G_1}\lambda_g=\sum_{g\in G}\lambda_g=0$, we have $u_\beta\in \mathbb{Q}$.
Now, we arrive at
\begin{equation*}
\sum_{g\in G_0}\lambda_g=\sum_{g\in G_0}\sum_{s\in S}\chi_g(s)=\sum_{s\in S}\sum_{g\in G_0}\chi_g(s)=\sum_{s\in S}\sum_{g\in G}\frac{1+\chi_a(g)}{2}\chi_g(s)=\left\{\begin{array}{cc}
n/2, & \mbox{ if $a\in S$,} \\
0, & \mbox{ otherwise}.
\end{array}
\right.
\end{equation*}
Thus, $\sum_{g\in G_0}\lambda_g\in \mathbb{Z}$ and so $\sum_{g\in G_1}\lambda_g\in \mathbb{Z}$ as well.
Let $d=|S|$. Thanks to (\ref{f-16}), we deduce that
\begin{equation*}
T\left(n/2\cdot d+\sum_{g\in G_1}\lambda_g\right)-n/2 \cdot u_\beta\in n/4+\mathbb{Z},
\end{equation*}
Thus we obtain that $T\in \mathbb{Q}$.
By (\ref{f-16}) again, we have $Td-u_0-u_\beta\in \mathbb{Z}$ which further implies that $u_0\in \mathbb{Q}$, and finally, we get that $\lambda_g\in \mathbb{Q}$ for every $g\in G$. Since $\lambda_g$ is an algebraic integer, we know that $\lambda_g\in \mathbb{Z}$. In other words, $\Gamma$ is an integral graph.
Meanwhile, since $T$ is rational, $e^{\imath t}$ is a root of unity.
Therefore, we have the following result.
\begin{thm}\label{lem-2}Let $G$ be an abelian group of an even order $n$ and let $S$ be a symmetric subset of $G$. Let $\Gamma={\rm Cay}(G,S)$ be a Cayley graph with the adjacency matrix $A$. Suppose that $\alpha$ and $\beta$ are two nonzero complex numbers. Then $\Gamma$ has $(\alpha,\beta)$-FR from a vertex $x$ to a vertex $y$ at some time $t$ if and only if the following conditions hold.
(1) $a=x-y$ is of order two;
(2) $\Gamma$ is an integral graph;
(3) for every $g\in G_0$, $e^{\imath t \lambda_g}=\alpha+\beta$; for every $g\in G_1$, $e^{\imath t \lambda_g}=\alpha-\beta$,
where $G_0$ and $G_1$ are defined by (\ref{f-12}). \end{thm}
\begin{proof}The necessity has been proved by the foregoing discussion. Next, we prove the sufficiency. Since for every $g\in G_0$, $e^{\imath t \lambda_g}=\alpha+\beta$, and for every $g\in G_1$, $e^{\imath t \lambda_g}=\alpha-\beta$, by (\ref{f-5}), we get that, for every $x\in G$,
\begin{equation*}
H(t)_{x,x}=\frac{1}{n}\sum_{z\in G}e^{\imath t \lambda_z}=\frac{1}{n}\left[\sum_{z\in G_0}(\alpha+\beta)+\sum_{z\in G_1}(\alpha-\beta)\right]=\alpha.
\end{equation*}
Meanwhile,
\begin{eqnarray*}
H(t)_{x,x+a}&=&\frac{1}{n}\sum_{z\in G}e^{\imath t \lambda_z}\chi_z(a)\\
&=&\frac{1}{n}\left[\sum_{z\in G_0}(\alpha+\beta)\chi_z(a)+\sum_{z\in G_1}(\alpha-\beta)\chi_z(a)\right]\\
&=&\frac{\alpha}{n}\sum_{z\in G}\chi_z(a)+\frac{\beta}{n}\left[\sum_{z\in G_0}\chi_z(a)-\sum_{z\in G_1}\chi_z(a)\right].
\end{eqnarray*}
Due to $a$ is of order two, and $\sum_{z\in G}\chi_z(a)=0$, we get
\begin{equation*}
H(t)_{x,x+a}=\frac{2\beta}{n}\sum_{z\in G_0}\chi_z(a)=\frac{2\beta}{n}\sum_{z\in G_0}1=\beta.
\end{equation*}
Moreover, for every $b\neq 0,a$, we have
\begin{eqnarray*}
H(t)_{x,x+b}&=&\frac{1}{n}\sum_{z\in G}e^{\imath t \lambda_z}\chi_z(b)\\
&=&\frac{1}{n}\left[\sum_{z\in G_0}(\alpha+\beta)\chi_z(b)+\sum_{z\in G_1}(\alpha-\beta)\chi_z(b)\right]\\
&=&\frac{\alpha}{n}\sum_{z\in G}\chi_z(b)+\frac{\beta}{n}\left[\sum_{z\in G_0}\chi_z(b)-\sum_{z\in G_1}\chi_z(b)\right]\\
&=&\frac{2\beta}{n}\sum_{z\in G_0}\chi_z(b)\\
&=&\frac{2\beta}{n}\sum_{z\in G}\frac{1+\chi_z(a)}{2}\chi_z(b)\\
&=&\frac{\beta}{n}\sum_{z\in G}\chi_z(a+b)\\
&=&0.
\end{eqnarray*}
Therefore, we have $|\alpha|^2+|\beta|^2=1$, and $H(t)=\alpha I_n+\beta Q$, where $Q=\sum_{g\in G}e_ge_{g+a}^T$ is a symmetric permutation matrix. By Lemma \ref{lem-1}, $\Gamma$ has $(\alpha,\beta)$-FR at time $t$.
\end{proof}
\begin{remark}
According to Definition \ref{f-1}, the existence of $(\alpha,\beta)$-FR is equivalent to the existence of $e^{\imath \theta}(\alpha,\beta)$-FR with $\theta,\alpha$ being real numbers and $\beta$ being a complex number. Based on Theorem \ref{lem-2}, we claim that $\theta$ is a rational number and $\beta$ is a pure complex number if $\Gamma={\rm Cay}(G,S)$ has $e^{\imath \theta}(\alpha,\beta)$-FR. Due to Lemma \ref{lem-1}, it is easy to check that $\beta$ is a pure complex number. By \eqref{f-13}, we obtain that
\begin{equation*}
e^{\imath t \lambda_g-\imath \theta}=\left\{\begin{array}{cc}
\alpha+\beta, & \mbox{ if $g\in G_0$},\\
\alpha-\beta, & \mbox{ if $g\in G_1$}.
\end{array}
\right.
\end{equation*}
Since $\alpha$ is a real number and $\beta$ is a pure complex number, we deduce that $e^{\imath t (|S|+\lambda_g)-2\imath \theta}=1$ with $g\in G_1$. Thanks to Items (2) and (3) in Theorem \ref{lem-2}, $\theta$ is a rational number.
\end{remark}
In what follows, we continue to investigate the conditions in Theorem \ref{lem-2}.
Taking an fixed element $g_1$ in $G_1$, by (\ref{f-13}), we have
\begin{equation}\label{f-18}
T(d-\lambda_g)\in \mathbb{Z} \mbox{ for every $g\in G_0$ and }T(\lambda_g-\lambda_{g_1})\in \mathbb{Z} \mbox{ for every $g\in G_1$}.
\end{equation}
Since $\Gamma$ is an integral graph, we define the following integers:
\begin{equation}\label{f-19}
M_0=\gcd(d-\lambda_g: g\in G_0), M_1=\gcd(\lambda_{g_1}-\lambda_g: g\in G_1), M=\gcd(M_0,M_1).
\end{equation}
(\ref{f-18}) is equivalent to
\begin{equation*}
TM_0, TM_1\in \mathbb{Z},
\end{equation*}
and thus
\begin{equation}\label{f-20}
T\in \frac{1}{M}\mathbb{Z}.
\end{equation}
Next, we show that $M$ is a divisor of $n=|G|$.
\begin{lem}\label{lem-M} Let $G$ be an abelian group of order $n$ and let ${\rm Cay}(G,S)$ be an integral graph. Then $M$ is a divisor of $n$, where $M$ is defined by (\ref{f-19}).
\end{lem}
\begin{proof}Write
\begin{equation*}
d=\lambda_g+M t_g \mbox{ for every $g\in G_0$ and $\lambda_g=\lambda_{g_1}+Ms_g$ for every $g\in G_1$, where $t_g,s_g\in \mathbb{Z}$}.
\end{equation*}
Take an element $z\in S$ with $z\neq 0,a$. Then
\begin{equation}\label{f-21}
\sum_{g\in G}\lambda_g\overline{\chi_g(z)}=\sum_{s\in S}\sum_{g\in G}\chi_g(s-z)=n.
\end{equation}
Meanwhile,
\begin{eqnarray}\label{f-22}
\nonumber \sum_{g\in G}\lambda_g\overline{\chi_g(z)}&=&\sum_{g\in G_0}(d-t_gM)\overline{\chi_g(z)}+\sum_{g\in G_0}(\lambda_{g_1}-s_gM)\overline{\chi_g(z)}\\
&=&(d+\lambda_{g_1})\sum_{g\in G_0}\overline{\chi_g(z)}-M\sum_{g\in G_1}(t_g+s_g)\overline{\chi_g(z)}.
\end{eqnarray}
Moreover,
\begin{equation}\label{f-23}
\sum_{g\in G_0}\overline{\chi_g(z)}=\sum_{g\in G}\frac{1+\chi_g(a)}{2}\overline{\chi_g(z)}=\sum_{g\in G}\chi_g(a-z)=0.
\end{equation}
(\ref{f-21})-(\ref{f-23}) imply that
\begin{equation}\label{f-24}
n/M=-\sum_{g\in G_1}(t_g+s_g)\overline{\chi_g(z)}.
\end{equation}
The LHS of (\ref{f-24}) is in $\mathbb{Q}$, while the RHS is an algebraic integer. Thus $n/M\in \mathbb{Z}$.
\end{proof}
By Lemma \ref{lem-M}, we have the following result.
\begin{cor}
Let $G$ be an abelian group of order $n$ and let ${\rm Cay}(G,S)$ be an integral graph. If ${\rm Cay}(G,S)$ has FR at some time $t$, then $e^{\imath t}$ is an $n$-th root of unity.
\end{cor}
\begin{proof}
It is a direct consequence of (\ref{f-20}) and Lemma \ref{lem-M}.
\end{proof}
According to (\ref{f-20}), we show the non-existence of FR on integral abelian Cayley graphs if $M=1$.
\begin{prop}Let $\Gamma={\rm Cay}(G,S)$ be an integral abelian Cayley graph. Let $M$ be defined by in (\ref{f-19}). If $M=1$, then $\Gamma$ cannot provide FR between any two vertices.\end{prop}
\begin{proof}If $M=1$, then by (\ref{f-20}), we know that $T\in \mathbb{Z}$ which means that $t=2\pi$. The desired result follows from the fact $H(t)=I_n$.\end{proof}
Thanks to Theorem \ref{lem-2}, if one wants to construct graphs admitting FR between two distinct vertices, the feasible graphs are integral and the difference of two vertices has even order. Using this fact, we propose a general construction of abelian Cayley graphs having FR, as the following corollary reveals. The proof of the following corollary is based on a direct computation. Of course, one can use Theorem \ref{lem-2} to get another proof of this result as well.
\begin{cor}\label{thm-1.61} Let $G=\mathbb{Z}_{2}^s\times H$ such that $s\geq 1$ and $H$ is an abelian group of order $\ell$. Let $\Gamma={\rm Cay}(G,S)$ be an integral graph and let $M$ be defined in \eqref{f-19}. If $\frac{|S|}{M}\notin \frac{1}{4}\mathbb{Z}$, then $\Gamma={\rm Cay}(G,S)$ has $\left(\cos\left(\frac{2\pi|S|}{M}\right),\imath\sin\left(\frac{2\pi|S|}{M}\right)\right)$-FR between $(x,y)$ and $(x+a,y)$ for every $(x,y)\in G$ and $0\neq a\in \mathbb{Z}_{2}^s$ at time $t=\frac{2\pi}{M}$.
\end{cor}
\begin{proof}
Denote vertices $(x,y)$ and $(x+a,y)$ by respective $\mu$ and $\nu$. It is well known that $\mathbb{Z}_{2}^s$ is isomorphic to the additive group of ${\mathbb F}_{2^s}$, the finite field with $2^s$ elements. Given $g\in{\mathbb F}_{2^s}$, an additive character of ${\mathbb F}_{2^s}$ is given by $\chi_{g}(h)=(-1)^{\operatorname{Tr}^s_1(gh)}$ with $h\in{\mathbb F}_{2^s}$ and $\operatorname{Tr}^s_1(\cdot)$ being the trace function mapping from ${\mathbb F}_{2^s}$ to ${\mathbb F}_2$. We refer readers to \cite[Chapter 5]{Lidl1997} for more details on additive characters. Below, we view $G$ as $G=\mathbb{F}_{2^s}\times H$.
By \eqref{f-5}, we obtain that
$$
H(t)_{\mu,\mu}=\frac{1}{2^s\ell}\sum_{z\in G} e^{\imath t\lambda_z}\ \ \hbox{and}\ \ H(t)_{\mu,\nu}=\frac{1}{2^s\ell}\sum_{z=(z_1,z_2)\in G} e^{\imath t\lambda_z}(-1)^{\operatorname{Tr}^s_1(az_1)}.
$$
Thanks to $|\{t\in {\mathbb F}_{2^s}: \chi_c(t)=0\}|=|\{t\in {\mathbb F}_{2^s}: \chi_c(t)=1\}|=2^{s-1}$, we deduce that
\begin{eqnarray*}
\left|H(2\pi/M)_{\mu,\mu}\right|^2+\left|H(2\pi/M)_{\mu,\nu}\right|^2&=&\frac{1}{2^{2s}\ell^2}\sum_{z\in G}\sum_{g\in G} e^{\imath 2\pi \frac{\lambda_z-\lambda_g}{M}}\\
&&+\frac{1}{2^{2s}\ell^2}\sum_{z=(z_1,z_2)\in G}\sum_{g=(g_1,g_2)\in G}e^{\imath 2\pi \frac{\lambda_z-\lambda_g}{M}}(-1)^{\operatorname{Tr}^s_1(az_1+ag_1)}\\
&=&\frac{1}{2^{2s}\ell^2}\sum_{z=(z_1,z_2)\in G}\sum_{g=(g_1,g_2)\in G}\left(1+(-1)^{\operatorname{Tr}^s_1(az_1+ag_1)}\right)\\
&=&1.
\end{eqnarray*}
By \eqref{f-81} and \eqref{f-13}, there exist a rational number $v$, a real number $\alpha$ and a pure complex number $\beta$ such that $e^{\imath \frac{2\pi}{M} |S|}=e^{\imath 2\pi v}(\alpha+\beta)$. It can be easily seen that $\alpha$ and $\beta$ are nonzero if $\frac{|S|}{M}+v\notin \frac{1}{4}\mathbb{Z}$. Putting $v=0$, it follows from $\frac{|S|}{M}\notin \frac{1}{4}\mathbb{Z}$ that $\Gamma={\rm Cay}(G,S)$ has $\left(\cos\left(\frac{2\pi|S|}{M}\right),\imath\sin\left(\frac{2\pi|S|}{M}\right)\right)$-FR at time $t=\frac{2\pi}{M}$.
\end{proof}
\begin{remark}
(1)By the proof of the above corollary, if there exist a rational number $v$, a real number $\alpha$ and a pure complex number $\beta$ such that $e^{\imath \frac{2\pi}{M} |S|}=e^{\imath 2\pi v}(\alpha+\beta)$ and $\frac{|S|}{M}+v\notin \frac{1}{4}\mathbb{Z}$, then the graph $\Gamma={\rm Cay}(G,S)$ defined in Corollary \ref{thm-1.61} has $e^{\imath 2\pi v}(\alpha,\beta)$-FR between $(x,y)$ and $(x+a,y)$ for every $(x,y)\in G$ and $0\neq a\in \mathbb{Z}_{2}^s$ at time $t=\frac{2\pi}{M}$. For simplicity, we state the case $v=0$ in Corollary \ref{thm-1.61}.
(2)Since we only discuss the case $\alpha\beta\neq 0$, the condition $\frac{|S|}{M}\notin \frac{1}{4}\mathbb{Z}$ (resp. $\frac{|S|}{M}+v\notin \frac{1}{4}\mathbb{Z}$) is needed in Corollart \ref{thm-1.61}. Removing the condition $\frac{|S|}{M}\notin \frac{1}{4}\mathbb{Z}$ (resp. $\frac{|S|}{M}+v\notin \frac{1}{4}\mathbb{Z}$), the graph $\Gamma={\rm Cay}(G,S)$ in Corollary \ref{thm-1.61} is said to have PST, FR or be periodic.
\end{remark}
Using Corollary \ref{thm-1.61}, we construct the following family of abelian Cayley graphs having FR.
\begin{prop}\label{thm-1.6}Let $p$ be a prime and let $r\geq 1$ be an integer. Let $G=\mathbb{Z}_{2}\times \mathbb{Z}_{p^r}\times H$ with $H$ being an abelian group of order $m$. Let $S$ be a symmetric subset in $G$ which is defined by
\begin{equation*}
S=\{(0,\ell,h): 1\leq \ell \leq p^r-1, \gcd(\ell,p)=1,h\in H\}\cup \{(1,0,0)\}.
\end{equation*}
If $p^{r-1}m\neq 1,2,4$, then $\Gamma={\rm Cay}(G,S)$ has FR between $(x,y,z)$ and $(x+1,y,z)$ for every $(x,y,z)\in G$ at the time $t=\frac{2\pi}{p^{r-1}m}$.
\end{prop}
\begin{proof}It is easy to see that the dual group of $G$ is
\begin{equation*}
\widehat{G}=\{\varphi_u\chi_v\psi_w: u\in \mathbb{Z}_2, v\in \mathbb{Z}_{p^r},w\in H\}
\end{equation*}
where $\varphi_u\chi_v\psi_w(x,y,z)=(-1)^{ux}\omega_{p^r}^{vy}\psi_w(z)$, and $\omega_v=e^{\imath \frac{2\pi}{v}}$ is a primitive $v$-th root of unity. Therefore, the eigenvalues of $\Gamma$ are
\begin{eqnarray*}
\lambda_{x,y,z}&=&\sum_{(u,v,w)\in S}\varphi_u\chi_v\psi_w(x,y,z)\\
&=&\sum_{1\leq \ell \leq p^r-1, \gcd(\ell,p)=1}\omega_{p^r}^{\ell y}\sum_{w\in H}\psi_w(z)+(-1)^x, \ \ (x,y,z)\in G.
\end{eqnarray*}
The summation
\begin{equation*}
\sum_{1\leq \ell \leq p^r-1, \gcd(\ell,p)=1}\omega_{p^r}^{\ell y}:=c(y,p^r)
\end{equation*}
is in fact a Ramanujan function \cite[Corollary 2.4]{McCarthy1986} and
\begin{equation*}
c(y,p^r)=\left\{\begin{array}{ll}
-p^{r-1}, & \mbox{ if $v_p(y)={r-1}$}, \\
p^{r-1}(p-1) & \mbox{ if $y=0$},\\
0, & \mbox{ otherwise}.
\end{array}
\right.
\end{equation*}
Here we identify the elements in $\mathbb{Z}_{p^r}$ as integers and $v_p(y)$ is the highest power of $p$ that divides $y$. Then we derive that
\begin{equation*}
\lambda_{x,y,z}=\left\{\begin{array}{ll}
(-1)^x+(-p^{r-1})m, & \mbox{ if $z=0$ and $v_p(y)={r-1}$}, \\
(-1)^x+p^{r-1}(p-1)m, & \mbox{ if $z=0$ and $y=0$}, \\
(-1)^x,& \mbox{ otherwise.}
\end{array}
\right.
\end{equation*}
Since $a=(x+1,y,z)-(x,y,z)=(1,0,0)$, it is obvious that
\begin{equation*}
G_0=\{(0,y,h):y\in \mathbb{Z}_{p^r},h\in H\}, \mbox{ and } G_1=\{(1,y,h):y\in \mathbb{Z}_{p^r},h\in H\}.
\end{equation*}
We have $M_0=M_1=M=p^{r-1}m$, which are defined by (\ref{f-19}). Due to $\frac{|S|}{M}=p-1+\frac{1}{p^{r-1}m}\notin \frac{1}{4}\mathbb{Z}$ and Corollary \ref{thm-1.61} or Theorem \ref{lem-2} directly, $\Gamma$ has $\left(\cos\left(\frac{2\pi}{p^{r-1}m}\right),\imath\sin\left(\frac{2\pi}{p^{r-1}m}\right)\right)$-FR between $(x,y,z)$ and $(x+1,y,z)$ for every $(x,y,z)\in G$ at the time $t=\frac{2\pi}{p^{r-1}m}$.
\end{proof}
The following example is used to illustrate Proposition \ref{thm-1.6}.
\begin{example}\label{exam-2} Let $G=\mathbb{Z}_2\times \mathbb{Z}_9$ and $S=\{(0,1),(0,2),(0,4),(0,5),(0,7),(0,8),(1,0)\}$. Then $\Gamma={\rm Cay}(G,S)$ has FR between $(0,0)$ and $(1,0)$ at time $t=2\pi/3$.\end{example}
\begin{proof}The eigenvalues of $\Gamma$ are as follows.
\begin{equation*}
\lambda_{x,y}=\omega_9^{y}+\omega_9^{2y}+\omega_9^{4y}+\omega_9^{5y}+\omega_9^{7y}+\omega_9^{8y}+(-1)^x, (x,y)\in G.
\end{equation*}
Thus,
\begin{equation*}
\lambda_{0,0}=7, \lambda_{0,3}=\lambda_{0,6}=-2, \lambda_{0,y}=1, y\in \mathbb{Z}_9^*:=\{1,2,4,5,7,8\}, \mbox{ and}
\end{equation*}
\begin{equation*}
\lambda_{1,0}=5, \lambda_{1,3}=\lambda_{1,6}=-4, \lambda_{1,y}=-1, y\in \mathbb{Z}_9^*.
\end{equation*}
It is obvious that
\begin{equation*}
G_0=\{(0,y): y\in \mathbb{Z}_9\}, G_0=\{(1,y): y\in \mathbb{Z}_9\}.
\end{equation*}
Thus, $M_0=M_1=M=3$. Moreover,
\begin{equation*}
e^{\imath \frac{2\pi}{3}\lambda_g}=\left\{\begin{array}{cc}
e^{\imath \frac{2\pi}{3}}, & g\in G_0, \\
e^{-\imath \frac{2\pi}{3}}, & g\in G_1.
\end{array}
\right.
\end{equation*}
Thus $\Gamma$ has $(\cos(2\pi/3), \imath \sin(2\pi/3))$-FR at time $t=2\pi/3$.
Indeed, by a Magma program, we find that
\begin{equation*}
H(2\pi/3)={\rm diag}\left(\underset{9 \ \ cpoies}{\underbrace{{\left(
\begin{array}{cc}
-\frac{1}{2} & \frac{\sqrt{3}}{2} \\
\frac{\sqrt{3}}{2} & -\frac{1}{2} \\
\end{array}
\right), \cdots, \left(
\begin{array}{cc}
-\frac{1}{2} & \frac{\sqrt{3}}{2} \\
\frac{\sqrt{3}}{2} & -\frac{1}{2} \\
\end{array}
\right)}
}}\right).
\end{equation*}
Thus $\Gamma$ has $(-\frac{1}{2},\frac{\sqrt{3}}{2})$-FR at time $t=2\pi/3$.
\end{proof}
In Proposition \ref{thm-1.6}, if $p^{r-1}m= 1,2\ \mbox{or}\ 4$, the Cayley graph $\Gamma={\rm Cay}(G,S)$ may have FR. We do not have a general result like Corollary \ref{thm-1.6} to guarantee the existence of FR on $\Gamma$ whenever $p^{r-1}m= 1,2\ \mbox{or}\ 4$. Using Theorem \ref{lem-2}, we can check the existence of FR on $\Gamma$ under these cases one by one. The following example shows how to verify it when $p=3$, $r=1$ and $m=1$.
\begin{example}Let $G=\mathbb{Z}_2\times \mathbb{Z}_3$ and $S=\{(0,1),(0,2),(1,0)\}$. Then the Cayley graph $\Gamma={\rm Cay}(G,S)$ has FR at time $t=2\pi/3$.\end{example}
\begin{proof}One can check that the spectra of $\Gamma$ are
\begin{equation*}
\lambda_{0,0}=3, \lambda_{0,1}=\lambda_{0,2}=0, \lambda_{1,0}=1, \lambda_{1,1}=\lambda_{1,2}=-2.
\end{equation*}
Meanwhile,
\begin{equation*}
G_0=\{(0,0),(0,1),(0,2)\}, G_1=\{(1,0),(1,1),(1,2)\}.
\end{equation*}
It is easy to check that $M=3$ and
\begin{eqnarray*}
e^{\imath \frac{2\pi}{3}\lambda_g}&=&\left\{\begin{array}{cc}
1, & g\in G_0, \\
\cos(2\pi/3)+\imath \sin(2\pi/3), & g\in G_1.
\end{array}
\right.\\
&=& \left\{\begin{array}{cc}
e^{\imath \frac{\pi}{3}}(\frac{1}{2}-\imath\frac{\sqrt{3}}{2}), & g\in G_0, \\
e^{\imath \frac{\pi}{3}}(\frac{1}{2}+\imath\frac{\sqrt{3}}{2}), & g\in G_1.
\end{array} \right.
\end{eqnarray*}
By Theorem \ref{lem-2}, $\Gamma$ has $e^{\imath \frac{\pi}{3}}\left(\frac{1}{2},-\imath\frac{\sqrt{3}}{2}\right)$ at time $t=2\pi/3$.
\end{proof}
In Corollary \ref{thm-1.61}, we propose a general construction of Cayley graphs over $G=\mathbb{Z}_{2}^s\times H$ having FR. Next, we discuss the existence of FR on Cayley graphs over $G=\mathbb{Z}_{2^s}\times H$.
\begin{cor}\label{thm-2.81} Let $G=\mathbb{Z}_{2^s}\times H$ such that $s\geq 1$ and $H$ is an abelian group of order $\ell$. Let $\Gamma={\rm Cay}(G,S)$ be an integral graph and let $M$ be defined in \eqref{f-19}. If $\frac{|S|}{M}\notin \frac{1}{4}\mathbb{Z}$, then $\Gamma={\rm Cay}(G,S)$ has $\left(\cos\left(\frac{2\pi|S|}{M}\right),\imath\sin\left(\frac{2\pi|S|}{M}\right)\right)$-FR between $(x,y)$ and $(x+2^{s-1},y)$ for every $(x,y)\in G$ at time $t=\frac{2\pi}{M}$.
\end{cor}
\begin{proof}
The proof of this corollary can be completed by the method analogous to that used in Corollary \ref{thm-1.61}
\end{proof}
At the end of this section, we propose another family of Cayley graphs admitting FR.
\begin{prop}\label{thm-2.8}Let $r_0,r_1,\cdots,r_s$ be positive integers and let $p_0=2, p_1,\cdots,p_s$ be distinct primes such that $\prod_{j=0}^sp_j^{r_j-1}\neq 1,2,4$ with $s\geq 1$. Let $G=\prod_{j=0}^s\mathbb{Z}_{p_j^{r_j}}$ and let $S$ be a subset of $G$ defined by
\begin{equation*}
S=\{(\ell_0,\ell_1,\cdots, \ell_s): 1\leq \ell_i\leq p_i^{r_i}-1, \gcd(\ell_i,p_i)=1, 0\leq i\leq s\}\cup \{(2^{r_0-1},0,\cdots,0)\}.
\end{equation*}
Let $\Gamma={\rm Cay}(G,S)$ be a Cayley graph. Then $\Gamma$ has FR between $(x_0,x_1,\cdots,x_s)$ and $(x_0+2^{r_0-1},x_1,\cdots,x_s)$ for every $(x_0,x_1,\cdots,x_s)\in G$ at time $2\pi/\prod_{i=0}^sp_i^{r_i-1}$.\end{prop}
\begin{proof}
Due to $a=(x_0+2^{r_0-1},x_1,\cdots,x_s)-(x_0,x_1,\cdots,x_s)=(2^{r_0-1},0,\cdots,0)$, it is easy to see that
\begin{equation*}
G_0=\{(2x,y_1,\cdots,y_s): x \in \mathbb{Z}_{2^{r_0}}, y_i \in \mathbb{Z}_{p_i^{r_i}}\}, G_1=\{(2x+1,y_1,\cdots,y_s): x \in \mathbb{Z}_{2^{r_0}}, y_i \in \mathbb{Z}_{p_i^{r_i}}\}.
\end{equation*}
For $\mathbf{x}=(x_0,x_1,\cdots,x_s)$, the corresponding eigenvalue of $\Gamma$ is
\begin{equation*}
\lambda_{\mathbf{x}}=(-1)^{x_0}+\sum_{\underset{1\leq \ell_j \leq p_j^{r_j}-1, \gcd(p_j,\ell_j)=1}{\ell_0,\ell_1,\cdots, \ell_s}}\prod_{j=0}^s\xi_{{p_j}^{r_j}}^{\ell_j x_j}.
\end{equation*}
Thus,
\begin{equation*}
\lambda_{\mathbf{x}}=\left\{\begin{array}{ll}
1+\prod_{i=0}^sp_i^{r_1-1}(p_1-1), & \mbox{ if $(x_0,x_1,\cdots,x_s)=\mathbf{0}$}, \\
1+A,& \mbox{ if $x_j\neq 0,v_{p_j}(x_j)=r_j-1,$ for some $ j\in I\subseteq \{0,\cdots,s\}$}, \\
(-1)^{x_0},& \mbox{ if there is some $x_i\neq 0, v_{p_i}(x_i)<{r_i-1}$}.
\end{array}
\right.
\end{equation*}
where $A=(-1)^{|I|}\prod_{j\in I}p_j^{r_j-1}\prod_{i\not\in I}p_i^{r_i-1}(p_i-1)$.
It follows that $M=\prod_{i=0}^sp_i^{r_i-1}$, where $M$ is defined by (\ref{f-19}). Due to $|S|/M\notin \frac{1}{4}\mathbb{Z}$ and Corollary \ref{thm-2.81}, $\Gamma$ has $\left(\cos (2\pi/M), \imath \sin(2\pi/M)\right)$-FR at time $2\pi/M$.
\end{proof}
\begin{remark}(1) It was shown in \cite{Tan_2019} that the order of the underlying group should be doubly-even if an abelian Cayley graph has perfect state transfer. However, for FR, Propositions \ref{thm-1.6} and \ref{thm-2.8} show that there are many more graphs having FR than that for PST on abelian Cayley graphs.
(2) Proposition \ref{thm-2.8} says that for every even number $n$ having a square divisor of an odd prime power, there is an abelian Cayley graph having FR with the underlying group of order $n$. \end{remark}
\section{Plateaued functions over abelian groups and FR on abelian Cayley graphs}\label{S3}
Let $G$ be an abelian group of order $n$ and let $\mathbb{C}^G$ be the set of complex-valued functions on $G$. Then $\mathbb{C}^G$ is an $n$-dimensional vector space. Given $f,g\in \mathbb{C}^G$, define a scalar product as
\begin{equation*}
\langle f, g\rangle=\sum_{x\in G}f(x)\overline{g(x)}.
\end{equation*}
Then $\mathbb{C}^G$ is a Hilbert space with $\{\frac{1}{\sqrt{n}}\chi_x: x\in G\}$ an orthnormal basis. Thus every function $f$ has the following decomposition.
\begin{equation*}
f(x)=\sum_{z\in G}\langle f,\frac{1}{\sqrt{n}}\chi_z\rangle \frac{1}{\sqrt{n}}\chi_z(x).
\end{equation*}
For any $f\in \mathbb{C}^G$, its Fourier transform is defined by
\begin{equation*}
\widehat{f}(\chi_z)=\sum_{x\in G}f(x)\overline{\chi_z(x)}=\langle f, \chi_z\rangle.
\end{equation*}
\begin{defn}\cite{LOGACHEV1997} Let $G$ be an abelian group of order $n$. A function $f\in \mathbb{C}^G$ is called bent on $G$ if for every $z\in G$, we have
\begin{equation*}
|\widehat{f}(\chi_z)|=\sqrt{n}.
\end{equation*}
\end{defn}
Bent functions over finite fields of even characteristic were invented and named in 1976 by Rothaus \cite{Rothaus1976}. As a generalization of bent functions, the so-called plateaued functions over finite fields of characteristic two were first introduced by Zheng and Zhang in \cite{Zheng2001}.
\begin{defn}An $n$-variable
Boolean function is said to be $r$-plateaued if the values of its Walsh transform (the Fourier transform of $(-1)^f$)
belong to the set $\{0,\pm 2^{\frac{n+r}{2}}\}$
for some fixed $r$.\end{defn}
Bent and plateaued functions have many applications in communications and cryptography, for a comprehensive survey on such functions, interested readers may refer to \cite{Mesnager2016}.
Let $G$ be an abelian group of order $n$ and let $e=\exp(G)$ be the largest order of the elements in $G$. Then the values of $\chi(z) (z\in G, \chi\in \hat{G})$ are in the cyclotomic field $\mathbb{Q}(\omega_e)$. The Galois group of the extension $\mathbb{Q}(\omega_e)/\mathbb{Q}$ is isomorphic to
\begin{equation*}
\mathbb{Z}_e^*=\{\ell: 1\leq \ell\leq e-1, \gcd(\ell, e)=1\}.
\end{equation*}
For every $\ell \in \mathbb{Z}_e^*$, the corresponding isomorphic in ${\rm Gal}(\mathbb{Q}(\omega_e)/\mathbb{Q})$ is determined by
\begin{equation*}
\sigma_\ell: \omega_e\mapsto \omega_e^\ell.
\end{equation*}
\begin{defn}A function $f\in \mathbb{C}^G$ is called a {\it class function} over $G$ if $f$ is integer-valued, namely, $f(x)\in \mathbb{Z}$ for every $x\in G$, and for every $\ell \in \mathbb{Z}_e^*$, it holds that
\begin{equation}\label{f-161}
f(\ell x)=f(x), \forall x\in G.
\end{equation}\end{defn}
\begin{lem} \label{lem-3.3}Let $G$ be an abelian group of order $n$ and let $f\in \mathbb{C}^G$ be a class function over $G$. Then for every $\chi\in \hat{G}$, we have
\begin{equation*}
\hat{f}(\chi)\in \mathbb{Z}.
\end{equation*}
\end{lem}
\begin{proof}Since $f$ is integer-valued,
\begin{equation*}
\hat{f}(\chi_g)=\sum_{x\in G}f(x)\chi_g(x)
\end{equation*}
is contained in the integer ring $\mathbb{Z}[\omega_e]$, and for every $\ell \in \mathbb{Z}_e^*$,
\begin{eqnarray*}
\sigma_\ell(\hat{f}(\chi_g))&=&\sum_{x\in G}f(x)\sigma_\ell(\chi_g(x))=\sum_{x\in G}f(x)(\chi_g(\ell x))\\
&=&\sum_{x\in G}f(\ell x)(\chi_g(\ell x))\ \ (\mbox{ since $f$ is a class function})\\
&=&\sum_{x\in G}f(x)\chi_g(x)\\
&=& \hat{f}(\chi_g).
\end{eqnarray*}
Therefore, $\hat{f}(\chi_g)\in \mathbb{Q}\cap \mathbb{Z}[\omega_e]=\mathbb{Z}$.
\end{proof}
As an analogue to plateaued functions over finite fields of even characteristic, we define plateaued functions over abelian groups as follows.
\begin{defn}\label{defn-3.4}Let $G$ be an abelian group of order $n$. Suppose that $p$ is a prime number. Let $f\in \mathbb{C}^G$ be a class function such that for every $\chi\in \widehat{G}$, the Fouerier transform of $f$ at $\chi$ satisfying
\begin{equation}\label{f-162}
\widehat{f}(\chi)\equiv k\pmod {p^r}
\end{equation}
for some fixed non-negative integers $k,r$. Then $f$ is called $p^r$-plateaued over $G$, where $r$ is the greatest integer $r$ satisfying (\ref{f-162}).\end{defn}
It is easy to see that the classical plateaued functions over finite fields with even characteristic are special cases of our plateaued functions over abelian groups.
\begin{example}Let $G=\mathbb{Z}_3\times \mathbb{Z}_3$ and $S$ be a subset of $G$ defined by
\begin{equation*}
S=\{(1,0),(2,0),(0,1),(0,2)\}.
\end{equation*}
Let $f$ be the characteristic function of $S$, namely, $f(x)=1$ if $x\in S$ and $0$, otherwise. Then $f$ is a $3$-plateaued function over $G$.
\end{example}
\begin{proof}Indeed, since for every $\ell\in \mathbb{Z}_3^*=\{1,2\}$, we have $\ell S:=\{\ell x: x\in S\}=S$. Thus $f$ is a class function. Moreover,
for every $g=(g_1,g_2)\in G$, we have
\begin{equation*}
\hat{f}(\chi_g)=\sum_{x\in G}f(x)\chi_g(x)=\omega_3^{g_1}+\omega_3^{-g_1}+\omega_3^{g_2}+\omega_3^{-g_2}.
\end{equation*}
Hence, $\hat{f}(\chi_g)\in \{4,-2,1\}$ and then $\hat{f}(\chi_g)\equiv 1 \pmod 3$. So that $f$ is a $3$-plateaued function over $G$.
\end{proof}
Our main result in this section is as follows.
\begin{thm}\label{thm-3.6} Let $H$ be an abelian group of order $n$ and $S_1$ be a subset of $H$ satisfying $0\notin S_1$ and $\ell S_1=S_1$ for all $\ell \in \mathbb{Z}_e^*$. Let
$G=\mathbb{Z}_2\times H$ and let $S$ be a subset of $G$ defined by
\begin{equation*}
S=(1,S_1)\cup (0,S_1)\cup \{(1,0)\}.
\end{equation*}
Assume that $p$ divides $d=|S_1|$.
Let $f$ be the characteristic function of $S_1$ and suppose that $f$ is an $p^r$-plateaued function over $H$ with $r\geq 1$. Then the Cayley graph $\Gamma={\rm Cay}(G,S)$ has FR at time $t=\pi/p^{r_0}$ between $(x+1,y)$ and $(x,y)$ for every $(x,y)\in G$, where $r_0=\min(r, v_p(d))\geq 1$.
\end{thm}
\begin{proof}For every $(x,y)\in G$, the corresponding eigenvalue of $\Gamma$ is
\begin{equation*}
\lambda_{x,y}=(-1)^x+(-1)^x\sum_{z\in S_1}\chi_y(z)+\sum_{z\in S_1}\chi_y(z)=(-1)^x+((-1)^x+1)\hat{f}(\chi_y).
\end{equation*}
Due to $a=(x+1,y)-(x,y)=(1,0)$, it is easy to see that
\begin{equation*}
G_0=\{(0,h):h\in H\}, \mbox{ and } G_1=\{(1,h):h\in H\}.
\end{equation*}
Thus
\begin{equation*}
\lambda_{x,y}=\left\{\begin{array}{cc}
1+2\hat{f}(\chi_y), & \mbox{ if $(x,y)\in G_0$}, \\
-1, & \mbox{ if $(x,y)\in G_1$}.
\end{array}
\right.
\end{equation*}
Therefore, for every $(0,y)\in G_0$, we have
\begin{equation*}
e^{\imath t \lambda_{0,y}}=e^{\imath \frac{\pi}{p^{r_0}}(\lambda_{0,y}-\lambda_{0,0}+\lambda_{0,0})}=e^{\imath \frac{\pi}{p^{r_0}}(2(\hat{f}(\chi_y)-\hat{f}(\chi_0))+1+2d)}.
\end{equation*}
Since $\hat{f}(\chi_y)-\hat{f}(\chi_0) \equiv 0 \pmod {p^r}$, and $d\equiv 0 \pmod {p^{v_p(d)}}$, we get
\begin{equation*}
e^{\imath t \lambda_{0,y}}=e^{\imath \frac{\pi}{p^{r_0}}}.
\end{equation*}
Meanwhile, we have
\begin{equation*}
e^{\imath t \lambda_{1,y}}=e^{-\imath \frac{\pi}{p^{r_0}}}.
\end{equation*}
Thus we get the desired result by Theorem \ref{lem-2}.
\end{proof}
Next we present an example to illustrate Theorem \ref{thm-3.6}.
\begin{example}Let $H=\mathbb{Z}_9$ and let $S_1$ be defined by
\begin{equation*}
S_1=\{1,2,4,5,7,8\}.
\end{equation*}
Let $G=\mathbb{Z}_2\times \mathbb{Z}_9$ and let $S$ be defined by
\begin{equation*}
S=\{(0,1),(0,2),(0,4),(0,5),(0,7),(0,8),(1,1),(1,2),(1,4),(1,5),(1,7),(1,8),(1,0)\}.
\end{equation*}
Then $\Gamma={\rm Cay}(G,S)$ has FR at time $t=\pi/3$.
\end{example}
\begin{proof}Let $f$ be the characteristic function of the set $S_1$. Then for every $y\in H$, we have
\begin{equation*}
\hat{f}(\chi_y)=\sum_{x\in H}f(x)\chi_x(y)=\omega_9^{y}+\omega_9^{2y}+\omega_9^{4y}+\omega_9^{5y}+\omega_9^{7y}+\omega_9^{8y}=c(y,9),
\end{equation*}
where $c(y,9)$ is a Ramanujan function. Thus
\begin{equation*}
\widehat{f}(\chi_0)=6, \widehat{f}(\chi_1)=\widehat{f}(\chi_2)=\widehat{f}(\chi_4)=\widehat{f}(\chi_5)=\widehat{f}(\chi_7)=\widehat{f}(\chi_8)=0, \widehat{f}(\chi_3)=\widehat{f}(\chi_6)=-3.
\end{equation*}
Therefore, $f$ is a $3$-plateaued function. Moreover $3|d$. By Theorem \ref{thm-3.6}, $\Gamma$ has FR at time $\pi/3$.
We have checked the results by a Magma programme, it shows that
\begin{equation*}
H(\pi/3)={\rm diag}\left(\underset{9 \ \ copies}{\underbrace{{\left(
\begin{array}{cc}
-\frac{1}{2} & \frac{\sqrt{3}}{2} \\
\frac{\sqrt{3}}{2} & -\frac{1}{2} \\
\end{array}
\right), \cdots, \left(
\begin{array}{cc}
-\frac{1}{2} & \frac{\sqrt{3}}{2} \\
\frac{\sqrt{3}}{2} & -\frac{1}{2} \\
\end{array}
\right)}
}}\right).
\end{equation*}
This completes the proof.
\end{proof}
Note that the transfer matrix at time $t=\pi/3$ in the above example is exactly the same as that in Example \ref{exam-2}. Combining these two examples, we know that for a given abelian group of even order, there may have different choices of the connection sets such that the resulting graphs have FR at a different time.
\section{Cubelike graphs}\label{S4}
Let $\mathbb{F}_{2^n}$ be the finite field with $2^n$ elements. A Cayley graph ${\rm Cay}(G,S)$ is called a cubelike graph if $G$ is the additive group of some finite field $\mathbb{F}_{2^n}$. Since $(\mathbb{F}_{2^n},+)$ is an elementary $2$-group, the classical Hamming graph is a special kind of cubelike graphs.
Note that every nonzero element in $\mathbb{F}_{2^n}$ has order two. Moreover, the map $x\mapsto x+a$ is an isomorphism of the Cayley graph ${\rm Cay}(G,S)$ for every $a\in G$. Thus, in this section, we can fix $a$ as an arbitrary nonzero element in $G$. Thus, in the sequel, we prescribe $a$ as $(1,0,\cdots,0)\in \mathbb{F}_2^n$. Meanwhile, we can assume that $a$ is contained in the connection set.
In this section, we show that many cubelike graphs admit FR. Our first result in this section is as follows.
\begin{thm}\label{thm-4.1}Let $\Gamma={\rm Cay}(\mathbb{F}_{2^n}, S)$ be a cubelike graph. Suppose that
\begin{equation*}
S=(0,S_0)\cup (1,S_1)\cup \{(1,\mathbf{0})\},
\end{equation*}
where $S_0,S_1\subseteq \mathbb{F}_2^{n-1}$, $\mathbf{0}\not\in S_1$. Denote $|S_0|=d_0$ and $|S_1|=d_1$. If $\min(v_2(d_0+d_1),v_2(d_0-d_1))\geq 3$ and $M\geq 8$, then $\Gamma$ exhibits FR at some time $t$, where $M$ is defined by (\ref{f-19}). \end{thm}
\begin{proof}The additive group of $\mathbb{F}_{2^n}$ is an elementary $2$-group isomorphic to $\mathbb{F}_2^n$, the additive group of the $n$-dimensional vector space over $\mathbb{F}_2$. Take $\mathbf{a}=(1,\mathbf{0}_{n-1})$, where $\mathbf{0}_{n-1}\in \mathbb{F}_2^{n-1}$. We can assume that $\mathbf{a}\in S$, for otherwise, we use an isomorphism of $\Gamma$ to replace an element in $S$ with $\mathbf{a}$. Then
\begin{equation*}
G_0=\{(0, \mathbf{x}): \mathbf{x}\in \mathbb{F}_2^{n-1}\}, \mbox{ and } G_1=\{(1, \mathbf{x}): \mathbf{x}\in \mathbb{F}_2^{n-1}\},
\end{equation*}
where $G_0$ and $G_1$ are defined by (\ref{f-12}). Write $S=(0,S_0)\cup (1,S_1)\cup \{\mathbf{a}\}$ ($0\not\in S_1$) and denote $|S_i|=d_i, i=0,1$. By the assumption that $|S|\equiv 1 \pmod 8$, we can assume that $d_0+d_1+1=1+2^\nu d'$, where $\nu\geq 3$. For every $(0,\mathbf{x})\in G_0$, the corresponding eigenvalue is
\begin{equation*}
\lambda_{(0,\mathbf{x})}=\sum_{\mathbf{s}_0 \in S_0}(-1)^{\mathbf{s}_0\cdot \mathbf{x}}+\sum_{\mathbf{s}_1 \in S_1}(-1)^{\mathbf{s}_1\cdot \mathbf{x}}+1
\end{equation*}
Particularly, $\lambda_{0,\mathbf{0}}=d_0+d_1+1$.
Meanwhile, for every $(1,\mathbf{x})\in G_1$, the corresponding eigenvalue is
\begin{equation*}
\lambda_{(1,\mathbf{x})}=\sum_{\mathbf{s}_0 \in S_0}(-1)^{\mathbf{s}_0\cdot \mathbf{x}}-\sum_{\mathbf{s}_1 \in S_1}(-1)^{\mathbf{s}_1\cdot \mathbf{x}}+(-1)
\end{equation*}
Particularly, $\lambda_{1,\mathbf{0}}=d_0-d_1-1$.
Since $M(\geq 8)$ is a divisor of $ |\mathbb{F}_{2^n}|=2^n$ by Lemma \ref{lem-M}, we can write $M=2^r$, $r\geq 3$.
For a different $\mathbf{x'}\in \mathbb{F}_2^{n-1}$, by the definition of the integer $M$, it is obvious that
\begin{equation*}
\lambda_{(0,\mathbf{x})}- \lambda_{(0,\mathbf{x'})}\equiv 0 \pmod {2^r}, \lambda_{(1,\mathbf{x})}- \lambda_{(1,\mathbf{x'})}\equiv 0 \pmod {2^r}.
\end{equation*}
Write $\min(r, v_2(d_0+d_1), v_2(d_0-d_1))=\kappa$. Then $\kappa\geq 3$.
Take $t=\frac{2\pi}{2^{\kappa}}$. Then
\begin{equation*}
e^{\imath t \lambda_{x,\mathbf{y}}}=\left\{\begin{array}{ll}
e^{\imath \frac{2\pi}{2^\kappa}(d_0+d_1+1)}=e^{\imath \frac{2\pi}{2^\kappa}}, & \mbox{ if $(x,\mathbf{y})\in G_0$}, \\
e^{\imath \frac{2\pi}{2^\kappa}(d_0-d_1-1)}=e^{-\imath \frac{2\pi}{2^\kappa}}, & \mbox{ if $(x,\mathbf{y})\in G_1$}.
\end{array}
\right.
\end{equation*}
Then we obtain the desired result by Theorem \ref{lem-2}.
\end{proof}
Below we present a simple example as a replenishment to Theorem \ref{thm-4.1}.
\begin{example}\label{exam-4.2} Let $G=\mathbb{F}_2^5$ and let the connection set $S$ be defined by
\begin{equation*}
S=\{(10000)\}\cup \{(0,s): s\in S_0\}\cup \{(1,s): s\in S_0\}
\end{equation*}
where
\begin{equation*}
S_0=\{(1100),(0011),(1011),(0111), (1101),(1110)\}.
\end{equation*}
Then $\Gamma={\rm Cay}(G,S)$ has FR at time $t=\pi/4$.
\end{example}
\begin{proof}Take $\mathbf{a}=(10000)$. Then
\begin{equation*}
G_0=\{(0vxyz): v,x,y,z \in \{0,1\}\}, G_1=\{(1vxyz): v,x,y,z \in \{0,1\}\}.
\end{equation*}
The eigenvalues of $\Gamma$ are
\begin{eqnarray*}
\lambda_{(uxyz)}&=&(-1)^u+(-1)^{v+x}+(-1)^{y+z}+(-1)^{v+y+z}+(-1)^{x+y+z}+(-1)^{v+x+z}\\
&&+(-1)^{v+x+y}+(-1)^{u+v+x}+(-1)^{u+y+z}+(-1)^{u+v+y+z}\\
&&+(-1)^{u+x+y+z}+(-1)^{u+v+x+z}+(-1)^{u+v+x+y}.
\end{eqnarray*}
Thus
\begin{equation*}
\lambda_{(00000)}=13, \lambda_{(0vxyz)}\in \{5,-3\}, \lambda_{(1vxyz)}=-1, v,x,y,z\in \{0,1\}.
\end{equation*}
It is obvious that
\begin{equation*}
\lambda_{\mathbf{x}}\equiv \left\{\begin{array}{ll}
5 \pmod 8, & \mbox{ if $\mathbf{x}\in G_0$}, \\
-1 \pmod 8, & \mbox{ if $\mathbf{x}\in G_1$}.
\end{array}\right.
,
\end{equation*}
Take $t=\pi/4$. Then
\begin{equation*}
e^{\imath t \lambda_\mathbf{x}}=\left\{\begin{array}{ll}
e^{\imath \frac{\pi}{4}}, & \mbox{ if $\mathbf{x}\in G_0$}, \\
e^{-\imath \frac{\pi}{4}}, & \mbox{ if $\mathbf{x}\in G_1$}.
\end{array}
\right.
\end{equation*}
This completes the proof.
\end{proof}
Our next result is based on bent functions and plateaued functions over $\mathbb{F}_2^n$. Before stating our constructions, we recall some basic properties about such kind of functions.
Let $f: \mathbb{F}_2^n\rightarrow \mathbb{F}_2$ be a nonzero Boolean function. Since the additive group $(\mathbb{F}_2^n,+)$ is isomorphic to the additive group of the finite field $\mathbb{F}_{2^n}$. Then $f$ viewed as a function on $\mathbb{F}_{2^n}$ has a (unique) trace expansion of the form:
\begin{equation*}
f(x)=\sum_{j\in \Gamma_n}{\rm Tr}^{o(j)}_1(a_jx^j)+\epsilon (1+x^{2^n-1}), \forall x\in \mathbb{F}_{2^n},
\end{equation*}
where $\Gamma_n$ is the set of integers obtained by choosing one element in each cyclotomic coset of 2
modulo $2^n-1$, $o(j)$ is the size of the cyclotomic coset of 2 modulo $2^n-1$ containing $j$, $a_j\in \mathbb{F}_{2^{o(j)}}$
and $\epsilon = wt(f)$ modulo 2 where $wt(f)$ is the Hamming weight of the image vector of $f$, that is, the
number of $x$ such that $f(x) = 1$. Denote ${\rm supp}(f)=\{x\in \mathbb{F}_{2^n}: f(x)=1\}$. It is obvious that the map $f\mapsto {\rm supp}(f)$ gives a one-to-one mapping from the set of Boolean functions to the power set of $\mathbb{F}_{2^n}$.
The Walsh-Hadamard transform of $f$ is defined by
\begin{equation*}
\widehat{f}_w(a)=\sum_{x\in \mathbb{F}_{2^n}}(-1)^{f(x)+{\rm Tr}^n_1(ax)}, \forall a\in \mathbb{F}_{2^n}.
\end{equation*}
\begin{defn}Let $n=2k$ with $k$ being a positive integer. A Boolean function $f$ is called bent if $\widehat{f}_w(a)\in \{\pm 2^{k}\}$ for all $a\in \mathbb{F}_{2^n}$.\end{defn}
It is well-known that bent function exists on $\mathbb{F}_{2^{2k}}$ for every $k$. Moreover, for every positive integer $k$, there exist many families of such functions. Semi-bent functions also exist in $\mathbb{F}_{2^n}$ for all integer $m\geq 1$, e.g., in \cite{Mesnager2016}.
Let $f: \mathbb{F}_2^n\rightarrow \mathbb{F}_2$ be a Boolean function. $S={\rm supp}(f)\subseteq \mathbb{F}_{2^n}$. For the cubelike graph $\Gamma={\rm Cay}(\mathbb{F}_{2^n},S)$, its eigenvalues are $\lambda_0=|S|$ and
\begin{equation}\label{f-e14}
\lambda_x=\sum_{z\in S}\chi_x(z)=\sum_{y\in \mathbb{F}_{2^m}}\frac{1-(-1)^{f(y)}}{2}\chi_x(y)=-\frac{1}{2}\widehat{f}_w(x), 0\neq x\in \mathbb{F}_{2^n}.
\end{equation}
For bent and semi-bent functions, the following results are known in the literature, e.g., on pages 72 and 422 in \cite{Mesnager2016}.
\begin{lem}\label{lem-4.4} Let $n=2k$ be a positive even integer. Let $f$ be a Boolean function on $\mathbb{F}_{2^n}$ and let $S={\rm supp}(f)$.
(1) If $f$ be a bent function, then $|S|=2^{n-1}\pm 2^{k-1}$. Moreover, define a function $\widetilde{f}$, called the dual of $f$, by
\begin{equation*}
\widehat{f}_w(x)=2^k(-1)^{\widetilde{f}(x)}.
\end{equation*}
Then $\widetilde{f}$ is also a bent function. As a consequence, the numbers of occurrences of $\widehat{f}_w(x)$ taking the values $\pm 2^k$ are $2^{n-1}\pm 2^{k-1}$ or $2^{n-1}\mp 2^{k-1}$.
(2) If $f$ is a semi-bent function, then $|S|\in \{2^{n-1}, 2^{n-1}\pm 2^{k-1}\}$. Moreover, we have the following table.
\begin{center}Table 1. Walsh spectrum
of semi-bent functions
$ f$ with $f(0)=0$
\begin{tabular}{|c|c|}
\hline
Value of $\widehat{f}(x),x\in \mathbb{F}_{2^m}$ & Frequency \\
\hline
$0$ & $2^{n-1}+2^{n-2}$ \\
\hline
$2^{k+1}$ & $2^{n-3}+2^{k-2}$ \\
\hline
$-2^{k+1}$ & $2^{n-3}-2^{k-2}$ \\
\hline
\end{tabular}
\end{center} \end{lem}
It is known also that if $f$ is a bent function, then so is its complementary function, i.e, $g=1+f$.
The following two constructions are actually corollaries of Theorem \ref{thm-3.6}.
\begin{cor}\label{bentconstruction}Let $n=2k\geq 4$ for a positive integer $k$ and $f$ be a bent (resp. semi-bent) function on $\mathbb{F}_2^n$. Denote $S_0=supp(f)$ and define
\begin{equation*}
S=\{(1\mathbf{0}_{n})\}\cup \{(0,s):\mathbf{s}\in S_0\}\cup \{(1,s): \mathbf{s}\in S_0\}\subset \mathbb{F}_2^{n+1}.
\end{equation*}
Then the cubelike graph $\Gamma={\rm Cay}(\mathbb{F}_2^{n+1},S)$ has FR at the time $t=\frac{\pi}{2^k}$ (resp. $t=\frac{\pi}{2^{k+1}}$ ).
\end{cor}
\begin{proof}We just give the proof for the case that $f$ is a bent function. The case for the semi-bent functions can be proved similarly.
Obviously,
\begin{equation*}
G_0=\{(0\mathbf{y}): \mathbf{y}\in \mathbb{F}_2^n\}, G_1=\{(1\mathbf{y}): \mathbf{y}\in \mathbb{F}_2^n\}.
\end{equation*}
It is easily seen that the eigenvalues of $\Gamma$ are
\begin{equation*}
\lambda_{(x\mathbf{y})}=(-1)^x+((-1)^x+1)\sum_{\mathbf{s}\in S_0}(-1)^{\mathbf{s}\cdot \mathbf{y}}, x\in \mathbb{F}_2, \mathbf{y}\in \mathbb{F}_2^n.
\end{equation*}
Thus, for every $(1\mathbf{y})\in G_1$, $\lambda_{1\mathbf{y}}=-1$. Moreover, for $\mathbf{y}\neq \mathbf{0}$, $(0\mathbf{y})\in G_0$, the corresponding eigenvalue is
\begin{equation*}
\lambda_{(0\mathbf{y})}=1-\hat{f}_w(\mathbf{y})=1\pm 2^k.
\end{equation*}
When $\mathbf{y}=\mathbf{0}_n$, the corresponding trivial eigenvalue is $\lambda_{(0\mathbf{0}_n)}=1+(2^n\pm 2^k)$.
One can check that $\lambda_{(0\mathbf{y})} \equiv 1+2^k \pmod {2^{k+1}}$. Take $t=\frac{\pi}{2^k}$. We have
\begin{equation*}
e^{\imath t \lambda_\mathbf{x}}=\left\{\begin{array}{cc}
e^{\imath \frac{\pi}{2^k}}, & \mbox{ if $\mathbf{x}\in G_0$}, \\
e^{-\imath \frac{\pi}{2^k}}, & \mbox{ if $\mathbf{x}\in G_1$}.
\end{array}
\right.
\end{equation*}
This completes the proof.
\end{proof}
Note that when $k=2$ in the above Corollary \ref{bentconstruction}, we can choose a bent function $f(x_1x_2x_3x_4)=x_1x_2+x_3x_4$, then we have $supp(f)=\{(1100),(0011),(1011),(0111), (1101),(1110)\}$. In this case, we get the graph in Example \ref{exam-4.2}.
We can also use plateaued functions to construct cubelike graphs that admit FR. The main idea is that presented in Theorem \ref{thm-3.6}. The detail of which is omitted.
\section{Concluding remarks}\label{S5}
Using the spectral decomposition of adjacency matrices of abelian Cayley graphs, we characterized abelian Cayley graphs having FR. This characterization indicates potential abelian Cayley graphs admitting FR. We established a general construction of integral graphs ${\rm Cay}(\mathbb{Z}_{2^s}\times H,S)$ having FR. Meanwhile, we constructed several new families of abelian Cayley graphs having FR. The notation of plateaued functions over finite fields was generalized to finite abelian groups. Based on plateaued functions, a new family of abelian Cayley graphs having FR was built. This paper and reference \cite{Tan_2019} show that bent, semi-bent and plateaued functions are powerful ingredients in the constructions of abelian Cayley graphs having PST or FR. It is important to study properties of plateaued functions over finite abelian groups. Another particularly interesting direction is to construct Cayley graphs over non-abelian groups having FR.
\section*{Acknowledgement}
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
Traditional recommendations are typically developed using content-based filtering~\cite{lops2011content, son2017content} or collaborative filtering approaches~\cite{basilico2004unifying, ramlatchan2018survey}, which predict the user's future interest based on past preferences.
However, because users' preferences change over time, modelling based on previous interests may not provide an accurate prediction.
In order to grasp shifting user interests, dynamic recommendation was developed as a practical technique to improve recommendation systems by supporting interactive processes~\cite{chen2018stabilizing, chen2020knowledge, zou2020pseudo, chen2021locality}.
Interactive recommendation is a process in which the system takes an optimal action
at each step to maximise the user's feedback reward.
Since Reinforcement Learning (RL) can learn from users'
interactive
feedback, it has been regarded as having the ability to address the dynamic evolution process of user feedback loops and as an effective approach for modelling dynamic recommendation system~\cite{dulac2015deep, zheng2018drn,chen2021survey}.
One of the main obstacles for the RL-based recommender system is that it will struggle to precisely grasp users' preferences and generate suitable recommendations only with limited interaction data.
In RL-based recommender systems, the agent will take action based on the current state of the environment, interact with it, and receive the reward.
The system will mistakenly presume that the user is uninterested in the item and return zero reward if there is no record of this circumstance, which will prevent the recommender system from accurately reflecting the users' actual preferences.
Therefore, the sparse interaction between users and items may limit the system's capabilities and
harm
user satisfaction.
There has been recent interest in attempting to incorporate causality into recommender systems to address the data sparsity~\cite{xiong2021counterfactual, yang2021top}.
From the standpoint of causality, building counterfactuals
provide an answer to the counterfactual question: "What would the interaction process be if we intervened on some parts of the observational data?"
Counterfactuals derived from intervention on observational data can be viewed as supplementing the circumstances not covered by the observational data.
Thus, modelling from both observational and counterfactual data distributions serves as a powerful data augmentation technique to assist the recommendation engine in understanding users' true preferences.
Recent works have made several initial attempts to alleviate the sparsity problem through data augmentation.
~\citet{zhang2021causerec} propose to measure the similarity between the representation of each item and the target item and replace the top half of items with the lowest similarity scores to obtain the positive counterfactual user sequence.
~\citet{wang2021counterfactual} design a sampler model that makes minor changes to the user's historical items on the embedding space to implement a counterfactual data augment framework.
However, these approaches are all based on the embedding space.
Although the state representation in RL has similarities to the embedding, they are conceptually different. In contrast to the embedding, the state representation is dynamic and affected by the agent's actions. State representation in an RL-based recommender system consists of user features, feedback and demographic information, and other information, whereas the standard embedding method only contains users' recent actions. As a result, RL-based dynamic recommendations in response to users' shifting interests are incompatible with existing counterfactual generating frameworks.
To address the above issues, we propose to train a counterfactual synthesis policy
to generate counterfactuals.
We identify the concepts of essential and trivial components in the state representation based on the different degrees of influence on the user's interest.
Our goal is to enable agents to identify optimal actions that only find and change the trivial components in a state.
We treat the modified state as an intervention on the original state to evaluate the causal effect of states on rewards. Since the rewards in a recommender system reflect the users' interests, we can indirectly measure the causal effect of states on users' interests through the effect on rewards. Thus, we identify the modified components in a state as trivial components if the modification has a weak causal effect on the reward.
The weak causal effect can be indicated if the intervened reward's distribution is similar to the initial distribution, suggesting a low impact on users' interest.
As a result, the counterfactual synthesis policy is learned by minimising the difference between observational and intervention reward distributions to fulfil the goal of replacing a state's trivial components.
The proposed counterfactual synthesis policy is model-agnostic, allowing it to be implemented in any RL algorithm and cooperate with the recommender policy to synthesise counterfactual interactive data during interaction.
This approach effectively handles the data sparsity problem by modelling both counterfactual and observational interaction data. Our main contributions are summarised as follows:
\begin{itemize}
\item We propose to generate counterfactuals based on users' dynamic interests in RL-based dynamic recommendations and model both the observation and counterfactual distribution to address data scarcity.
\item We design a novel model-agnostic counterfactual synthesis policy and provide two learning methods. Our policy can be employed in any RL algorithm to cooperate with the recommender policy and generate counterfactuals during the interaction process.
\item We theoretically analyze the identification of causal effects in the recommender system and introduce an effective reward to guide our agents to detect and replace the trivial components in a state.
\item We conduct experiments on online simulation and offline datasets to demonstrate that the counterfactual synthesis policy can be applied in several RL algorithms and significantly improve the performance of the recommender policy.
\end{itemize}
\section{PRELIMINARIES}
\subsection{RL-based Dynamic Recommendation}
The interactive process between users and a recommendation system achieved by RL can be formally described as training an agent that interacts with an environment, which follows a Markov Decision Process (MDP)~\cite{sutton2018reinforcement}.
Specifically, the agent interacts with the environment in each step of a discrete-time sequence $t = 0, 1, 2, ..., n$. In each interaction at time $t_i$, the agent initially receives the state representation $S_t \in \mathcal{S}$ from the environment and chooses an action $A_t \in \mathcal{A}(S_t)$ based on the state $S_t$.
The interaction with the environment then yields a reward $R_{t+1} \in \mathcal{R}$ back to the agent, and the environment will enter into the next state $S_{t+1}$. Formally, the process described above can be formulated by a tuple $(\mathcal{S}, \mathcal{A}(S_t), \mathcal{R}, \mathcal{P}, \gamma)$, where:
\begin{itemize}
\item $\mathcal{S}$: the set of all states, including the initial and terminal state. State representations contain some information of the environment for the agent to make decisions.
\item $\mathcal{A}(S_t)$: the set of available actions in the state $S_t$.
\item $\mathcal{R}:\mathcal{S} \times \mathcal{A} \to \mathbb{R}$ is the set of possible rewards related to user feedback.
\item $\mathcal{P}$: $\mathcal{S} \times \mathcal{A} \times \mathcal{S}\to \mathbb{R}$ are the state-transition probabilities.
\item $\gamma$: discount factor that satisfies $0 \leq \gamma \leq 1$.
\end{itemize}
In RL,
a policy ($\pi: \mathcal{S} \to \mathcal{A}$) is a mapping from perceived states to actions that the agent would take when in those environment states. Each interaction between the agent and the environment will yield an immediate reward $R_t$. The learning agent aims to select the action to maximise discounted return, which is the sum of discounted rewards it receives over the episode:
\begin{equation}
\begin{aligned}
G_t := R_{t+1} + \gamma R_{t+2} + \gamma ^2 R_{t+3} + \gamma ^3 R_{t+4} + \cdots = \sum_{n=1}^{\infty} \gamma^n R_{t+n},
\end{aligned}
\end{equation}
where $T$ is the final time step of an episode.
\begin{figure*}[h]
\centering
\begin{subfigure}[b]{0.41\linewidth}
\centering
\includegraphics[width=\linewidth]{SCM_a.pdf}
\caption{}
\label{SCM_a}
\end{subfigure}
\begin{subfigure}[b]{0.16\linewidth}
\centering
\includegraphics[width=\linewidth]{SCM_b.pdf}
\caption{}
\label{SCM_b}
\end{subfigure}
\begin{subfigure}[b]{0.265\linewidth}
\centering
\includegraphics[width=\linewidth]{SCM_c.pdf}
\caption{}
\label{SCM_c}
\end{subfigure}
\caption{Causal diagram for MDP. The red edges indicate where we can intervene on. (a) An general MDP instance; (b) one-step MDP to express the possible state changes; (c) Intervening on the action by the $A_t$ to generate counterfactual state $S_c$.}
\Description{.}
\label{SCM}
\end{figure*}
\subsection{Structural Causal Models}
The causal model describes a system by random variables that can be divided into two separate sets: endogenous variables and exogenous variables.
The value of an endogenous variable is determined by the state of the other variables in the system. And in contrast, the value of exogenous variables is determined by factors outside the causal system, meaning that it is independent of other variables in the system~\cite{halpern2020causes}.
A standard approach to modelling the causal relationships of interacting variables is using structural equations.
For a complex system, Structural Causal Modeling (SCM) is a practical way to describe the causal relationships between variables with a set of structural equations.
In an SCM, causal relationships are generated by functions that compute variables from other variables. Formally, we refer to these functions as assignments and define SCM as follows~\cite{peters2017elements}:
\vspace{1mm}\noindent\textbf{Definition 1. (Structural Causal Models)}.
\textit{A Structural Causal Model (SCM) $\mathcal{M} := (S, P_N)$ is associated with a directed acyclic graph (DAG) \( \mathcal{G} \), which consists of a collection $S$ of $n$ structural assignments} :
\begin{equation}
\begin{aligned}
X_i := f_i(\boldsymbol{PA}_i, U_i), \:\: i = 1,...,n,
\end{aligned}
\end{equation}
where $V = \{X_1,...,X_n\}$ is a set of endogenous variables, and $\boldsymbol{PA}_i \subseteq \{X_1,...,X_n\} \setminus \{X_i\}$ represent parents of $X_i$, which are also called direct causes of $X_i$. And $U = \{U_1,...,U_n\}$ are noise variables, determined by unobserved factors. We assume that noise variables are jointly independent. Correspondingly, $P_N$ is the joint distribution over the noise variables. Each structural equation $f_i$ is a causal mechanism that determines the value of $X_i$ based on the values of $\boldsymbol{PA}_i$ and the noise term $U_i$.
Since the definition of SCM requires its underlying graph be acyclic, each SCM $\mathcal{M}$ models a unique distribution over the endogenous variables $V = \{X_1,...,X_n\}$. Intuitively, SCMs with different causal structures have different distributions. Through intervention on the causal structures, we can obtain intervention distributions that are usually different from the observational one.
We refer the definition of intervention as follows~\cite{peters2017elements}:
\vspace{1mm}\noindent\textbf{Definition 2. (Intervention)}.
\textit{Given an SCM $\mathcal{M} := (S, P_N)$ over $X$, an intervention $\mathcal{I}$ is defined as replacing one or several structural assignments of the SCM $\mathcal{M}$.
Assume we replace the assignment for $X_k$
by the following expression:}
\begin{equation}
\begin{aligned}
X_k := \widetilde f_k(\widetilde {\boldsymbol {PA}_k}, \widetilde N_k).
\end{aligned}
\end{equation}
Then we say that variable $X_k$ have been intervened on, resulting in a new SCM $\widetilde {\mathcal{M}}$. The corresponding distribution is changed from the observational distribution $P_X^{\mathcal{M}}$ to the intervention distribution $P_X^{\widetilde {\mathcal{M}}}$, expressed as:
\begin{equation}
\begin{aligned}
P_X^{\widetilde {\mathcal{M}}} := P_X^{\mathcal{M}; do(X_k := \widetilde f_k(\widetilde {\boldsymbol {PA}_k}, \widetilde N_k))},
\end{aligned}
\end{equation}
where the operator $do(X_k := \widetilde f_k(\widetilde {\boldsymbol {PA}_k}, \widetilde N_k))$ denotes the intervention that we use to replace the assignment $X_k$. Note that the new noise variables $\widetilde N_k$ are also required to be jointly independent with the original noise variables $N_k$ in the SCM $\widetilde {\mathcal{M}}$.
Another approach to modifying the SCM is to maintain the causal relationship between the variables, but change the distribution of the noisy variables in the causal structure.
In a recommendation system, for example, the counterfactual would deal with the question: "What does the user want to buy if he/she has previously clicked on a different item?". This counterfactual query is based on observed data (the same list of items) but considers a different interventions distribution (clicked a different item). Answering the counterfactual question can be regarded as the process of simulating the causal effect of modifying the actual course of history with minimal change of the causal mechanisms~\cite{pearl_2009}.
\vspace{1mm}\noindent\textbf{Definition 3. (Counterfactual)}.
\textit{Given a SCM $\mathcal{M} := (S, P_N)$ over $X$, a counterfactual is defined as replacing the distribution of noise variables by the noise distribution based on the condition $X=x$, where $x$ are some observations. This can be expressed as:}
\begin{equation}
\begin{aligned}
\mathcal{M}_{X=x} := (S, P_N^{\mathcal{M}|X=x}),
\end{aligned}
\end{equation}
where $ P_N^{\mathcal{M}|X=x}$ denotes the distribution of the noise variable $N$ when we observe the variable $X=x$ in the SCM $\mathcal{M}$.
\section{Methodology}
\subsection{Problem Formulation}\label{Formulation}
To learn a counterfactual generation policy, we start by formalizing MDP in the semantics of SCMs. The endogenous variables $V$ in SCMs contain the states $\mathcal{S}$, available actions $\mathcal{A}(S_t)$
and the rewards $\mathcal{R}$. The exogenous variables are the noise variables. The causal mechanisms in SCMs are made up of three components: policy $\pi$, state-transition probabilities $P\in\mathcal{P}$ and reward function $f_R$.
The causal relationship corresponding to the above variables can be illustrated in~\Cref{SCM_a}.
We assume no confounders in the environment.
The SCMs for MDP can be formulated by the following functions:
\begin{equation}
\begin{aligned}
S_{t+1} := P(S_t, A_t),\:
A_t := \pi_t(S_t),\:
R_{t+1} := f_R(S_t, A_t),
\end{aligned}
\end{equation}
\begin{figure*}[h]
\centering
\includegraphics[width=\linewidth]{train.pdf}
\caption{An illustration of the proposed counterfactual synthesis policy. The left part is the schematic of how to train the counterfactual synthesis policy. The right part is how to apply the counterfactual synthesis policy to other RL algorithms for data augmentation.}
\Description{.}
\label{train}
\end{figure*}
Our work decomposes the state $S_t \in \mathcal{S}$ into two disjoint components, \textit{essential components} $S_t^{ess}$ and \textit{trivial components} $S_t^{tri}$, to identify their different levels of influence on learning users interest representations. We denote the decomposition of state as $S_t = S_t^{ess} \oplus S_t^{tri}$.
As shown in~\Cref{SCM_b},
both the essential and trivial components may be changed when transitioning from state $S_t$ to $S_{t+1}$.
Considering users' dynamic preferences, we assume that not all historical interaction data has a substantial causal effect on inferring users' preferences. That is, only the most recent $m$ interactions are more relevant to users' interests.
Since the rewards in a recommender system reflect users' interests, we measure the causal effect on users' interests through rewards. Formally, we identify the essential components of the state $S_t$ containing essential information of the user's interest.
Changes to the essential components will result in significant changes to the rewards as well as to the items recommended.
In contrast, the trivial components are the part of the state that are less important for representing the user's interest. Thus
changing the trivial components have a weak causal effect on the reward in the SCMs. Given the identification of two components, the essence of our work is to find and replace the trivial components in a state.
\Cref{SCM_c} shows that we only change the trivial components in the state $S_t$ to generate counterfactual state $S_c$ by intervening on the action by $A_t$, donated as $do(A(S_t):=A_c)$.
\subsection{Counterfactual Synthesis Policy}
The reinforcement learning method handles problems by considering the interaction of a goal-directed agent with an uncertain environment. It specifies that the agent maximises the total reward it receives over time and adjusts its policy based on its experience. The policy also provides guidance for the agent to choose the optimal action in a given state to maximise overall reward.
Our goal is to allow the agent to change its policy such that the policy can guide the agent to discover the action to only replace the trivial components in a state.
Formally, the agent chooses an action $A_c$ as the intervention on the action $A_t$, which can be formulated with $do$ calculus as $do(A_t:=A_c)$. According to the Peal's rules of $do$ calculus~\cite{pearl_2009}, the probability distribution for state $S_{t+1}$ induced after intervention can be calculated by:
\begin{equation}
\label{backdoor}
\begin{aligned}
P^{\mathcal{M}; do(A_t := A_c)}(S_{t+1})
&= \sum_{S_t}P(S_{t+1}|do(A_t), S_t) P(S_t|do(A_t)) \\
&= \sum_{S_t}P(S_{t+1}|A_c, S_t) P(S_t) \\
&= E_{S_t}P(S_{t+1}|A_c, S_t),
\end{aligned}
\end{equation}
which is a special case of the back-door adjustment formula. Hence, we assume that the state variable $S_t$ in the SCM defined in~\Cref{Formulation} satisfies the back-door criterion such that it is sufficient for identifying $P^{\mathcal{M}; do(A_t := A_c)}(S_{t+1})$.
\vspace{1mm}\noindent\textbf{Assumption 1. (Back-Door)}.
\textit{For the SCM $\mathcal{M}$ defined in~\Cref{Formulation} with the corresponding DAG $\mathcal{G}$ shown in~\Cref{SCM_a},
the variable $S_t$ satisfies the back-door criterion with regard to the pair of variables $(A_t, S_{t+1})$ because it meets the following criteria:
\begin{itemize}
\item There is no descendant of $A_t$ in $S_t$.
\item All paths containing an arrow into $A_t$ between $A_t$ and $S_{t+1}$ are blocked by $S_t$.
\end{itemize}
}
With Assumption 1, the causal effect of $A_t$ on $S_{t+1}$ is identifiable and is given by the~\Cref{backdoor}. We consider the state $S_{t+1}$ after intervention as the counterfactual state $S_c$ of the state $S_t$, where only the trivial components have been affected, that is:
\begin{equation}
\begin{aligned}
S_c = S_t^{ess} \oplus S_c^{tri} \: \text{and} \: S_c \sim E_{S_t}P(S_{t+1}|A_c, S_t)
\end{aligned}
\end{equation}
Since the agent always learns to maximize its reward, we propose to set up the rewards to evaluate whether only the trivial components have been replaced after intervention on the action.
Given the SCMs defined in~\Cref{Formulation}, we regard the counterfactual state $S_c$ as an intervention on the state $S_t$ and evaluate the counterfactual state by calculating the causal effect on the reward. Formally, we perform intervention on the state $S_t$ via $do(S_t:=S_c)$, and the $do$ calculus provides us:
\begin{equation}
\label{eq10}
\begin{aligned}
P^{\mathcal{M}; do(S_t:=S_c)}(R_{t+1})
&= \sum_{A_t}P(R_{t+1}|do(S_t), A_t) P(A_t|do(S_t)) \\
&= \sum_{A_t}P(R_{t+1}|S_c, A_t) P(A_t|S_c),
\end{aligned}
\end{equation}
from which we can identify the causal effect of $do(S_t:=S_c)$ on the reward $R_{t+1}$.
If the intervened probability distribution of reward is similar to the original distribution, substituting $S_t$ with $S_c$ has a minor causal effect on reward, indicating that it also has a minor influence on learning the user’s interest.
To this end, the reward will take the measurement of the distance between two probability distributions into account. We adopt the KL-Divergence to measure how the intervened reward probability distribution $P^{\mathcal{M}; do(S_t:=S_c)}(R_{t+1})$ is different from the original distribution $P(R_{t+1})$:
\begin{equation}
\begin{aligned}
&D_{KL}(P^{\mathcal{M}; do(S_t:=S_c)}(R_{t+1}) \Vert P^{\mathcal{M}}(R_{t+1}))\\
&= \sum_{r \in R} P^{\mathcal{M}; do(S_t:=S_c)}(R_{t+1}) \log \Big( \frac{P^{\mathcal{M}; do(S_t:=S_c)}(R_{t+1})}{P^{\mathcal{M}}(R_{t+1})} \Big)\\
&\medmath{= \sum_{r \in R} \Big( \sum_{A_t}P(R_{t+1}|S_c, A_t) P(A_t|S_c)\Big) \log \Big( \frac{\sum_{A_t}P(R_{t+1}|S_c, A_t) P(A_t|S_c)}{P(R_{t+1})} \Big)}
, \label{eq:kl}
\end{aligned}
\end{equation}
and the reward will be written as:
\begin{equation}
\begin{aligned}
R_{t+1} \leftarrow R_{t+1} + \frac{1}{D_{KL}(P^{\mathcal{M}; do(S_t:=S_c)}(R_{t+1}) \Vert P^{\mathcal{M}}(R_{t+1})) + \epsilon}
, \label{eq:reward}
\end{aligned}
\end{equation}
where $\epsilon$ is a small constant to prevent the denominator from being zero. In order to maximise the total reward, the agent will select the optimal action to make the intervened probability distribution similar to the original one. Thus, the agent will adjust its policy to achieve only replacing the trivial components in a state.
\subsection{Model-agnostic Learning and Integration}
\begin{algorithm}[ht]
\SetKwInOut{Input}{input}
\Input{Initial counterfactual critic and actor parameters: $\theta_{\mu^c}$ and $\theta_{\phi^c}$, recommendation critic and actor parameters: $\theta_{\mu^{re}}$ and $\theta_{\phi^{re}}$, replay buffer $D_c$ and $D_{re}$\;
}
\caption{Training and Application Algorithm for Counterfactual Synthesis Policy with
DDPG}
\For{episode = 1, G}{
Receive initial observation state $s_1$\;
\For{t = 1, T} {
\textbf{Stage 1: Train $\pi_{re}$}\;
Agent following $\pi_{re}$ interacts with the environment to generate transition $(s_t, a_t, s_{t+1}, r_{t+1})$\;
Store transition $(s_t, a_t, s_{t+1}, r_{t+1})$ in $D_{re}$\;
\textbf{Stage 3: Counterfactual Generation}\;
\If{counterfactual synthesis policy $\pi_c$}{
Agent following $\pi_c$ observe state $s_t$ and select action $a_c$\;
Execute action $a_c$ and observe counterfactual state $s_c$\;
Agent following $\pi_{re}$ observe counterfactual state $s_c$ and execute action $a_t$\;
Observe counterfactual reward $r_{c, t+1}$\;
Store transition $(s_c, a_t, s_{t+1}, r_{c,t+1})$ in $D_{re}$\;
}
Sample minibatch of $\mathcal{N}$ transitions from $D_{re}$ and update the model by using DDPG algorithm\;
}
\textbf{Stage 2: Train $\pi_c$}\;
\If{average episode reward > $\epsilon_1$}{
Store the current recommendation policy\;
\While{r < $\epsilon_2$}{
Agent following $\pi_{re}$ observe the current state $s$ and select action $a$\;
Execute action $a$ and observe reward $r_o$\;
Agent following $\pi_{c}$ observe the current state $s$ and select action $a'$\;
Execute action $a'$ and observe counterfactual state $s'$\;
Agent following $\pi_{re}$ observe counterfactual state $s'$, execute action $a$ and observe intervention reward $r_i$\;
Calculate the reward $r$ based on the~\cref{eq:reward}\;
Store transition $(s, a', s', r)$ in $D_{c}$\;
Sample minibatch of $\mathcal{N}$ transitions from $D_{c}$ and update counterfactual model by using DDPG\;
}
Store the current counterfactual policy\;
$\epsilon_1 = \epsilon_1 + \Delta_1$\;
$\epsilon_2 = \epsilon_2 + \Delta_2$\;
}
}
\end{algorithm}
Given the above-described reward function, it is possible to develop a policy that meets our goal of identifying a counterfactual state for the current state in which only the trivial components have been replaced.
The key to establishing the above-described reward function is obtaining the reward probability distribution of the observational data and the intervened reward probability distribution.
To this end, we utilize an additional policy to help extract both the observational and intervened reward probability distribution.
Our proposed counterfactual synthesis policy is a model-agnostic approach due to its key point on the reward function.
Thus, it can be easily implemented in current RL-based algorithms to achieve the training process and perform data augmentation.
We design two strategies for implementing the architecture as mentioned above: the first one is learning the counterfactual synthesis policy assisted by the expert demonstrations. The second one is joint training of the counterfactual synthesis policy and recommendation policy, in which the recommendation policy also serves as the distribution construction policy.
\subsubsection{Learning with Expert Demonstrations}\label{pre}
We introduce a pre-trained policy as the expert policy that uses external knowledge to obtain the observational and intervened reward distribution.
The pre-trained policy can be learned using any RL-based algorithm, and so does our policy.
The idea is motivated by the policy distillation~\cite{rusu2015policy}, in which the student policy is learned by minimizing the divergence between each teacher policy and itself over the dataset of the teacher policy.
Consider an expert recommendation policy that has learned the knowledge about user interest. It can guide the agent to select an optimal action to get positive feedback from users.
By interacting with the environment, the agent following the expert policy $\pi_e$ will construct the reward distribution over the observational data that reflects user interests. When retaining the same actions, the expert policy $\pi_e$ can also assemble the reward distribution under the intervention of $S_t$ with the~\Cref{eq10}.
The overall schema of learning with expert demonstrations architecture is illustrated in the left part of~\Cref{train}.
We use one actor-critic framework to learn the counterfactual synthesis policy $\pi_{c}$ and load the expert policy $\pi_e$ in an additional agent. The process starts from the expert agent interacting with the environment.
Depending on the current state $S_t$ at time $t$, the agent following the expert policy $\pi_e$ takes the action $A_t$ and interacts with the environment to obtain the reward $R_{t+1}$.
Meanwhile, the counterfactual synthesis agent will apply intervention on the action under the same environment whose current state is $S_t$ to generate a counterfactual state.
By replacing the action $A_t$ with the action $A_c$ taken by our training agent, the environment would enter into the counterfactual state $S_c$.
For the expert policy, the generated counterfactual state $S_c$ can be regarded as an intervention on the state $S_t$. Under this intervention, the expert policy can construct the intervened reward distribution and evaluate the causal effect of this intervention on the user interest.
Specifically, we estimate the reward by putting the agent with expert policy $\pi_e$ in the environment, of which the current state is counterfactual $S_c$, to receive the reward $R_c$.
With the observational and intervention reward distribution, the counterfactual synthesis policy is learned by minimizing the KL-Divergence between two distribution as formulated in~\Cref{eq:kl}:
\begin{equation}
\label{eq12}
\begin{aligned}
\medmath{\min_{\theta} \sum_{r \in R} \Big( \sum_{A_t}P(R_{t+1}|S^{\theta}_c, A_t) P(A_t|S^{\theta}_c)\Big) \log \Big( \frac{\sum_{A_t}P(R_{t+1}|S^{\theta}_c, A_t) P(A_t|S^{\theta}_c)}{P(R_{t+1})} \Big)}
.
\end{aligned}
\end{equation}
Once learned, the counterfactual synthesis policy can be used in any RL-based dynamic recommendation to collaborate with the recommendation policy for counterfactual synthesis.
The procedure follows the right part in~\Cref{train}.
Formally, based on the current state $S_k$, the trained policy $\pi_c$ is used to find and replace the trivial components in $S_k$ to synthesize counterfactual state $S_{c, k}$.
The recommendation policy $\pi_{re}$ receives the counterfactual state $S_{c, k}$ but perform the action $A_k$ based on the state $S_k$ to get the reward $R_{c, k+1}$.
Then the counterfactual transition $(S_{c,k}, A_k, S_{k+1}, R_{c, k+1})$ can be put into the replay buffer.
The generated counterfactual can provide additional information about the user's interests for the recommendation policy, as it is based on changing states. This can also be regarded as exploration for recommendation policy while retain the current level of exploitation.
\subsubsection{Joint Training}
The goal of joint training is to combine the training processes for the recommendation policy $\pi_{re}$ and the counterfactual synthesis policy $\pi_{c}$. To do this, the training procedure is divided into three stages. The conventional recommendation policy training process, which can be based on any RL algorithm, is the first stage. We begin by training the recommendation policy until we store a policy with an average episode reward greater than a certain threshold.
The stored policy can be used to build the observation and intervention reward distribution. The second stage, training the counterfactual synthesis policy, can then begin. This stage adheres to the approach outlined in~\Cref{pre} as well as the learning purpose mentioned in~\Cref{eq12}.
The training for policy $\pi_c$ will end when the average episode reward of policy $\pi_c$ reaches a particular threshold. Then the third stage is to apply the learned policy $\pi_c$ to the first stage to construct the counterfactual together with the policy $\pi_{re}$.
In this approach, the recommendation policy aids policy $\pi_c$ training, while the trained policy $\pi_c$ provides counterfactuals to supplement the transition in the recommendation policy's reply buffer, assisting the recommendation policy in comprehending the users' interests.
We raise the threshold each time we finish training the policy $\pi_c$. If the recommended policy meets the new threshold, we will use it to restart the second stage to improve the policy $\pi_c$.
The above-described three stages can use any RL algorithm as an underlying framework. We take the DDPG as an example and present the process in Algorithm 1, in which the training objective can be indicated as minimising the loss function:
\begin{align}
\medmath{L(\theta_\mu, \mathcal{D}) = \mathop{E}_{(s, a, s', r) \sim \mathcal{D}} \Big[\Big(\big(r+\gamma(\mu_{\theta^{targ}_\mu}(s', \phi_{\theta^{targ}_\phi}(s'))\big) - \mu_{\theta_\mu}\big(s, a\big) \Big)^2 \Big],}
\end{align}
where $\mathcal{D}$ is a set of mini-batch of transitions $(s, a, s', r)$ for $s \in S$, $a\in A(s)$, $r \in R$, and $s' \in S^+$ ($S^+$ is $S$ plus a terminal state). $\theta_\mu$ and $\theta_\phi$ are parameters for the critic and actor network, respectively. And $\mu_{\theta^{targ}_\mu}$ represent the target critic network.
\section{Experiments}
\subsection{Experimental Setup}
\subsubsection{Data}
We conduct experiments and evaluate our model in both {\it online} and {\it offline} manners. A public simulation
platform, VirtualTaobao~\cite{shi2019virtual}, is used for online evaluation. The benchmark datasets MovieLens-100k
and MovieLens-1M
are used for offline evaluation.
\vspace{1mm}\noindent\textbf{VirtualTaobao} mimics a real-world online retail environment for recommender systems. It is trained using hundreds of millions of genuine Taobao data points, one of China's largest online retail sites.
The VirtualTaobao simulator provides a "live" environment by generating customers and generating interactions, in which the agent may be tested with virtual customers and the recommendation system.
It uses the GAN-for-Simulating Distribution (GAN-SD) technique with an extra distribution constraint to produce varied clients with static and dynamic features. The dynamic attributes represent changing interests throughout an interactive process. It also employs the Multi-agent Adversarial Imitation Learning (MAIL) technique to concurrently learn the customers' policies and the platform policy to provide the most realistic interactions possible.
\vspace{1mm}\noindent\textbf{MovieLens}~\cite{harper2015movielens}. MovieLens-100k and MovieLens-1M are stable benchmark datasets based on user ratings on watching movies on the MovieLens website collected during different periods. Ratings are given on a 5-star scale, with each user having at least 20 ratings. Each user is assigned five features, whereas each movie has 23 features with 19 different genres.
\subsubsection{Baselines and evaluation metrics}
The proposed Counterfactual Synthesis Policy is model-agnostic that can be employed in various popular RL algorithms.
Although some priors works also approach counterfactual reasoning by generating counterfactual sequences, their works are designed for supervised learning, which is different from RL.
Existing RL-based recommendation methods do not have unified state representations, which cannot produce a fair comparison.
Therefore, we mainly focus on the following RL algorithms as baselines:
\begin{itemize}
\item \textbf{Deep Deterministic Policy Gradient (DDPG)}~\cite{lillicrap2015continuous}. DDPG is an off-policy method for environments with continuous action spaces. DDPG employs a target policy network to compute an action that approximates maximisation to deal with continuous action spaces.
\item \textbf{Soft Actor Critic (SAC)}~\cite{haarnoja2018soft}. SAC is an off-policy maximum entropy Deep Reinforcement Learning approach that optimises a stochastic policy. It employs the clipped double-Q method and entropy regularisation that trains the policy to maximise a trade-off between expected return and entropy.
\item \textbf{Twin Delayed DDPG (TD3)}~\cite{fujimoto2018addressing}: TD3 is an algorithm that improves on baseline DDPG performance by incorporating three key tricks: learning two Q-functions instead of one, updating the policy less frequently, and adding noise to the target action.
\end{itemize}
In terms of evaluation measures, click-through rate is the primary indicator used by VirtualTaobao. For dataset evaluation, three widely used numerical criteria are utilised: Precision, Recall, and Accuracy.
\subsection{Overall Comparison}
\begin{figure*}[h]\label{Comparison}
\centering
\begin{subfigure}[b]{0.33\linewidth}
\centering
\includegraphics[width=\linewidth]{ddpg.png}
\caption{}
\label{a}
\end{subfigure}
\begin{subfigure}[b]{0.33\linewidth}
\centering
\includegraphics[width=\linewidth]{sac.png}
\caption{}
\label{b}
\end{subfigure}
\begin{subfigure}[b]{0.33\linewidth}
\centering
\includegraphics[width=\linewidth]{TD3.png}
\caption{}
\label{c}
\end{subfigure}
\caption{Overall comparison result between the baselines, baselines with counterfactual synthesis policy learned with expert demonstrations and joint training of baselines and our policy: (a) DDPG as the baseline; (b) SAC as the baseline; (c) TD3 as the baseline.}
\end{figure*}
{\bf Online Experiment.} The overall comparison results conducted on the VirtualTaobao platform are depicted in~\Cref{Comparison}. In a nutshell, we find that both our approaches of learning through expert demonstrations and joint training achieve significant improvements over the chosen baselines.
Among the baselines, DDPG achieves the best performance. It obtains a stable policy around the 70,000 episodes but suffers more considerable variance than others.
The learning pace of SAC is slow. At the end of 100,000 episodes, SAC does not finish learning and does not reach a plateau. One probable explanation is that SAC utilizes a stochastic policy that introduces extra noise to the agent.
TD3 initially receives a better policy but does not maintain it. The downward trend and the fluctuation may be ascribed to the delayed update parameter mechanism.
Applying a policy learned through expert demonstrations to DDPG shows a similar rising tendency as DDPG, shown in~\Cref{a}. However, with our policy, DDPG rises dramatically with each ascent. This may be due to the fact that the counterfactual transactions generated by our policy assist the recommendation policy in better learning the dynamic interests of users.
The counterfactual synthesis policy trained by expert demonstrations assists DDPG in rising more steadily than others and achieving the highest CTR.
The joint training method helps DDPG rapidly grow and discovers a good policy around 60,000 episodes. However, we can also observe a fluctuation in the last section.
One of the possible reasons for the different performances of the two methods is that the counterfactual synthesis policy is frozen under the expert demonstrations setting while continuously optimising under the joint training scenario.
As shown in~\Cref{b}, the performance of SAC has a significant improvement when using our policy learned with either approach. The joint training procedure even assists SAC in reaching a plateau of roughly 90,000 episodes. Although the expert demonstration method drops dramatically near the 80,000 episodes, it rapidly returns and continues to increase.
Because the TD3 employs a special strategy to force agent to conduct exploration at the beginning of training, ~\Cref{c} shows that all three lines start with high CTRs. We can also observe that the expert demonstrations approach initially achieves a very high CTR.
Moreover, TD3 utilising our strategy learned from the two approaches displays a similar tendency. They all descend to a relatively stable level after a defined number of steps at the start, which can be attributed to the fact that TD3 changes the policy less frequently and introduces noise into the target action for target policy smoothing.
\begin{table*}
\caption{Performance comparisons of baselines, baselines with counterfactual synthesis policy learned with expert demonstrations and joint training of baselines and our policy on the MovieLens-1M dataset and the Amazon Fashion dataset. "-E" indicates that the counterfactual synthesis policy is learned with expert demonstration and "-J" indicates a joint training approach. The best results
are highlighted in bold and the second best ones use the symbol *.}
\label{Offline}
\begin{tabular}{lccc|ccc}
\hline
\multirow{2}{*}{} & \multicolumn{3}{c|}{MovieLens-100k} & \multicolumn{3}{c}{MovieLens-1M} \\
& Recall & Precision & Accuracy & Recall & Precision & Accuracy \\ \hline
DDPG & 0.6979±0.2251 & 0.5019±0.0071 & 0.6338±0.1487 & 0.7207±0.0632 & 0.4225±0.006 & 0.7198±0.0716 \\
DDPG-E & \textbf{0.8361±0.0698} & 0.5105±0.0048* & 0.7285±0.0462* & 0.7994±0.0189* & 0.4324±0.0031* & \textbf{0.8035±0.0171} \\
Improvement & 19.81\% & 1.70\% & 14.94\% & 10.92\% & 2.36\% & 11.63\% \\
DDPG-J & 0.7755±0.002* & \textbf{0.5111±0.0009} & \textbf{0.7468±0.0178} & \textbf{0.8033±0.0266} & \textbf{0.4329±0.0042} & 0.7846±0.0285* \\
Improvement & 11.13\% & 1.83\% & 17.83\% & 11.46\% & 2.47\% & 9.00\% \\ \hline
SAC & 0.6899±0.15 & 0.499±0.0079 & 0.6266±0.1002 & 0.7149±0.0438 & 0.4103±0.0086 & 0.7016±0.0557 \\
SAC-E & \textbf{0.7849±0.0861} & 0.5102±0.0025* & \textbf{0.6951±0.0559} & 0.7974±0.0616* & \textbf{0.4257±0.0084} & \textbf{0.7751±0.0215} \\
Improvement & 13.768\% & 2.248\% & 10.932\% & 11.538\% & 3.767\% & 10.476\% \\
SAC-J & 0.7795±0.0041* & \textbf{0.5118±0.0017} & 0.6912±0.0604* & \textbf{0.8055±0.0493} & 0.4222±0.0088* & 0.7663±0.0323* \\
Improvement & 12.992\% & 2.573\% & 10.302\% & 12.679\% & 2.899\% & 9.210\% \\ \hline
TD3 & 0.6822±0.1352 & 0.5001±0.0043 & 0.623±0.092 & 0.7034±0.0546 & 0.4201±0.0055 & 0.7121±0.0585 \\
TD3-E & \textbf{0.7706±0.0785} & 0.5068±0.0045* & 0.6827±0.0521* & \textbf{0.7885±0.0047} & 0.4315±0.0029* & \textbf{0.8077±0.0107} \\
Improvement & 12.954\% & 1.349\% & 9.584\% & 12.097\% & 2.713\% & 13.426\% \\
TD3-J & 0.7541±0.0253* & \textbf{0.5074±0.0051} & \textbf{0.698±0.0443} & 0.7711±0.0087* & \textbf{0.4338±0.0083} & 0.7805±0.0139* \\
Improvement & 10.535\% & 1.472\% & 12.049\% & 9.617\% & 3.261\% & 9.602\% \\ \hline
\end{tabular}
\end{table*}
\vspace{0.2cm}
\noindent{\bf Offline Experiment.} The comparison results of employing our policy with baselines are listed in~\Cref{Offline}. On both datasets, the statistical results show that using our policy learned with any approach clearly outperforms the compared underlying models, DDPG, SAC, and TD3.
In a nutshell, The results demonstrate the efficacy of our model-agnostic counterfactual synthesis policy in enhancing any RL-based recommendation model. It is worth mentioning that expert demonstration learning and joint training approaches generate comparable progress across all three baselines.
We attribute the improved prediction result to the fact that our policy assists the recommendation policy in understanding users' interests.
The improved performance of RL-based recommendation systems in both online and offline experiments demonstrates the generality of our policy.
\subsection{Hyper-parameter Study}
We examine the number of hidden sizes to investigate how they affect the performance of counterfactual synthesis policies. Because the DDPG using our policy learned through expert demonstrations achieves the best results, we will utilise it as the baseline for the trial.
We set the hidden size number in the range $[64, 128, 256]$, and present the comparison result in~\Cref{parameter}.
\begin{figure}[h]
\centering
\includegraphics[width=0.58\linewidth]{fig3.png}
\caption{Performance comparison on the hidden size of 64, 128 and 256.}
\label{parameter}
\end{figure}
The graph demonstrates that the best performance is obtained when the hidden size is set to 128. Either a too small or too large number would harm the performance of our counterfactual synthesis policy.
Although the standard deviation of the result with 64 hidden sizes is small, it has a slower convergence speed and worse performance, which may be attributed to too few neurons in the hidden layers to collect information adequately.
A too large number of hidden sizes leads to a high standard deviation. One possible explanation is that using too many neurons in the hidden layers causes overfitting, making it difficult to converge.
\section{Related Work}
\vspace{1mm}\noindent\textbf{RL-based interactive recommendation.}
Since RL trains an agent that is able to learn from interaction trajectories, it has shown great effectiveness in modelling interactive recommendation processes.
~\citet{mahmood2007learning} modify the Markov Decision Process (MDP) to achieve an adaptive interaction recommender system using Reinforcement Learning (RL) techniques.
~\citet{zheng2018drn} apply a DQN structure and Dueling Bandit Gradient Descent method in reinforcement learning framework to address the dynamic nature of news recommendation.
To handle the drawback of DQN in large and dynamic action space scenarios, ~\citet{zhao2017deep} build the list-wide recommendation model with the capability of dealing with large and dynamic item space upon the Actor-Critic framework.
~\citet{chen2020knowledge} utilize the knowledge graphs on the actor-critic network to improve the stability and quality of the critic network for interactive recommendation.
\vspace{1mm}\noindent\textbf{Counterfactual for Recommendation.}
Causality and counterfactual reasoning have received increased attention in a variety of fields.
Many researchers leverage counterfactual reasoning in recommender systems to release bias issues.
~\citet{schnabel2016recommendations} propose a paradigm for dealing with selection bias in recommendation system learning through the use of methodologies from causal inference.
They provide unbiased estimators for evaluating recommendation systems using propensity-based methodologies and present an Empirical Risk Minimization (ERM) framework for training recommendation systems under selection bias.
~\citet{liu2020general} focus on knowledge distillation and propose a simple framework KDCRec for extracting information from uniform data to overcome bias problems in recommender systems.
Some works attempt to correct the exposure bias by connecting classical recommendation models with causality~\cite{bonner2018causal, wang2020causal}.
A rising number of works use counterfactual reasoning to provide explanations, improve model interpretability, and learn a robust representation~\cite{tan2021counterfactual, madumal2020explainable, zhang2021causerec}.
~\citet{zhang2021causerec} use similarity function to identify the subsets in the behaviour sequence that can be replaced to generate counterfactual sequences, which help to learn a robust user representation for sequential recommendation.
Another line of work focuses on data augmentation of training samples for data-scarce problems.
~\citet{wang2021counterfactual} propose to generate counterfactual sequences by finding replacement items in the embedding space for sequential recommendation models.
~\citet{lu2020sample} focus on the learning of a causal mechanism by using a GAN-like adversarial framework for counterfactual data augmentation.
\section{Conclusion}
In this paper, a novel counterfactual synthesis policy has been proposed based on the casual view of MDP.
The linking reward with the divergence between the observational and intervening reward distributions can guide the agent to find a replacement state with minimal causal effect on users' interest.
The counterfactual synthesis policy is simple to implement in various RL frameworks and works with the recommender agent.
During the interaction between the recommender agent and the environment, our agent provides counterfactual data that takes into account the dynamic users' preferences.
Results show that the counterfactual synthesis policy performs well on different RL frameworks and achieves a considerable improvement for all compared baselines.
\balance
\bibliographystyle{ACM-Reference-Format}
|
\subsection{MD simulations}
All bilayers containing random mixtures of DPPC, DLiPC, and CHOL were prepared using the CHARMM-GUI web-based graphical interface.\cite{Lee2016}
The simulations of the DPPC bilayers, the DLiPC bilayers, and the PL/CHOL bilayers were run for 1~$\mu$s. The initial 50~ns of each trajectory were discarded to assure the exclusive analysis of equilibrium data. The simulations were run at temperatures from 290~K to 350~K in 10~K steps. Bilayer compositions and trajectory lengths are listed in table \ref{tab:MDlist} in the SI.
All simulations were conducted using the CHARMM36 force-field\cite{Klauda2010,Lim2012a} and Gromacs~2020 software package\cite{Pall2015,Abraham2015}. The systems were equilibrated using the 6-step established CHARMM-GUI parameter set, gradually decreasing restraints on lipid head group positions and chain dihedral angles (see table \ref{tab:equilibration_protocol}). The TIP3P model was used as water model.\cite{Jorgensen1983} The Nos\'e-Hoover algorithm was used to maintain the temperature with a coupling constant of 1~ps, coupling bilayer and solvent separately.\cite{Nose1984} To maintain the pressure at 1~bar the Parrinello-Rahman barostat was used with a semiisotropric coupling scheme, a coupling constant of 5~ps and compressibility of 4.5x10$^{-5}$ bar$^{-1}$.\cite{Parrinello1981} Particle mesh Ewald electrostatics were used with a real-space cutoff of 1.2~nm.\cite{Darden1993} The Lennard-Jones potential was shifted to zero between 1.0 and 1.2~nm, with a cutoff of 1.2~nm and the nonbonded interaction neighbor list was updated every 20 steps with a cutoff of 1.2~nm.
\subsection{Parameter definitions and software}
Handling and analysis of MD data heavily relied on the MDAnalysis package for Python \cite{Michaud-Agrawal2011b,Gowers2016}.
All molecular visualizations in this work were created using VMD \cite{Humphrey1996}.
The radial distribution functions (RDFs) were calculated using the Gromacs analysis tools \cite{Abraham2015}. All RDFs were calculated individually for each leaflet using in-plane distances and if not stated otherwise were averaged. For the PL molecules, the phosphor atoms and, for CHOL, the hydroxyl oxygen atoms were used as reference positions.
The lipid chain order parameter S is defined as $S = \left< 1.5\cos^2\theta - 0.5 \right> $, where $\theta$ denotes the angle between the vector spanned by every second carbon atom in a lipid chain and a the bilayer normal (z-axis). This definition of calculating the order parameter differs from the commonly used S$_\mathrm{CH}$ order parameter, as we do not consider hydrogen atoms in this calculation. Thus $S=-0.5$ is the lowest order state, while $S=1.0$ is the highest order state (linear chains). This has the advantage of easier comparability between the parameters of saturated and unsaturated PLs.
\subsection{Lateral bilayer geometry of binary PL/CHOL mixtures}
\begin{figure}[bt]
\centering
\includegraphics[width=0.9\textwidth]{./figures/rdf_pl.pdf}
\caption{Radial distribution functions of PL-PL (top) and PL-CHOL (bottom) lateral distances from simulations of DPPC/CHOL or DLiPC/CHOL (with CHOL concentrations from 0 to 30\%) averaged over simulations at temperatures above the phase transition. For the PL head groups phosphor atoms (dashed lines), for the CHOL head groups the hydroxy oxygen atoms and for the PL tails the eighth carbon atoms (solid lines) were used as reference positions. Shades of blues (tail level) or reds (head group level) indicate the CHOL concentration, with darker colors representing higher CHOL concentrations. }
\label{fig:rdfs_pl}
\end{figure}
First, we analyse radial distribution functions (RDFs) which describe the local structure and are the basis to setting up the geometry of the lattice. Here, we show in figure \ref{fig:rdfs_pl} the RDFs of PL-PL and PL-CHOL lateral distances at the head group level (P-P and P-O atom distances) and at the acyl chain level (C8-C8 and C8-O atom distances) averaged over the respective simulations above the phase transition temperatures. The RDFs are used for a consistent definition of nearest neighbors. Note that the PL-CHOL RDFs are finite at r=0 because the hydroxy oxygens of CHOL lie in a different bilayer height than the C8 or P atoms and can thus overlap.
The RDFs of PL-PL distances show profound differences comparing head group and chain level. The chains are well-structured, up to distances of 1.5~nm (second neighbor shell, see \cite{Leeb2018a}). The structuring is stronger with increasing CHOL concentration, especially for DPPC. In contrast, the CHOL concentration exhibits no influence on the head group structure (figure \ref{fig:rdfs_pl} top) and accordingly we find that the chain level is more sensitive to changes of the bilayer composition.
The DPPC-CHOL RDFs (figure \ref{fig:rdfs_pl} bottom) show very similar behavior compared to the head and tail levels of the PL-PL. Generally, the structuring between DPPC and CHOL is weaker than between DPPC and DPPC. The chain level RDFs also show increased structuring by CHOL, while CHOL does not affect the head level structure. Interestingly, the maxima of the DPPC-CHOL chain level RDFs roughly coincide with the DPPC chain level maxima, indicating that CHOL takes the position of an acyl chain. The RDFs of DLiPC-CHOL, on the other hand, show no structuring in the chain region.
The investigation of the different structuring behavior of the head and tail levels through the RDFs showed that, contrary to the chain region, the head group structures are independent of the concentration of CHOL. Accordingly, a consistent neighbor definition using a static distance cutoff is possible employing the head groups as reference positions for, e.g., lattice models. Based on the first minimum of PL-PL and PL-CHOL RDFs we use a distance cutoff of 1~nm between the head group atoms to define nearest neighbors in this work.
\begin{figure}[bt]
\centering
\includegraphics[width=0.9\textwidth]{./figures/NofS.pdf}
\caption{Top: Number of nearest DPPC (left) and DLiPC (right) neighbors of a PL as a function of its order parameter, separately evaluating reference lipids with 0, 2, and 4 nearest CHOL neighbors. Bottom: The respective number of nearest DPPC (left) and DLiPC (right) neighbors around CHOL. The line style indicates the CHOL concentration of the respective PL/CHOL bilayers.}
\label{fig:NofS}
\end{figure}
Now we can quantify the number of nearest neighbors of PLs or CHOL. The average number of PL neighbors of the respective PL with 0, 2, or 4 CHOL neighbors as a function of its order parameter is shown in figure~\ref{fig:NofS} top. The PL neighborhood around CHOL is shown in figure \ref{fig:NofS} bottom. Note that in our definition, the number of nearest neighbors only reflects the adjacent DPPC and DLiPC lipids, respectively, but {not} CHOL.
The number of neighbors is not affected by the lipid's order. We find roughly 4.5 and 3.5 neighbors for DPPC and DLiPC, respectively. Interestingly, the number of nearest neighbors hardly depends on the number of CHOL molecules around DPPC; the average number of PL neighbors is decreased by, on average, 0.5 PL neighbors comparing PLs with 0 and 4 CHOL neighbors, regardless of the overall CHOL concentration. From this observation, we can obtain three conclusions: (1) The influence of CHOL on the lateral PL head group structure is weak, reflecting the well-known condensing effect. (2) The residual dependence mainly depends on the local concentration which allows the use of the same neighbor function for all CHOL concentrations. (3) The neighbor function hardly depends on the order parameter which further simplifies the impact of the neighbor function.
The PL neighborhood around CHOL resembles the neighborhood of DPPC and DLiPC, except that CHOL has, on average, one additional PL neighbor. This increase may be related to the fact that there exists a finite probability that the projection of the CHOL hydroxy oxygen atom on the membrane plane may be close to the projection of the PL head-group. However, all three conclusions, drawn for the PL-PL neighborhood, hold here as well.
\begin{figure}[bt]
\centering
\includegraphics[width=0.25\textwidth]{./figures/lattice.pdf}
\caption{Square lattice representation with CHOL on a sublattice (yellow dots) with void sites. PLs reside on the main lattice (cyan). The shades of cyan indicate the different order parameters that a PL can exhibit.}
\label{fig:squarelattice}
\end{figure}
Using the presented results, we are in the position to find a lattice geometry, based on the lipid head group region, that can incorporate the average lipid neighborhood. The PL head group region is not affected by the presence of CHOL. Therefore, PLs and CHOL should reside on separate square lattices. In that framework, the CHOL lattice is a sub-lattice with void sites, enabling the simulation of different CHOL concentrations. A minimum version of the resulting lattice model could be formulated in the way depicted in figure \ref{fig:squarelattice}.
The slight deviation from an optimal number of 4 neighbors on a square lattice can be handled by including an appropriate scaling factor $N/4$ ($N$: total number of neighbors for the specific CHOL environment) for the individual Pl-PL interaction functions, thereby correctly accounting for the total interaction. This minor correction was already successfully used in previous work \cite{Hakobyan2017} and brings in a very minor mean-field character into the theoretical description.
From the RDF plots of the chain level it is obvious, that the number of PL (chain) neighbors around PL increases with increasing CHOL concentration. Thus, it would no longer be possible to define neighbor functions of chains without taking into account the overall CHOL concentration.
\subsection{Cholesterol-related interaction functions}
Next, we investigate the characteristics of the PL-CHOL and CHOL-CHOL interactions. The interaction profile of a PL with CHOL as a function of the PL order S$_{PL}$ is shown in figure \ref{fig:Eplc} top. Equivalent to the discussion of the PL-PL interaction functions, we also calculated the interaction for a given number $N_c$ of CHOL neighbors of the respective PL-CHOL pair (figure \ref{fig:Eplc} middle) and calculated the differences between the interaction energies from different CHOL concentrations (figure \ref{fig:Eplc} bottom).
Similarly to the PL-PL interactions, the strength of the PL-CHOL pair interaction decreases with each additional CHOL neighbor (figure \ref{fig:Eplc} middle) and, for fixed value of $N_c$ the interaction becomes stronger with increasing CHOL interaction. Both effects are counterbalanced in figure \ref{fig:Eplc} top.
For DPPC and for all CHOL concentrations and $N_c$-values the strength of the DPPC-CHOL interaction energy displays a maximum for S$_{PL}$ $\approx$ 0.65. Thus, the optimal PL chain order for interaction with CHOL is not an all-trans but a slightly disordered chain. The DLiPC-CHOL interaction displays a monotonous dependence on S$_{PL}$ . However, this may be related to the fact that for the unsaturated DLiPC-lipid no order parameters beyond 0.7 are accessible. Interestingly, DLiPC shows a stronger dependence on the order parameter than DPPC which results in a stronger interaction by approx. 4~kJ/mol at S$_{PL}$ =0.6. The differences between profiles of different CHOL concentrations lie between 0 and -5~kJ/mol and, importantly, hardly depend on S$_{PL}$ (figure \ref{fig:Eplc} bottom). In analogy to the PL-PL interaction, the variation with order parameter is less than 2kJ/mol.
\begin{figure}[tbhp]
\centering
\includegraphics[width=0.9\textwidth]{./figures/eplc_panel.pdf}
\caption{Top: DPPC- and DLiPC-CHOL interaction energies as a function of the PL order parameter S$_{PL}$. The energies were derived from PL/CHOL mixtures with CHOL concentrations ranging from 10 to 30\%. Middle: DPPC- and DLiPC-CHOL interaction energies when separated by the number of the pair’s nearest CHOL neighbors (N$_c$). Bottom: PL pair interaction energy difference between energies from simulations with X=10, 20\% and 30\% CHOL.}
\label{fig:Eplc}
\end{figure}
The CHOL-CHOL interaction does not directly affect the PL order parameter, but it determines how favorable an agglomeration of CHOL is from an energetic point of view. We show the CHOL pair interaction as a function of the pair's CHOL neighbors N$_c$ in DPPC and DLiPC bilayers and the overall CHOL concentration in figure \ref{fig:Ecc_Nc}. Qualitatively, we see the same tendency as for the other two interactions: stronger interaction for less CHOL neighbor and, as a much smaller effect, higher CHOL content. In DPPC mixtures, the concentration dependence is slightly higher with approximately 2~kJ/mol per 10\% CHOL. In DLiPC mixtures, the concentration dependence is roughly half as strong as in DPPC. The interaction decreases with 1 to 1.5~kJ/mol per N$_c$, both in the DPPC and DLiPC mixtures, and the overall interaction strength between CHOL is slightly higher in DLiPC mixtures.
\begin{figure}[tbhp]
\centering
\includegraphics[width=0.4\textwidth]{./figures/Ecc_Nc.pdf}
\caption{CHOL-CHOL pair interaction energy as a function of the number of their CHOL neighbors N$_c$ derived from simulations of DPPC/CHOL (solid) and DLiPC/CHOL (dashed) mixtures. The CHOL concentration is 10\% (orange), 20\% (green), and 30\% (red).}
\label{fig:Ecc_Nc}
\end{figure}
\section{Comparison with MD results}
\subsection{Single PL-CHOL mixture: from the lattice model to MD results }
Here we discuss for interesting properties of CHOL-containing membranes, as obtained from MD simulations, how the results can be interpreted from the properties of the neighbor and interaction functions.
\begin{figure}[tbhp]
\centering
\includegraphics[width=0.5\textwidth]{./figures/mean_order.pdf}
\caption{Average order parameter as a function of temperature and total CHOL concentration for DPPC/CHOL and DLiPC/CHOL mixtures. The CHOL concentrations are 0\%, 10\%, 20\%, 30\% (from below to above).}
\label{fig:averageS}
\end{figure}
We start with the evaluation of the order parameters in the different CHOL-containing bilayers, shown in figure \ref{fig:averageS}. The average order of the DLiPC-containing bilayers increases linearly with CHOL concentration and decreasing temperature. The CHOL ordering is roughly twice as strong for DPPC as it is for DLiPC. For DPPC, we see a phase transition from gel to the liquid phase at around 320~K. The properties of the PL-CHOL interaction function suggest a possible explanation. In the relevant range of order parameters below S$_{PL}$ $\approx 0.7$ the PL-CHOL interaction strength increases with increasing order parameter S$_{PL}$ . Thus, addition of CHOL and the concomitant increasing relevance of the PL-CHOL interaction for the overall energy balance prefers higher PL order parameters. Since the interaction functions are only determined for temperatures in the liquid phase, we restrict our discussion to this regime.
However, a closer analysis of the MD results reveals several important features, for which a fresh perspective is provided based on the results of this work. (1) Why is the order parameter for the CHOL-free case higher for DPPC ($\approx 0.35$) than for DLiPC ($\approx 0.2$)? The PL-PL interaction functions suggest that the enthalpic gain upon ordering is not stronger than for DLiPC (compare, e.g., the increase of energy between order parameters of 0 and 0.4 in figure \ref{fig:Eplpl} which is approx. a factor of 1.5 larger for DLiPC). Although this effect is slightly counterbalanced by the fact that the number of nearest PL neighbors is smaller by a factor of $\approx0.75$ for DLiPC bilayers compared to the DPPC bilayers, these considerations do not explain the significant difference between the order parameters of DPPC vs. DLiPC. Thus, this observation suggests that the difference in chain entropy must be used to explain this difference. Indeed, due to the disordered nature of DLiPC one may envisage that for DLiPC there is a significant entropy penalty to reach more ordered states. Among others, this also reflected by the observation that for DLiPC order parameters S$_{min}$ above 0.5 are hardly visited whereas for DPPC order parameters up to 0.8 can be observed. (2) Why is the impact of CHOL on the chain order much larger for DPPC than for DLiPC? Again, this cannot be explained by the specific form of the PL-CHOL interaction function since, if at all, it would predict a stronger impact of CHOL on the DLiPC order. However, the chain entropy argument, already used in (1), would again explain the additional suppression of states with higher order parameters for DLiPC. In contrast, the enthalpic gain at higher order parameters upon addition of CHOL is less suppressed by the chain entropy, which naturally favors smaller order parameters. (3) Why is the impact of CHOL on the DPPC order particularly strong at $T =320$~K? We envision that this effect is related to the critical effects close to the transition to the gel state. As shown in figure \ref{fig:pscd} the fluctuations of the order parameter for pure DPPC is particularly large at $T =320$~K. In analogy to the general system behavior at critical points (e.g. for the Ising model close to $T_c$), the system is particularly susceptible to external perturbations such as the addition of CHOL.
Finally, one may wonder whether this insight may help to rationalize the emergence of rafts for the ternary mixtures, i.e. the presence of a DLiPC-domain with low order parameters and a DPPC-CHOL-domain with higher order parameters. We have seen that the impact of CHOL on DLiPC domains is very small (see figure \ref{fig:averageS}). As a consequence there is hardly any gain in free energy due to the mutual interaction. In contrast, for DPPC-CHOL domains the DPPC order parameter is strongly increased which implies a significant free energy gain as compared to the pure DPPC domain. Thus, these observations suggests that the free energy is lower if CHOL forms the liquid ordered phase together with DPPC, which is exactly what happens upon raft formation.
\begin{figure}[tbhp]
\centering
\includegraphics[width=0.9\textwidth]{./figures/pscd_c0.pdf}
\caption{Order parameter distributions of simulations of pure DPPC or DLiPC bilayers close to the phase transition temperature of DPPC ($\approx 320$~K). }
\label{fig:pscd}
\end{figure}
As another observable we characterize the agglomeration of CHOL. It is reflected by the average number of CHOL neighbors around CHOL which we extract from the CHOL-CHOL RDF for distances smaller than 1 nm. When starting from the RDF, this value is automatically normalized by the number of nearest CHOL neighbors in case of a statistical distribution of CHOL molecules. The results are shown in figure \ref{fig:nr_cholchol}. For all temperatures and concentrations, an agglomeration of CHOL is observed. For concentrations of 10\%, the enrichment is even 4 to 5 times larger than for the case of a homogeneous CHOL distribution. The CHOL-CHOL enrichment is slightly higher in mixtures with DLiPC, and, with increasing concentration, the agglomeration effect becomes smaller. This may be explained via a simplified view of the bilayers. Envisioning the bilayer as a lattice with a fixed number of sites for PLs and CHOL (e.g., the formulation in figure \ref{fig:squarelattice}), the space for CHOL gets filled upon increasing CHOL concentrations, such that the local concentration becomes similar to the bulk concentration and becomes unity when all CHOL spaces are filled. Within the lattice model this would be the case for 50\% CHOL concentration.
Naturally, the significant interaction between adjacent CHOL molecules is likely to drive this agglomeration effect which in the considered range of temperatures is basically T-independent. Since this interaction is stronger for the DLiPC-CHOL systems, it naturally emerges that the agglomeration effect is somewhat larger for that case. Closer inspection of the PL-CHOL interaction shows that for fixed overall concentration it (approximately) shows a linear dependence on the number $N_c$ of CHOL neighbors. Thus, the overall contribution of the PL-CHOL interaction to the total interaction energy does not depend on the arrangement of CHOL molecules.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.9\textwidth]{./figures/nr_cholchol_binary.pdf}
\caption{Number of CHOL neighbors of CHOL divided by the number of CHOL neighbors that would arise if CHOL is distributed homogeneously such that values larger than unity indicate CHOL agglomeration. The data is shown for concentrations of 10, 20, and 30\% (yellow, green, and red) in mixtures with DPPC (left) and DLiPC (right).}
\label{fig:nr_cholchol}
\end{figure}
\subsection{PL-PL related interaction functions}
The interaction function between lipids, which form the key aspect of the lattice model, is influenced by many factors (e.g., variation of lipid head groups, chain length, and saturation). These factors all influence the inherent order of a lipid. In previous work\cite{Keller2021}, we evaluated the interaction of a PL molecule with its environment and expressed the {\it total} interaction as a function of the PL's order parameter $S$. Indeed, a significant dependence on $S$ was observed. Here we go one step further and characterize the interaction between the individual adjacent lipid pairs with indices $i$ and $j$ as a function of the pair's chain order parameters $S_i$ and $S_j$. In particular, we are interested in the effect of CHOL on this interaction which should reflect CHOL's unique ability to alter the phase behavior in lipid mixtures. The interaction functions are determined as averages over the temperature ranges from 330 to 350~K (DPPC) and 290~K to 350~K (DLiPC), respectively.
The interaction energy matrix $E(S_i, S_j)$ of a pair $ij$, shown in figure \ref{fig:interaction2d}, is necessarily symmetric. The interaction increases with increasing order parameters, and the CHOL concentration does not change the general features of the energy matrix. Following our previous work, we aim to express the interaction matrix as a 1D function where the argument depends on both order parameters. This approach is convenient from a statistical perspective (less noisy) and allows a more direct view on the dependence of the interaction energy on the order parameter. Of course, when actually approaching level II (employing \textit{neighbor} and \textit{enthalpy} functions to get the \textit{entropy}) of the lattice modeling, one might still have the option to work with the full 2D interaction matrix. We now have multiple options to simplify the interaction description (i.e., reducing the two-dimensional to a one-dimensional function). In previous work \cite{Hakobyan2017} we chose the pair's average order parameter S$_{avg}$ for PL mixtures. For this purpose we calculated the weighted average for given average order parameter. Here we propose another scheme, using the minimum order parameter S$_{min}$ averaging over horizontal/vertical entries in the interaction matrix and show that this choice is superior to our previous one.
\begin{figure*}[tbhp]
\centering
\includegraphics[width=0.9\textwidth]{./figures/energy_2d.pdf}
\caption{Interaction energy of a pair of lipids i and j of DPPC (top) or DLiPC (bottom) as a function of the order parameters S$_i$ and S$_j$ in bilayers of DPPC/CHOL or DLiPC/CHOL at CHOL concentrations from 0 to 30\%.}
\label{fig:interaction2d}
\end{figure*}
\begin{figure}[tbhp]
\centering
\includegraphics[width=0.9\textwidth]{./figures/eplpl_compare_s_schemes.pdf}
\caption{Standard deviation of the pair interaction energies if the minimal order parameter of the lipid pair is considered (Smin, horizontal/vertical averaging of \ref{fig:interaction2d}) or if the average order parameter of the pair is considered (Savg, diagonal averaging). Blue, yellow, green and red indicate CHOL concentrations of 0, 10, 20 and 30\%, respectively.}
\label{fig:s_scheme_var}
\end{figure}
An optimal choice minimizes the error of using the 1D variant as compared to the complete 2D interaction matrix. Each CHOL concentration is analyzed separately. Specifically, we calculated for each value of the respective order parameter (S$_{avg}$ or S$_{min}$ ) the weighted variance, see figure \ref{fig:s_scheme_var}, where its square root, i.e. the standard deviation, is displayed. It shows that in the relevant regime of positive order parameters, S$_{min}$ is a better choice than S$_{avg}$ . Furthermore, when weighting each value with the occurrence probabilities of the individual order parameters, one obtains an overall quality (figure \ref{fig:s_scheme_var_bar}) which also shows the superiority of the S$_{min}$ -scheme, in particular for DPPC.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.9\textwidth]{./figures/eplpl_panel.pdf}
\caption{Top: PL pair interaction energies as a function of the pair’s minimal order parameter and derived from binary PL/CHOL mixtures with CHOL concentrations from 0 to 30\%. Middle: The respective PL pair interaction energies as a function of the pair’s minimal order parameter when separated by the number of the pair’s nearest CHOL neighbors. N$_c$ considers the total number of CHOL molecules around the pair. Bottom: PL pair interaction energy difference between energies from simulations with 30\% CHOL and 0\%, 10\%, and 20\%.}
\label{fig:Eplpl}
\end{figure}
In figure \ref{fig:Eplpl} top we show the interaction enthalpy between the PL-PL pairs as a function of S$_{min}$ for different CHOL concentrations. In agreement with previous work on DPPC bilayers without CHOL \cite{Hakobyan2017}, the interaction of DPPC molecules is stronger for chains with high order (figure \ref{fig:Eplpl} top). The interaction energy reaches a plateau at around 60~kJ/mol for S$_{min}$ smaller than a threshold value that depends on the CHOL concentration. The DLiPC-DLiPC pair interactions behave quite similarly. Importantly, the PL-PL interaction becomes weaker for higher CHOL concentration. The interpretation is straightforward: since one PL also wants to optimize its interaction with adjacent CHOL molecules (particularly relevant for high CHOL overall concentration), the efficiency of its interaction with the adjacent PL is reduced.
To get more specific information about the effect of CHOL on the PL-PL interaction, we additionally record the number of direct CHOL neighbors N$_c$ and display the PL-PL interaction in dependence of the overall CHOL concentration {\it and} the value of N$_c$. Interestingly, with this finer resolution the main dependence is with respect to N$_c$. Each additional CHOL neighbor decreases the interaction energy. In contrast, the explicit dependence on the overall concentration in this representation is quite weak (figure \ref{fig:Eplpl} middle) and displays a slightly stronger interaction strength for higher CHOL concentration. This may reflect the residual global effect of a generally tighter packing in bilayers of high CHOL content. These observations imply that the initially observed dependence on the CHOL concentration (figure \ref{fig:Eplpl} top), reflecting a weighted average of the distribution of N$_c$, is just a consequence of the typically larger number of direct CHOL neighbors for a PL with high chain order.
These results also imply that the influence of CHOL is mainly local.
To quantify the non-local contributions, we compare the interaction energies at 30\% CHOL concentration with those at lower concentrations (figure \ref{fig:Eplpl} bottom). Here we see that the dependence on S$_{min}$ is very weak (approx. 2 KJ/mol in the relevant regime of order parameters). Energies that are independent of the configuration and thus do not provide driving forces of chain ordering are irrelevant for the membrane lattice model. Thus, the dependence on the overall CHOL content can be neglected. This rationalizes the choice of identical interaction parameters in the whole range of CHOL concentrations. We would like to add that the dependence on the overall CHOL concentration is even smaller for DLiPC. This holds for the absolute energy shift as well as for the dependence on S$_{min}$ .
The interaction functions were determined as an average from simulations at 330~K, 340~K, and 350~K for DPPC (290 K to 350 K for DLiPC), presupposing the energies H(S) are not temperature-dependent. To evaluate the temperature dependence of respective interactions, we calculated the difference of the interaction functions from 330~K and 350~K. This energy difference for DPPC and DLiPC pair interactions is depicted in figure \ref{fig:deltaEplpl_Tdepscheme} in the SI. The differences between the energies at 330~K and 350~K lie below 2~kJ/mol. Only the interaction energy for the unlikely case of 30\% CHOL and 0~CHOL neighbors shows a higher temperature dependence. Consequently, we infer that the temperature dependence of the interaction as a function of S$_{min}$ and N$_c$ is basically temperature independent in the investigated temperature range.
\begin{figure}[tbhp]
\centering
\includegraphics[width=0.95\textwidth]{./figures/epl_bintern.pdf}
\caption{Left and middle: DPPC and DLiPC pair interaction difference between energies derived from binary (PL/CHOL) and ternary (DPPC/DLiPC/CHOL) lipid mixtures as a function of the pair’s minimal order parameter. Right: DPPC-DLiPC pair interaction energies as a function of the pair’s minimal order parameter derived from DPPC/DLiPC/CHOL mixtures with CHOL concentrations from 0 to 30\%.}
\label{fig:Epl_bintern}
\end{figure}
Finally, we investigate whether the PL-PL interaction profiles depend on the bilayer composition. Therefore, we determined the energy difference between respective PL interactions in binary DPPC/CHOL or DLiPC/CHOL and ternary mixtures of DPPC/DLiPC/CHOL, depicted in figure \ref{fig:Epl_bintern} left. The absolute PL-PL interaction functions are shown in figure \ref{fig:Epl_bintern} right.
For small order parameters the absolute interaction values differ by about 2 kJ/mol between profiles from ternary and binary bilayers whereas they are basically identical in the limit of large order parameters. The interaction in the binary mixtures is stronger for DPPC (decreased packing in ternary mixtures) and weaker for DLiPC (increased packing in ternary mixtures). Thus, the energy gain for ordering DPPC acyl chains is slightly weakened in the ternary mixtures. However, due to the smallness of these effects we may state the PL-PL interaction is only slightly affected when going from the binary to the ternary bilayer. Thus the present results suggest that the interaction functions, derived from pure PL mixtures, can be, to a good approximation, used also in more complex mixtures.
\section{Computational details}
\begin{table*}[htbp]
\caption{Simulation lengths, temperatures and compositions of the utilized MD simulations.}
\begin{tabular}{ccccc}
System & Temperatures (K) & Lengths ($\mu$s) & \#PL & \#CHOL \\
\hline
DPPC & 290-350 & 1 & 350 & 0 \\
DLiPC & 290-350 & 1 & 316 & 0 \\
DPPC/DLiPC & 290-350 & 1 & 166/166 & 0 \\
DPPC/CHOL10\% & 290-350 & 1 & 342 & 38 \\
DPPC/CHOL20\% & 290-350 & 1 & 304 & 76 \\
DPPC/CHOL30\% & 290-350 & 1 & 280 & 120 \\
DLiPC/CHOL10\% & 290-350 & 1 & 306 & 34 \\
DLiPC/CHOL20\% & 290-350 & 1 & 280 & 70 \\
DLiPC/CHOL30\% & 290-350 & 1 & 266 & 114 \\
DPPC/DLiPC/CHOL10\% & 290-350 & 1 & 162/162 & 36 \\
DPPC/DLiPC/CHOL20\% & 290-350 & 1 & 148/148 & 74 \\
DPPC/DLiPC/CHOL30\% & 290-350 & 1 & 140/140 & 120 \\
\end{tabular}
\label{tab:MDlist}
\end{table*}
\begin{table*}[htbp]
\caption{Ensembles and force constants used during the equilibration steps on head group positions and chain dihedrals.}
\begin{tabular}{ccccc}
step & time / ps & ensemble & P atom z pos. (kJ/mol) & chain dih. ang. (kJ/mol) \\
\hline
1 & 25 & NVT & 1000 & 1000 \\
2 & 25 & NVT & 1000 & 400 \\
3 & 25 & NVT & 400 & 200 \\
4 & 100 & NpT & 200 & 200 \\
5 & 100 & NpT & 40 & 100 \\
6 & 100 & NpT & 0 & 0 \\
\end{tabular}
\label{tab:equilibration_protocol}
\end{table*}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.9\textwidth]{./figures/eplpl_compare_s_schemes_sums_bar.pdf}
\caption{Sum of variances per order parameter bin (see figure \ref{fig:s_scheme_var}) weighted by the respective order parameter distributions for the S$_{min}$ and S$_{avg}$ scheme.}
\label{fig:s_scheme_var_bar}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.9\textwidth]{./figures/DeltaEplpl_Nc_Smin_Tdep.pdf}
\caption{Pair interaction difference between energies from simulations of 330~K and 350~K at CHOL concentrations of 0, 10, 20, and 30~mol\%.}
\label{fig:deltaEplpl_Tdepscheme}
\end{figure}
\end{suppinfo}
\section{Introduction}
\input{./documents/introduction}
\section{Computational details}
\input{./documents/methods}
\section{Results and discussion}
\input{documents/results/bilayer_geometry}
\input{documents/results/pl_interactions}
\input{documents/results/chol_interactions}
\input{documents/results/compare_with_MD}
\section{Conclusion}
\input{documents/conclusion}
\section{Acknowledgment}
\input{documents/acknowledgment}
\typeout{}
|
\section{Introduction \label{sec:Introduction}}
Nowadays, streaming data from multiple data sources or streams have become more and more common in public health surveillance. Rapid \textit{hotspot} detection for the streaming data across different spatial regions over time is important and can provide valuable information to the decision-makers to mitigate the risk as soon as possible. In this paper, we define the \textit{hotspot} as the structured outliers that are persistent after a certain time point. Besides, we also assume that full observation is not possible due to the limited sampling resources.
A motivating example for this research is to monitor the number of COVID-19 confirmed cases for different spatial regions in the United States. Monitoring the entire population is not possible due to the limited testing resources or labels. Here, we would like to focus on adaptively distributing the testing resources such that the hotspot can be detected as soon as possible. Please see Section \ref{sec:Motivation} for a more detailed description of the COVID-19 monitoring problem.
Generally speaking, hotspot detection in spatio-temporal data can be treated as a change-point detection problem. Take the COVID-19 monitoring as an example. There are three types of challenges for the hotspot detection problem: 1) High-dimensionality: the number of sites or locations to be monitored is often quite large. 2) Spatial sparsity of the hotspot: the affected counties are often quite sparse in the high-dimensional space. It implies that the counties undergoing significant changes in the number of confirmed cases should be small at certain time points. 3) Temporal consistency: the hotspots should last for a reasonably long period of time.
Besides the aforementioned challenges, there are also additional challenges in sequential change detection related to the limited sensing resource as follows: 1) Lack of access to the fully observational data: existing research for high dimensional sequential sparse change detection focuses on a fully observable process, i.e., at each sampling time point, all the variables can be observed for analysis. However, in reality, it is infeasible to acquire full measurements for the entire population. 2) Partial sampling with count data observation: existing research often focuses on the continuous observational data and often assumes that a location is either observed or not observed. However, in many examples, such as the COVID-19 monitoring case, the decision variables are not binary, which implies that you can distribute more test kits to a particular region to obtain a more accurate estimation of the percentage of the infected person. In such cases, the distribution of the sampling resources is much more complicated than simply deciding whether a region should be observed or not. 3) Adaptive decision-making to balance the exploration and exploitation: On one hand, we would distribute more tests in the particular region that has been detected with a high infection rate. On the other hand, other unobserved regions may be potentially at risk as well, and some testing resources should be distributed in these regions as well.
Overall, the trade-off between exploration and exploitation is a fundamental challenge in multi-arm bandit (MAB) and reinforcement learning. Recently, Such a balance is also an important aspect to be considered in adaptive resource allocation problems for change point detection, e.g., in \cite{xu2021optimum, zhang2019partially, guo2020partially, zhang2020bandit}. Take Figure \ref{figure-US infection rate} as an example, the infection rates in Florida and Louisiana were higher than in other states on Sep 13, 2020. However, the resource allocation algorithm should distribute enough samples randomly to all states to discover such a fact (i.e., exploration) before distributing more tests to these two states specifically (i.e., exploitation) for final confirmation and change point detection.
\begin{figure}
\centering
\includegraphics[scale=.6]{US_Infection_Rate.png}
\caption{Infection Rate on Sep 13, 2020 in the United States}
\label{figure-US infection rate}
\end{figure}
This paper aims to develop an adaptive resource allocation method to automatically distribute the number of testing resources into multiple regions for the quickest change point and hotspot detection. More specifically, we will focus on the case where the data follows a binomial distribution, where the total number of tests for each region $i$ (i.e., $n_i$) depends on the number of testing resources distributed to region $i$. Overall, we propose to borrow the idea from the Upper Confidence Bound (UCB) method in solving the MAB problem to automatically distribute the testing resources to balance the exploration and exploitation. Bayesian statistics are used to update the estimation of the infection rate online with the consideration of uncertainty. A "reward" function is proposed based on the CUSUM statistics such that the change point can be identified as soon as possible.
Intuitively, the algorithm should work adaptively. Typically, the testing resources should be distributed equally to different regions in the early stages (i.e., exploration). If some regions with potentially high infection rates are observed, more testing resources should be distributed in these regions for confirmation (i.e., exploitation).
Furthermore, for resource allocation, we will borrow the idea from MAB. MAB aims to sequentially allocate a limited set of resources between competing "arms" to maximize their expected gain, where each arm’s reward function is unknown. MAB provides a principled way to balance exploration and exploitation. To apply MAB for change point detection under the sampling constraint, we propose to use MAB to combine the CUSUM statistics into the reward function in the MAB problem. However, different from \cite{zhang2020bandit, guo2020partially}, the "arm" does not refer to whether a single sensor should be observed or not, but the combinations of the number of testing resources (e.g., testing kits) should be distributed to each particular region. Finally, an integer programming algorithm is proposed to optimize such reward function for resource allocation.
The rest of the paper is organized into the following parts. In Section \ref{sec:Motivation}, we describe the motivation example of the problem in detail.
In Section \ref{sec:Literature-Review}, we review some related methods in change point detection. In Section \ref{sec:Proposed-Methodology}, we introduce the formulation and algorithm of the proposed method in detail. In Section \ref{sec:Simulation}, we use the simulation to study the behavior and some properties of the algorithm. A threshold is also generated to be applied to the case study. In Section \ref{sec:Case Study}, we study the case of county-level daily positive cases in Washington State (WA). Section \ref{sec:Conclusion} summarises the result and contributions of this study.
\section{Motivation Example}\label{sec:Motivation}
Since the initial outbreak of the novel coronavirus in early January 2020, the COVID-19 pandemic has rapidly spread across the world and brought enormous disruption to the economy and society. Various emergency measures, such as social distancing, school closures, and economic shutdowns, have been taken by different countries to control the first wave of the pandemic. To balance saving lives and the economy, a condition-based phased approach has since been adopted in the USA, where the strictness of public health measures is set to adapt according to the present epidemic condition. The success of such a phased approach critically depends on the accurate assessment of the current and near-future status of the pandemic.
Much recent research has been focusing on modeling and prediction of the spread of the COVID-19 case report data. For example, \citet{JIANG2020} studied the pandemic from a statistical point of view. The piece-wise linear trend model has been applied to model the positive case growth and the self-normalization-based method was applied to detect the change points and predict cases in the future.
\citet{tariq2020real} monitored pandemic data in Singapore and estimated the effective reproduction number $R_t$. They also took advantage of different local clusters and some international transmission. Despite various studies to forecast the positive cases and deaths, \citet{luo2020predictive} pointed out that it would be hard to predict the future of the pandemic. The direction of the mutant of a virus is random and, thus, hard to predict whether it will be more contagious or not. For example, the Delta and Omicron variant of COVID-19 caused a remarkable increase of positive cases in the US starting in June 2021 and Jan 2022.
Despite the prediction and modeling of the COVID-19 spread, many recent works focus on developing a monitoring framework to detect the pandemic outbreak while taking advantage of a more scientific resource allocation method to distribute the test kits and detect the outbreak as soon as possible.
\citet{scobie2021monitoring monitored and demonstrated the importance of the vaccine in terms of the reduction of the incidence rate ratios (IRRs).}
\citet{astley2021global} used the data from The University of Maryland Global COVID Trends and Impact
Survey (UMD-CTIS) in cooperation with Facebook to monitor the pandemic.
In general, this algorithm can monitor the COVID-19-related daily spatio-temporal cases with limited testing resources, and in this paper, we call this problem the adaptive partially observed spatio-temporal hotspot detection problem.
However, most of these works assume that the data is passively collected but do not focus on how to adaptively distribute the testing resources. \citet{chatzimanolakis2020optimal} discussed the optimal resources allocation problem in a different stage to achieve maximum information gain.
To better understand the COVID-19 status, different types of testing resource is typically distributed in different regions. Centers for Disease Control and Prevention (CDC) has classified the testing for COVID-19 into the following two categories: 1) \textbf{Diagnostic testing} is intended to identify current infection in individuals and is performed when a person has signs or symptoms consistent with COVID-19, or is asymptomatic, but has recently known or suspected exposure to COVID-19. 2) \textbf{Screening tests} are recommended for unvaccinated people to identify those who are asymptomatic and do not have known, suspected, or reported exposure to COVID-19. Screening helps to identify unknown cases so that measures can be taken to prevent further transmission.\cite{covid192022}
Overall, the screening test is very useful in testing employees in a workplace setting or universities, testing a person before or after travel, or randomly distributed test in some underdeveloped areas to identify unknown cases so that measures can be taken to prevent further transmission. In this paper, we will focus on hotspot detection for the case report data with limited screening test resources. For simplicity, we assume that the screening test is the only available source of information, and limited testing resource is available.
In this paper, we will use the real COVID-19 test report data from Johns Hopkins University Center for Systems Science and Engineering (JHU CSSE) (\citet{dong2020interactive}). The dataset is available on \url{https://github.com/CSSEGISandData/COVID-19}.
More specifically, we will use the confirmed COVID-19 cases from all 39 counties in WA.
The source of daily positive cases in WA is the Department of Health (\url{https://www.doh.wa.gov/Emergencies/COVID19}). The time-series data is updated daily around 23:59 (UTC). We will use the data from Jan 23, 2020, to Sep 13, 2020, a total of 235 days, as an illustration. On each day, the confirmed cases are recorded in all counties in the United States. We aim to identify hotspots in such a region and see if we can discover the hotspot with a much smaller number of screening test resources.
In this study, we will focus on the distribution of the number of screening testing resources for efficient hotspot detection. With this technique, we are able to conduct a real-time online monitoring process and actively emphasize the regions potentially at risk.
First, we showed the infection rates in some example counties in Figure \ref{Total-WA-cases}. From this figure, we can see that the positive rates of Yakima County started to rise around day 70, and increased dramatically. Furthermore, we have added a figure of the proportion of infected cases in some of the time points during the spread of the COVID-19 data in Figure \ref{fig:data-summary}. We can see that the infection rates on day 100 are not very high for all counties. The infection rates in some of the counties started to rise.
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth]{WA_infection_rates.png}
\caption{Example of infection rates in WA}
\label{Total-WA-cases}
\end{figure}
\begin{figure}[H]
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=\linewidth]{Proportion-100.png}
\caption{Infection proportion on day 100: May 1 2020}
\label{Proportion-Example-100}
\end{subfigure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=\linewidth]{Proportion-200.png}
\caption{Infection proportion on day 200: Aug 9 2020}
\label{Proportion-Example-200}
\end{subfigure}
\caption{Comparison of county-wise infection rate in WA on different dates}
\label{fig:data-summary}
\end{figure}
\section{Literature Review \label{sec:Literature-Review}}
In this Section, we will present a literature review on the change point detection methodology that can be potentially applied to the proposed spatio-temporal problem, including change point detection on the high dimensional data stream, change point detection with sampling control, and change-point detection on count data.
\subsection{Change Point Detection for multiple data streams}
Originally, the change point detection is first proposed to detect some sudden changes in the \textit{univariate data stream} with the shortest detection delay. The CUSUM statistics is proposed by \citet{page1954continuous, page1955test} to minimize the supremum average detection delay \cite{moustakides2009numerical}. Then, the Shiryaev\textendash Roberts procedure was introduced by \citet{shiryaev1961problem} in a Bayesian decision framework. Please see \cite{truong2020selective} for a review paper of the existing change point detection methods. However, as mentioned, these methods cannot be applied to problems with multiple data streams.
To extend the change-point detection method for \textit{multiple data streams}, \citet{mei2010efficient} extended the CUSUM statistics to multiple data streams by the summation of the CUSUM statistics for each data stream. \citet{xie2013sequential} compared the mixed procedure method they developed with some generalized log-likelihood ratio-based procedures to detect the change point in multiple but unknown data streams under the assumption that the data are normally distributed with known variance, and the change is a mean shift with unknown post-change mean. There are some works to extend these methods to multiple change-point detections with abrupt changes, such as \cite{zhang2021dynamic,avanesov2018change}, which considered the functional subspace learning and exhaustively search of the change point. \citet{enikeeva2019high} studied the change point detection when the number of dimensions and the length of each data stream goes to infinity based on the maximum of scan statistic and linear test statistic. However, these methods typically assume that the data follow a normal distribution or the data is fully observed.
\subsection{Change Point Detection for Partially Observed Data with Sampling Control}
Another line of research focuses on detecting anomalies for the \textit{spatio-temporal data}. For example, \citet{das2019spatio} applied some tests to study the change point of winter maximum, minimum, and average temperature from $6$ meteorological stations in 102 year time period. \citet{chen2020s} proposed a new $S^3T$ statistic to detect change points in spatio-temporal data with mean shift or covariance structure change. \citet{zhao2019composite} developed a likelihood-based algorithm to solve the multiple change detection in non-stationary spatio-temporal processes by splitting the data into piecewise stationary processes. \citet{knoblauch2018spatio} developed a Bayesian online change-point detection algorithm to model the process between change points in spatio-temporal data. \citet{zhang2018multiple} studied the change point detection problem when there are system-inherent variations, and sensor feature differences and is able to identify the local change point in each cluster. Again, these methods focus on the fully observed data and do not provide a method to iteratively update the sampling resources.
To generalize the change point detection for \textit{partially observed data}, some works focus on the change point detection by assuming that only partial data streams at each time point can be observed due to limited sampling resources. Such setting has been studied in social network monitoring \cite{farajtabar2015back}, biological network \cite{raue2009structural}, and manufacturing systems \cite{liu2015adaptive}. To deal with partially observed data, \citet{xie2012change} developed a change point detection method by combining subspace manifold learning, and multiscale analysis to solve the missing data problem. \citet{dubey2021online} studied the online change point detection problem in the network where the pattern of the missing data of the network is heterogeneous. \citet{corradin2022bayesian} studied multiple change-point detections in multivariate time series with missing values. The conditional distribution of each data point given the rest data is derived to handle the missing value problem. However, these methods do not actively control which subsets of data to measure, which hinders the detection performance.
Recently, many existing works have focused on the change-point detection problem with sampling control, where the sensor(s) to be observed in each epoch are determined adaptively based on the current and historical observation. For example, \citet{liu2015adaptive} proposed to combine CUSUM statistics and a top-$r$ strategy to select the data streams to be observed for online change detection. \citet{xu2021multi} studied the problem when there is only one sensor to be observed at each time and an adaptive sampling strategy based on the CUSUM statistic and showed that it is asymptotically optimal in the sense of minimizing the detection delay. \citet{zhang2019partially} developed an online learning algorithm for anomaly detection for independent data streams using the combinatorial MAB approach. \citet{zhang2020bandit} studied the change point detection under sampling control by the combination of Thompson Sampling and a top-$r$ local Shiryaev-Robert statistic to raise a global alarm.
\citet{guo2020partially} proposed a Bayesian spike-and-slab prior distribution to decompose the data into the smooth background and sparse anomaly and focus on identifying the anomaly by selecting the sensors to observe using Thompson sampling.
However, all aforementioned methods focus on change detection under the continuous observation data and assume each sensor is either fully observed or not, which cannot be used for more complex resource allocation problems for anomaly detection.
\subsection{Change Point Detection on Counting Data}
Much existing research has been focusing on extending the change point detection method for \textit{count data or binary data}. For example, \citet{hinkley1970inference} studied the change point detection problem for binary variables based on the asymptotic distribution of the likelihood ratio test statistics. \citet{yu2013change} developed the CUSUM statistics to study the change point detection problem in the binomial thinning process and applied it to the 2001-2012 influenza data.
\citet{jiang2011weighted} proposed a weighted CUSUM control chart for monitoring Poisson processes with varying sample sizes.
\citet{gut2002truncated} studied the change point detection problem in renewal counting data with the discrete-time frame, where the term `partially observed' meant the discrete-time frame. \citet{ellenberger2021exact} developed a binary segmentation method to solve the change point detection problem when the sample size of the binomial distribution is small. However, these methods typically assume that the sample size of the binomial distribution is pre-determined, which cannot be used to solve the resource allocation problem.
\section{Proposed Methodology \label{sec:Proposed-Methodology}}
In this Section, we will explore the adaptive sampling method on discrete data with the binomial distribution.
In Section \ref{subsec:Problem-Formulation}, we will review the overall problem formulation. In Section
\ref{subsec:ProposedAlgorithm}, we will introduce the procedure of the proposed algorithm. In Section
\ref{subsec:Tuning-Parameter-Selection}, we will introduce the tuning parameter selection procedure for the proposed algorithm.
\subsection{Problem Formulation \label{subsec:Problem-Formulation} }
We assume that the data is $X_{k,t}$, where $t=1,2,\ldots$ is the time index and $k=1,2,\ldots,K$ is the region index.
In the COVID-19 monitoring example, $X_{k,t}$ can be the number of newly observed confirmed cases in region $k$ and time $t$.
We further assume that all $X_{k,t}$ prior-change follows a binomial distribution $f_k^0$ as $\text{Bin}(c_{k,t},p)$, with the number of test $c_{k,t}$ and the probability $p$. Furthermore, we assume that after the change $t>t_0$, only one hotspot region (i.e., $k \in \mathcal{H}$) is impacted by the change. In contrast, we assume that the regions affected by the changes follow another post-change distribution $f_{k}^1$ where $f_{k}^1$ is another binomial distribution as $\text{Bin}(c_{k,t},q)$, with the number of tests as $c_{k,t}$ with the probability $q$, as shown in Eq. (\ref{eq: Binomial}). The superscript $0$ means in control, and the superscript $1$ means out of control.
\begin{equation}
X_{k,t} \sim \begin{cases}
\text{Bin}(c_{k,t},p), & t\le t_0 \text{ or } t > t_0, k \notin \mathcal{H} \\
\text{Bin}(c_{k,t},q), & t> t_0, k \in \mathcal{H}
\end{cases}
\label{eq: Binomial}
\end{equation}
Finally, we assume that there are limited sampling resources. The constraint we assume is that limited testing kits $C$ are distributed to all the regions at each time $t$ with the constraint $\sum_k c_{k,t}=C$.
For simplicity, we assume that the actively distributed test kits are the only source of observation data.
The objective of this paper is to sequentially distribute the testing resources $c_{k,t}$ for region $k$ at time $t$ for faster change point detection.
\subsection{Proposed Adaptive Resources Allocation CUSUM for Binomial Count data (ARA-CUSUM) Method \label{subsec:ProposedAlgorithm}}
In this Section, we first introduce the overall step of the algorithm and then discuss three major components of the proposed methods as follows: monitoring statistics updates, planning for resource allocation, and change-point detection decision. The overall framework is shown in Figure \ref{fig: flow chart}.
The detailed steps are discussed as follows.
\begin{enumerate}
\item \textbf{Monitoring statistics update}: We propose to update the monitoring statistics at each time $t$ for the quickest change-point detection. This procedure will be discussed in detail in Section \ref{subsubsec:Statistics}.
\item \textbf{Planning for adaptive sampling}: We would like to distribute the testing resource for the next time point based on previous results by the UCB bandit algorithm. The detailed planning formulation and the optimization algorithm are discussed in Section \ref{subsubsec:Planning} and Section \ref{subsec:Optimization-algorithm}.
\item \textbf{Change point detection decision}: Finally, the alarm is raised when the largest CUSUM statistics for all regions exceeds some threshold. Finally, the corresponding region that triggers such change can also be used for hotspot identification.
\end{enumerate}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{flow_chart.png}
\caption{Flow chart for the algorithm}
\label{fig: flow chart}
\end{figure}
\subsubsection{Monitoring Procedure \label{subsubsec:Statistics} }
In this Section, we will first introduce the definition of CUSUM statistics and then derive the monitoring procedure for online change detection.
\begin{defn}
Suppose in a sequential data, $x_n$ is the score of $n$th observation. The CUSUM statistics \citet{page1954continuous} is defined by
\begin{equation}
S_0=0, ~~ S_n = \max\{0, S_{n-1}+x_n\},(n\ge 1)
\end{equation}
\end{defn}
\citet{moustakides1986optimal} suggested that the CUSUM statistic is mini-max optimal in the sense of detection delay \cite{lorden1971procedures}.
\begin{equation}
D(T)=\sup_{\nu\ge 0}\text{ess}\sup \mathbb E_\nu [(T-\nu)^+|X_1,\ldots,X_\nu]
\end{equation}
with the constraint of in-control average run length $\mathbb E_\infty[T]\ge\gamma>0$.
In change point detection, the score $x_n$ for each observation can be the log-likelihood ratio, which is $x_n=\log\frac{f^1(X_n)}{f^0(X_n)}$, where $X_n$ is the $n$th observed data, $f^0$ is the pdf of $X_n$ before change and $f^1$ is the pdf of $X_n$ after change.
In the proposed formulation, $X_{k,t}$ is defined as the observed positive cases in time $t$ and region $k$. Inspired by the CUSUM statistics definition, the CUSUM statistics $W_{t}^{k}$ is formulated as:
\begin{equation}
W_{k,t}=\max\{W_{k,t-1},0\}+\log\frac{f^1_{k}(X_{k,t})}{f^0_{k}(X_{k,t})},\label{eq-CUSUM}
\end{equation}
where $f^0_{k} \sim \text{Bin} (c_{k,t}, p)$ and $f^{1}_{k} \sim \text{Bin}(c_{k,t}, q)$ represent the distribution of null and alternative hypothesis, respectively. The initial value $W_{k,0}$ is set to be $0$ for all $k$. Following the CUSUM statistics definition, we define the CUSUM statistics $W_{k,t}$ in the proposed formulation for each county $k$ and time $t$ in Proposition \ref{prop-CUSUM}.
\begin{prop}
\label{prop-CUSUM} Suppose that at time $t$, the number of tests distributed at region $k$ is $c_{k,t}$,
the number of positive cases collected is $X_{k,t}$. For each region $k$, for $k=1,2,\ldots,K$, the
CUSUM statistics $W_{k,t}$ can be computed for each time $t$ and region $k$ can be derived as
\begin{equation}
W_{k,t}=\max\{W_{k,t-1},0\}+c_{k,t}\log\frac{1-q}{1-p}+X_{k,t}\log\frac{q(1-p)}{p(1-q)}\label{eq-Update-Single}
\end{equation}
\end{prop}
The proof is detailed in Appendix \ref{proof-prop-CUSUM}.
Here, we will raise the alarm when the largest of the CUSUM statistics $\max_{k} W_{k,T}$ exceeds some threshold $c$ at some time $T$. Finally, we will discuss the procedure of selecting the threshold $c$ in Section \ref{subsec:Tuning-Parameter-Selection}.
\subsubsection{Planning Objective Function for Resource Allocation \label{subsubsec:Planning}}
In this Section, we present an efficient algorithm for resource allocation at the next time point
for hotspot detection. Overall, a good resource allocation algorithm for change detection should achieve a good balance
between exploration and exploitation. Here, exploration in the problem formulation implies the test
kits should be distributed to the regions that were not measured before, and exploitation means
that when the change occurs, the planning will focus more on the region with the change for the quickest change detection.
To better tackle the exploration-exploitation trade-off problem, we apply the UCB as the target function to be optimized. \citet{auer2002using} suggested that the UCB criteria for the MAB problem maximize the final reward.
This method takes advantage not only of the mean reward $\mu$ but also the standard deviation of the reward $\sigma$.
In this case, we should not only focus on the region leading to the larger reward but also the action with larger variability, which has the potential to lead to a larger reward. Specifically, in this hotspot detection application, for the true hotspot region with a larger infection rate, we may want to continue to distribute more test kits to the particular region the next day to encourage further exploitation. However, when a region has not been observed enough in previous days, it typically has a larger variance, and some testing resources will be distributed in the regions with large variability the next time to encourage further exploration.
Given that the log-likelihood ratio $\log\frac{f^1_{k}(X_{k,t})}{f^{0}_k(X_{k,t})}$ for change point detection is often a good quantification of how likely the change may happen in a particular region $k$, here we propose to use the sum of UCB of the likelihood ratio in the CUSUM statistics as the reward function. For the convenience of notation, we denote the log-likelihood ratio in the CUSUM statistics in each region $k$ by
\begin{equation}
\Delta r_{k,t}=c_{k,t}\log\frac{1-q}{1-p}+X_{k,t}\log\frac{q(1-p)}{p(1-q)}\label{eq-Increment-Single}
\end{equation}
Therefore, the UCB function of the summation of the likelihood ratio statistics can be expressed as
\begin{equation}
f(\mathbf{c}_{t})=\sum_{k=1}^{K}\left[E[\Delta r_{k,t}]+\sqrt{Var(\Delta r_{k,t})}\right]\label{eqUCB}
\end{equation}
Here, the goal is to find the best resource allocation planning method to optimize the resource distribution $\mathbf{c}_{t}=(c_{1,t},\ldots,c_{K,t})$ at time $t$. However, one of the challenges is that the $X_{k,t}$ is the number of positive testing results at time $t$, which is a random variable at time $t$.
In Section \ref{subsubsec:Bayesian},
we will derive the Bayesian updating procedure to update the posterior distribution of $X_{k,t}$, which can be used estimate the $E[\Delta r_{k,t}]$ and $Var(\Delta r_{k,t})$.
\subsubsection{Bayesian Updating of $X_{k,t}$ \label{subsubsec:Bayesian} }
To estimate the distribution of $X_{k,t}$, we assume that it follows a binomial distribution of $\text{Bin}(c_{k,t},p_{k})$, where $p_{k}$ is the true infection rate to be estimated. Therefore, an online estimation of the rate is needed for better planning and resource allocation. Here, we propose a weighted Bayesian update method to estimate the infection rate in Proposition \ref{prop-posterior}.
The reason to consider such a weighted updating procedure is that the samples in recent time points are typically more likely to represent the current state of the system and should be used to better detect the change. Therefore, more weight should be put into the recent time points. For simplicity, we propose to use a exponential decayed weight $\lambda_t = w^{T-t}$ for each sample $X_{k,t}$.
Finally, the posterior distribution $p_{k,t}$ can be derived as
\begin{equation}
P(p_{k}|X_{k,1},\cdots, X_{k,t}) \propto P(p_{k}) \prod_{t=1}^T P(X_{k,t}|p_{k})^{\lambda_t},
\end{equation}
where $P(p_{k})$ is the prior distribution of the infection rate, which is assumed to follow a Beta distribution as $\text{Beta}(a,b)$.
$P(p_{k}|X_{k,1},\cdots, X_{k,t})$ is the posterior distribution of $p_k$, which can be derived in Proposition \ref{prop-posterior}.
\begin{prop}
\label{prop-posterior} Suppose that the data $(c_{k,t}, X_{k,t})$ of the $k^{th}$ region until time $T$ is collected, where $c_{k,t}$ is the number of total tests distributed to region $k$ and time $t$ and $X_{k,t}$ is the daily positive cases. We further assume the prior distribution of $p_{k}$ is $\mathrm{Beta}(a,b)$,
the weighted posterior distribution of $P(p_{k}|X_{k,1},\cdots, X_{k,T})$ can be derived as
\begin{equation}
\mathrm{Beta}(a+\sum_{t=1}^{T}X_{k,t}w^{T-t},b+\sum_{t=1}^{T}(c_{k,t}-X_{k,t})w^{T-t})\triangleq \mathrm{Beta}(\alpha_{k,t},\beta_{k,t})\label{Distribution}
\end{equation}
\end{prop}
The proof is given in Appendix \ref{proof-prop-posterior}.
Here, the updated posterior distribution of $p_{k}|X_{k,1},\cdots, X_{k,T}$ can be used to calculate the uncertainty in $X_{k,t}$
to derive the UCB function in the planning method.
\subsubsection{Optimization Algorithm for Planning \label{subsec:Optimization-algorithm}}
By plugging in the weighted posterior distribution of $P(p_{k}|X_{k,1},\cdots, X_{k,T})$ into the UCB reward function in (\ref{eqUCB}), we will derive the objective function to be optimized and propose an integer programming algorithm to optimize the number of test kits $c_{k,T+1}$ in each region $k$ at time $T+1$.
\begin{prop}
\label{prop-planning-single} Suppose $X_{k,t}\sim \mathrm{Bin}(c_{k,t},p_{k})$ and the posterior distribution $p_{k}|X_{k,1},\cdots, X_{k,T} \sim \mathrm{Beta}(\alpha_{k,t},\beta_{k,t})$.
The target function $f_t(\mathbf{c}_{T+1})=\sum_{k=1}^{K}\left[E[\Delta r_{k,T+1}]+\sqrt{Var(\Delta r_{k,T+1})}\right]$ can be derived as
\begin{footnotesize}
\begin{align}
f_t(\mathbf{c}_{T+1}) &= \sum_{k=1}^{K} f_{k,t} (c_{k,T+1})\nonumber\\ &=\sum_{k=1}^{K}\left[\frac{\alpha_{k,t}}{\alpha_{k,t}+\beta_{k,t}} c_{k,T+1}+\sqrt{c_{k,T+1}\frac{\alpha_{k,t}\beta_{k,t}}{(\alpha_{k,t}+\beta_{k,t})(\alpha_{k,t}+\beta_{k,t}+1)}(\frac{c_{k,T+1}}{\alpha_{k,t}+\beta_{k,t}}+1)}\right]\label{eq-Target-Independent}
\end{align}
\end{footnotesize}
\end{prop}
The proof is shown in Appendix \ref{proof-prop-planning-single}.
Therefore, Algorithm \ref{alg:planning} is proposed to determine the best resource allocation $(c_{1,T+1},\ldots,c_{K,T+1})$ at time $T+1$ as follows:
\begin{align}
\hat{c}_{1,T+1},\ldots,\hat{c}_{K,T+1} & =\arg\max f_t(\mathbf{c}_{T+1})\label{eq-Planning-Independent}\\
& s.~t.~\sum_{k=1}^{K}c_{k,T+1}\le C,c_{k,T+1}\ge0,k=1,\ldots,K\nonumber
\end{align}
\begin{figure}
\centering
\includegraphics[scale=0.5]{reward_function_example.png}
\caption{Example of reward functions with different estimated infection rate}
\label{fig: reward example}
\end{figure}
Here, we would like to give a brief discussion on the reward function $f_{k,t} (c_k)$ defined in \eqref{eq-Target-Independent}. An example of the function $f_{k,t} (c_k) $ with different infection rates are given in Figure \ref{fig: reward example}.
The expectation part and standard deviation part is larger for regions with a higher estimated infection rate. However, there is a diminishing return effect on this reward function, which implies that as more testing resources are distributed in the highly infected regions, the increase of the reward function will become smaller. At a certain point, it may be less rewarding to distribute additional tests to the highly infected region compared to the low infected region. To demonstrate, Figure \ref{fig: reward example} shows that $f_5(x)>f_k(x)$ for $k=1,2,3,4$ and $x\in[1,50]$, but $f_k'(x)>f_5'(y)$ for some $k=1,2,3,4$ and $x,y\in[1,50]$.
To optimize problem \eqref{eq-Planning-Independent}, a greedy algorithm is proposed. The idea of the greedy algorithm is to distribute the testing kits one by one to the region that can result in the largest target increment until all the tests are properly distributed. Given the concavity of $f_{t}(\mathbf{c})$, such a greedy approach can guarantee the global optimum. This property will be discussed in Section \ref{prop-convex}. The greedy algorithm is designed as Algorithm \ref{alg:planning}.
\LinesNumberedHidden{
\begin{algorithm}[t]
\caption{Optimal Resources Planning by the Greedy Algorithm}
\label{alg:planning}
\SetAlgoLined
\Parameter{$\alpha_{k,t}$, $\beta_{k,t}$, $k=1,\ldots,K$}
\KwResult{$c_{k}$, $k=1,\ldots,K$}
\textbf{Step 1:} Initialize $c_{k}=0$ for $k=1,\ldots,K$\\
\textbf{Step 2:} \For{$i=1$ \KwTo $C$}{
\textbf{Step 2.1:} \For{$k=1$ \KwTo $K$}{
Estimate the increment in region $k$: $\Delta_k=f_k(c_k+1)-f_k(c_k)$.
}
\textbf{Step 2.2:} Let $k_m=\arg\max_k \Delta_k$. $c_{k_m}=c_{k_m}+1$
}
\end{algorithm}}
Finally, the entire process can be shown in Algorithm 2. Overall, the steps of the algorithm are described here: 1) On day one, we have an initial test distribution. If the prior distribution is set the same for all regions, uniform distribution can be used. The confirmed number of $X_{k,1}$ will be collected on Day 1. 2) The collected dataset $X_{k,t}$ can be used to update the posterior distribution of the rate $p_k$ by \eqref{Distribution}. 3) Such distribution information can be plugged into \eqref{eq-Planning-Independent} to optimize the test distribution of the next day. Such a procedure can be estimated recursively until the current time $T$.
\LinesNumberedHidden{
\begin{algorithm}
\caption{Proposed ARA-CUSUM Algorithm}
\label{alg: process}
\SetAlgoLined
\Parameter{$a$, $b$, $w$, $c$}
\KwResult{$W_{k,t}$,$k=1,\ldots,K$,$t=1,2,\ldots$}
\textbf{Step 1:} Initialize $c_{k,1}$, $X_{k,1}$ and $W_{k,1}$ for $k=1,\ldots,k$\;
\textbf{Step 2:} \For{$t=1,2,\ldots$}{
\textbf{Step 2.1: Bayesian Update:}
Set $\alpha_{k,t}=a+\sum_{i=1}^{t}X_{k,i}w^{t-i}$; $\beta_{k,t}=b+\sum_{i=1}^{t}(c_{k,i}-X_{k,i})w^{t-i}$\;
\textbf{Step 2.2: Resource Allocation:} Solve $c_{k,t+1}$ in the optimization problem \eqref{eq-Planning-Independent} by Algorithm \ref{alg:planning} \;
\textbf{Step 2.3: Monitoring Statistics Update and Detection:} Update $W_{k,t+1}$ by (\ref{eq-Update-Single}) \;
\textbf{Step 2.4:} \uIf{$W_{k,t+1}>c$
for some $k$}{ Raise alarm at time $t+1$ and county $k$\; } }
\end{algorithm}}
\subsubsection{Property of the Optimization Algorithm \label{subsec:Property}}
In this subsection, we would like to prove that the greedy approach proposed in Section \ref{subsec:Optimization-algorithm} can obtain the global optimum.
\begin{lemma}\label{lemma-concavity}
The target function \eqref{eq-Target-Independent} is the sum of $K$ concave functions and is therefore concave.
\end{lemma}
The target function in each region can be expressed as $f_{k,t}(x)=a x+\sqrt{b x+c x^2}$. One can easily verify that the second order derivative of $f_{k,t}(\cdot)$ is negative when $b,c>0$.
\begin{prop}\label{prop-convex}
The algorithm \ref{alg:planning} can obtain the global optimum for the optimization problem \eqref{eq-Planning-Independent}.
\end{prop}
\begin{proof}
From Lemma \ref{lemma-concavity}, $f_{k,t}(\cdot)$ is an increasing and strictly concave function. For any fixed time $t$, we denote $g_k(x)=f_{k,t}(x)-f_{k,t}(x-1)$ be the increment of the reward function to distribute one more tests on region $k$. $g_k(\cdot)$ will be a positive decreasing function for any $k$. Then the total reward function given the test distribution $c_{1,t},\ldots,c_{K,t}$ is
$$
f_{t}(c_{1,t},\ldots,c_{K,t})=\sum_{k=1}^K \left(\sum_{x_k=1}^{c_{k,t}}g_k(x_k)\right)
$$
Now consider the increment parameter matrix
$$
\begin{pmatrix}
g_1(1) & g_1(2) & \cdots & g_1(C)\\
g_2(1) & g_2(2) & \cdots & g_2(C)\\
\vdots & \vdots & \ddots & \vdots \\
g_K(1) & g_K(2) & \cdots & g_K(C)
\end{pmatrix}
$$
In Algorithm \ref{alg:planning}, the first iteration will choose the largest value in the first column. Since each row in the matrix is decreasing, the first iteration will also choose the largest value in the matrix. Each of the remaining iterations will choose the largest value in the matrix that has not been chosen. Therefore, Algorithm \ref{alg:planning} chooses the top $C$ largest values in the matrix. And because the target function can be written by the sum of $C$ values in the matrix, Algorithm \ref{alg:planning} gives the optimal solution to the optimization problem \eqref{eq-Planning-Independent}
\end{proof}
\subsection{Tuning Parameter Selection \label{subsec:Tuning-Parameter-Selection}}
There are several tuning parameters that need to be determined for the algorithm. The key parameters are prior distribution parameter $a,b$, weight $w$ and threshold $c$. In this Section, we will discuss the effect of these tuning parameters and how to decide them.
\begin{itemize}
\item \textbf{Prior Distribution}: There are two parameters $a,b$ in the prior distribution. The prior distribution reflects the prior knowledge about the infection rate. Given that the infection rate for the null hypothesis is $p$, we can also set the ratio to be $p$, i.e., $\frac{a}{a+b}=p$, assuming no change happens. Since $p$ is a preset constant, we only need to determine $a$. When $a$ is large, the influence of prior distribution is large, and the difference in the infection rate estimation from the posterior distribution will be smaller. Thus, the difference in test distribution will not be very large. In the simulation, we will choose $a+b=0.5\times C$.
\item \textbf{Weight $w$}: The weight is used in Bayesian update to get the posterior distribution of $X_{k,t}$. Weight allows the algorithm to take not only the information from one day before but also the previous days. A large weight will make the test distribution in each region smooth across different days. Therefore, when occasionally the observed infection rate is high, the algorithm will only distribute a reasonably larger amount of testing kits on the next day rather than some extreme testing kit distribution. In the simulation, we choose $w=0.3$.
\item \textbf{Total Resources $C$}: This parameter determines the total number of tests to be distributed in all regions. Typically, $C$ is pre-determined by the resource constraint.
\item \textbf{Threshold $c$}: This parameter determines whether an alarm should be raised in each region every day. Higher threshold $c$ leads to high in-control average run length ($ARL_0$) and higher out-of-control $ARL_1$. Typically, we would like to maintain a preset $ARL_0$, which determines the threshold $c$. Overall, for a preset $C, p, q$, a simulation study can be used to determine the best $C$ with certain $p$ and $q$.
We will determine this parameter by the simulation work and use a preset average run length.
\end{itemize}
\section{Simulation \label{sec:Simulation}}
In this Section, we will start with the simulation setup in Section \ref{subsec:simulation setup} and
then discuss the result evaluation in Section \ref{subsec:result}. The balance of exploration and exploitation is discussed in Section \ref{subsec: exploration and exploitation}.
\subsection{Simulation Setup \label{subsec:simulation setup}}
In this Section, we will evaluate the proposed methodology using a simulation study. To generate the simulation dataset, we will simulate a true infection rate for each county in the WA as an example.
There are $39$ counties in WA, i.e., $K=39$.
Before the distribution change, the true infection proportion is set to be $p=0.01$ for all counties. We will compare the detection delay for different infection proportions where the out-of-control infection rates are $q=0.025,0.03,0.04,0.05$ in the first county after the distribution change. We further assume that the total number of tests is $3900$, where each county takes an average of $100$ tests per day. The weighted update parameter $w$ is set to be $0.3$ and the prior distribution parameter $a,b$ are set as the $50\%$ of the total tests to make the prior distribution as
$a=3900\times50\%\times0.01=19.5$, $b=3900\times50\%\times0.99=1930.5$.
The process of the simulation is described as follows:
\begin{enumerate}
\item The test distribution $c_{k,t}$ in each county determined from the previous iteration can be used.
\item Before the change point $t<t_0$, we will sample the observed positive cases by $X_{k,t}\sim \text{Bin}(c_{k,t},p)$.
\item After the change point $t>t_0$, we will sample the observed positive cases in the first county by the out-of-control infection rate $q$ and $X_{1,t}\sim \text{Bin}(c_{1,t},q)$. We will sample the rest counties still by $X_{k,t}\sim \text{Bin}(c_{k,t},p)$
\end{enumerate}
Finally, we will follow Algorithm \ref{alg:planning} for the proposed algorithm. The CUSUM statistics are updated, and the test distribution of the next day is determined by the proposed method or benchmark methods. When studying the average run length and detection delay, we will stop the process when the alarm is raised.
In this simulation study, we will compare the proposed algorithm with the following two benchmark methods.
\begin{enumerate}
\item \textbf{Even Distribution}: The first benchmark (denoted as Even) is a simple approach, which always evenly distributes the tests to all the counties. This even distribution will focus on exploration of the testing resources but do not have the power to exploit for a specific region.
\item \textbf{Top-R Distribution}: The second benchmark (denoted as top-R') is inspired by the Top-R strategy proposed in \cite{liu2015adaptive}. We first evenly distribute all tests evenly into $20$ batches. For in total $3900$ testing kits, it averages out $195$ testing kits per batch. We then select the top $20$ regions with the largest CUSUM statistics in the previous day. Here, we select the top $20$ regions, given it is a little more than half of the whole regions.
\end{enumerate}
\subsection{Evaluation of the Detection Delay and Precition \label{subsec:result}}
To compare this algorithm with the benchmarks, we will first need to determine the thresholds for the proposed method and benchmark methods. Here, binary search can be used to decide the optimal threshold for each benchmark method to maintain the in-control average run length (ARL) as 200 for all benchmark methods with 1000 replications.
To evaluate the proposed methods, we use the following three metrics: 1) the out-of-control average run length ($\mathrm{ARL}_1$): $\mathrm{ARL_1}$ is defined as the average detection delay after the change has happened. Overall, we aim to reduce the ARL when keeping the same $ARL_0 = 200$; 2) Detection precision (DP): DP is evaluated as the proportion in the $1000$ iterations that the alarm is raised in the correct county. Overall, we would aim to increase the DP so that we can identify the county that is accountable for the change more accurately.
3) the standard deviation of run-length (SDRL): SDRL is evaluated by computing the standard deviation with $1000$ iterations. Overall, a smaller SDRL would imply a more robust detection with smaller variability for detection delay.
These three metrics for the proposed method and the benchmark methods (namely the evenly distributed and the Top-R distributed) are evaluated in Table \ref{table-result-1}.
From Table \ref{table-result-1}, we can conclude that ARL is the smallest for the different $q$ values. Especially, when $q$ is small (i.e., 0.025 and 0.03), the detection delay can even be greatly reduced (i.e., 7.893 for "proposed", 14.885 for "Even", and 10.385 for "Top-R" for $q=0.025$). For larger $q$, the differences between the proposed methods and the second best method, such as "Top-R" are similar, given it becomes easy to detect all the methods. However, the proposed method still outperforms others in this case. Finally, we find that the SDRL of the proposed method is also the smallest, especially when $q$ is small. This implies that the proposed methods can also make the detection much more robust by reducing the standard deviation of the detection delay.
\begin{table}
\caption{ The result of detection delay and detection precision when $ARL_0 \approx 200$}
\centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Metric & Methods & $q=0.025$ & $q=0.03$ & $q=0.04$ & $q=0.05$\tabularnewline
\hline
\multirow{3}{*}{$\mathrm{ARL_1}$} & Proposed & $\mathbf{7.893}$ & $\mathbf{4.958}$ & $\mathbf{3.388}$ & $\mathbf{2.863}$\tabularnewline
\cline{2-6}
& Even & $14.885$ & $7.93$ & $4.408$ & $3.299$\tabularnewline
\cline{2-6}
& Top-R & $10.385$ & $5.688$ & $3.443$ & $2.891$\tabularnewline
\hline
\multirow{3}{*}{DP} & Proposed & $\mathbf{0.918}$ & $\mathbf{0.932}$ & $0.938$ & $0.939$\tabularnewline
\cline{2-6}
& Even & $0.892$ & $0.926$ & $\mathbf{0.939}$ & $\mathbf{0.943}$\tabularnewline
\cline{2-6}
& Top-R & $0.904$ & $0.918$ & $0.936$ & $0.937$\tabularnewline
\hline
\multirow{3}{*}{SDRL} & Proposed & $\mathbf{4.67}$ & $\mathbf{2.47}$ & $\mathbf{1.31}$ & $\mathbf{1.01}$\tabularnewline
\cline{2-6}
& Even & $12.81$ & $5.48$ & $2.28$ & $1.45$\tabularnewline
\cline{2-6}
& Top-R & $10.556$ & $3.72$ & $1.63$ & $1.12$\tabularnewline
\hline
\end{tabular}
\label{table-result-1}
\end{table}
\subsection{Exploration and Exploitation}\label{subsec: exploration and exploitation}
As formerly mentioned, a good change point detection with a sampling control algorithm should balance both exploration and exploitation behavior. In this simulation, exploration means that we will need to explore all counties, given that any county can undergo large changes at certain times. Exploitation means that we will need to distribute more resources to the counties that have been detected or have the potential to be the hotspot. Figure \ref{Simu_Distribution} and Figure \ref{Simu_County} give a more intuitive understanding of the exploration and exploitation behavior of this proposed algorithm from the following simulation study setup. In the first $500$ days, all the counties are in control with the same infection rate $p=0.01$. From day $501$ to day $1000$, only the first county, Adams County is out of control with the infection rate $q=0.05$, while the rest counties stay the same with $p=0.01$. Figure \ref{Simu_Distribution} shows the density and the boxplot of the number of tests in control. Figure \ref{Simu_County} shows the median number of test kits in each county under different conditions, in control and out of control.
Here are some behaviors of the algorithm.
When there is no distribution change, the distributions of the number of tests in each county are similar, see Figure \ref{Simu_Distribution} and Figure \ref{Simu_County_IC}. This shows that the proposed method is able to explore all the counties undergoing potential changes.
However, if a county has a higher infection rate on the current day, it is likely to have higher tests the next day. If the sampled positive proportion is small, the increment of CUSUM statistics will be likely to be $0$. If the distribution change happens in that county, the CUSUM test statistic will be likely to grow dramatically and we will continue to focus on that county further. See Figure \ref{Simu_County_OC}, when Adams County is out of control, the median number of the distributed tests of Adams County is significantly larger than the number of other counties. In this case, the CUSUM test statistic of that Adams County is more likely to increase dramatically, which is a good behavior of exploitation.
\begin{figure}[H]
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=\linewidth]{In_control.png}
\caption{Test Distributions Density In Control}
\label{Simu_Distribution_IC}
\end{subfigure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=\linewidth]{boxplot_IC.png}
\caption{Boxplot of Number of Tests In Control}
\label{Simu_boxplot_IC}
\end{subfigure}
\caption{Visualization of Test Distributions In Control}
\label{Simu_Distribution}
\end{figure}
\begin{figure}[H]
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=\linewidth]{usmap_in_control.png}
\caption{Median Number of Tests In Control}
\label{Simu_County_IC}
\end{subfigure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=\linewidth]{usmap_out_of_control.png}
\caption{Median Number of Tests Out of Control}
\label{Simu_County_OC}
\end{subfigure}
\caption{Median Number of Tests}
\label{Simu_County}
\end{figure}
\section{Case Study}\label{sec:Case Study}
In this case study, we would like to evaluate the proposed methods to the real case study. The data we are working on is the county-level
daily positive cases, and the parameters are $K=39$, $p=0.01$, $q=0.05$, $a=19.5$, $b=1930.5$ and
$w=0.3$. The average number of tests in each county is $100$.
\subsection{Detection of the Out-of-control County}
The alarms were raised in Yakima county on Jun 19, 2020. The comparison figure of the test statistics is in Figure \ref{Test-Statistics}. Starting from day $135$, the infection rate of Yakima county started to show a dominating trend over the other counties. Therefore, the algorithm emphasizes on Yakima county to seek for potential risk. But the true infection rate is not large enough to produce a positive increment to the CUSUM statistics. Thus, we will see this oscillating trend. When the infection rate of Yakima county increases consistently, the CUSUM statistics will increase dramatically. Finally on June 21, 2021, the CUSUM statistics in Yakima County exceeds the threshold of $6.5$ and the alarm is raised.
\begin{figure}[H]
\centering
\includegraphics[width=0.6\linewidth]{Case_Study.png}
\caption{Comparison of the Test Statistics}
\label{Test-Statistics}
\end{figure}
\subsection{Behavior of Sampling Resource Distribution}
The plots in Figure~\ref{Test-Distribution-Case} are examples of the test distribution of this case study before and after the alarm is raised. This shows similar behavior as the simulation study. The test distribution before the alarm is close to an even distribution to explore all the counties for the risks. The test distribution after the alarm focuses on Yakima county and consistently monitors this county to show good behavior of exploitation.
\begin{figure}[H]
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{Test_Distribution_100.png}
\caption{Example of test distribution in control}
\label{Test-Case-IC}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{Test_Distribution_200.png}
\caption{Example of test distribution out of control}
\label{Test-Case-OC}
\end{subfigure}
\caption{Example of test distribution in WA}
\label{Test-Distribution-Case}
\end{figure}
\section{Conclusion}\label{sec:Conclusion}
In conclusion, the proposed method is able to handle the challenge of exploration and exploitation trade-offs and limited resources. In the simulation work, we show that the adaptive sampling method can beat the two benchmarks either in the sense of detection delay or detection precision. In the case study in WA, the alarm was raised in Yakima county. This algorithm works well when there is only one out-of-control region. When there are multiple out-of-control regions, this algorithm will focus on the first region with a higher out-of-control risk. If there are multiple potential regions at risk, one can start the algorithm from the beginning again when an alarm is raised and remove the alerted region. In the future, we can address this concern and detect multiple out-of-control regions in a time series manner.
Some other future work includes taking geometric distance and graphing into consideration and solving the detection problem when there are multiple out-of-control regions.
\bibliographystyle{plain}
|
\section{Introduction}
Transformer-based large pretrained Language Models, such as GPT3 and T5~\cite{transformer, gpt3, t5}, have been widely used as few-shot learners in many NLP tasks. Recent work even finds these models can achieve state-of-the-art performance in arithmetic and symbolic reasoning~\cite{ scratchpad,cot}.
Although these models exhibit surprisingly impressive capabilities in complex arithmetic reasoning tasks, such as MultiArith~\cite{multiarith} and GSM8k~\cite{gsm8k}, it has also been pointed out that they tend to make certain calculation errors and perform significantly worse when the number of math operations increases in equations~\cite{cot}. \citet{gpt3} find that GPT3 displays strong proficiency in 2-digit arithmetic addition, but struggles in arithmetic addition on numbers with more than three digits. \citet{investigating2021} also observe that the fine-tuned T5 model can not correctly add or subtract arbitrarily long numbers. Larger models might perform better on the testing data, but worse on numbers that are longer than the training data (out-of-distribution, OOD) \cite{investigating2021}. However, even with the largest T5 model they experimented, the out-of-distribution (OOD) accuracy is not as high as the in-distribution accuracy, and increasing the training data does not improve OOD generalization beyond a critical amount.
\begin{figure}[!t]
\centering
\includegraphics[width=0.4\textwidth]{emnlp2022-latex/EMNLP-2022/figure/intro-addition-example.png}
\caption{Examples of addition: the baseline setting (top) and Scratchpad \cite{scratchpad} with intermediate computation steps (bottom). A similar method with more detailed demonstration is introduced in \cite{Recchia-2021}. }
\label{fig:intro-addition-example}
\end{figure}
\begin{figure*}[!t]
\centering
\includegraphics[width=0.98\textwidth]{emnlp2022-latex/EMNLP-2022/figure/intro-repeat-2.jpg}
\caption{The horizontal axis is the number of digits and the vertical axis is the accuracy. The prompts for GPT3 consist of 4 examples. The T5 models are trained on 1-5 digits of up to 2,000 examples and each training example consists of random numbers in the format of \texttt{2 4 1}. In-dist: in-distribution. Out-of-dist.: out-of-distribution (OOD). In-distribution
refers to training on up to k-digit numbers and
testing on up to k-digit numbers while out-of-
distribution refers to training on up to k-digit numbers and testing on numbers with more digits. $\alpha$ indicates the repetition level of the examples. An example $x_1\cdots x_n$ with $n$ digits are sampled with the next digit probability $p(x_{i+1}|x_{i})= \alpha$, when $x_{i+1}=x_{i}$; otherwise, $(1-\alpha)/9$. Larger $\alpha$ indicates a higher repetition level.
}
\label{fig:intro-repeat}
\end{figure*}
Figure~\ref{fig:intro-addition-example} shows two possible addition exemplars for LMs. Addition can be considered as a basic arithmetic operation and a simple symbolic manipulation task. The scratchpad version gives more details on how humans do basic arithmetic. \citet{scratchpad} show that with more fine-grained demonstrations, the accuracy of addition can be improved dramatically with fine-tuning. Yet, it still can not achieve 100\% on OOD data, even with thousands of training data points provided. Figure \ref{fig:intro-repeat} shows the performance of GPT-3 and T5 on addition using the scratchpad version of training data. The problem becomes more severe when there are repeating digits in the addition operands.
As the performance drops with repeating digits, we suspect that LMs might not handle the repeating symbols well. Figure \ref{fig:intro-repeat} illustrates the performance of GPT-3 and T5 on the copy task, one of the simplest symbolic manipulation operations. GPT-3 and T5 still can not perform well on OOD. We further do a preliminary experiment where a T5 model is fine-tuned using the data containing repeating numbers of up to 80 digits, T5 still can not achieve 100\% in-distribution accuracy on long repeating digits. The results indicate that there are two problems intervening: Transformers are not good at handling repeating symbols and OOD generalization. The repeating symbols can also be a problem even for in-distribution data.
Why do large pretrained LMs that can do complex language generation fail on a simple symbolic manipulation task?
In this paper, we investigate the potential causes behind this phenomenon. We believe that overcoming the aforementioned limitations is of critical importance for the future application of Transformer-based LMs to reasoning-intensive tasks. What are the necessary steps to take to significantly improve or even approach 100\% accuracy on these simple but fundamentally important induction tasks? We examine a set of possible mitigation solutions including fine-grained computation steps, positional markers, and LMs with callable programs.
Since incorporating computation steps improves the OOD generalization in arithmetic addition~\cite{scratchpad}, one possible direction is to provide more fine-grained computation steps in the fine-tuning data or the few-shot prompt. However, it may not be sufficient to alleviate the problem of repeating numbers. When a human does addition, the position of each digit is used to differentiate the repeating digits. However, the self-attention mechanism in the Transformer may not tell which ``1'' is referred to in the input. This prompts us to explore using positional markers to differentiate the important tokens. Using these two methods to augment the reasoning process, we find that the performance of pretrained LMs still can not reach satisfying results. Then we resort to a method where the copy operation is implemented as a primitive function and explore whether the LM can further boost its performance.
We experiment with three symbolic manipulation tasks: copying, reversing, and addition. Experimental results show that although generalization in these symbolic manipulation tasks is straightforward for humans, it is still challenging for LMs, and none of these mitigation methods fully solves the problems. In the end, we introduce LMs with tutor which demonstrates every single step of teaching, pinpointing where these digits come from. LMs with tutor is able to deliver 100\% accuracy in situations of OOD and repeated symbols. In this design, LMs are used to generate actions that mimic operations in multiple tape Turing machines, rather than the intermediate results. These actions generate the intermediate results on tapes. We hope this could shed light on the capability of Transformer-based LMs in addition to providing large training datasets or scaling up the size of these models.
To conclude, our main contributions are:
\begin{itemize}
\item We identify a set of simple symbolic manipulation tasks and uncover the limitations of the LMs in arithmetic and symbolic induction.
\item We examine a set of potential techniques including positional markers, fine-grained computation steps, and LMs with callable programs. Though these techniques could mitigate the limitations of the LMs, none of them can completely solve the generalization problem.
\item Finally, we demonstrate that LMs with tutor is able to deliver 100\% accuracy in situations of OOD and repeated symbols. Our analysis could inspire new thoughts to overcome the limitation of LMs in symbolic manipulation.
\end{itemize}
\section{Related Work}
Previous research in using LMs for symbolic induction improves the model performance in the following three directions.
\textbf{Large Pretrained Language Models}: \citet{gpt3} show that GPT3 exhibits strong proficiency on 2-digit addition and subtraction using simply few-shot prompting, without any task-specific training. Furthermore, the larger the LM, the better the performance. Following GPT3, \citet{palm} further scale the Transformer-based LMs to a 540-billion parameter model, called Pathways Language Model (PaLM). Same as \citet{gpt3}, \citet{palm} find that scaling the LMs consistently results in better arithmetic reasoning ability with few-shot prompting. However, the reasoning ability of the large LMs is still limited. GPT3 struggles with 3-digit arithmetic and with direct prompting, even 540B PaLM can not achieve high performance on complex tasks requiring multi-step reasoning. Therefore \citet{cot} propose to augment the large pretrained LMs with the following prompting method.
\textbf{Chain-of-Thought Prompting}: This prompting method provides a few chain-of-thought demonstrations, which is a series of intermediate reasoning steps, as exemplars in the prompting. Therefore, given a complex reasoning task, the model is allowed to calculate the intermediate results step-by-step before generating the final answer. With chain-of-thought prompting, a complex reasoning task is decomposed into a list of simple operations and LMs can
derive these operations one by one. \citet{kim-etal-2022-ept} adopt faithful explanations that accurately represent the reasoning process behind solving a math word problem. \citet{cot} show that combining chain-of-thought prompting and a sufficiently large LM, 540B PaLM, can significantly improve the LMs' reasoning ability on complex tasks, such as math word problems.
\textbf{Fine-tuning with Large Training Datasets}:
Instead of few-shot prompting, another direction is to fine-tune large LMs with a sufficient amount of training data. \citet{investigating2021} fine-tune T5 with different ways of representing numbers, but even with the best-performing representation, the fine-tuned model can not achieve as good accuracy on out-of-distribution testing examples as in-distribution testing examples. \nop{Here in-distribution refers to training on up to k-digit numbers and testing on up to k-digit numbers while out-of-distribution refers to training on up to k-digit numbers and testing on numbers with more digits.} \citet{scratchpad} propose to use Scratchpad to improve the out-of-distribution accuracy. Scratchpad combines step-by-step reasoning with fine-tuning. The training examples include the intermediate steps of an algorithm in target, so the model is trained to generate not only the final answer, but also the intermediate steps, which is similar to chain-of-thought, but requires more training data. \citet{scratchpad} show that using the training data augmented with intermediate steps significantly improves the model performance, but even with 100k augmented training examples for the addition task, the fine-tuned 1B LM still does not perform well on out-of-distribution addition.
Our work is also related to \citet{ntm}, which extends the capabilities of Recurrent Neural Networks to two simple symbolic manipulation tasks, copy and sort, by augmenting the model with external memory resources. Instead of using hundreds of thousands of training examples, we focus on large pretrained LMs with few-shot prompting or fine-tuning settings in this work.
\section{Observations}
\label{sec:observation}
We first analyze the difficulty of generalizing the copy operation, one of the most fundamental, simplest symbolic manipulation operations.
We start by copying random numbers. For GPT3, we augment each testing example with the few-shot prompt as shown in Figure \ref{fig:gpt3_copy_prompt}.
\begin{figure}[!h]
\centering
\includegraphics[width=0.4\textwidth]{emnlp2022-latex/EMNLP-2022/figure/GPT3_copy_prompt.pdf}
\caption{The prompt for GPT3 on the copy task.}
\label{fig:gpt3_copy_prompt}
\end{figure}
We also fine-tune a T5 model for copying. The training data follows the same format as above and consists of random numbers of up to 5 digits.
We first evaluate the prompted GPT3 and fine-tuned T5 on copying random numbers of up to 80 digits ($\alpha=0.1$ in Figure~\ref{fig:intro-repeat}). GPT3 achieves nearly 100\% accuracy on all the testing examples of up to 80 digits when $\alpha$ is 0.1.
The finetuned T5 does not generalize well beyond 7 digits, and it achieves nearly 100\% accuracy on all the testing examples of 7 digits, except for a few error cases as follows:\\
\texttt{
``input'': copy: $\cdots$ 9 8 9 8 9 4 $\cdots$ \\
``pred'': $\cdots$ 9 8 9 4 $\cdots$ \\
``input'': copy: $\cdots$ 6 0 6 0 6 5 $\cdots$ \\
``pred'': $\cdots$ 6 0 6 5 $\cdots$ \\
}
A common feature of these error cases is that they all contain consecutive repeating numbers and the model tends to mistakenly skip part of them or over-replicate them.
\begin{figure}[!t]
\centering
\includegraphics[width=0.47\textwidth]{emnlp2022-latex/EMNLP-2022/figure/observation-f.png}
\caption{An illustration of GPT3 and T5 performance on copying repeating numbers. }
\label{fig:copy-repetitive-baselines}
\end{figure}
Therefore we further do a copying task where the testing examples consist only of repeated digits, such as \texttt{copy: 2 2 2 2 2}.
The results are shown in Figure~\ref{fig:copy-repetitive-baselines}. Although T5 performs well at copying random numbers of up to 7 digits, its accuracy at copying 5-digit repeating numbers drops below 80\%. Similarly, the prompted GPT3 can not generalize to copying more than 30 repeating digits. Both GPT3 and T5 are Transformer-based LMs, which use the self-attention mechanism. When copying numbers, the models are required to use self-attention to locate the next digit to copy. When copying random numbers without repeated digits, it usually would be sufficient to locate the next digit by comparing the previous few digits. However, when copying repeated digits, this mechanism no longer works since all the previous digits are the same. Instead, the model needs to locate the next digit either by counting the repetitive digits or by remembering the previous position. Therefore, the results suggest that the Transformer-based LMs, such as GPT3 and T5, might have limited capability of locating in symbolic manipulation.
\section{Mitigation Methods}
\label{sec:method}
In order to mitigate the limitations, we examine a few potential solutions.
\subsection{Positional Markers}
\label{subsec:method-implicit}
We first explore possible methods to mitigate the problem of repeating numbers. We introduce two types of positional markers: implicit positional markers and explicit ones.
As stated above, LMs tend to make mistakes when the input contains repeating numbers. When humans deal with repeating numbers in basic arithmetic, we usually use indices to distinguish those digits in the input. Positional encoding in large LMs is closely related to this human practice. Most Transformer-based LMs encode the positional information into positional vectors and add each of them to the corresponding word vector. Although large LMs have already incorporated positional encoding in the model architecture (Figure~\ref{fig:deberta}), results in Figure~\ref{fig:intro-repeat} indicate that the positional encoding commonly used in large LMs may not be sufficient to locate each repeating digit effectively.
Instead of representing each token by the sum of its contextual token embedding and the position embedding, DeBERTa~\cite{deberta} represents each token with a token embedding and a position embedding, respectively, and the attention weights are computed using disentangled matrices based on both embeddings, respectively (Figure~\ref{fig:deberta}). In other words, the self-attention in DeBERTa is disentangled. With the disentangled relative position embeddings, the attention scores between tokens depend not only on the content but also on the relative position between the tokens, so the disentangled relative position embeddings act as implicit position markers within DeBERTa, which might make it easier for the model to learn the latent position relationship in the training data of the symbolic manipulation tasks.
\begin{figure}[!t]
\centering
\includegraphics[width=0.47\textwidth]{emnlp2022-latex/EMNLP-2022/figure/deberta.png}
\caption{An illustration of standard Transformer attention (left) and DeBERTa disentangled attention (right). }
\label{fig:deberta}
\end{figure}
Although DeBERTa uses disentangled attention mechanism, it was not originally introduced to enhance the locating capability of LMs, so no pretraining task was specifically proposed for training the position embeddings in DeBERTa. This may potentially lead to limited generalization ability of DeBERTa on the induction tasks requiring accurate locating.
Rather than relying on implicit positional markers, another, more straightforward approach is to add explicit positional markers in the input for the model. For example, the input string \texttt{2 2 2} is augmented with positional markers \texttt{A, B, C, $\cdots$}. We explore two methods of adding explicit positional markers:\\
\textbf{Ordered marker}: The markers are inserted into the input in order. \texttt{2 2 2} $\rightarrow$ \texttt{A 2 B 2 C 2}\\
\textbf{Random marker}: The markers are inserted into the input in random order. \texttt{2 2 2} $\rightarrow$ \texttt{E 2 X 2 J 2}
With the explicit positional markers, each repeating \texttt{2} becomes different for the model. When doing symbolic manipulation, the Transformer-based LMs can easily locate the digit by recognizing the explicit positional markers.
Essentially, adding explicit positional markers breaks the repeating numbers into a non-repeating input sequence. This method is also related to pointer networks \cite{ptrnet}, which uses attention as a pointer to select the position indexes of the input tokens as the output. A hybrid pointer-generator network can also be leveraged to copy number from the source text, while retaining the ability to produce new numbers through the generator \cite{see-etal-2017-get}. Compared with implicit markers, explicit markers provide more direct and clearer location information in text format. However, similar to the implicit positional markers, whether using the explicit positional markers can generalize to arbitrary length or unseen markers is still questionable.
\subsection{Fine-grained Computation Steps}
\label{subsec:method-cot}
We then explore possible methods to alleviate the OOD generalization problem.
One observation is that the complexity of addition with long digits is larger than that of the 1-digit addition. Thus, the model should be given more computation time on the task when the numbers are large. The fine-tuned T5 and prompted GPT3 mentioned above, however, is required to generate the answer with a fixed amount of computation, so one possible direction to mitigate this limitation is to allow the model to operate step-by-step instead of generating the answer in one forward pass. For example, in k-digit addition, the model is allowed to break it down into k simple 1-digit addition and the model is allowed to generate k intermediate addition results to get the final answer.
Generating fine-grained computation steps can potentially alleviate the generalization problem, but may not contribute to the locating capability of the Transformer-based LMs. To mitigate the locating problem, we add positional markers to scratchpad~\cite{scratchpad} (Figure \ref{fig:gpt3_addition_prompt}):
\begin{figure}[!h]
\centering
\includegraphics[width=0.48\textwidth]{emnlp2022-latex/EMNLP-2022/figure/GPT3_addition_prompt.pdf}
\caption{The prompt for GPT3 on the addition task. We use \faHandPointRight[regular] and \faHandPointRight ~to denote optional different markers as described in Section \ref{subsec:method-implicit} if they are applied.}
\label{fig:gpt3_addition_prompt}
\end{figure}
We also experiment a more comprehensive scheme where each number in the demonstration is associated with an explicit positional marker or reference marker. A reference marker refers the positional marker where the following number is copied from as shown in Figure \ref{fig:addition_reference_marker}.
\begin{figure}[!h]
\centering
\includegraphics[width=0.46\textwidth]{emnlp2022-latex/EMNLP-2022/figure/t5_addition_marker_reference.pdf}
\caption{The demonstration of comprehensive scheme for addition problem, where position marker is marked red and reference marker is marked in green.}
\label{fig:addition_reference_marker}
\end{figure}
Through our experiments, we found that although these markers can help achieve higher accuracy for in domain data. It does not help much for OOD data. It clearly indicates the limitation of Transformers and pre-trained language models in induction. In the following discussion, we will shed some insights on how to eliminate such limitation.
\subsection{LM with callable Programs}
\label{subsec:method-callable}
If both implicit positional markers and explicit positional markers do not generalize well in a symbolic reasoning task, then an alternative is to combine LMs with callable programs to replace the basic symbolic operations when possible, since callable programs do not have the generalization problem. For example, when combined with the fine-grained computation steps in the addition task, the convert, add, or combine operations can be considered callable programs. When the LM generates the text sequence \texttt{add(1,5)}, the callable function \texttt{add} will be invoked and return the result in text: \texttt{carry C: 0, result 6}.
Following the example in Section~\ref{subsec:method-cot}, with callable functions, the prompt format is as follows:
\begin{figure}[!h]
\centering
\includegraphics[width=0.48\textwidth]{emnlp2022-latex/EMNLP-2022/figure/gpt3_with_callable_program_v2.pdf}
\caption{The prompt for GPT3 on the addition task with callable programs. \faHandPointRight[regular] and \faHandPointRight ~are positional markers. Different callable programs (convert, add and combine) are marked in different colors, and the results they returned are underlined with the corresponding color.}
\label{fig:gpt3_callable_program}
\end{figure}
Given a testing example, the prompted GPT3 first generates the solution step by step. During the process, the results of the function calls will be appended to the generated result to be used in the following steps.
Callable programs can be viewed as decomposing a complex task to smaller, simpler jobs. The remaing issue is to learn chaining these smaller jobs together to complete the task.
Callable programs can guarantee the correctness of output given correct input for a given job. However, even augmented with callable programs, LMs may still suffer from the locating problem since the callable programs rely on LMs to decide which token to copy (Figure~\ref{fig:gpt3_callable_program_error}). Unfortunately, LMs cannot guarantee the correctness of this copy action.
\begin{figure}[!h]
\centering
\includegraphics[width=0.48\textwidth]{emnlp2022-latex/EMNLP-2022/figure/callable-error-example.jpg}
\caption{An error example of GPT3 with callable functions. The error is highlighted.}
\label{fig:gpt3_callable_program_error}
\end{figure}
\subsection{LM with Tutor}
\label{subsec:method-simulator}
Scratchpad \cite{scratchpad} ignores the visual process when an elementary school tutor visually illustrates how to perform addition step by step: Pinpointing where each digit in the output sequence comes from, adding single digits together and iterating. It turns out that these details and abstractions are important in order to simplify the learning process and help kids learn how to do addition in a few shots.
A tutor shows every single step visually and sometimes calls an already learned sub-module to complete a task. In this way, the hypothesis space between two consecutive steps can be dramatically simplified; hence the chance of learning a correct model can be improved.
Taking copy as an example. Instead of providing a training example: \texttt{copy: 1 1 1 2 2 2 result: 1 1 1 2 2 2}, we need to demonstrate where the first \texttt{1}, the second \texttt{1}, the third \texttt{1} in the output sequence comes from, which exactly imitates the finest action a human could do to perform such an operation. Suppose there is a cursor placed at the beginning of the input sequence, a ``rmov'' operation moves the cursor one token to the right. A ``cpy'' operation copies a single digit to the output sequence. An ``end'' operation checks if the marker reaches the end of the sequence. ``T'' and ``F'' represent true and false respectively. We assume all these actions are unitary and have been learned. Then a possible action sequence to complete the copy operation is as follows: \\
\texttt{rmov, end=F, cpy, rmov, end=F, cpy, \ldots, rmov, end=T.}\\
\begin{figure}[!t]
\centering
\includegraphics[width=0.37\textwidth]{emnlp2022-latex/EMNLP-2022/figure/pattern_rec.png}
\caption{An illustration of doing copy with pattern matching.}
\label{fig:pattern}
\end{figure}
This fine-grained action sequence accurately describes the whole copy operation. Certainly, there are other ways to perform copying. For example, instead of using a cursor, one can use a pattern match to perform the copy operation (Figure~\ref{fig:pattern}).
We suspect that the copy operation learned from Transformer is following this pattern-matching approach, which is error-prone when the pattern has repeating symbols and when the long pattern is out-of-distribution. Positional markers do not help either as it seems unable to handle the OOD generalization problem.
If we take the action sequence ``rmov, end=F, \ldots'' to train a Transformer for copying, the hypothesis space is simplified, thus making it possible to find the simplest model that can simulate the whole action sequence. This is related to imitation learning \cite{alvinn,imitationross}. Although there is no guarantee that Transformer can definitely find the correct model, the chance is much higher. One can also relate the setting with a multiple tape Turing machine where state transition is conducted among the positions of tape heads and read/write operations. The Transformer is trained to learn such state transition, thus completing the programming of a Turing machine.
As for the addition operation, a similar action sequence can be obtained to simulate how humans tutor kids do addition at an early age (Figure \ref{fig:simulator}). Let ``lmov'' denote moving the cursor one token to the left. The ``add'' operation adds three single digits together, one from each of the two operands and the third one from the carry digit, appends the result to the output, and updates the carry digit. Assume ``add'' is a callable program as kids have learned how to do single digits addition. Suppose the cursor starts from the end of the operands. The entire action sequence looks like the following. \\
\texttt{lmov, end=F, add, lmov, end=F, add, \ldots, lmov, end=T.}
\begin{figure}[!t]
\centering
\includegraphics[width=0.47\textwidth]{emnlp2022-latex/EMNLP-2022/figure/simulator-2.png}
\caption{An illustration of the LM with Tutor method. With the tutor (right), the LM (left), or just a transformer generates an action sequence that simulates how humans do arithmetic addition.}
\label{fig:simulator}
\end{figure}
The main difference between the tutor and the Scratchpad method \cite{scratchpad} is the abstract callable function and detailed action sequence. The action sequence includes all the state transitions needed to complete the task. It perfectly overcomes the OOD issue and does not require many training examples in order to achieve 100\% accuracy.
While there is a great effort to enlarge Transformer-based LMs such as PALM \cite{palm} and Minerva \cite{minerva}, to improve the performance in symbolic and logical reasoning, our result reveals that it might be necessary to demonstrate the action sequence with reasonable abstraction to the Transformer to leverage its full strength.
The action sequence ``rmov, end=F, ...'' can also be viewed as an algorithm execution instance. It is well known that if one step is missing in an algorithm, it will most likely not produce the expected output. In order to fill the gap in case one step is missed during demonstration, we have to rely on training examples of that step and learn it, which will likely incur errors. As the number of missing steps increases, errors will accumulate and eventually make learning more difficult and hard to generalize. To make learning in symbolic reasoning easier, it will be important to have a detailed action sequence as we have demonstrated in Figure~\ref{fig:simulator}.
\section{Experiments}
In this section, we conduct experiments on three different problems including copying, addition, and another basic symbolic manipulation operation, reverse. We illustrate the limitation of LMs in symbolic and arithmetic induction and the improvement that could be achieved by the introduced mitigation methods.
\begin{figure*}[!t]
\centering
\includegraphics[width=0.97\textwidth]{emnlp2022-latex/EMNLP-2022/figure/results-f2.png}
\caption{Experimental results. (a): results of copying repeating numbers. (b)(c): results of reversing the list. (d)(e)(f): results on arithmetic addition. The x-axis is the number of digits or number of items.}
\label{fig:results}
\end{figure*}
\subsection{Copy Operation}
\label{subsec:exp-copy}
Copying is the most basic operation used in symbolic manipulation and arithmetic reasoning. We experiment with the following methods:\\
\textbf{GPT3}: We use the prompt as shown in Section~\ref{sec:observation}.\\
\textbf{DeBERTa / T5}: The training example is as follows: \texttt{copy: 1 2 3 4 result: 1 2 3 4}\\
\textbf{T5 + ordered marker}: The training data is augmented with explicit positional markers. \texttt{copy: A 1 B 2 C 3 result: A 1 B 2 C 3} \\
\textbf{T5 + random marker}: Same as above, but the augmented positional markers are in random order. \texttt{copy: E 1 A 2 F 3 result: E 1 A 2 F 3} \\
\textbf{T5 / GPT3 + tutor}: The training and testing examples are as described in Section~\ref{subsec:method-simulator}.
We experiment with the T5-base (220M) model, DeBERTa-base (140M) model, and GPT3 text-davinci-002. The models are initiated with the pretrained parameters and further fine-tuned on the training data. For GPT3 or T5 with tutor, the training data consists of 15 examples of up to 5 digits. For all the other T5 models and DeBERTa, the training data consists of 2,000 random numbers of up to 5 digits. We evaluate all the models on copying repeating numbers of up to 80 digits. The results are illustrated in Figure~\ref{fig:results}(a).
As shown in Figure~\ref{fig:results}(a), GPT3 achieves 100\% accuracy on the in-distribution testing data (1-5 digits) but the fine-tuned T5 achieves 78\% accuracy on the 5-digit repeating numbers although they are in-distribution. Augmented with random or ordered positional markers, the T5 models achieve 100\% in-distribution accuracy, and so does using implicit positional markers (DeBERTa). This suggests that both implicit positional markers and explicit positional markers may help with the locating capability of LMs. However, using explicit positional markers, either ordered or random, the model exhibits significantly better generalization to OOD testing data whereas DeBERTa fails on OOD data. GPT3 exhibits better OOD generalization than T5 with positional markers but it does not generalize well beyond 30 digits. Both \textit{T5 + tutor} and \textit{GPT3 + tutor} keeps 100\% accuracy when the number of digits increases.
\begin{figure}[!th]
\centering
\includegraphics[width=0.45\textwidth]{emnlp2022-latex/EMNLP-2022/figure/GPT3_addition_wo_step_prompt.pdf}
\caption{The prompt for GPT3 on the addition task without intermediate steps.}
\label{fig:gpt3_addition_wo_step}
\end{figure}
\subsection{Addition}
\label{subsec:exp-addition}
For arithmetic addition, we experiment with the following methods:\\
\textbf{GPT3}: The exemplars are shown in Figure \ref{fig:gpt3_addition_wo_step}.\\
\textbf{GPT3 + coarse-grained steps}: The exemplar is similar to that in Figure~\ref{fig:gpt3_addition_prompt}, but the instructions for the result combination and the computation of the carry digit and step result are omitted.\\
\textbf{GPT3 + fine-grained steps (+ ordered marker)}: The exemplar we use is as shown in Figure~\ref{fig:gpt3_addition_prompt}. \\
\textbf{GPT3 + callable programs}: The exemplar is shown in Figure~\ref{fig:gpt3_callable_program}.\\
\textbf{DeBERTa / T5}: The training data follows the format of the exemplar for GPT3. \\
\textbf{DeBERTa / T5 + fine-grained steps}: The training data used in this setting follow the format as the exemplar in \textit{GPT3 + fine-grained steps}.\\
\textbf{T5 + ordered / random marker}: The training example is augmented with ordered or random markers. For example, \texttt{question: G 1 C 1 + G 2 C 5 result: G 3 C 6}.\\
\textbf{T5 + fine-grained steps + ordered / random marker}: The training data in this setting follow a similar format as the exemplar in \textit{GPT3 + fine-grained steps + ordered marker}, but the positional markers can be in random order.\\
\textbf{T5 / GPT3 + tutor}: The training and testing examples are as described in Section~\ref{subsec:method-simulator}.
The model settings are the same as in the above copy experiments. For LMs with tutor, the training data or prompt consists of 15 examples of up to 5 digits. In other settings, the training data consists of 1,000 examples of 1-5 digit addition and for GPT3, the prompt includes 4 examples. We evaluate all the models on the addition of up to 30 digits. The results are shown in Figure~\ref{fig:results}(d)(e)(f).
As shown in Figure~\ref{fig:results}(d), both coarse-grained and fine-grained computation steps contribute to the in-distribution performance of GPT3, and using finer-grained steps achieves larger performance gains on both in-distribution data and OOD data. The performance is further boosted with explicit positional markers. Experiments on T5 (Figure \ref{fig:results}(e)(f)) also show the effectiveness of using explicit positional markers, with or without fine-grained computation steps, indicating that the explicit positional markers might make it easier for LMs to learn the induction in the arithmetic reasoning tasks.
Similar to the results on the copying task, both DeBERTa and \textit{DeBERTa + fine-grained steps} achieve near 100\% in-distribution accuracy but 0\% OOD accuracy, suggesting that the relative position embedding of DeBERTa might have limited OOD generalization ability.
On T5, incorporating fine-grained computation steps does not improve the OOD performance as significantly as on GPT3 (Figure~\ref{fig:results}(f)). The reason might be that fine-tuning T5 tends to overfit more easily than prompting GPT3.
Unsurprisingly, \textit{GPT3 + callable programs} achieves much better OOD generalization. However, its OOD performance still degrades as the number of digits increases. Same as in the copy experiments, \textit{LMs + tutor} keeps 100\% accuracy on all the experimented numbers of digits.
\subsection{Reverse List}
Besides copying and addition, we also experiment with reversing. Reversing is similar to copying. Both require replicating the items in the input, but reversing might be more challenging than copying in the terms of locating. In copying, the distance between each source digit and the replicated digit is the same for each digit in the number. However, when reversing, the distance between the source item and the replicated item keeps increasing during the generation.
For this problem, we experiment with the following methods: \\
\textbf{GPT3}: The prompt without any intermediate steps is used, as shown in Figure \ref{fig:gpt3_reverse_wo_step}.
\begin{figure}[!h]
\centering
\includegraphics[width=0.45\textwidth]{emnlp2022-latex/EMNLP-2022/figure/GPT3_reverse_wo_step.pdf}
\caption{The prompt for GPT3 on the reverse task without intermediate steps.}
\label{fig:gpt3_reverse_wo_step}
\end{figure}
\noindent
\textbf{DeBERTa / T5}: \texttt{reverse the list: bike, apple, book result: bike, cat, pen} \\
\textbf{GPT3 / DeBERTa / T5 + fine-grained steps}: The training example for T5 and the exemplar for GPT3 are shown in Figure \ref{fig:gpt3_reverse_with_fine_grained_step}.
\begin{figure}[!h]
\centering
\includegraphics[width=0.45\textwidth]{emnlp2022-latex/EMNLP-2022/figure/GPT3_reverse_fine_grained_step.pdf}
\caption{The prompt for GPT3 on the reverse task with fine-grained steps.}
\label{fig:gpt3_reverse_with_fine_grained_step}
\end{figure}
\noindent
\textbf{T5 + ordered marker}: The list items are augmented with the ordered positional markers in the input. \texttt{reverse the list: A bike, B cat, C pen result: pen, cat, bike}. \\
\textbf{T5 / GPT3 + tutor}: The training and testing examples are very similar to that for the copy task. The only difference is the direction of the move operation. ``rmov'' in the copy task is replaced by ``lmov'' here.
The model settings are the same as in the above experiments and the training data consists of examples of 1-5 items, which are randomly sampled from a predefined list of single-token nouns. For LMs with tutor, the training data or prompt consists of 15 examples of up to 5 items. In other settings, the training data consists of 1,000 examples for T5, and for GPT3, each prompt includes 4 examples. We evaluate all the models on reversing the list of up to 30 items. The results are shown in Figure~\ref{fig:results}(b)(c).
Although GPT3 can generalize to 80 digits on copying random numbers (Figure~\ref{fig:intro-repeat}), it does not generalize well beyond 20 items on reversing, which suggests that reversing might require stronger locating capability than copying. This problem also occurs on DeBERTa and T5. When tested on the OOD data, the models tends to generate only a sublist of the input. Using fine-grained steps (Figure~\ref{fig:results}(b)) or positional markers, whether implicit or explicit (Figure~\ref{fig:results}(c)), does not significantly improve the generalization of the experimented models. The reason might be the increasing distance between the source item and the replicated item as stated above. Again, \textit{LMs + tutor} maintains 100\% accuracy throughout the experiments.
\subsection{Discussion}
\label{subsec:discussion}
From the experimental results, we observe that fine-grained computation steps may improve the LM's induction ability on the arithmetic reasoning tasks and the granularity of the steps has an impact on the performance improvement. Finer-grained computation steps may contribute to larger performance improvement.
Positional markers, whether implicit or explicit, improves LMs' in-distribution performance on all the symbolic manipulation tasks in our experiments. However, We find that augmented with the relative position embeddings, DeBERTa tends to face more severe over-fitting than T5 during fine-tuning. In the reversing experiment, using the T5 model without pretrained parameters, the fine-tuned model can not achieve a good in-distribution performance after 200k optimization steps. However, the DeBERTa model without pretrained parameters achieves 100\% in-distribution accuracy within only 2k optimization steps while the OOD accuracy drops, indicating that it has overfitted within 2k optimization steps. In other words, the relative position embeddings in DeBERTa significantly improve the model's capacity of positions, which improves in-distribution performance on simple symbolic manipulation tasks, but may not generalize well on OOD data. Compared with the implicit positional markers (relative position embeddings in DeBERTa), explicit positional markers might have better OOD generalization ability. However, incorporating symbolic manipulation tasks in the LM pretraining stage might alleviate this problem, so incorporating implicit positional markers can still be a possible direction of improving the LM's performance on reasoning tasks requiring locating ability.
Using LM with callable programs exhibits strong OOD performance on addition, suggesting that the LMs' ability to perform simple symbolic operations, such as copying, splitting, and combining, can be critical for improving their performance on reasoning tasks. How to further improve the LMs' performance on more complex reasoning tasks in this direction is left for future work.
\section{Conclusion}
In this work, we explore the limitations of pretrained LMs on arithmetic reasoning and symbolic manipulation. We experiment with three simple symbolic manipulation tasks and show that improving the locating and induction capability of LMs can be important for further improving their performance on induction tasks. Our method that combines abstraction and finest-grained step-by-step tutoring demonstrates its potential to generalize correctly, shedding light on possible directions orthogonal to scaling up LMs for future work in this area.
|
\section{Introduction}
The Casimir effect for two parallel conducting plates in a vacuum state that generates an attractive force was firstly discussed in Ref.~\cite{Casimir1948}. While the effect was experimentally confirmed \cite{Sparnaay:1958wg}, the theoretical discussion was extended for various models (see, e.g., Refs.~\cite{Lutken1984, Valuyan2009,Fosco:2008vn, Seyedzahedi2010, Fosco2022, Shahkarami2011,Moazzemi2007, DePaola1999,Boyer1968}). Such an effect is a manifestation of the quantum field theory with appropriate boundary conditions. Thus, the choice of the boundary condition plays an important role in the Casimir effect. The essential aspect of discussing the Casimir effect is not only placed on the type of the boundary conditions but also on the field.
Discussing the boundary condition for a Dirac fermion is nontrivial because the Dirac equation is a first-order differential equation. To discuss the Casimir effect for the Dirac field, several authors in Refs.~\cite{Saghian2012, Cruz2018, Elizalde2002, Bellucci2009} used the boundary condition in the MIT bag model \cite{Chodos11974, Chodos21974, KJohnson1975} (see also Refs.~\cite{Alberto1996, Alberto2011} for confinement system); it guarantees the vanishing flux or the normal probability density at the boundary surface. However, this boundary condition leads to a discontinuity of the axial-vector current at the boundary surface that breaks its chiral symmetry. An alternative way to address this issue is to introduce the chiral bag model in the presence of the pion field \cite{Chodos1975}.
A more general form of the boundary condition in the MIT bag model that includes the chiral angle is the so-called chiral MIT boundary conditions \cite{Theberge1980, Lutken1984, Jaffe1989}. Using this boundary, one can investigate the interaction between the particle and boundary surface that may change the spin orientation depending on the chiral angle \cite{Nicolaevici2017}. Thus, the roles of the chiral angle in boundary conditions for a Dirac field may give essential features (see, e.g., Refs.~\cite{Ambrus:2015lfr, Chernodub:2016kxh, Chernodub:2017ref, Chernodub:2017mvp, Rohim2021}). Ref.~\cite{Rohim2021} showed that the particle's energy in the confinement system also depends on the chiral angle. There is another general form of the boundary condition in the MIT bag model, e.g., the self-adjoint boundary condition, as used by Ref.~\cite{Sitenko:2014kza} to discuss the Casimir effect by including the background magnetic field (see also Ref.~\cite{Donaire}).
This study investigates the Casimir effect of a Dirac field confined between two parallel plates. The properties of both plates are described by chiral MIT boundary conditions \cite{Lutken1984}.
We propose the general solution for the Dirac equation in such a system following the arguments in Refs.~\cite{Alberto1996, Alberto2011, Nicolaevici2017}. We write the mass of a particle as a function of position and include an analysis of the spin orientation by distinguishing its field components. Along with the mentioned procedure, we discuss not only the Casimir energy but also the Casimir pressure. Compared with Ref.~\cite{Lutken1984}, they applied the chiral MIT boundary conditions for the massless fermion case. In this paper, we apply the boundary conditions for the case of a massive fermion particle. From the viewpoint of the boundary conditions type, the present paper could be an extension of the earlier study on the Casimir energy by the authors in Ref.~\cite{Saghian2012} (see also Refs.~\cite{Cruz2018, Bellucci2009}), where they have discussed the nonchiral case.
The rest of this paper is organized as follows. In Sec.~\ref{PhysicalSetup}, we introduce the general setup for our model. In Sec.~\ref{GeneralFeatures}, we discuss the features under the chiral MIT boundary conditions focusing on the discrete momenta and change of spin orientations. In Sec.~\ref{CasimirEnergy}, we investigate the Casimir effect of a massive fermion under this boundary condition. Section \ref{Summary} is devoted to summary and conclusion. In Appendix~\ref{detaildiscretemomenta}, we provide the complementary derivations for discrete momenta. Throughout this paper, we will use natural units $c=\hbar=1$.
\section{Physical setup}
\label{PhysicalSetup}
We consider a free massive Dirac fermion confined between two parallel plates in $3+1$-dimensional Minkowski spacetime background. The first plate is placed at $x_3=0$ while the second one is placed at $x_3=\ell$ (see Fig.~\ref{setup}).
Both plates are in parallel with $(x_1,x_2)$-plane.
\begin{figure}[tbp]
\centering
\includegraphics[width=.5\textwidth]{physicalsetup.pdf}
\caption{\label{setup} Physical setup of a Dirac fermion confined between two parallel plates. We use $L$ and $\ell$ to represent the plate's size and distance, respectively. In the present paper,
we assume the limit $L\rightarrow \infty$ approximately.}
\end{figure}
In such a system, the action of a Dirac field with mass $m$ is given by
\begin{eqnarray}
S=\int d^4x \bar \Psi(i\gamma^\mu \partial_\mu-m) \Psi,
\label{DiracAction}
\end{eqnarray}
where $\bar \Psi\equiv\Psi^\dagger \gamma^0$ is the Dirac adjoint and $\gamma^\mu$ are the gamma matrices
in the Dirac representation given by
\begin{eqnarray}
\gamma^0= \begin{pmatrix}
I & 0\\
0 &-I
\end{pmatrix}
~~{\rm and}~~
\gamma^i= \begin{pmatrix}
o & \sigma_i\\
-\sigma_i &0
\end{pmatrix},
~~ i=1,2,3~,
\end{eqnarray}
with $I$ is $2\times 2$ identity matrix and $\sigma_i$ are Pauli matrices. The above gamma matrices satisfy anti-commutation relation $\lbrace \gamma^\mu,\gamma^\nu\rbrace=2\eta^{\mu\nu}$ with $\eta^{\mu\nu}={\rm diag.}(1,-1,-1,-1)$.
Taking the variation of the action in Eq.~(\ref{DiracAction}) leads to the following Dirac equation,
\begin{eqnarray}
(i\gamma^\mu \partial_\mu-m) \Psi=0.
\label{DiracEq}
\end{eqnarray}
To propose the specific form of the Dirac field in the region between two parallel plates that satisfies the Dirac equation (\ref{DiracEq}), we follow the arguments used in Refs.~\cite{Alberto1996, Alberto2011, Nicolaevici2017} as follows. (i) The particle mass depends on its position, which is originally described by the MIT bag model for hadron \cite{Theberge1980, Lutken1984, Jaffe1989}. Namely, inside the region between two parallel plates, the mass of a Dirac field is finite and becomes infinite at both plates \cite{Alberto1996, Alberto2011}. Under this condition, the Dirac field outside the confinement area vanishes. (ii) The form of a Dirac field consists of two-component fields associated with their spin orientations \cite{Nicolaevici2017}. Based on the arguments above, the proposal for a positive-frequency of the massive Dirac field in the region between two parallel plates is given by
\begin{eqnarray}
&&\psi^{(+)}_{{\bm k}s}(t,{\bm x})=e^{-iE t} \psi^{(+)}_{{\bm k} s}({\bm x}),
\label{TotalDiracwavetime}\\
&&\psi^{(+)}_{{\bm k} s}({\bm x})= B
\begin{pmatrix}\chi_R\\
r\hat {\bm k}_R\cdot{\bm \sigma} \chi_R
\end{pmatrix} e^{i{\bm k}_R \cdot {\bm x}}+C
\begin{pmatrix}\chi_L\\
r\hat {\bm k}_L\cdot{\bm \sigma} \chi_L
\end{pmatrix} e^{i{\bm k}_L \cdot {\bm x}},
\label{TotalDiracwave}
\end{eqnarray}
where $B$ and $C$ are complex coefficients, $s$ represents the spin orientation, $r=|{\bm k_{R(L)}}|/(m+E)$, and $E = \sqrt{k^2_1+k^2_2+k^2_3+m^2}$ is the energy of a Dirac field. The two-component spinors $\chi_{R(L)}$ satisfy the normalized condition of $\chi^\dagger_{R(L)}\chi_{R(L)}=1$ with subscripts $``R"$ and $``L"$ corresponding to the right and left of the Dirac field components, respectively. We also use the notations ${\bm k}_R=(k_1,k_2,k_3)$ and ${\bm k}_L=(k_1,k_2,-k_3)$ to represent the spatial momentum of the right- and left-moving field components. Note that we have distinguished the two-component spinor for each component of the Dirac field in Eq.~(\ref{TotalDiracwave}) because their spin orientations may depend on the boundary condition
\cite{Nicolaevici2017}. The corresponding form of the negative-frequency for the massive Dirac field can be obtained by taking the charge conjugation of the above positive-frequency Dirac field as
\begin{eqnarray}
\psi^{(-)}_{{\bm k}s}(t,{\bm x})=\psi^{(+)\rm C}_{{\bm k}s}(t,{\bm x})=i\gamma^2\psi^{(+)*}_{{\bm k}s}(t,{\bm x}).
\end{eqnarray}
In our model, the properties of both plates are described by chiral MIT boundary conditions given as \cite{Lutken1984}
\begin{eqnarray}
iN_\mu\gamma^\mu\psi
=e^{-i\Theta\gamma^5}\psi,
\label{chiralMITboundaryconditions}
\end{eqnarray}
where $N_\mu$ is an inward normal unit four-vector perpendicular to the boundary surface, $\Theta\in [0,2\pi)$ denotes the chiral angle, and $\gamma^5\equiv i\gamma^0\gamma^1\gamma^2\gamma^3$. The above boundary condition guarantees the vanishing of the normal probability current density at the boundary surface for any chiral angles \cite{Jaffe1989}
\begin{eqnarray}
N_\mu J^\mu(\equiv N_\mu \bar \psi\gamma^\mu \psi)=0.
\end{eqnarray}
In the nonchiral case $\Theta=0$, the boundary condition (\ref{chiralMITboundaryconditions}) reduces to that given in the MIT bag model.
\section{Features under boundary conditions}
\label{GeneralFeatures}
In this section, we investigate two features of a massive fermion confined between two parallel plates under chiral MIT boundary conditions \cite{Lutken1984}. Namely, we discuss how the boundary condition affects the structure of the discrete momenta and the changes of the spin orientation following the procedure in Refs.~\cite{Nicolaevici2017, Rohim2021}. However, our system proceeds with the general momentum (see also Refs.~\cite{Saghian2012, Cruz2018, Bellucci2009}).
\subsection{Discrete momenta}
The boundary condition at the first plate $x_3=0$ is given by
\begin{eqnarray}
i\gamma^3\psi|_{x_3=0} =e^{-i\Theta\gamma^5}\psi |_{x_3=0},
\end{eqnarray}
which is obtained from Eq.~(\ref{chiralMITboundaryconditions}) with the inward normal unit four-vector given as $N_\mu=(0,0,0,1)$.
It can be rewritten into two equivalent equations as follows
\begin{eqnarray}
&&i(\sigma_3+\sin\Theta I)\chi_2|_{ x_3=0}-\cos\Theta\chi_1|_{ x_3=0}=0,
\label{chiral10}\\
&&i(\sigma_3-\sin\Theta I)\chi_1|_{ x_3=0}+\cos\Theta\chi_2|_{ x_3=0}=0,
\label{chiral20}
\end{eqnarray}
where $\chi_1$ and $\chi_2$ are the upper and lower two-components Dirac field,
respectively. Applying the boundary conditions (\ref{chiral10}) or (\ref{chiral20}) to the positive frequency Dirac field (\ref{TotalDiracwave}), we obtain the relation of coefficients $B$ and $C$ at the first plate as
\begin{eqnarray}
\big [i(\sigma_3+\sin\Theta I) r\hat {\bm k}_R\cdot {\bm \sigma}-\cos\Theta I\big] B \chi_R =- \big [i(\sigma_3+\sin\Theta I) r\hat {\bm k}_L\cdot {\bm \sigma}-\cos\Theta I\big] C \chi_L.
\label{relation01}
\end{eqnarray}
Below, we show that this relation is useful for investigating the discrete momenta and the changes of the spin orientation at the first plate.
We next employ the boundary condition at the second plate to know the behavior of discrete momenta. At the surface of the second plate $ x_3=\ell$, the inward normal unit four-vector reads $N_\mu=(0,0,0,-1)$. Then, the corresponding boundary condition is given by
\begin{eqnarray}
-i\gamma^3\psi|_{x_3=\ell} =e^{-i\Theta\gamma^5}\psi |_{x_3=\ell},
\end{eqnarray}
that leads to the following two equivalent equations
\begin{eqnarray}
&&i(-\sigma_3+\sin\Theta I)\chi_2|_{ x_3=\ell}-\cos\Theta\chi_1|_{ x_3=\ell}=0,
\label{chiral1L}\\
&&i(-\sigma_3-\sin\Theta I)\chi_1|_{ x_3=\ell}+\cos\Theta\chi_2|_{ x_3=\ell}=0.
\label{chiral2L}
\end{eqnarray}
Applying boundary conditions~(\ref{chiral1L}) or (\ref{chiral2L}) to the positive-frequency Dirac field (\ref{TotalDiracwave}), we have the relation as follows
\begin{align}
&\big [i(-\sigma_3+\sin\Theta I) r\hat {\bm k}_R\cdot {\bm \sigma}-\cos\Theta I\big] e^{ik_3\ell} B \chi_R \nonumber\\
&~~~~~~~~~~~~~~~~~~~~=- \big [i(-\sigma_3+\sin\Theta I) r\hat {\bm k}_L\cdot {\bm \sigma}-\cos\Theta I\big] e^{-ik_3\ell} C \chi_L.
\label{relationL1}
\end{align}
Note that two-component Dirac spinors $\chi_R$ and $\chi_L$ in Eq.~(\ref{relationL1}) will be the same as those in Eq.~(\ref{relation01}) when the spin orientations are consistently reflected; namely, the reflected spin orientation at the second plate is the same as the incident spin orientation at the first plate (see Ref.~\cite{Rohim2021} for one-dimensional case).
Utilizing relations given in Eqs.~(\ref{relation01}) and (\ref{relationL1}), we are able to derive the contraint for the momentum as follows
\begin{eqnarray}
m\ell\cos\Theta\sin(k_3 \ell)+k_3\ell\cos(k_3\ell)= 0.
\label{DiscreteMomenta}
\end{eqnarray}
For the detailed derivation, see Appendix~\ref{detaildiscretemomenta}. The solution for Eq.~(\ref{DiscreteMomenta}) shows that the momentum $k_3$ must be discrete and
depends on the chiral angle. It also shows that in our system, momenta $k_1$ and $k_2$ remain to continue because the boundary conditions appear only at the $x_3$-axis in parallel with $(x,y)$-plane, as mentioned in the previous section.
Next, we define $k'_{3n}\equiv k_3\ell$ with $n=1,2,3,\cdots$ to denote the solution for Eq.~(\ref{DiscreteMomenta}). In comparison to the solution for the nonchiral case, we here have a factor of $``\cos\Theta"$ that contributes to determining the structure of the discrete momenta. However,
when the chiral angle takes values as $\Theta=\pi/2, 3\pi/2$, the discrete momenta have a nontrivial solution for all mass $m$ as follows \cite{Rohim2021},
\begin{eqnarray}
k'_{3n}={(2n-1)\pi\over 2},
\label{solnontrivialK3}
\end{eqnarray}
which is the same as the discrete momenta solution in the massless case $m=0$.
As has been discussed in the previous work \cite{Rohim2021}, there are two cases in analyzing the solution for Eq.~(\ref{DiscreteMomenta}).
For the case of light mass
$m'(\equiv m\ell)\ll 1$, the solution of the discrete momenta is approximately given by
\begin{eqnarray}
k'_{3n, l}\simeq{(2n-1)\pi\over 2}+ {2 m'\cos\Theta\over (2n-1)\pi}.
\end{eqnarray}
We note that the first term of the above solution covers the solution of Eq.~(\ref{solnontrivialK3}) while the second term gives the correction to the discrete momenta $k'_{3n}$
depending on the mass $m$ and the chiral angle $\Theta$.
Whereas in the case of heavy mass $m'\gg 1$, the solution of the discrete momenta is approximately given by
\begin{eqnarray}
k'_{3n, h}\simeq n\pi-{n\pi\over m'\cos\Theta}.
\label{k3nh}
\end{eqnarray}
In Eq.~\eqref{k3nh}, the first term covers the discrete momenta for the Schr\"{o}dinger equation in an infinite potential well under the Dirichlet boundary condition \cite{Rohim2021}. Since $m'$ depends on $\ell$, the case of $m'\ll 1$ and $m'\gg 1$ correspond to small and large distances between two plates, respectively. In other words, one can write the distance of two plates as a function of the Compton wavelength that determines whether the confinement system approaches ultra- or non-relativistic limits \cite{Alberto1996, Alberto2011}.
\subsection{Change of spin orientation}
The two-component spinor $\chi_{R}$ can be connected to $\chi_{L}$ using a rotation operator in spin space, which is determined by the boundary condition.
Once we know the structure of one of them, we will obtain complete information on the spin orientations for both Dirac field components (\ref{TotalDiracwavetime}) and (\ref{TotalDiracwave}).
The rotation operator in spin space is given by \cite{Nicolaevici2017}
\begin{eqnarray}
{\cal U} =e^{i\varsigma }\left[\cos({\varphi \over 2})I-i\sin({\varphi \over 2})
\hat {\bm n} \cdot{\bm \sigma}\right],
\label{rotoperatorspinspace}
\end{eqnarray}
where $e^{i\varsigma }$ denotes a pure phase, $\varphi $ is the rotation angle, and $
\hat {\bm n} $ represents the unit rotation axis generated by the reflection with the plate.
At the first plate, from the relation given in Eq.~(\ref{relation01}), we have $\chi_R={\cal U}^{(1)}_R \chi_L$
with ${\cal U}^{(1)}_R$ is the rotation operator in spin space generated by the reflection with the first plate given as
\begin{align}
{\cal U}^{(1)}_{R}={C\over B}{\cos\Theta \big[(k'^2_{3n}/\ell^2+m E_n+m^2)I+i(k'_{3n}/\ell)(k_2\sigma_1-k_1 \sigma_2+(m+E_n)\tan\Theta\sigma_3)\big]\over (m+E_n)(ik'_{3n}/\ell-m\cos\Theta)},
\label{rotU1}
\end{align}
where we have used the discrete momenta $k'_{3n}$ and the eigen energies
\begin{eqnarray}
E_n=\sqrt{k^2_1+k^2_2+\bigg({k'_{3n}\over \ell}\bigg)^2+m^2}.
\label{Discreteenergy}
\end{eqnarray}
Taking the correspondence between the obtained rotation operator (\ref{rotU1}) and the general formulation in Eq.~(\ref{rotoperatorspinspace}), we have
\begin{eqnarray}
&&e^{i\varsigma^{(1)} }\cos({\varphi^{(1)} \over 2})I
={C\over B}{\cos\Theta (k'^2_{3n}/\ell^2+m E_n+m^2)I\over (m+E_n)(ik'_{3n}/\ell-m\cos\Theta)}
,\\
&&e^{i\varsigma ^{(1)}}\sin({\varphi^{(1)} \over 2})\hat {\bm n}^{(1)} \cdot{\bm \sigma}
=-{C\over B}{\cos\Theta (k'_{3n}/\ell) \big[ k_2\sigma_1-k_1 \sigma_2+(m+E_n)\tan\Theta\sigma_3\big]\over (m+E_n)(ik'_{3n}/\ell-m\cos\Theta)}
,
\end{eqnarray}
which lead to the expression of the
rotation angle and its rotation axis at the first plate as follows
\begin{eqnarray}
&&\tan({\varphi^{(1)}\over 2})=-{(k'_{3n}/\ell)\sqrt{(2m^2+2mE_n+k'^2_{3n}/\ell^2
)\tan^2\Theta+{\bm k}^2_\perp\sec^2\Theta}\over (k'^2_{3n}/\ell^2+m E_n+m^2)},
\label{rotangle1}\\
&&\hat {\bm n}^{(1)}={k_2\hat x_1-k_1 \hat x_2+(m+E_n)\tan\Theta\hat x_3 \over \sqrt{(2m^2+2mE_n+k'^2_{3n}/\ell^2
)\tan^2\Theta+{\bm k}^2_\perp\sec^2\Theta}},
\label{rotaxis1}
\end{eqnarray}
respectively\footnote{The rotation angle and rotation axis can also be written as $\tan({\varphi^{(1)}/2})=-\tan \Theta (k'_{3n}/\ell)\sqrt{2m^2+2mE_n+k'^2_{3n}/\ell^2
+{\bm k}^2_\perp \cot^2\Theta}/ (k'^2_{3n}/\ell^2+m E_n+m^2)$ and $\hat {\bm n}^{(1)}={k_2\hat x_1-k_1 \hat x_2+(m+E_n)\tan\Theta\hat x_3 /(\tan\Theta \sqrt{2m^2+2mE_n+k'^2_{3n}/\ell^2
+{\bm k}^2_\perp\cot^2\Theta}})$, respectively. These expressions reduce to those in the previous work of Ref.~\cite{Rohim2021} in the case of suppressed perpendicular momentum ${\bm k}_\perp=0$. As has been mentioned in Ref.~\cite{Rohim2021}, one may choose the opposite sign of the above rotation axis. As the consequence, the rotation angle has an opposite sign to the above one. This condition holds not only for the first plate but also the second plate.}. It can be seen that the roles of the pure phase $e^{i\varsigma ^{(1)}}$ can be cancelled out from the rotation angle and its axis.
In the case of ${\bm k}_\perp\neq 0$, the reflected spin orientation changes for arbitrary chiral angles (both chiral and nonchiral cases). However, for the case of ${\bm k}_\perp=0$, it does not change for $\Theta=0, \pi$ (see, e.g., Refs.~\cite{Cruz2018, Bellucci2009} for the discussion on the relation of $\chi_R$ and $\chi_L$ under nonchiral boundary condition, where they used $B=C$).
At the second plate, from the relation provided in Eq.~(\ref{relationL1}), we have $\chi_L={\cal U}^{(2)}_L \chi_R,$ where ${\cal U}^{(2)}_L$ is the rotation operator generated by the reflection with the second plate given by
\begin{align}
{\cal U}^{(2)}_{L}={B e^{2ik'_{3n}}\over C}
{\cos\Theta \big[(k'^2_{3n}/\ell^2+m E_n+m^2)I-i(k'_{3n}/\ell)(k_2\sigma_1-k_1 \sigma_2+(m+E_n)\tan\Theta\sigma_3)\big]\over (m+E_n)(ik'_{3n}/\ell-m\cos\Theta)}.
\label{rotU2}
\end{align}
Again, taking the correspondence between the obtained rotation operator (\ref{rotU2}) and general formulation Eq.~(\ref{rotoperatorspinspace}), we obtain
\begin{align}
&e^{i\varsigma^{(2)} }\cos({\varphi^{(2)} \over 2})I
={B e^{2ik'_{3n}}\over C}{\cos\Theta (k'^2_{3n}/\ell^2+mE_n+m^2)I\over (m+E_n)(i(k'_{3n}/\ell)-m\cos\Theta)}
,\\
&e^{i\varsigma ^{(2)}}\sin({\varphi^{(2)} \over 2})\hat {\bm n}^{(2)} \cdot{\bm \sigma}
={B e^{2ik'_{3n}}\over C}{\cos\Theta (k'_{3n}/\ell) \big[ k_2\sigma_1-k_1 \sigma_2+(m+E_n)\tan\Theta\sigma_3\big]\over (m+E_n)(ik'_{3n}/\ell-m\cos\Theta)}.
\end{align}
Then, it is straightforward to show that the rotation angle and the rotation axis are given by
\begin{eqnarray}
&&\tan({\varphi^{(2)}\over 2})=-{(k'_{3n}/\ell)\sqrt{(2m^2+2mE_n+k'^2_{3n}/\ell^2
)\tan^2\Theta+{\bm k}^2_\perp\sec^2\Theta}\over (k'^2_{3n}/\ell^2+mE_n+m^2)},
\label{rotangle2}\\
&&\hat {\bm n}^{(2)}={-k_2\hat x_1+k_1 \hat x_2-(m+E_n)\tan\Theta\hat x_3 \over \sqrt{{\bm k}^2_\perp\sec^2\Theta+(2m^2+2mE_n+k'^2_{3n}/\ell^2
)\tan^2\Theta}},
\label{rotaxis2}
\end{eqnarray}
respectively\footnote{Similar to the result at the first plate, the rotation angle and the rotation axis at the second plate can also be written in the form of $\tan({\varphi^{(2)}/2})=-\tan \Theta k_{3n}\sqrt{2m^2+2mE_n+k'^2_{3n}/\ell^2
+{\bm k}^2_\perp \cot^2\Theta}/ (k'^2_{3n}/\ell^2+m E_n+m^2)$ and $\hat {\bm n}^{(2)}={-k_2\hat x_1+k_1 \hat x_2-(m+E_n)\tan\Theta\hat x_3 /(\tan\Theta \sqrt{2m^2+2mE_n+k'^2_{3n}/\ell^2
+{\bm k}^2_\perp\cot^2\Theta}})$, respectively. In the case of suppressed perpendicular momentum, they reduce to those of Ref.~\cite{Rohim2021}.}.
From the above results, one can show that $\tan({\varphi^{(1)}\over 2})=\tan({\varphi^{(2)}\over 2})$ and $\hat{\bm n}^{(1)}=-\hat{\bm n}^{(2)}$, which means the reflection at the first plate generates the same rotation angle as at the second plate, but their rotation axis are in the opposite direction. Recalling that at the first and second plates, we have $\chi_R={\cal U}^{(1)}\chi_L$ and $\chi_L={\cal U}^{(2)}\chi_R$, respectively,
which implies ${\cal U}^{(1)}{\cal U}^{(2)}={\cal U}^{(2)}{\cal U}^{(1)}=I$
for the allowed momenta (\ref{DiscreteMomenta}) under
the reflection in a consistent way.
\section{Casimir energy}
\label{CasimirEnergy}
In this section, we calculate the Casimir energy of a massive Dirac fermion confined between two parallel plates under the chiral MIT boundary conditions.
In the presence of this boundary condition, the vacuum energy is given by,
\begin{eqnarray}
E_0=-{L^2\over 2\pi^2}\int^{\infty}_{-\infty}dk_1\int^{\infty}_{-\infty}dk_2\sum^\infty_{n=1}\sqrt{k^2_1+k^2_2+\bigg({k'_{3n}\over \ell}\bigg)^2+m^2},
\label{VacuumE}
\end{eqnarray}
where $L^2$ is the surface area of the plate and the discrete momenta $k'_{3n}\equiv k_{3}\ell$ satisfies the momentum constraint in Eq. (\ref{DiscreteMomenta}). In the above expression, we have used the eigen energies of a Dirac field given in Eq. (\ref{Discreteenergy}).
In fact, the vacuum energy in Eq. (\ref{VacuumE}) is divergent. Then, to solve this problem, one can use
the Abel-Plana like summation \cite{Romeo2000} following the procedure used in Refs.~\cite{Bellucci2009, Cruz2018} as follows,
\begin{eqnarray}
\sum^{\infty}_{n=1} {\pi f(k'_{3n})\over \big(1 -{\sin(2 k'_{3n})\over 2 k'_{3n}}\big)} =-{\pi m'\cos\Theta f(0) \over 2(m'\cos\Theta+1)}+\int^\infty_0 dz f(z) -i\int^\infty_0 dt {f(it)-f(-it)\over {t+m'\cos\Theta\over t-m'\cos\Theta}e^{2t}+1}.
\label{AbelPlanaLike}
\end{eqnarray}
Considering the denominator of the left-hand side of Eq.~(\ref{AbelPlanaLike}), from the momentum constraint given in Eq.~(\ref{DiscreteMomenta}), we may replace it with the following relation
\begin{eqnarray}
1 -{\sin(2 k'_{3n})\over 2 k'_{3n}}= 1+{m' \cos\Theta\over (m'\cos\Theta)^2+k'^2_{3n}}.
\end{eqnarray}
Then, the above vacuum energy can be rewritten as follows
\begin{align}
E_0=-{L^2\over 2\pi^3 \ell }\int^{\infty}_{-\infty}dk_1\int^{\infty}_{-\infty}dk_2\bigg(-{\pi m'\cos\Theta f(0) \over 2(m'\cos\Theta+1)}+\int^\infty_0 dz f(z) -i\int^\infty_0 dt {f(it)-f(-it)\over {t+m'\cos\Theta\over t-m'\cos\Theta}e^{2t}+1}\bigg),
\label{VacE}
\end{align}
where the function $f(z)$ is defined as
\begin{eqnarray}
f(z)=\sqrt{(k^2_1+k^2_2)\ell^2+z^2+m'^2} \bigg(1+{m' \cos\Theta\over (m'\cos\Theta)^2+z^2}\bigg).
\label{fz}
\end{eqnarray}
In the next step, we can separate the above vacuum energy in Eq.~(\ref{VacE}) into three terms as follows,
\begin{eqnarray}
E_0=\ell E^{(0)}_0+2 E^{(1)}_0+\Delta E_0.
\label{threetermvac}
\end{eqnarray}
$E^{(0)}_0$ in the first term of Eq.~(\ref{threetermvac}) reads
\begin{eqnarray}
E^{(0)}_0=-{L^2\over 4\pi^3}\int^{\infty}_{-\infty}dk_1\int^{\infty}_{-\infty}dk_2 \int^{\infty}_{-\infty}
dk_3 \sqrt{k^2_1+k^2_2+k^2_3+m^2},
\end{eqnarray}
which represents the vacuum energy in the absence of the boundary conditions. The contribution of $E^{(1)}_0$ in the second term of Eq.~(\ref{threetermvac}) is given by
\begin{align}
E^{(1)}_0=-{L^2\over 4\pi^3}\int^{\infty}_{-\infty}dk_1\int^{\infty}_{-\infty}dk_2 \bigg[-{\pi\over 2}\sqrt{k^2_1+k^2_2+m^2}+m\cos\Theta\int^{\infty}_0 dz {\sqrt{z^2+k^2_1+k^2_2+m^2}\over m^2\cos^2\Theta+z^2}\bigg],
\end{align}
which is the vacuum energy in the presence of one plate only and it does not contribute to the Casimir force. Then,
the Casimir energy is related to the last term of Eq.~(\ref{threetermvac}), $E_{\rm Cas.}\equiv \Delta E_0$, and it is given as
\begin{eqnarray}
E_{\rm Cas.}={iL^2\over 2\pi^3 \ell }\int^{\infty}_{-\infty}dk_1\int^{\infty}_{-\infty}dk_2\int^\infty_0 dt {f(it)-f(-it)\over {t+m'\cos\Theta\over t-m'\cos\Theta}e^{2t}+1}.
\end{eqnarray}
By using Eq.~(\ref{fz}), the above Casimir energy can be rewritten as,
\begin{eqnarray}
E_{\rm Cas.}&=&{iL^2 \ell \over 2\pi^3 }\int^{\infty}_{-\infty}dk_1\int^{\infty}_{-\infty}dk_2\int^\infty_0 du {u-m\cos\Theta \over (u+m\cos\Theta)e^{2\ell u}+u-m\cos\Theta}
\nonumber\\
&&
\times \bigg(1+{m\cos\Theta\over \ell m^2\cos\Theta-\ell u^2}\bigg) \bigg[\sqrt{(iu)^2+k^2_1+k^2_2+m^2}-\sqrt{(-iu)^2+k^2_1+k^2_2+m^2}\bigg],\nonumber\\
\label{Casseparate}
\end{eqnarray}
where we have changed the variable,
$t=\ell u$. It is convenient to separate the integration over $u$ in Eq.~(\ref{Casseparate}) into two intervals. The first interval that is from $0$ to $\sqrt{k^2_1+k^2_2+m^2}$ vanishes, whereas the second interval that is from $\sqrt{k^2_1+k^2_2+m^2}$ to $\infty$ remains.
Then, the Casimir energy reads
\begin{eqnarray}
E_{\rm Cas.}&=&{-L^2 \over \pi^3 }\int^{\infty}_{-\infty}dk_1\int^{\infty}_{-\infty}dk_2\int^\infty_{\sqrt{k^2_1+k^2_2+m^2}} du \sqrt{u^2-k^2_1-k^2_2-m^2} \nonumber\\ & & \times
\bigg({ \ell(u-m\cos\Theta)-m\cos\Theta/(m\cos\Theta+u)
\over (u+m\cos\Theta)e^{2\ell u}+(u-m\cos\Theta)}\bigg).
\end{eqnarray}
To further proceed with the above integration, we then use the following formula \cite{Bellucci2009}
\begin{align}
&\int_{-\infty}^{\infty} d{\bm k}_\perp\int^\infty_{\sqrt{{\bm k}^2_\perp+c^2}} dz(z^2-{\bm k}^2_\perp-c^2)^{(s+1)/2} f(z)\nonumber\\
&~~~~~~~~~~~~~~~~~~ ={\pi^{p/2}\Gamma((s+3)/2)\over \Gamma((p+s+3)/2)}\int_c^\infty dx (x^2-c^2)^{(p+s+1)/2}f(x).
\end{align}
Then, the Casimir energy becomes
\begin{equation}
E_{\rm Cas.}={-L^2 \over \pi^3}{\pi\Gamma(3/2)\over \Gamma(5/2)}
\int^\infty_m dx
(x^2-m^2)^{3/2}
\bigg({\ell(x-m\cos\Theta)-m\cos\Theta/(m\cos\Theta+x)
\over (x+m\cos\Theta)e^{2\ell x}+(x-m\cos\Theta)}\bigg).
\end{equation}
We next rewrite the following factor
\begin{eqnarray}
{\ell(x-m\cos\Theta)-m\cos\Theta/(m\cos\Theta+x)
\over (x+m\cos\Theta)e^{2\ell x}+(x-m\cos\Theta)}=-{1\over 2}{d\over dx}\ln{\bigg(1+{x-m\cos\Theta\over x+m\cos\Theta}e^{-2\ell x}\bigg)}
\end{eqnarray}
and introduce a new variable $y=\ell x-m'$. Then, the Casimir energy now reads
\begin{eqnarray}
E_{\rm Cas.}={-L^2 \over \pi^2 \ell^3 } {2\over 3} \int^\infty_0 dy (y^2+2y m')^{3/2}
{-1\over 2}{d\over dy}\ln{\bigg(1+{y+ m'(1-\cos\Theta)\over y+ m'(1+\cos\Theta)}e^{-2(y+m')}\bigg)}.\nonumber\\
\label{intbypart}
\end{eqnarray}
Performing integration by parts in Eq.~(\ref{intbypart}), we have a simpler form of the Casimir energy as
\begin{eqnarray}
E_{\rm Cas.}={-L^2 \over \pi^2 \ell^3 }\int^\infty_0 dy (y+m') \sqrt{y(y+2 m')}\ln{\bigg(1+{y+ m'(1-\cos\Theta)\over y+ m'(1+\cos\Theta)}e^{-2(y+m')}\bigg)}.
\label{intCasimirEnergy}
\end{eqnarray}
\begin{figure}[tbp]
\centering
\includegraphics[width=.49\textwidth]{FigCasEnergy1a.pdf}
\hfill
\includegraphics[width=.49\textwidth]{FigCasEnergy1b.pdf}
\caption{\label{CasimirEnergyplot1} The Casimir energy per plate area. The left panel shows the Casimir energy as a function of the chiral angle $\Theta$ and dimensionless parameter $m'(\equiv m\ell)$. In the right panel, the Casimir energy is demonstrated as a function of the parameter $m'$ for various values of the chiral angle. We used the fixed plates' distance $\ell=1/5~\mu\text{m}$ and plate's size $L^2=1~\text{cm}^2$. The curves show that in the massless case, the Casimir energy for all chiral angles gives the same value, which is analytically given in Eq.~(\ref{masslessCasimir}). They also show that the chiral angle $\Theta=\pi$ gives maximum contribution. The role of the chiral angle changes depending on the mass $m$ and plates' distance $\ell$.}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[width=.49\textwidth]{FigCasEnergy1e.pdf}
\hfill
\includegraphics[width=.49\textwidth]{FigCasEnergy1f.pdf}
\caption{\label{CasimirEnergyplot2} The ratio of the Casimir energy for an arbitrary chiral case to the nonchiral case. The left panel shows the curves as a function of the parameter $m'$ for several values of the chiral angle. The right panel shows the curves as a function of the chiral angle for the given parameter $m'$.}
\end{figure}
In what follows, we will discuss the Casimir energy based on Eq.~(\ref{intCasimirEnergy}). The left panel of Fig.~\ref{CasimirEnergyplot1} depicts the behavior of the Casimir energy as the function of the chiral angle $\Theta$ and the parameter $m'(\equiv m\ell)$, while the right panel demonstrates the Casimir energy as the function of the parameter $m'$ for several values of the chiral angle. In both panels, we have used the fixed plates' distance. In general, one can see that the Casimir energy has symmetrical shapes with respect to the chiral angle $\Theta=\pi$.
In the case of the light mass, $m'\ll 1$, the right panel of Fig.~\ref{CasimirEnergyplot1} is approximately given by the linear function, as shown below. While in the case of the heavy mass, $m'\gg 1$, the Casimir energy tends to zero. In the massless case, Fig.~\ref{CasimirEnergyplot1} shows that the Casimir energy gives the same value for any chiral angle; it is explicitly given by
\begin{eqnarray}
E_{\rm Cas.}={-L^2 \over \pi^2 \ell^3 } \int^\infty_0 dy y^2\ln (1+e^{-2y})=-{7 L^2\pi^2\over 2880 \ell^3},
\label{masslessCasimir}
\end{eqnarray}
which is consistent with that of Ref.~\cite{KJohnson1975}. Both right and left panels show that the maximum contribution is given by the chiral angle $\Theta=\pi$, while the minimum contribution is given by $\Theta=0$ (nonchiral case).
Figure~\ref{CasimirEnergyplot2} plots the ratio of the Casimir energy for an arbitrary chiral case to the nonchiral case. The curves show that the ratio is maximum when the chiral angle takes the value $\Theta=\pi$. This ratio increases as the parameter $m'$ becomes larger.
Since the Casimir energy is symmetrical as the function of the chiral angle, the ratio $E_{\rm Cas.}(m',\Theta)/E_{\rm Cas.}(m',0)$ is also symmetrical with respect to $\Theta=\pi$ (see the right panel of Fig.~\ref{CasimirEnergyplot2}).
\begin{figure}[tbp]
\centering
\includegraphics[width=.49\textwidth]{FigCasEnergy1c.pdf}
\hfill
\includegraphics[width=.49\textwidth]{FigCasEnergy1d.pdf}
\caption{\label{CasimirEnergyplot3} The Casimir energy per plate area as a function of $\ell$. The left panel corresponds to the dependence of the Casimir energy on the mass $m$ for the nonchiral case. The right panel corresponds to the dependence of the Casimir energy on the chiral angle $\Theta$ with fixed mass $m=1$ eV and plate's size $L^2=1~\text{cm}^2$.}
\end{figure}
Figure~\ref{CasimirEnergyplot3} plots the Casimir energy as a function of the plates' distance $\ell$. The curves in the left panel show that the Casimir energy of a massive Dirac fermion approaches the massless case as the reduction of its mass. From the right panel, one can see that the Casimir energy in the nonchiral case decays faster than in the chiral case. Meanwhile, the slowest decaying occurs in the case of the chiral angle $\Theta=\pi$.
We next turn to consider the analytical evaluation of the Casimir energy in Eq.~(\ref{intCasimirEnergy}).
In the case of $m'\ll 1$,
the Casimir energy for arbitrary chiral angle approximately reduces to
\begin{eqnarray}
E_{\rm Cas.}&\simeq&{-L^2 \over \pi^2 \ell^3 } \int^\infty_0 dy \bigg[ y^2\ln (1+e^{-2y})-\bigg({2y {e^{-2y}} (y+\cos\Theta )\over (1+e^{-2y})}-2y\ln (1+e^{-2y})\bigg)m'\bigg]\nonumber\\
&=&-{7 L^2\pi^2\over 2880 \ell^3} \bigg(1-{120 m' \cos\Theta \over 7\pi^2}\bigg).
\label{CasimirEnergymless1}
\end{eqnarray}
Compared to the massless case (\ref{masslessCasimir}), the above expression has an additional term that contributes to the Casimir energy. We also note that this term shows the Casimir energy dependence on the chiral angle. In the case of $m' \gg 1$, the Casimir energy (\ref{intCasimirEnergy}) is approximately given by
\begin{eqnarray}
E_{\rm Cas.} &\simeq& {-L^2 \over \pi^2 \ell^3 } \int^\infty_0 dy\nonumber\\
&&\times \bigg[ \bigg( {\sqrt{2m'^3}y^{1/2} (1-\cos\Theta)\over 1+\cos\Theta}+\sqrt{m'\over 2}{y^{3/2}(5+8\cos\Theta-5\cos^2\Theta)\over 2(1+\cos\Theta)^2}\bigg)e^{-2(y+m')}\bigg]\nonumber\\
&=&{-L^2 m'^{1/2}e^{-2m'} \over 64 (1+\cos\Theta)^2\pi^{3/2} \ell^3 }\bigg[{16 m'\sin^2\Theta} +{3(5+8\cos\Theta-5\cos^2\Theta)}\bigg].
\label{CasimirEnergymgreater1}
\end{eqnarray}
In this case, the Casimir energy
decays and convergent go to zero as the increase of the parameter $m'$.
From the above Casimir energy, one can investigate the Casimir force $F_{\rm Cas.}=-d E_{\rm Cas.}/d\ell$ as well as the Casimir pressure $P_{\rm Cas.}=F_{\rm Cas.}/L^2$. From Eqs.~\eqref{CasimirEnergymless1} and \eqref{CasimirEnergymgreater1}, one can obtain the following Casimir pressure,
\begin{align}
P_{\rm Cas.} =\left\{
\begin{array}{ll}
-{7 \pi^2 \over 960 \ell^4}\bigg(1-{80 m' \cos\Theta \over 7\pi^2}\bigg), ~~ &\text{for $m'\ll 1$},\\
-{m'^{1/2}e^{-2m'} \over 64 (1+\cos\Theta)^2\pi^{3/2} \ell^4 }\\
\times\bigg[{ 8( 4m'+3)m'\sin^2\Theta} +(6 m'+{5\over 2})(5+8\cos\Theta-5\cos^2\Theta)\bigg], ~~ &\text{for $m'\gg 1$}.
\end{array}
\right.
\label{CasimirPressure}
\end{align}
The above Casimir pressures with respect to the chiral angle have a similar behavior to those of the Casimir energy. In other word, the attractive Casimir force in the chiral case is always stronger than that in the nonchiral case.
\section{Summary and conclusion}
\label{Summary}
We have studied the behavior of a massive Dirac fermion confined between two parallel chiral plates. In our setup, the plates are represented by chiral MIT boundary conditions \cite{Lutken1984}, where the condition of vanishing probability current density at the boundary surfaces is satisfied for arbitrary chiral angles. The Dirac field inside the confinement area consists of two-component fields associated with their spin orientations. We discussed the general discrete momenta and the changes of the spin orientations under the boundary conditions. The result shows that only momentum $k_3$ (parallel component to the boundary surface) is discretized depending on the mass, plates' distance, and the chiral angle. We also found that, in the case of non-zero perpendicular momenta, the spin orientation changes for arbitrary chiral angle. In the case of suppressed perpendicular momenta, we found that these features reduce to those of the previous work \cite{Rohim2021} for a confinement system in a one-dimensional box.
We have also investigated the effect of the chiral angle on the Casimir energy of a massive fermion. To obtain the Casimir energy, we calculated the vacuum energy in the presence of the boundary condition. Unfortunately, this vacuum energy is divergent. To solve this issue, we adopt the Abel-Plana like summation \cite{Romeo2000}, as previously used in Refs.~\cite{Cruz2018, Bellucci2009}.
The obtained vacuum energy consists of three parts (see Eq.~\eqref{threetermvac}), i.e., (i) the vacuum energy in the absence of the boundary conditions, (ii) the vacuum energy in the presence of a single plate, (iii) the Casimir energy. We notice that the second term of the vacuum energy in Eq.~(\ref{threetermvac}) for a single plate is not relevant to the Casimir force. The Casimir energy is defined by taking the differences between the vacuum energy of a Dirac field in the presence of the boundary conditions to that in the absence of one. We investigated the behavior of the Casimir energy as well as its attractive force using numerical analysis.
The result shows that the Casimir energy of a massive fermion can be written as a function of the chiral angle. It is symmetrical with the maximum contribution $\Theta=\pi$; the Casimir energy in the chiral case is always higher than that in the nonchiral case.
Besides the above results, we also found that the behavior of the Casimir energy depends on the values of the Dirac fermion's mass. In the analysis, we investigated two approximation cases, i.e., light and heavy masses. In the case of heavy fermion mass, the Casimir energy converges to zero as the mass increases. While in the case of light fermion mass, the Casimir energy converges to that of the massless fermion as the mass decreases. We found that, in both cases, the roles of the chiral angle become weaker. For the case of the massless fermion, the Casimir energy gives the same value for all chiral angles. In this case, we recover the expected result by Ref.~\cite{KJohnson1975}. For future work, it will be interesting to study a similar setup by including a background such as magnetic field (c.f., Ref.~\cite{Sitenko:2014kza}).
\acknowledgments
This work was started when A.~R. stayed at Kyushu University, Japan as an Academic Researcher (Post-Doctoral fellow) and was completed during he was the Post-Doctoral program at the National Research and Innovation Agency (BRIN), Indonesia. A.~R. would like to thank the Theoretical Astrophysics Laboratory of Kyushu University for their kind hospitality.
|
\section{Introduction}
~~~~ It is well known that the piston problem is an initial-boundary value problem in fluid dynamics. As shown in [6, 23-24], a thin and long tube with a piston at one end and open at the other end, which is initially filled with static gas, then any movement of the piston will cause the corresponding movement of the gas in the tube. Determining the state of the gas and the propagation of the nonlinear waves in the tube is called a piston problem.
In [4, 7] the authors took the one-dimensional piston problem as a model to analyze the occurrence and the motion of the basic nonlinear waves in the isentropic fluid, which shows the importance of the piston problem. In this paper, we study the one-dimensional piston problem of the isentropic Euler equations [2, 10, 14, 17, 20, 26]
\begin{equation}\label{key}
\left\{\begin{array}{ll} \rho_{t}+(\rho u)_{x}=0,\\
(\rho u)_{t}+(\rho u^{2}+P)_{x}=0,
\end{array}\right . \tag{1.1}
\end{equation}
where $\rho $, $u$ and $P$ are the density, velocity and scalar pressure of certain fluid flow located at position $x$ and time $t$, respectively.
The system (1.1) has been extensively used in numerous fields due to it's important physical significance. Brenier [2] study the solutions with concentration to the Riemann problem for the isentropic Chaplygin gas dynamics system. Guo, Sheng and Zhang [11] solved completely this problem, where the delta shock wave solutions were constructed. Roughly speaking, the delta shock wave solution is a solution such that at least one of the variables contains Dirac delta function [3, 5, 21]. In addition, Wang [22] obtained the solutions to the Riemann problem for system (1.1) with the generalized Chaplygin gas, and the formation of delta shock wave was analyzed. Sheng, Wang and Yin [18] clarified the limit behavior of these solutions as the pressure vanishes. Yang and Wang [25] further studied the system (1.1) for the modified Chaplygin gas, and the limit behavior of constructed Riemann solutions was analyzed as the pressure vanishes. They also analyzed and identified two kinds of occurrence mechanism on the phenomenon of concentration and the formation of delta shock wave in vanishing pressure limit of solutions to the modified Chaplygin gas equations [26].
As for piston problem, there are numerous excellent papers for related equations and results. Qu, Yuan and Zhao [16] studied the high Mach number limit of the piston problem for the full isentropic Euler equations of polytropic gas. They also study the one-dimensional piston problem for (1.1) of Chaplygin gas [17], which might be used as an approximate model of polytropic gases in aerodynamics. In their work, a Radon measure solution is constructed when the piston moves forward at sonic or supersonic speed, and the existence of a shock wave solution is proved as the piston moves forward at subsonic speed, as well as an integral weak solution when the piston recedes from the gas with any constant speed. Fan et.al [10] paid attention to
the generalized Chaplygin gas, which substantial difference with Chaplygin gas lies in that its two characteristic fields are genuinely nonlinear, and
obtained the similar results for the piston problem of the generalized Chaplygin Euler equations. Ding [8] established the global stability of large shock waves to the piston problem for the compressible Magnetohydrodynamics under small BV perturbations of initial data and the piston speed.
However, the above works have been focused on the cases that the pressure equation only contains one parameter. It is necessary to investigate the piston problem, in which the pressure equation contains more than one parameter.
In 2002, the modified Chaplygin gas (MCG) was first proposed by Benaoum [1] to describe the current accelerated expansion of the universe, which is given by
\begin{equation}\label{key}
P = A\rho - \frac{B}{\rho^{\alpha}}, ~~~~~~ 0 < \alpha \leq 1, \tag{1.2}
\end{equation}
where $A, B > 0$ are two small positive constants, with $A$ representing an ordinary fluid that is subject to linear positive pressure and $B$ relating pressure to some positive power of the inverse for the energy density. It is obvious to see that, when $B = 0$ in (1.2), the pressure is transformed into the standard equation of state of perfect fluid.
When $A = 0$, (1.2) reduces into the generalized Chaplygin gas
\begin{equation}\label{key}
P = - \frac{B}{\rho^{\alpha}}, ~~~~~~ 0 < \alpha \leq 1, \tag{1.3}
\end{equation}
which was extensively studied to compute the lifting force in aerodynamics and analyze dark energy related mathematical problems. Furthermore, as $\alpha = 1$, (1.3) represents the Chaplygin gas,
\begin{equation}\label{key}
P = - \frac{B}{\rho}, \tag{1.4}
\end{equation}
which owns a negative pressure and occurs in certain theories of cosmology.
Thus, both the Chaplygin and generalized Chaplygin gases are the special cases for the modified Chaplygin gas. Generally, the modified Chaplygin gas can describe the universe to a large extent and explain the mysterious nature of the dark energy and dark matter (see [9, 13]). See the detailed results in [12, 19] and the references cited therein.
In the present paper, we want to consider the piston problem for the system (1.1) with modified Chaplygin gas (MCG)
\begin{equation}\label{key}
P = A\rho - \frac{B}{\rho^{\alpha}}, ~~~~~~ 0< \alpha < 1. \tag{1.5}
\end{equation}
In comparison with
Chaplygin and generalized Chaplygin gases [10, 15, 17], the modified Chaplygin gas has two parameters and can be regard as a combination of the standard fluid and generalized Chaplygin gas, which is more complicated and general; Meanwhile, the analysis of solvability for the piston problem of (1.1) and (1.5) is more complex, especially for the receding case. Recall that the local sound speed of the modified Chaplygin gas is given by
\begin{equation}\label{key}
c = \sqrt{P'(\rho)} = \sqrt{A+\frac{B \alpha}{\rho^{\alpha+1}}}. \tag{1.6}
\end{equation}
Then, the Mach number $M_{0}$ is defined by
\begin{equation}\label{key}
M_{0} = \frac{ |v_{0}| }{c_{0}},\tag{1.7}
\end{equation}
where $v_{0}$ is the move speed of piston, $c_{0}$ is as in (1.6) for $\rho = \rho_{0}$. The piston are said to be subsonic, sonic and supersonic when $M_{0} < 1$, $M_{0} = 1$, $M_{0} > 1$, respectively. In this paper, we discover that there always exists a shock wave connecting two constant states as the piston rushes into the gas at any constant speed, which quite differs from the results in [10, 17]. Moreover, we prove strictly that only one rarefaction wave is a physically feasible solution, which appears in front of the piston as the piston moves backward from the gas. Finally, we analyze the limiting behavior of piston solutions to the system (1.1) and (1.5) as the modified Chaplygin gas pressure converges to generalized Chaplygin gas.
The organization of this paper is as follows. In Section 2, we formulate the piston problem of modified Chaplygin gas equations, and shift the coordinates system to move with the piston by Galilean transformation. In Section 3, we briefly recall the results for the piston problem of (1.1) with (1.3) as the piston rushes to the gas. In Section 4, the piston solutions to the modified Chaplygin Euler equations of one-dimensional isentropic fluid flow are analyzed and proven in detail as two parameters $A, B>0$, and two main theorems for the piston problem of (1.1) and (1.5) are presented. In Section 5, when the pressure fades away and $A \rightarrow 0 $, the formation mechanism of singularity is rigorously discussed in the limiting process, and a Radon measure solution is constructed and proven to be a limiting solution of the system (1.1) and (1.5), which is similar with the piston solutions of (1.1) and (1.3).
\section{The piston problem of Modified Chaplygin gas}
~~~~ In this section, we describe the piston problem of modified Chaplygin gas in more details.
Let the x-axis be the tube, $x = v_{0}t$ is the movement curve of the piston. Suppose the gas fills the domain $\left\lbrace x <0 \right\rbrace $ initially and its velocity and density are $u=0$, $\rho=\rho_{0}$, that is
\begin{equation}\label{key}
U_{0} = (\rho, u) |_{t=0} = (\rho_{0}, 0). \tag{2.1}
\end{equation}
Then the time-space domain is given by
\begin{equation}\label{key}
\Omega_{t} = \left\lbrace (t, x) : x< v_{0}t, ~~~t>0\right\rbrace. \tag{2.2}
\end{equation}
Moreover, the piston is subject to the usual impermeable condition
\begin{equation}\label{key}
\rho u(t, x) = 0 ~~~~~ on~~~ x = v_{0}t. \tag{2.3}
\end{equation}
The goal of piston problem is to find a solution of (1.1) and (1.5) in the domain $\Omega_{t}$, satisfying (2.1) and (2.3). For the convenience of treating the piston problem, we use Galilean transformation to shift the coordinates to move with the piston:
\begin{equation}\label{key}
\left\{\begin{array}{ll}
t' = t,\\
x' = x - v_{0}t,\\
\rho'(t', x') = \rho(t', x'+v_{0}t'),\\
u'(t', x') = u(t', x'+v_{0}t')-v_{0},\\
P'(t', x') = P(t', x'+v_{0}t').
\end{array}\right. \tag{2.4}
\end{equation}
\noindent
It is easy to verify that the equations in (1.1) are invariant under (2.4), and the domain $\Omega_{t}$ is reduced to $\Omega' = \left\lbrace (t', x') : x'< 0, t'>0\right\rbrace .$ For convenience of statement, we drop all the primes "$'$" and write $\Omega'$ by $\Omega = \left\lbrace (t, x) : x< 0, t>0\right\rbrace $ without confusion. Then the initial condition becomes
\begin{equation}\label{key}
(\rho, u)|_{t=0} = (\rho_{0}, -v_{0}), \tag{2.5}
\end{equation}
and the boundary condition becomes
\begin{equation}\label{key}
\rho u(t, x) = 0 ~~~~~ on~~~ x = 0. \tag{2.6}
\end{equation}
\noindent
In what follows, we perform the non-dimensional linear transformations of independent and dependent variables, which corresponds to some similarity laws in physics [17]:
\begin{equation}\label{key}
\tilde{t} = \frac{t}{T},~~~~
\tilde{x} = \frac{x}{L},~~~~
\tilde{\rho} = \frac{\rho}{\rho_{0}},~~~~
\tilde{u} = \frac{\sqrt{2}u}{\left| v_{0}\right|},~~~~
\tilde{P} = \frac{2P}{\rho_{0}v_{0}^{2}}, \tag{2.7}
\end{equation}
where $T$ and $L > 0$ are constants with $\frac{L}{T} = \frac{v_{0}}{\sqrt{2}} $. Inserting (2.7) into (1.1) and simplifying, we can obtain that
\begin{equation}\label{key}
\left\{\begin{array}{ll} \tilde{\rho}_{\tilde{t}} + (\tilde{\rho} \tilde{u})_{\tilde{x}} = 0,\\
(\tilde{\rho} \tilde{u})_{\tilde{t}} + (\tilde{\rho} \tilde{u}^{2}+ \tilde{P})_{\tilde{x}} = 0,
\end{array}\right. \tag{2.8}
\end{equation}
which, together with (2.7), implies that $\tilde{\rho}$, $\tilde{u}$, $\tilde{P}$ still satisfies (1.1) and hence (1.1) is invariant under (2.7), that is to say, the $\rho$, $u$ and $P$ in (1.1) are equivalent to $\tilde{\rho}$, $\tilde{u}$, $\tilde{P}$ in (2.8). So, from (2.7), in the following we shall take initial data as
\begin{equation}\label{key}
\rho_{0} = 1,~~ v_{0} = \pm \sqrt{2}. \tag{2.9}
\end{equation}
Furthermore,
by using (1.7) and $c_{0} = \sqrt{A+\frac{B \alpha}{\rho_{0}^{\alpha+1}}}$, we can obtain
\begin{equation}\label{key}
A \rho_{0}^{\alpha+1} + B \alpha= \frac{\rho_{0}^{\alpha+1} v_{0}^{2}}{M_{0}^{2}}, ~~~~~~
B = \frac{\rho_{0}^{\alpha+1} v_{0}^{2}-A\rho_{0}^{\alpha+1} M_{0}^{2}}{\alpha M_{0}^{2}}.
\tag{2.10}
\end{equation}
Then, (1.5) becomes
\begin{equation}\label{key}
P(\rho) = A\rho - \frac{\rho_{0}^{\alpha+1} v_{0}^{2}-A\rho_{0}^{\alpha+1} M_{0}^{2}}{\alpha M_{0}^{2} \rho^{\alpha}}. \tag{2.11}
\end{equation}
Combining (2.7) with (2.11) and replacing $\frac{\rho_{0}}{\rho}$ by $\frac{1}{\tilde{\rho}}$, we obtain
\begin{equation}\label{key}
\tilde{P}(\tilde{\rho}) = \frac{2A \tilde{\rho}}
{ v_{0}^{2}} - \frac{2(\rho_{0}^{\alpha+1} v_{0}^{2} - A\rho_{0}^{\alpha+1} M_{0}^{2} )}
{\alpha M_{0}^{2} \rho_{0}^{\alpha+1} v_{0}^{2} \tilde{\rho}^{\alpha}}. \tag{2.12}
\end{equation}
Since $\tilde{P}$, $\tilde{\rho}$ in (2.12) are equivalent to corresponding $P$, $\rho$ in (2.11), we conclude that $\rho_{0} = 1$, $ v_{0} = \pm \sqrt{2}$, which further verifies the consistency of initial data in (2.9). For simplicity of writing, we drop the tildes hereafter. Then, combining (2.9) and (2.10), one can obtain
\begin{equation}\label{key}
A + B\alpha = \frac{2}{M_{0}^{2}}. \tag{2.13}
\end{equation}
In summary, the initial data are defined by
\begin{equation}\label{key}
\rho_{0} = 1,~~ v_{0} = \pm \sqrt{2},~~ P_{0} = A-B = \frac{2}{ M_{0}^{2}} -B(1+\alpha). \tag{2.14}
\end{equation}
\section{Piston solutions to system (1.1) and (1.3)}
~~~~ For our discussion later, let us briefly recall some results, especially the construction of Radon measure solutions, on
the piston problem of isentropic Euler equations (1.1) with generalized Chaplygin gas (1.3).
First, we restate the piston problem as
\begin{equation}\label{key}
\left\{\begin{array}{ll} \rho_{t}+(\rho u)_{x}=0,\\
(\rho u)_{t}+(\rho u^{2} - \frac{B}{\rho^{\alpha}})_{x}=0,\\
(\rho, u)|_{t=0} = (\rho_{0}, -v_{0}),\\
\rho u(t, x) = 0 ~~~~~ on~~~ x = 0,\\
\rho_{0} = 1, ~~~~~v_{0} = \pm 1,~~~~~P_{0} = - \frac{1}{\alpha M_{0}^{2}}.\\
\end{array}\right. \tag{3.1}
\end{equation}
It is easy to know that the piston problem for (1.1) with Chaplygin gas (1.4) is as in (3.1) for $\alpha = 1$. Notice that the Radon measure solution is not a delta shock, since the singular part is supported on the boundary of the domain. Now, we recall some definitions about Radon measure solution [17], which will be employed in this paper.
A Radon measure $m$ on the upper plane $[0, \infty) \times R$ could act on the compactly supported continuous functions
\begin{equation}\label{key}
<m, \phi> = \int_{0}^{\infty}\int_{R}\phi(t,x)m(dxdt), \tag{3.2}
\end{equation}
where the test function $\phi \in C_{0}([0, \infty)\times R)$. The standard Lebesgue measure $\mathit{L}^{2}$ is one example of Radon measure on $R^{2}$. The other example is the following Dirac measure supported on a curve [17].
\noindent \textbf{Definition 3.1.}~~Let $\mathcal{L}$ be a Lipschitz curve given by $x=x(t)$ for $t$ $ \in$ $ [0,T)$, and $w_\mathcal{L}(t) \in L_{loc}^{1}(0,T)$. The Dirac measure $w_{\mathcal{L}}\delta_{\mathcal{L}}$ supported on $\mathcal{L} \subset R^{2} $ with weight $w_{\mathcal{L}}$ is defined by
\begin{equation}\label{key}
<w_{\mathcal{L}}\delta_{\mathcal{L}}, \phi> = \int_{0}^{T}\phi(t, x(t)) w_\mathcal{L}(t) \sqrt{x'(t)^2+1}dt,~~~~~~\forall \phi \in C_{0}(R^2). \tag{3.3}
\end{equation}
With this definition, a Radon measure solution can be introduced to construct the solution of (1.1).
Recall that for two measures $\mu$ and $\nu$, the standard notation $\mu \ll \nu$ means $\nu$ is nonnegative and $\mu$ is absolutely continuous with respect to $\nu$. Now we could formulate the piston problem rigorously by introducing the following definition of measure solutions.
\noindent \textbf{Definition 3.2.}~~ For fixed $0<M_{0}\leq \infty$, let $ \tau, m, n, \wp$ be Radon measures on $\overline{\Omega} = \overline{\{(t, x): x<0,~~ t>0\}}$, and $w_{p}$ a locally integrable nonnegative function on $[0, \infty)$. Then $(\tau, u, w_{p})$ is called a measure solution to the piston problem (3.1) or (3.1) for $\alpha = 1$, provided that
\noindent i) $m\ll \tau$, $n\ll m$, and they have the same Radon-Nikodym derivative $u$; namely
\begin{equation}\label{key}
u \triangleq \dfrac{m(dxdt)}{\tau(dxdt)} = \dfrac{n(dxdt)}{m(dxdt)};
\tag{3.4}
\end{equation}
ii) For any $\phi \in C_{0}^{1}(R^{2})$, there hold
\begin{equation}\label{key}
<\tau, \partial t \phi> + <m, \partial x \phi> + \int_{-\infty}^{0}\rho_{0}\phi(0, x)dx = 0, \tag{3.5}
\end{equation}
\begin{equation}\label{key}
<m, \partial t \phi> + <n, \partial x \phi> +<\wp, \partial x \phi> - <w_{p}\delta_{\{x=0, ~t \geq 0\} }, \phi>
+ \int_{-\infty}^{0}(\rho_{0} u_{0})\phi(0, x)dx = 0, \tag{3.6}
\end{equation}
iii) If $\tau \ll \mathit{L}^2$ with derivative $\rho(t, x)$ in a neighborhood of $(t, x) \in [0, \infty) \times (-\infty, 0]$, and $\wp \ll L^{2}$ with derivative $P(t, x)$ there, then $L^{2}$-a.e. there holds
\begin{equation}\label{key}
P = -\frac{1}{\alpha \rho^{\alpha}}\frac{1}{M_{0}^{2}} ~~~~~or~~~~~P = -\frac{1}{\rho}\frac{1}{M_{0}^{2}}. \tag{3.7}
\end{equation}
and in addition, the classical entropy condition holds for discontinuities of functions $\rho, u$ near $(t, x)$.
Since the Radon measure solutions exist only when the piston moves toward the generalized Chaplygin gas (or Chaplygin gas), we just present the corresponding results for the proceeding piston problem in (3.1) and (3.1) for $\alpha = 1$. For the results and relevant proofs of the receding cases, readers can refer to [10, 17], we omit here.
\noindent \textbf{Theorem 3.1.} ~~~ For $v_{0} = -1$, as the piston rushes into the generalized Chaplygin gas (or Chaplygin gas) at Mach number $M_{0} \in (0, \sqrt{\alpha^{-1}})$~ (or $0< M_{0} < 1$),
there exists a shock wave solution of (3.1) (or (3.1) for $\alpha = 1$) connecting states $V_{0} = (1, 1)$ and $V_{1} = (\rho_{1}, 0)$ in the domain $\Omega$, while if Mach number $M_{0} \geq \sqrt{\alpha^{-1}}$ ~(or $M_{0}\geq1$), the Radon measure solution defined above is proved to be a reasonable solution for (3.1) (or (3.1) for $\alpha = 1$) in the domain $\Omega$.
\section{Piston solutions to system (2.8) and (1.5)}
~~~~ In this section, we study the piston problem (2.8), (2.6) and (2.14) for both cases that the piston rushes into and recedes from the MCG (1.5) at a constant speed. The global existence of solution to piston problem is established and clearly described.
\subsection{Proceeding piston problem}
~~~~Since a shock wave appears ahead of the piston when it pushes into the gas, we first give the definition of the integral weak solution for problem (2.8), (2.6) and (2.14).
\noindent \textbf{Definition 4.1.}~~ We say $\left(\rho, u \right)\in L^{\infty}(\left[ 0,\infty\right) \times\left( -\infty,0\right] ) $ is an integral solution to the problem (2.8), (2.6) and (2.14), if for any $\phi \in C^{1}_{0}(R^{2})$, there hold
\begin{equation}\label{key}
\left\{\begin{array}{ll} \int_{\Omega}(\rho \partial t \phi+\rho u \partial x \phi)dxdt + \int_{-\infty}^{0}\rho_{0}(x)\phi(0,x)dx = 0,\\
\int_{\Omega}(\rho u \partial t \phi ++( \rho u^{2}+P) \partial x \phi)dxdt -
\int_{0}^{\infty}P(t,0) \phi(t,0)dt
+ \int_{-\infty}^{0}\rho_{0}(x)u_{0}(x) \phi(0,x)dx = 0. \tag{4.1}
\end{array}\right .
\end{equation}
Noticing that the problem (2.8), (2.6) and (2.14) is a Riemann problem with boundary conditions for fixed $M_{0} \in (0, \infty)$. Since the equations (2.8), the initial data (2.14) as well as the boundary condition (2.6) are invariant by the transformation
$$
(x, t)\rightarrow (\mu x, \mu t) ~~~~~for ~~~\mu \neq 0.
$$
Thus, we can construct a piecewise constant
self-similar solution $U(x, t) = V(\frac{x}{t}) $ to connect two states $(\rho_{0}, -v_{0})$ and $(\rho_{1}, 0)$, which is in the form
\begin{equation}\label{key}
U(x, t) = V\left( \frac{x}{t}\right) =
\left\{\begin{array}{ll}
V_{0} = (1, \sqrt{2}),~~~~~ -\infty \leq \frac{x}{t} < \sigma,\\
V_{1} = (\rho_{1}, 0),~~~~~ \sigma < \frac{x}{t} \leq 0, \tag{4.2}
\end{array}\right .
\end{equation}
where $V_{0}$ and $V_{1}$ are subject to the Rankine-Hugoniot condition
\begin{equation}\label{key}
\left\{\begin{array}{ll}
\sigma(\rho_{1} - \rho_{0}) = \rho_{1}u_{1} - \rho_{0}u_{0},~~~~\\
\sigma(\rho_{1}u_{1} - \rho_{0}u_{0}) = \rho_{1}u_{1}^{2} + P_{1} - \rho_{0}u_{0}^{2} - P_{0} .
\end{array}\right. \tag{4.3}
\end{equation}
\noindent
In view of $\rho_{0} = 1$, $u_{0} = \sqrt{2}$, $u_{1} = 0$, it follows from $(4.3)_{1}$ that
\begin{equation}\label{key}
\sigma = -\frac{\sqrt{2}}{\rho_{1}-1}. \tag{4.4}
\end{equation}
\noindent
Note that $\sigma <0$ requires that $\rho_{1} >1$. Substituting $\sigma$ into $(4.3)_{2}$ gives
\begin{equation}\label{key}
P_{1} = \frac{2}{\rho_{1}-1} + 2 + P_{0}
= \frac{2}{\rho_{1}-1}+2+\frac{2}{M_{0}^2}-B(1+\alpha). \tag{4.5}
\end{equation}
\noindent
Replacing $P_{1}$ in (4.5) by $P_{1} = \left( \frac{2}{M_{0}^2}-B\alpha \right) \rho_{1}
-\frac{B}{\rho_{1}^\alpha}$, then we obtain
\begin{equation}\label{key}
\frac{2(\rho_{1}-1)}{M_{0}^2}=\frac{2\rho_{1}}{\rho_{1}-1}
+B\alpha(\rho_{1}-1)+B\left( \frac{1}{\rho_{1}^\alpha}-1\right) . \tag{4.6}
\end{equation}
From (2.13), we infer that $A$ and $B$ have the form of $\frac{x}{M_{0}^2}$, in which the $x$ is unknown. Without loss of generality, we assume $x$ is a constant independent of $M_{0}$ as
$A$ and $B$ are constants. Then, we know $AM_{0}^2$ and $BM_{0}^2$
are constants independent of $M_{0}$. Multiplying both sides of the equation (4.6) by $M_{0}^2$, we have
\begin{equation}\label{key}
2\left( \rho_{1}-1\right) = \frac{2\rho_{1}M_{0}^2}{\rho_{1}-1}
+BM_{0}^2\alpha(\rho_{1}-1)+BM_{0}^2\left(\frac{1}{\rho_{1}^\alpha}-1 \right) . \tag{4.7}
\end{equation}
By a simple calculation, (4.7) is equivalent to
\begin{equation}\label{key}
M_{0}^2 = \frac{\left[ (2-BM_{0}^2\alpha)(\rho_{1}-1)-BM_{0}^2\left(\frac{1}{\rho_{1}^\alpha}-1 \right) \right] (\rho_{1}-1)
}{2\rho_{1}}= \frac{\left[ AM_{0}^2(\rho_{1}-1)-BM_{0}^2\left(\frac{1}{\rho_{1}^\alpha}-1 \right) \right] (\rho_{1}-1)
}{2\rho_{1}}, \tag{4.8}
\end{equation}
In what follows, we define a continuous function $f(\rho) = \frac{\left[ AM_{0}^2(\rho-1)-BM_{0}^2\left(\frac{1}{\rho^\alpha}-1 \right) \right] (\rho-1)
}{2\rho}$ with respect to $\rho$. Then, we can get $f(1)=0$ and $f(\infty)=\infty$. According to the intermediate value theorem of continuous function, there exists a $\rho_{1} > 1$, such that $0 < f(\rho_{1}) < \infty$.
Furthermore, we can calculate
\begin{equation}\label{key}
\begin{split}
f'(\rho) &= \frac{1}{2}\left\lbrace \left( AM_{0}^2+\alpha BM_{0}^2 \frac{1}{\rho^{\alpha+1}} \right) \left(1-\frac{1}{\rho} \right)+
\left[ AM_{0}^2(\rho-1)-BM_{0}^2\left( \frac{1}{\rho^{\alpha}}-1\right)
\right] \frac{1}{\rho^{2}} \right\rbrace \\&=\frac{1}{2} \left[ AM_{0}^2 \left(1-\frac{1}{\rho^{2}}\right) +\alpha BM_{0}^2 \left( \frac{1}{\rho^{\alpha+1}}-\frac{1}{\rho^{\alpha+2}}
\right)+BM_{0}^2
\left( \frac{1}{\rho^2}-\frac{1}{\rho^{\alpha+2}}
\right)
\right] .
\end{split}\tag{4.9}
\end{equation}
From (4.9), it is easy to infer $f'(\rho)>0$ for $\rho>1$, i.e., $f(\rho)$ is always a monotone and increasing function for $\rho>1$,
which guarantees the uniqueness of the shock wave solution. Thus, we solve constructively the existence and uniqueness of the shock wave solution for the proceeding piston problem (2.8), (2.6) and (2.14). In other word, if (2.13) holds, for any $A$ and $B$,
there alway have a shock wave solution connecting states $V_{0} $ and $V_{1}$ for $0<M_{0} <\infty$, which is quite different from the results in [10, 17].
Therefore, we have the following result.
\noindent \textbf{Theorem 4.1.} ~~ For $v_{0} = -\sqrt{2}$ and any fixed $A$ and $B$, as the piston rushes into the MCG at arbitrary Mach number $M_{0} \in (0, \infty)$, there always exists a unique shock wave solution of the system (2.8), (2.6) and (2.14) connecting states $V_{0} = (1, \sqrt{2})$ and $V_{1} = (\rho_{1}, 0)$ in the domain $\Omega$ due to the first term of (1.5), which is quite different from the results of Chaplygin gas and generalized Chalygin gas.
\noindent \textbf{Remark 4.1.}~~~For any $A$ and $B$, we have demonstrated the global existence and uniqueness of shock wave solution in the one-dimensional piston problem for (2.8), (2.6) and (2.14) as the piston rushes into the MCG for $0<M_{0} <\infty$. We discover that, the results of (2.8), (2.6) and (2.14) are mathematically different to the isentropic Euler equations with pure or generalized Chaplygin gas, in which a Radon measure solution shall be considered to deal with concentration of mass on the piston when the Mach numbers satisfy $M_{0}\geq1 $ or $M_{0}\geq \sqrt{\alpha^{-1}}$. In essence, the difference is caused by the first term of (1.5). In Section 5, we will analyze the formation of singularity on the surface of the piston as $A \rightarrow 0$, which shows that a Radon measure solution is the limiting piston solution to the modified Chaplygin gas equations as $A \rightarrow 0$.
\subsection{Receding piston problem}
~~~~Let $U(x, t) = (\rho(x, t), u(x, t))$. The system (1.1) with (1.5) have two eigenvalues
\begin{equation}\label{key}
\lambda_{1}(U) = u - \sqrt{A + \frac{B \alpha}{\rho^{\alpha+1}}}, ~~~~~ \lambda_{2}(U) = u + \sqrt{A + \frac{B \alpha}{\rho^{\alpha+1}}}, \tag{4.10}
\end{equation}
with the corresponding right eigenvectors
\begin{equation}\label{key}
\vec{r}_{1} = \left( \sqrt{A + \frac{B \alpha}{\rho^{\alpha+1}}}, -\rho\right)^{T},~~~~~
\vec{r}_{2} = \left( \sqrt{A + \frac{B \alpha}{\rho^{\alpha+1}}}, \rho\right)^{T}, \tag{4.11}
\end{equation}
satisfying
\begin{equation}\label{key}
\triangledown \lambda_{i}(U) \cdot\vec{r}_{i} = \frac{2A\rho^{\alpha}+B\alpha(1-\alpha)}{2\rho^{\alpha+1} \sqrt{A + \frac{B \alpha}{\rho^{\alpha+1}}}} \neq 0 ~~~ (i=1, 2). \tag{4.12}
\end{equation}
Thus, we can see that system (1.1) with (1.5) is strictly hyperbolic and both characteristic fields are genuinely nonlinear when A, B > 0, in which the associated waves are either rarefaction waves or shock waves [26].
\noindent
Now, we can as well construct a solution of the form $U(t,x) = V\left( \frac{x}{t}\right) $. For any fixed $0<M_{0}<\infty$, we suppose the solution is composed of two constant states $V_{0} = (1, -\sqrt{2})$, $V_{1} = (\rho_{1}, 0)$, and a rarefaction wave $V_{m}$ connecting them:
\begin{equation}\label{key}
U(x, t) = V\left( \frac{x}{t}\right) =
\left\{\begin{array}{ll}
V_{0} ,~~~~~~~~~~~~~~~ -\infty \leq \frac{x}{t} < \lambda_{i}(V_{0}),\\
V_{m}\left(\frac{x}{t} \right) ,~~~~~~~~~~ \lambda_{i}(V_{0}) \leq \frac{x}{t} < \lambda_{i}(V_{1}),\\
V_{1},~~~~~~~~~~~~~~~~~~ \lambda_{i}(V_{1}) < \frac{x}{t} \leq 0,
\end{array}\right. ~~~ (i=1, 2). \tag{4.13}
\end{equation}
\noindent If the piston recedes from the gas, then $v_{0} >0$ and $\rho_{1} < \rho_{0}$. For the first family rarefaction wave $R_{1}(U)$, we can deduce the self-similar solution as follow:
\begin{equation}\label{key}
\left\{\begin{array}{ll}
\eta = \frac{x}{t} = \lambda_{1}(U) = u - \sqrt{A + \frac{B \alpha}{\rho^{\alpha+1}}},\\
u -\frac{2}{\alpha+1}
\sqrt{A+\frac{B\alpha}{\rho^{\alpha+1}}} + \frac{\sqrt{A}}{\alpha+1}
\ln\left( 2\sqrt{A}\rho^{\alpha+1}
\left( \sqrt{A+\frac{B\alpha}{\rho ^{\alpha+1}}} + \sqrt{A}\right) +B\alpha \right) =\\
u_{0} -\frac{2}{\alpha+1}
\sqrt{A+\frac{B\alpha}{\rho_{0} ^{\alpha+1}}} + \frac{\sqrt{A}}{\alpha+1}
\ln\left( 2\sqrt{A}\rho_{0} ^{\alpha+1}
\left( \sqrt{A+\frac{B\alpha}{\rho_{0} ^{\alpha+1}}} + \sqrt{A}\right) +B\alpha \right), ~~\rho_{1} < \rho < \rho_{0}, \\
\lambda_{1}(V_{0}) \leq \lambda_{1}(U) \leq \lambda_{1}(V_{1}),
\end{array}\right.\tag{4.14}
\end{equation}
where $U(t, x) = V(\frac{x}{t})$. Setting $W_{0} =u_{0} -\frac{2}{\alpha+1}
\sqrt{A+\frac{B\alpha}{\rho_{0} ^{\alpha+1}}} + \frac{\sqrt{A}}{\alpha+1}
\ln\left( 2\sqrt{A}\rho_{0} ^{\alpha+1}
\left( \sqrt{A+\frac{B\alpha}{\rho_{0} ^{\alpha+1}}} + \sqrt{A}\right) +B\alpha \right)$ and $N = \sqrt{A+\frac{B\alpha}{\rho ^{\alpha+1}}} $, then we have
\begin{equation}\label{key}
u= \eta+N, ~~~~~~~ \frac{1}{\rho^{\alpha+1}}=\frac{N^{2}-A}{B\alpha}.\tag{4.15}
\end{equation}
Inserting $u$ into $(4.14)_{2}$, we obtain
\begin{equation}\label{key}
\eta + \frac{(\alpha-1)N}{\alpha+1}+\frac{\sqrt{A}}{\alpha+1}\left[ \ln B\alpha+ \ln \frac{N+\sqrt{A}}{N-\sqrt{A}} \right] = W_{0}. \tag{4.16}
\end{equation}
Further, (4.16) can be rewritten as
\begin{equation}\label{key}
\frac{(\alpha-1)N}{\alpha+1}+\frac{\sqrt{A}}{\alpha+1}\left[ \ln \frac{N+\sqrt{A}}{N-\sqrt{A}} \right] = W_{0}-\eta-\frac{\sqrt{A}\ln B\alpha}{\alpha+1}. \tag{4.17}
\end{equation}
It is too difficult to obtain completely explicit solution $\rho(\eta)$ from (4.17) due to the structure of implicit equation. Next, we turn to prove the monotonicity of $\rho(\eta)$, which will play an important role in the following discussion. Firstly, we calculate
\begin{equation}\label{key}
N' =\left[ \sqrt{A+\frac{B\alpha}{\rho ^{\alpha+1}}}\right]' = -\frac{(\alpha+1)B\alpha
\rho\prime(\eta)}{2 \sqrt{A+\frac{B\alpha}
{\rho ^{\alpha+1}}}\rho^{\alpha+2}} = -\frac{(\alpha+1)B\alpha
\rho\prime(\eta)}{2N\rho^{\alpha+2}}, \tag{4.18}
\end{equation}
and
\begin{equation}\label{key}
\left[ \ln \frac{N+\sqrt{A}}{N-\sqrt{A}} \right]' = \left[ \ln\left( N+\sqrt{A}\right) - \ln\left( N-\sqrt{A}\right)\right]' = -\frac{2N'\sqrt{A}}{\left( N+\sqrt{A}\right)\left( N-\sqrt{A}\right)}. \tag{4.19}
\end{equation}
Then, substituting $(4.18)$ into $(4.19)$, we have
\begin{equation}\label{key}
\left[ \ln \frac{N+\sqrt{A}}{N-\sqrt{A}} \right]' = \frac{(\alpha+1)B\alpha\sqrt{A}\rho\prime(\eta)}{N(N-\sqrt{A})
(N+\sqrt{A})\rho^{\alpha+2}}.\tag{4.20}
\end{equation}
Next, performing implicit differentiation with respect to $\eta$ in (4.17) and reorganizing, we have
\begin{equation}\label{key}
\frac{-B\alpha(\alpha-1)\rho \prime(\eta)}{2N\rho^{\alpha+2}}
+ \frac{A B\alpha \rho \prime(\eta)}{N(N-\sqrt{A})
(N+\sqrt{A})\rho^{\alpha+2}} = -1. \tag{4.21}
\end{equation}
Further simplification based on (4.15) yields
\begin{equation}\label{key}
\frac{-(\alpha-1)\left( N^{2}-A\right) \rho \prime(\eta)}{2N\rho}
+ \frac{ A \rho \prime(\eta)}{N\rho} = -1. \tag{4.22}
\end{equation}
By a simple calculation, (4.22) is equivalent to
\begin{equation}\label{key}
\frac{\left[ (\alpha+1)A+(1-\alpha) N^{2}\right] \rho \prime(\eta)}{2N\rho} = -1. \tag{4.23}
\end{equation}
Eventually, we can get
\begin{equation}\label{key}
\rho \prime(\eta) = \frac{-2N \rho(\eta)}{(\alpha+1)A+
(1-\alpha)N^{2}}. \tag{4.24}
\end{equation}
Since $\rho(\eta)>0$, $N= \sqrt{A+\frac{B\alpha}{\rho ^{\alpha+1}}}>0$ and $0<\alpha<1$, we can deduce
\begin{equation}\label{key}
\rho \prime(\eta)<0, \tag{4.25}
\end{equation}
which means that $\rho(\eta)$ of the first family rarefaction wave $R_{1}(U(x, t))$
is a monotone decreasing function of $\eta$.
Similarly,
from $(4.14)_{1}$, we have $\sqrt{A+\frac{B\alpha}{\rho ^{\alpha+1}}} = u-\eta$. Then, substituting it into $(4.14)_{2}$ immediately yields
\begin{equation}\label{key}
\frac{(\alpha-1)u}{\alpha+1}+\frac{\sqrt{A}}{\alpha+1}\ln
\frac{u-\eta+\sqrt{A}}{u-\eta-\sqrt{A}} = W_{0}-\frac{2\eta}{\alpha+1}-\frac{\sqrt{A}\ln B\alpha}{\alpha+1}. \tag{4.26}
\end{equation}
In view of state $V_{1}=(\rho_{1}, 0)$, $u(\eta)$ satisfies the boundary condition
\begin{equation}\label{key}
u(\eta_{0})=0. \tag{4.27}
\end{equation}
Then, a combination of (4.26) and (4.27) leads to
\begin{equation}\label{key}
\frac{\sqrt{A}}{\alpha+1}\ln
\frac{-\eta_{0}+\sqrt{A}}{-\eta_{0}-\sqrt{A}} = W_{0}-\frac{2\eta_{0}}{\alpha+1}
-\frac{\sqrt{A}\ln B\alpha}{\alpha+1}. \tag{4.28}
\end{equation}
With a simple calculation, (4.28) is equivalent to
\begin{equation}\label{key}
\frac{\eta_{0}-\sqrt{A}}{\eta_{0}+\sqrt{A}}
= \exp\left({ \frac{(\alpha+1)W_{0}-2\eta_{0}-\sqrt{A}\ln B\alpha}
{\sqrt{A}}} \right) . \tag{4.29}
\end{equation}
where
\begin{equation}\label{key}
\begin{split}
W_{0}& =u_{0} -\frac{2}{\alpha+1}
\sqrt{A+\frac{B\alpha}{\rho_{0} ^{\alpha+1}}} + \frac{\sqrt{A}}{\alpha+1}
\ln\left( 2\sqrt{A}\rho_{0} ^{\alpha+1}
\left( \sqrt{A+\frac{B\alpha}{\rho_{0} ^{\alpha+1}}} + \sqrt{A}\right) +B\alpha \right) \\
&= -\sqrt{2} - \frac{2\sqrt{A+B \alpha}}{\alpha+1} + \frac{\sqrt{A}}{\alpha+1} \ln \left[2\sqrt{A} \left(\sqrt{A+B \alpha}+\sqrt{A}\right) + B\alpha \right]\\
& = -\sqrt{2} - \frac{2\sqrt{2}}{(\alpha+1)M_{0}} +\frac{\sqrt{A}}{\alpha+1}\ln\left(\frac{2\sqrt{2A}M_{0}+AM_{0}^{2}+2}{M_{0}^{2}} \right).
\end{split}. \tag{4.30}
\end{equation}
Taking (4.30) into (4.29), one obtains
\begin{equation}\label{key}
\frac{\eta_{0}-\sqrt{A}}{\eta_{0}+\sqrt{A}}
= \exp \left( \frac{-\sqrt{2}(\alpha+1)-\frac{2\sqrt{2}}{M_{0}}}{\sqrt{A}}\right)\cdot \left( \frac{\sqrt{2}+\sqrt{A}M_{0}}{\sqrt{2}-\sqrt{A}M_{0}}\right)\cdot \exp\left( \frac{-2\eta_{0}}{\sqrt{A}}\right)
. \tag{4.31}
\end{equation}
\noindent Let $C = \exp \left( \frac{-\sqrt{2}(\alpha+1)-\frac{2\sqrt{2}}{M_{0}}}{\sqrt{A}}\right)\cdot \left( \frac{\sqrt{2}+\sqrt{A}M_{0}}{\sqrt{2}-\sqrt{A}M_{0}}\right)$ and $Q_{0} = \frac{\eta_{0}}{\sqrt{A}}$. It is easy to verify
\begin{equation}\label{key}
\begin{split}
C& = \exp \left( \frac{-\sqrt{2}(\alpha+1)-\frac{2\sqrt{2}}{M_{0}}}{\sqrt{A}}\right)\cdot \left( \frac{ \left(\sqrt{2}+\sqrt{A}M_{0}\right)^2}{2-AM_{0}^2
}\right) \\
&= \exp \left( \frac{-\sqrt{2}(\alpha+1)-\frac{2\sqrt{2}}{M_{0}}}{\sqrt{A}}\right)\cdot \left( \frac{ \left(\sqrt{2}+\sqrt{A}M_{0}\right)^2}{\alpha BM_{0}^2
}\right)>0.
\end{split}
. \tag{4.32}
\end{equation}
Then, (4.31) can rewritten as
\begin{equation}\label{key}
\exp (-2Q_{0}) = \frac{Q_{0}-1}{C(Q_{0}+1)}. \tag{4.33}
\end{equation}
In what follows, we define
\begin{equation}\label{key}
f(Q) = \exp (-2Q) - \frac{1}{C} + \frac{2}{C(Q+1)}. \tag{4.34}
\end{equation}
\noindent Then, it follows from (4.33) that there exists a solution $Q_{0} \in (-\infty, 0)$ satisfies $f(Q_{0}) = 0$, which can be verified by the graph relationships between the exponential function and inverse proportional function. Since (4.33) is a transcendental equation, it is difficult to find an explicit analytical solution $\eta_{0}$, we turn to study the monotonicity of function (4.34) to demonstrate the existence of a rarefaction wave solution.
Simple calculations show
\begin{equation}\label{key}
f'(Q) =-2\left( \exp(-2Q)+\frac{1}{C(Q+1)^2}\right)< 0.\tag{4.35}
\end{equation}
\noindent Considering $V_{0} = (1, -\sqrt{2})$, from $(4.14)_1$,
we have $$\eta_{-\sqrt{2}} = \lambda_{1}(V_{0}) = \sqrt{2}\left( -1 - M_{0}^{-1}\right). $$
\noindent
For $Q_{-\sqrt{2}} = \frac{\eta_{-\sqrt{2}}}{\sqrt{A}} $, it follows from (4.34) that
\begin{equation}\label{key}
\begin{split}
f(Q_{-\sqrt{2}})&= \exp \left(\frac{2\sqrt{2}\left(1+ \frac{1}{M_{0}} \right) }{\sqrt{A}} \right) -
\exp \left(\frac{\sqrt{2}(\alpha+1)+\frac{2\sqrt{2}}{M_{0}} }{\sqrt{A}} \right) \frac{(\sqrt{2}-\sqrt{A}M_{0})(\sqrt{2}+\sqrt{2}M_{0}
+\sqrt{A}M_{0})}
{(\sqrt{2}+\sqrt{A}M_{0})(\sqrt{2}+\sqrt{2}M_{0}-\sqrt{A}M_{0})}
\\
&=\exp\left( \frac{\sqrt{2}+\frac{\sqrt{2} }{M_{0}}}{\sqrt{A}}
\right) \left[ \frac{\exp\left( \frac{\sqrt{2}+\frac{\sqrt{2} }{M_{0}}}{\sqrt{A}}
\right) (2+2M_{0}-AM_{0}^2+\sqrt{2A}M_{0}^2)
-\exp\left( \frac{\sqrt{2}\alpha+\frac{\sqrt{2} }{M_{0}}}{\sqrt{A}}
\right) (2+2M_{0}-AM_{0}^2-\sqrt{2A}M_{0}^2)
}
{(\sqrt{2}+\sqrt{A}M_{0})(\sqrt{2}+\sqrt{2}M_{0}-\sqrt{A}M_{0})}
\right] \end{split} \notag
\end{equation}
Since $\sqrt{A}M_{0}<\sqrt{A+B\alpha}M_{0}=\sqrt{2}$, we have
\begin{equation}
f(Q_{-\sqrt{2}})>\frac{\exp\left( \frac{\sqrt{2}+\frac{\sqrt{2} }{M_{0}}}{\sqrt{A}}
\right) 2\sqrt{2A}M_{0}^2}
{(\sqrt{2}+\sqrt{A}M_{0})(\sqrt{2}+\sqrt{2}M_{0}-\sqrt{A}M_{0})}>0,\tag{4.36}
\end{equation}
i.e., $f(Q_{-\sqrt{2}}) > f(Q_{0})$. From the above discussion, it follows that
\begin{equation}\label{key}
V_{m}\left( \frac{x}{t} \right) = (\rho(\eta), u(\eta)),\\~~~~~~
\eta_{-\sqrt{2}} \leq \eta \leq \eta_{0},\\~~~~~~
0 < \rho(\eta_{0}) \leq \rho(\eta) \leq \rho(\eta_{-\sqrt{2}}). \tag{4.37}
\end{equation}
Therefore, the first family rarefaction wave $R_{1}(U)$ is a solution of (2.8), (2.6) and (2.14) connecting $V_{0} = (1, -\sqrt{2})$, $V_{1} = (\rho_{1}, 0)$ in the domain $\Omega$.
Analogously, for the second family rarefaction wave $R_{2}\left( U(x, t)\right) $, we can deduce the self-similar solution as follow:
\begin{equation}\label{key}
\left\{\begin{array}{ll}
\xi = \frac{x}{t} = \lambda_{2}(U) = u + \sqrt{A + \frac{B \alpha}{\rho^{\alpha+1}}},\\
u +\frac{2}{\alpha+1}
\sqrt{A+\frac{B\alpha}{\rho^{\alpha+1}}} - \frac{\sqrt{A}}{\alpha+1}
\ln\left( 2\sqrt{A}\rho^{\alpha+1}
\left( \sqrt{A+\frac{B\alpha}{\rho ^{\alpha+1}}} + \sqrt{A}\right) +B\alpha \right) =\\
u_{0} +\frac{2}{\alpha+1}
\sqrt{A+\frac{B\alpha}{\rho_{0} ^{\alpha+1}}} - \frac{\sqrt{A}}{\alpha+1}
\ln\left( 2\sqrt{A}\rho_{0} ^{\alpha+1}
\left( \sqrt{A+\frac{B\alpha}{\rho_{0} ^{\alpha+1}}} + \sqrt{A}\right) +B\alpha \right), ~~\rho_{1} < \rho < \rho_{0}, \\
\lambda_{2}(V_{0}) \leq \lambda_{2}(U) \leq \lambda_{2}(V_{1}).
\end{array}\right.\tag{4.38}
\end{equation}
By the same analysis of the first family rarefaction wave, we can get $$\xi_{-\sqrt{2}} = \lambda_{2}(V_{0}) = \sqrt{2}\left( -1 + M_{0}^{-1}\right),$$
and
\begin{equation}\label{key}
\frac{(\alpha-1)N}{\alpha+1}+\frac{\sqrt{A}}{\alpha+1}\left[ \ln \frac{N+\sqrt{A}}{N-\sqrt{A}} \right] = \xi-W_{1}-\frac{\sqrt{A}\ln B\alpha}{\alpha+1}. \tag{4.39}
\end{equation}
where
\begin{equation}\label{key}
\begin{split}
W_{1} &= u_{0} +\frac{2}{\alpha+1}
\sqrt{A+\frac{B\alpha}{\rho_{0} ^{\alpha+1}}} - \frac{\sqrt{A}}{\alpha+1}
\ln\left( 2\sqrt{A}\rho_{0} ^{\alpha+1}
\left( \sqrt{A+\frac{B\alpha}{\rho_{0} ^{\alpha+1}}} + \sqrt{A}\right) +B\alpha \right)\\
&=
-\sqrt{2} + \frac{2\sqrt{2}}{\left(\alpha+1 \right) M_{0}} - \frac{\sqrt{A}}{\left(\alpha+1 \right) }\ln\left(\frac{2\sqrt{2A}M_{0}+AM_{0}^{2}+2}{M_{0}^{2}} \right).
\end{split}
\tag{4.40}
\end{equation}
Differentiating the both sides of (4.39) with respect to $\xi$, we obtain
\begin{equation}\label{key}
-\frac{(\alpha-1)\left(N^{2}-A \right)\rho'(\xi) }{2N\rho}
+\frac{A\rho'(\xi)}{N\rho}
=1. \tag{4.41}
\end{equation}
Then, rearranging (4.41) leads to
\begin{equation}\label{key}
\rho'(\xi) = \frac{2N\rho}{2A-\left(\alpha-1 \right)\left(N^2-A \right) }
= \frac{2N\rho}{A+\left(1-\alpha\right)N^2+\alpha A}>0, \tag{4.42}
\end{equation}
which means that the density of the second family rarefaction wave $R_{2}(U(x, t))$ is monotonic increasing with respect to $\xi$.
Based on $(4.38)_1$, taking $u=\xi-\sqrt{A + \frac{B \alpha}{\rho^{\alpha+1}}}$ into $(4.38)_2$ to deduce
\begin{equation}\label{key}
\frac{(\alpha-1)u}{\alpha+1}-\frac{\sqrt{A}}{\alpha+1}\ln
\frac{u-\xi-\sqrt{A}}{u-\xi+\sqrt{A}} = W_{1}-\frac{2\xi}{\alpha+1}
+\frac{\sqrt{A}\ln B\alpha}{\alpha+1}. \tag{4.43}
\end{equation}
In view of state $V_{1} = (\rho_{1}, 0)$, $u(\xi) $ satisfies the boundary condition $u(\xi_{0}) = 0$. Inserting $u(\xi_{0})$ into (4.43) to obtain
\begin{equation}\label{key}
\frac{\xi_{0}+\sqrt{A}}{\xi_{0}-\sqrt{A}}
= \exp\left({ \frac{2\xi_{0}-(\alpha+1)W_{1}-\sqrt{A}\ln B\alpha}
{\sqrt{A}}} \right) . \tag{4.44}
\end{equation}
It follows from (4.44) and (4.40) that
\begin{equation}\label{key}
\frac{\xi_{0}+\sqrt{A}}{\xi_{0}-\sqrt{A}}
= \exp \left( \frac{\sqrt{2}(\alpha+1)-\frac{2\sqrt{2}}{M_{0}}}{\sqrt{A}}\right)\cdot \left( \frac{2\sqrt{2A}M_{0}+AM_{0}^2+2}{\alpha BM_{0}^2}\right)\cdot \exp\left( \frac{2\xi_{0}}{\sqrt{A}}\right)
. \tag{4.45}
\end{equation}
Let $C_{1} = \exp \left( \frac{\sqrt{2}(\alpha+1)-\frac{2\sqrt{2}}{M_{0}}}{\sqrt{A}}\right)\cdot \left( \frac{2\sqrt{2A}M_{0}+AM_{0}^2+2}{\alpha BM_{0}^2}\right)$ and $Q=\frac{\xi}{\sqrt{A}}$. Then, we define $g(Q) = \exp (2Q) - \frac{1}{C_{1}} - \frac{2}{C_{1}(Q - 1)}$ with $g(Q_{0})=g\left( \frac{\xi_{0}}{\sqrt{A}}\right) = 0$. Simple calculations lead to
\begin{equation}\label{key}
g'(Q) = 2\exp(2Q)+\frac{2}{C_{1}\left(Q-1 \right)^2 }> 0,\tag{4.46}
\end{equation}
and
\begin{equation}\label{key}
\begin{split}
g(Q_{-\sqrt{2}})&=g\left(\frac{\xi_{-\sqrt{2}}}{\sqrt{A}} \right)=\exp\left( \frac{-\sqrt{2}+\frac{\sqrt{2} }{M_{0}}}{\sqrt{A}} \right) \\
&\times
\left[\exp\left( \frac{-\sqrt{2}+\frac{\sqrt{2} }{M_{0}}}{\sqrt{A}}
\right)- \frac{
\exp\left( \frac{-\sqrt{2}\alpha+\frac{\sqrt{2} }{M_{0}}}{\sqrt{A}}
\right)\left( \frac{\sqrt{2}}{M_{0}}-\sqrt{2}+\sqrt{A}\right) (\sqrt{2}-\sqrt{A}M_{0})
}
{\left( \frac{\sqrt{2}}{M_{0}}-\sqrt{2}-\sqrt{A}\right) (\sqrt{2}+\sqrt{A}M_{0})}
\right]
\end{split}\tag{4.47}
\end{equation}
If $\frac{\sqrt{2}}{M_{0}}-\sqrt{2}-\sqrt{A} >0$, we have
\begin{equation}\label{key}
g(Q_{-\sqrt{2}})<\exp\left( \frac{-\sqrt{2}+\frac{\sqrt{2} }{M_{0}}}{\sqrt{A}}\right)
\left[ \frac{\exp\left( \frac{-\sqrt{2}+\frac{\sqrt{2} }{M_{0}}}{\sqrt{A}}\right) (-2\sqrt{2A}M_{0})}
{\left( \frac{\sqrt{2}}{M_{0}}-\sqrt{2}-\sqrt{A}\right) (\sqrt{2}+\sqrt{A}M_{0})}
\right]
<0.
\tag{4.48}
\end{equation}
which infers $\xi_{-\sqrt{2}} < \xi_{0}$ and $\rho(\xi_{-\sqrt{2}}) < \rho(\xi_{0})$, contradicting with the requirement $\rho_{0}>\rho_{1}$.
\noindent Similarly, if $\frac{\sqrt{2}}{M_{0}}-\sqrt{2}-\sqrt{A} <0$, we can obtain
\begin{equation}\label{key}
\begin{split}
g(Q_{-\sqrt{2}}) >0
\end{split}. \tag{4.49}
\end{equation}
which means $\xi_{-\sqrt{2}} > \xi_{0}$, contradicting with the requirement
$\lambda_{2}(V_{0}) \leq \lambda_{2}(U) \leq \lambda_{2}(V_{1})$. Hence, the second family rarefaction wave $R_{2}(U(x, t))$ is not a physical solution as the piston recedes from the gas.
Consequently, we have the following result.
\noindent \textbf{Theorem 4.2.} \label{{Theorem 4.2}}~~ For $v_{0} = \sqrt{2}$ and any fixed $A$ and $B$, as the piston recedes from the MCG for any $M_{0} \in (0, \infty)$, only one physically relevant rarefaction wave solution will be present in front of the piston to connect two states $V_{0} = (1, -\sqrt{2})$ and $V_{1} = (\rho_{1}, 0)$ in the domain $\Omega$, and the concentration will never occur.
\noindent \textbf{Remark 4.2.}~~ The results obtained here, to some extent, not only extend the results and proofs in [10, 17], in which the pressure state only contains one parameter, but also cover the results obtained in [10] for generalized Chaplygin gas as $A \rightarrow 0$, which will be analyzed and identified in Section 5.
\section{Limit behavior of the piston solutions as A $\rightarrow$ 0}
~~~~ This section discusses the limit behavior of piston solutions to (2.8) and (1.5) as the pressure (1.5) approaches the generalized Chaplygin gas (1.3), i.e., $A \rightarrow 0$.
Through the analysis in Section 2, taking the limit $A \rightarrow 0$, the parameter B in (2.10) approaches to
\begin{equation}\label{key}
B = \frac{2}{\alpha M_{0}^{2}}, \tag{5.1}
\end{equation}
and the pressure in (2.11) tends to
\begin{equation}\label{key}
P(\rho) = -\frac{2}{\alpha M_{0}^{2}\rho^{\alpha }}. \tag{5.2}
\end{equation}
Both (5.1) and (5.2) play an important role in forming singularity as the piston moves forward to the MCG and $A\rightarrow 0$. From (5.1) and (5.2), similar with the proof of the theorem 4.1, we can obtain
\begin{equation}\label{key}
\lim\limits_{A \rightarrow 0} M_{0}^{2} = {\alpha}^{-1} \left(1-\frac{1}{\rho_{1}} \right) \left(1-\frac{1}{\rho_{1}^{\alpha}} \right) . \tag{5.3}
\end{equation}
From (5.3), it is easy to discover, if $0<M_{0}< \sqrt{\alpha^{-1}}$, it follows that $\rho_{1}>1$, which shows that there exists an integral weak solution in front of the piston. While $M_{0}\geq \sqrt{\alpha^{-1}}$, it contradicts with $\rho_{1}>1$, i.e., the piston solution cannot be constructed by using shock waves.
Then, we can construct a Radon measure solution based on definition 3.2 for the case $M_{0}\geq \sqrt{\alpha^{-1}}$,
\begin{equation}\label{key}
\tau = I_{\Omega}L^{2} + w_{\rho}(t)\delta_{\{x=0,~ t \geq 0\} },
~~~m = \sqrt{2} I_{\Omega}L^{2} , ~~~n = 2 I_{\Omega}L^{2} , ~~~\wp = -\frac{2}{\alpha M_{0}^2}I_{\Omega}L^{2},\tag{5.4}
\end{equation}
where $I_{\Omega}$ is the characteristic function of $\Omega$. A Radon measure solution is a generalization of an ordinary shock wave. Informally speaking, it is a kind of discontinuity, on which at least one of the state variables may develop an extreme concentration in the form of a weighted Dirac delta function with the discontinuity as its support. It is more compressive than an ordinary shock wave in the sense of distributions.
Substituting (5.4) into (3.5), we infer that
\begin{equation}\label{key}
\begin{split}
0 &= <\tau, \partial t \phi> + <m, \partial x \phi> +
\int_{-\infty}^{0} \rho_{0} \phi(0, x)dx \\&= \int_{\Omega} \partial t \phi dxdt + \int_{0}^{\infty} w_{\rho}(t) \partial t \phi(t, 0)dt
+ \sqrt{2} \int_{0}^{\infty} \int_{-\infty}^{0} \partial x \phi dxdt
+ \int_{-\infty}^{0} \phi(0, x)dx
\\&= \int_{-\infty}^{0} \phi(t, x)|_{t=0}^{\infty} dx
+ w_{\rho}(t) \phi(t, 0)|_{t=0}^{\infty}
- \int_{0}^{\infty} w'_{\rho}(t) \phi(t, 0)dt +
\sqrt{2} \int_{0}^{\infty} \phi(t, x)|_{x=-\infty}^{0} dt + \int_{-\infty}^{0} \phi(0, x)dx
\\&= \int_{0}^{\infty} (\sqrt{2} - w'_{\rho}(t))\phi(t, 0)dt
-w_{\rho}(0)\phi(0, 0).
\end{split} \tag{5.5}
\end{equation}
Due to the arbitrariness of $\phi$, we have
\begin{equation}\label{key}
\left\{\begin{array}{ll}
w'_{\rho}(t) = \sqrt{2}, ~~~t\geq0,\\
w_{\rho}(0) = 0.
\end{array}\right. \tag{5.6}
\end{equation}
It follows that
\begin{equation}\label{key}
w_{\rho}(t) = \sqrt{2}t. \tag{5.7}
\end{equation}
Similarly, substituting (5.4) into (3.6), we have
\begin{equation}\label{key}
\begin{split}
0 &= <m, \partial t \phi> + <n, \partial x \phi> +
<\wp, \partial x \phi> - <w_{p}(t)\delta_{\{x=0,~ t \geq 0\} }, \phi>
+ \int_{-\infty}^{0} \rho_{0}u_{0} \phi(0, x)dx
\\&= \sqrt{2} \int_{0}^{\infty} \int_{-\infty}^{0}
\partial t \phi dxdt +
2 \int_{0}^{\infty} \int_{-\infty}^{0}
\partial x \phi dxdt -
\frac{2}{\alpha M_{0}^2} \int_{0}^{\infty} \int_{-\infty}^{0}
\partial x \phi dxdt-
\int_{0}^{\infty} w_{p}(t) \phi(t, 0)dt
+ \sqrt{2} \int_{-\infty}^{0} \phi(0, x)dx
\\&= \int_{0}^{\infty} \left[ 2-w_{p}(t) - \frac{2}{\alpha M_{0}^2}\right] \phi(t, 0)dt.
\end{split} \tag{5.8}
\end{equation}
Since the arbitrariness of $\phi$ and $M_{0}\geq \sqrt{\alpha^{-1}}$, then
\begin{equation}\label{key}
w_{p}(t) = 2 - \frac{2}{\alpha M_{0}^2} \geq 0. \tag{5.9}
\end{equation}
In summary,
we conclude that, as $A \rightarrow 0$ and $M_{0} \geq \sqrt{\alpha^{-1}}$, the shock wave and all gas between the shock wave and the piston adhere to the piston and then form a concentration of mass like a Dirac measure, which indicates that
the radon measure solution (5.4) is a reasonable solution for the limiting piston problem of modified Chaplygin gas as $A \rightarrow 0$ and $M_{0} \geq \sqrt{\alpha^{-1}}$.
For the receding case, as $A \rightarrow 0$, (4.10), (4.11) have changed to
\begin{equation}\label{key}
\lambda_{1}(U) = u - \sqrt{\frac{B \alpha}{\rho^{\alpha+1}}}, ~~~~~ \lambda_{2}(U) = u + \sqrt{\frac{B \alpha}{\rho^{\alpha+1}}}, \tag{5.10}
\end{equation}
with the corresponding right eigenvectors
\begin{equation}\label{key}
\vec{r}_{1} = \left( \sqrt{\frac{B \alpha}{\rho^{\alpha+1}}}, -\rho\right)^{T},~~~~~
\vec{r}_{2} = \left( \sqrt{ \frac{B \alpha}{\rho^{\alpha+1}}}, \rho\right)^{T}, \tag{5.11}
\end{equation}
satisfying
\begin{equation}\label{key}
\triangledown \lambda_{i}(U) \cdot\vec{r}_{i} \neq 0, ~~~ ~(i=1, 2,~~~~ 0<\alpha<1). \tag{5.12}
\end{equation}
Then (4.14) has degenerated into
\begin{equation}\label{key}
\left\{\begin{array}{ll}
\eta = \frac{x}{t} = \lambda_{1}(U) = u - \sqrt{\frac{B \alpha}{\rho^{\alpha+1}}},\\
u -\frac{2}{\alpha+1}
\sqrt{\frac{B\alpha}{\rho^{\alpha+1}}} =
u_{0} -\frac{2}{\alpha+1}
\sqrt{\frac{B\alpha}{\rho_{0} ^{\alpha+1}}} , ~~\rho_{1} < \rho < \rho_{0}, \\
\lambda_{1}(V_{0}) \leq \lambda_{1}(U) \leq \lambda_{1}(V_{1}).
\end{array}\right.\tag{5.13}
\end{equation}
We can observe that the self-similar solution above is totally the same as the Eq. (3.8) in [10]. So, performing the similar analysis in [10], we also obtain only the first rarefaction wave $R_{1}\left( U(x,t)\right) $ is a physical solution. Readers can refer to [10] for more details.
In particular, when $\alpha=1$, we get $ \triangledown \lambda_{i}(U) \cdot\vec{r}_{i} = 0 ~ (i=1, 2)$, which means that both the characteristic fields are linearly degenerate and the elementary waves only involve contact discontinuities, i.e., the rarefaction waves coincide with that of the shock waves in the physical plane. Replacing $\rho_{0}, u_{0} $ by $\rho_{0} = 1, u_{0} = -\sqrt{2}$ and the $(4.3)_{1}$,
we have
\begin{equation}\label{key}
\sigma = \frac{\sqrt{2}}{\rho_{1}-1}. \tag{5.14}
\end{equation}
Then, by $(4.3)_{2}$, (5.2) and (5.14), we have
\begin{equation}\label{key}
(2+P_{0})\rho_{1}^{2}-2P_{0}\rho_{1}+P_{0} = 0. \tag{5.15}
\end{equation}
For $M_{0} \in (0, \infty)$ and the non-negativeness of $\rho_{1}$, it follows from (5.15) that
\begin{equation}\label{key}
\rho_{1} = \frac{2P_{0}+2\sqrt{2}\sqrt{-P_{0}}}{4+2P_{0}}
=\frac{1}{1+M_{0}}<1, \tag{5.16}
\end{equation}
i.e., we prove the existence of rarefaction wave solutions when the piston recedes from the MCG and $A \rightarrow 0$.
Therefore,
as $A \rightarrow 0$ and $0<\alpha \leq 1$, the limit of solution of
piston problem (2.8) and (1.5) is
similar with that of generalized or pure Chaplygin Euler equations [10, 17].
\sec{\Large\bf Acknowledgments}
This work is supported by the Minnan Normal University (Grant No. KJ2021020) and the Department of Education, Fujian Province (Grant No. JAT210254).
\sec{\Large\bf Author Declarations}
\subsec{\large\bf Conflict of Interest}
The authors have no conflicts to disclose.
\sec{\Large\bf Data Availability}
Data sharing is not applicable to this article as no new data were created or analyzed in this study.
|
\section{Introduction}\label{sec:intro}
While deep neural networks (DNNs) have been successful in several applications, their performance degrades under distributional shifts between the training data and test data~\cite{quinonero2008dataset}. This distributional shift hinders DNNs from being widely deployed in many risk-sensitive applications, such as autonomous driving, medical imaging, and mobile health care, where new types of test data unseen during training could result in undesirable disasters. For instance, Tesla Autopilot has caused 12 ``deaths'' until recently~\cite{bachman_capulet_2022}.
To address this problem, \textit{test-time adaptation} (TTA)
aims to adapt DNNs to the target/unseen domain with only unlabeled test data streams, without any additional data acquisition or labeling cost.
Recent studies reported that TTA is a promising, practical direction to mitigate distributional shifts~\cite{bnstats, bnstats2, tent, lame, cotta}.
\begin{figure}[t]
\centering
\begin{minipage}[t]{0.48\textwidth}
\begin{subfigure}[t]{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{fig/real_dist1.pdf}
\caption{KITTI dataset.}
\vfill
\includegraphics[width=1\linewidth]{fig/real_dist2.pdf}
\caption{HARTH dataset.}
\end{subfigure}
\caption{Illustration of test sample distributions along the time axis from real-world datasets: (a) autonomous driving (KITTI~\cite{kitti}) and (b) human activity recognition (HARTH~\cite{harth}). They are temporally correlated.}\label{fig:real_dist}
\end{minipage}
\hspace{0.1cm}
\begin{minipage}[t]{0.48\textwidth}
\vspace{-1.78cm}
\centering
\includegraphics[width=\linewidth]{fig/performance_overview.pdf}
\caption{Average classification error (\%) of existing TTA methods and our method (NOTE{}) on CIFAR10-C~\cite{cifarc}. The error rates significantly increase under the non-i.i.d.\ setting compared with the i.i.d.\ setting. Lower is better.}\label{fig:performance_overview}
\end{minipage}
\end{figure}
Prior TTA studies typically assume (implicitly or explicitly) that a target test sample $\mathbf{x}_{t}$ at time $t$ and the corresponding ground-truth label $y_{t}$ (unknown to the learner) are independent and identically distributed (i.i.d.) following a target domain, i.e.,
$(\mathbf{x}_{t}, y_t)$ is drawn independently from a time-invariant distribution $P_{\mathcal{T}}(\mathbf{x}, y)$. However, the distribution of online test samples often changes across the time axis, i.e., $(\mathbf{x}_t, y_t) \sim P_{\mathcal{T}}(\mathbf{x}, y\,|\,t)$ in many applications; for instance, AI-powered self-driving car's object encounter will be dominated by cars while driving on the highway, but less dominated by them on downtown where other classes such as pedestrians and bikes are visible.
In human activity recognition, some activities last for a short term (e.g., a fall down), whereas certain activities last longer (e.g., a sleep). Figure~\ref{fig:real_dist} illustrates that some data distributions in the real world, such as autonomous driving and human activity recognition, are often temporally correlated. Considering that most existing TTA algorithms simply use an incoming batch of test samples for adaptation~\cite{bnstats, bnstats2, tent, cotta}, the model might be biased towards these imbalanced samples under the temporally correlated test streams. Figure~\ref{fig:performance_overview} compares the performance of the state-of-the-art TTA algorithms under the i.i.d.\ and non-i.i.d.\footnote{We use the terms \emph{temporally correlated} and \emph{non-i.i.d.}\ interchangeably in the context of test-time adaptation.} conditions. While the TTA methods perform well under the i.i.d.\ assumption, their errors increase under the non-i.i.d.\ case.
Adapting to temporally correlated test data results in overfitting to temporal distributions, which in turn harms the generalization of the model.
Motivated by this, we present a NOn-i.i.d.\ TEst-time adaptation scheme, \textit{NOTE}, that consists of two components: (a) Instance-Aware Batch Normalization (IABN) and (b) Prediction-Balanced Reservoir Sampling (PBRS). First, we propose a novel normalization layer, IABN, that eliminates the dependence on temporally correlated data
for adaptation while being robust to distribution shifts. IABN detects out-of-distribution instances \textit{sample by sample} and corrects via instance-aware normalization. The key idea of IABN is synthesizing Batch Normalization (BN)~\cite{pmlr-v37-ioffe15} with Instance Normalization (IN)~\cite{ulyanov2016instance} in a unique way; it calculates how different the learned knowledge (BN) is from the current observation (IN) and corrects the normalization by the deviation between IN and BN. Second, we present PBRS that resolves the problem of overfitting to non-i.i.d.\ samples by mimicking i.i.d.\ samples from non-i.i.d.\ streams. By utilizing predicted labels of the model,
PBRS aims for both time-uniform sampling and class-uniform sampling
from the non-i.i.d.\ streams, and stores the `simulated' i.i.d.\ samples in memory. With the i.i.d.-like batch in the memory, PBRS enables the model to adapt to the target domain without being biased to temporal distributions.
We evaluate NOTE{} with state-of-the-art TTA baselines~\cite{bnstats, bnstats2, pl, onda, lame, cotta} on multiple datasets, including common TTA benchmarks (CIFAR10-C, CIFAR100-C, and ImageNet-C~\cite{cifarc}) and real-world non-i.i.d.\ datasets (KITTI~\cite{kitti}, HARTH~\cite{harth}, and ExtraSensory~\cite{extrasensory}). Our results suggest that NOTE{} not only significantly outperforms the baselines under non-i.i.d.\ test data, e.g., it achieves a 21.1\% error rate on CIFAR10-C which is on average 15.1\% lower than the state-of-the-art method~\cite{lame}, but also shows comparable performance even under the i.i.d.\ assumption, e.g., 17.6\% error on CIFAR10-C where the best baseline~\cite{cotta} achieves 17.8\% error. Our ablative study demonstrates the individual effectiveness of IABN and PBRS and further highlights their synergy when jointly used.
Finally, we summarize the key characteristics of NOTE{}. First,
NOTE{} is a batch-free inference algorithm (requiring a single instance for inference), different from the state-of-the-art TTA algorithms~\cite{bnstats, bnstats2, tent, lame, cotta} where a batch of test data is necessary for inference to estimate normalization statistics (mean and variance).
Second, while some recent TTA methods leverage augmentations to improve performance at the cost of additional forwarding passes~\cite{sun2020test, cotta}, NOTE{} requires only a single forwarding pass.
NOTE{} updates only the normalization statistics and affine parameters in IABN, which is, e.g., approximately 0.02\% of the total trainable parameters in ResNet18~\cite{7780459}.
Third, NOTE{}'s additional memory overhead is negligible.
It merely stores predicted labels of test data to run PBRS.
These characteristics make NOTE{} easy to apply to any existing AI system and particularly,
are beneficial in latency-sensitive tasks such as autonomous driving and human health monitoring.
\section{Background}\label{sec:background}
\subsection{Problem setting: test-time adaptation with non-i.i.d.\ streams}
\noindent\textbf{Test-time adaptation.} Let $\mathcal{D_S} = \{\mathcal{X}^{\mathcal{S}}, \mathcal{Y}\}$ be the data from the source domain and $\mathcal{D_T} = \{\mathcal{X^T}, \mathcal{Y}\}$ be the data from the target domain to adapt to. Each data instance and the corresponding ground-truth label pair $(\mathbf{x}_i, y_i) \in \mathcal{X^S} \times \mathcal{Y}$ in the source domain follows a probability distribution $P_{\mathcal{S}}(\mathbf{x}, y)$. Similarly, each target test sample and the corresponding label at test time $t$, $(\mathbf{x}_{t}, y_t) \in \mathcal{X^T} \times \mathcal{Y}$, follows a probability distribution $P_{\mathcal{T}}(\mathbf{x}, y)$ where $y_t$ is unknown for the learner. The standard covariate shift assumption in domain adaptation is defined as $P_{\mathcal{S}}(\mathbf{x}) \neq P_{\mathcal{T}}(\mathbf{x})$ and $P_{\mathcal{S}}(y|\mathbf{x}) = P_{\mathcal{T}}(y|\mathbf{x})$~\cite{quinonero2008dataset}. Unlike traditional domain adaptation that uses $\mathcal{D_S}$ and $\mathcal{X^T}$ collected beforehand for adaptation, test-time adaptation (TTA) continually adapts a pre-trained model $f_\theta(\cdot)$ from $\mathcal{D_S}$ by utilizing only $\mathbf{x}_{t}$ obtained at test time $t$.
\noindent\textbf{TTA on non-i.i.d. streams.} Note that previous TTA mechanisms typically assume that each target sample $(\mathbf{x}_{t}, y_t) \in \mathcal{X^T} \times \mathcal{Y}$ is independent and identically distributed (i.i.d.) following a time-invariant distribution $P_{\mathcal{T}}(\mathbf{x}, y)$. However, the data obtained at test time is non-i.i.d.\ in many scenarios. By non-i.i.d., we refer to distribution changes over time, i.e., $(\mathbf{x}_t, y_t) \sim P_{\mathcal{T}}(\mathbf{x}, y\,|\,t)$, which is a practical setting in many real world applications~\cite{ZHU2021371}.
\subsection{Batch normalization in TTA}\label{sec:background:bn_tta}
Batch Normalization (BN)~\cite{pmlr-v37-ioffe15} is a widely-used training technique in deep neural networks as it reduces the internal covariant shift problem. Let $\mathbf{f}\in \mathbb{R}^{B\times C \times L}$ denote a batch of feature maps in general, where $B$, $C$, and $L$ denote the batch size, the number of channels, and the size of each feature map, respectively. Given the statistics of the feature maps for normalization, say mean ${\boldsymbol{\mu}}$ and variance ${\boldsymbol{\sigma}}^2$, BN is \emph{channel-wise}, i.e., ${\boldsymbol{\mu}}, {\boldsymbol{\sigma}}^2 \in \mathbb{R}^C$ and computes:
\begin{equation}
\mathrm{BN}(\mathbf{f}_{:,c,:}; {\boldsymbol{\mu}}_c, {\boldsymbol{\sigma}}^2_c) := \gamma\cdot\frac{\mathbf{f}_{:,c,:} - {\boldsymbol{\mu}}_c}{\sqrt{{\boldsymbol{\sigma}}_c^2 + \epsilon}} + \beta,
\end{equation}
where $\gamma$ and $\beta$ are the affine parameters followed by the normalization, and $\epsilon > 0$ is a small constant for numerical stability.
Although a conventional way of computing BN in test-time is to set ${\boldsymbol{\mu}}$ and ${\boldsymbol{\sigma}}^2$ as those estimated from \emph{training} (or source) data, say $\bar{{\boldsymbol{\mu}}}$ and $\bar{{\boldsymbol{\sigma}}}^2$, the state-of-the-art TTA methods based on adapting BN layers~\cite{bnstats, bnstats2, tent, cotta} instead use the statistics computed directly from the recent test batch to de-bias distributional shifts at test-time, i.e.:
\begin{equation}
\hat{{\boldsymbol{\mu}}}_c := \frac{1}{BL} \sum_{b,l} \mathbf{f}_{b,c,l} \text{, \ and \ \ } \hat{{\boldsymbol{\sigma}}}^2_c := \frac{1}{BL} \sum_{b,l} (\mathbf{f}_{b,c,l} - \hat{{\boldsymbol{\mu}}}_c)^2.
\end{equation}
This practice is simple yet effective under distributional shifts and is thus adopted in many recent TTA studies~\cite{bnstats, bnstats2, tent, cotta}. Based on the test batch statistics, they often further adapt the affine parameters via entropy minimization of the model outputs~\cite{tent} or update the entire parameters with self-training~\cite{cotta}.
\section{Method}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{fig/tta_main.pdf}
\caption{An overview of the proposed methodology: Instance-Aware Batch Normalization (IABN) and Prediction-Balanced Reservoir Sampling (PBRS). IABN aims to detect non-i.i.d.\ streams} and in turn corrects the normalization for inference. PBRS manages data in a time- and prediction-uniform manner from non-i.i.d.\ data streams and gradually adapts IABNs with the balanced data afterwards.
\label{fig:tta_main}
\end{figure}
In the same vein as previous work~\cite{bnstats, bnstats2, tent}, we focus on adapting BN layers in the given model to perform TTA, and this includes essentially two approaches: {(a) re-calibrating} (or adapting) channel-wise statistics for normalization (instead of using those learned from training), and (b) adapting the affine parameters (namely, $\gamma$ and $\beta$) after the normalization with respect to a certain objective based on test samples, e.g., the entropy minimization of model outputs~\cite{tent}.
Under scenarios where test data are temporally correlated, however, na\"ively adapting to the incoming batch of test samples~\cite{bnstats, bnstats2, tent, cotta} could be problematic for both approaches: the batch is now more likely to (a) {remove instance-wise variations} that are actually useful to predict $y$, i.e., the ``contents'' rather than ``styles'' through normalization, and (b) include a bias in $p(y)$ rather than uniform, which can negatively affect the test-time adaptation objective such as entropy minimization.
We propose two approaches to tackle each of the failure modes of adapting BN
under temporal correlation. Our method consists of two components: (a) Instance-Aware Batch Normalization (IABN)~(Section \cref{sec:iabn}) to overcome the limitation of BN under distribution shift and (b) Prediction-Balanced Reservoir Sampling (PBRS)~(Section \cref{sec:pbrs}) to combat with the temporal correlation of test batches. Figure~\ref{fig:tta_main} illustrates the overall workflow of NOTE{} with IABN and PBRS.
\subsection{Instance-Aware Batch Normalization}\label{sec:iabn}
As described in Section~\cref{sec:background:bn_tta}, recent TTA algorithms rely solely on the test batch to re-calculate BN statistics. We argue that this common practice does not successfully capture the feature statistics to normalize the feature map {$\mathbf{f}\in \mathbb{R}^{B\times C \times L}$} under temporal correlation in the test batch $B$. In principle, standardizing a given feature map {$\mathbf{f}_{:,c,:}$} by the statistics $\hat{{\boldsymbol{\mu}}}_c, \hat{{\boldsymbol{\sigma}}}^2_c$ computed across $B$ and $L$ is posited on premise that averaging information across $B$ can marginalize out {uninformative} instance-wise variations {for predicting $y$}. Under temporal correlation in $B$, however, this assumption is no longer valid, and averaging across $B$ may not fully de-correlate useful information in {$\mathbf{f}_{:,c,:}$} from ${\boldsymbol{\mu}}_c$ and ${\boldsymbol{\sigma}}_c^2$.
In an attempt to bypass such an ``over-whitening'' effect of using $\hat{{\boldsymbol{\mu}}}_c$ and $\hat{{\boldsymbol{\sigma}}}^2_c$ in test-time under temporal correlation, we propose
correcting normalization statistics on a \emph{per-sample basis}: specifically, instead of completely switching from the original statistics of $(\bar{{\boldsymbol{\mu}}}, \bar{{\boldsymbol{\sigma}}}^2)$ into $(\hat{{\boldsymbol{\mu}}}_c, \hat{{\boldsymbol{\sigma}}}^2_c)$, our proposed \emph{Instance-Aware Batch Normalization} (IABN) considers the \emph{instance-wise} statistics $\tilde{{\boldsymbol{\mu}}}, \tilde{{\boldsymbol{\sigma}}}^2 \in \mathbb{R}^{B,C}$ of $\mathbf{f}$ similarly to Instance Normalization (IN) \cite{ulyanov2016instance}, namely:
\begin{equation}
\tilde{{\boldsymbol{\mu}}}_{b,c} := \frac{1}{L} \sum_{l} \mathbf{f}_{b,c,l} \text{\ and \ \ } \tilde{{\boldsymbol{\sigma}}}^2_{b,c} := \frac{1}{L} \sum_{l} (\mathbf{f}_{b,c,l} - \tilde{{\boldsymbol{\mu}}}_{b,c})^2\text{.}
\end{equation}
We assume that $\tilde{{\boldsymbol{\mu}}}_{b,c}$ and $\tilde{{\boldsymbol{\sigma}}}^2_{b,c}$ follow the \emph{sampling distribution} of a sample size $L$ in $\mathcal{N}(\bar{{\boldsymbol{\mu}}}, \bar{{\boldsymbol{\sigma}}}^2)$ as the population. Then the corresponding variances for the sample mean $\tilde{{\boldsymbol{\mu}}}_{b,c}$ and the sample variance $\tilde{{\boldsymbol{\sigma}}}^2_{b,c}$ can be calculated as:
\begin{equation}
s^2_{\tilde{{\boldsymbol{\mu}}},c} := \tfrac{\bar{{\boldsymbol{\sigma}}}^2_c}{L} \text{\ \ and \ \ } s^2_{\tilde{{\boldsymbol{\sigma}}}^2, c} := \tfrac{2\bar{{\boldsymbol{\sigma}}}^4_c}{L-1}.
\end{equation}
IABN corrects $(\bar{{\boldsymbol{\mu}}}, \bar{{\boldsymbol{\sigma}}}^2)$ only in cases when $\tilde{{\boldsymbol{\mu}}}_{b,c}$ (and $\tilde{{\boldsymbol{\sigma}}}^2_{b,c}$) significantly differ from $\bar{{\boldsymbol{\mu}}}_c$ (and $\bar{{\boldsymbol{\sigma}}}^2_c$). {Specifically,} we propose to use the following statistics for TTA:
\begin{multline}
{{\boldsymbol{\mu}}}_{b,c}^{\tt IABN} := \bar{{\boldsymbol{\mu}}}_c + \psi(\tilde{{\boldsymbol{\mu}}}_{b,c} - \bar{{\boldsymbol{\mu}}}_c; \alpha{s_{\tilde{{\boldsymbol{\mu}}},c}}) \text{, \ and \ \ } {({{\boldsymbol{\sigma}}}_{b,c}^{\tt IABN})^2} := \bar{{\boldsymbol{\sigma}}}^2_c + \psi(\tilde{{\boldsymbol{\sigma}}}^2_{b,c} - \bar{{\boldsymbol{\sigma}}}^2_c; \alpha{s_{\tilde{{\boldsymbol{\sigma}}}^2, c}}), \\ \text{ where } \psi(x; \lambda) = \begin{cases}
x - \lambda, & \text{ if } x > \lambda \\
x + \lambda, & \text{ if } x < -\lambda \\
0, & \text{ otherwise }
\end{cases} \text{\ \ \ is the soft-shrinkage function.}
\end{multline}
$\alpha\geq 0$ is the hyperparameter of IABN that determines the confidence level of the BN statistics. A high value of $\alpha$ relies more on the learned statistics (BN), while a low value of $\alpha$ is in favor of the current statistics measured from the instance. Finally, the output of IABN can be described as:
\begin{equation}
\mathrm{IABN}(\mathbf{f}_{b,c,:}; \bar{{\boldsymbol{\mu}}}_c, \bar{{\boldsymbol{\sigma}}}^2_c; \tilde{{\boldsymbol{\mu}}}_{b,c}, \tilde{{\boldsymbol{\sigma}}}^2_{b,c}) := \gamma\cdot\frac{\mathbf{f}_{b,c,:} - {{\boldsymbol{\mu}}}_{b,c}^{\tt IABN}}{\sqrt{({{\boldsymbol{\sigma}}}_{b,c}^{\tt IABN})^2 + \epsilon}} + \beta.
\end{equation}
Observe that IABN becomes IN and BN when $\alpha=0$ and $\alpha=\infty$, respectively. If one chooses too small $\alpha\geq 0$, IABN may remove useful features, e.g., styles, of input (as with IN), which can degrade the overall classification (or regression) performance~\cite{nam2018batch}. Hence, it is important to choose an appropriate $\alpha$. Nevertheless, we found that a good choice of $\alpha$ is not too sensitive
across tested scenarios, where we chose $\alpha=4$ for all experiments.
This way, IABN can be robust to distributional shifts without the risk of eliminating crucial information to predict $y$.
\subsection{Adaptation via Prediction-Balanced Reservoir Sampling}\label{sec:pbrs}
Temporally correlated distributions lead to an undesirable bias in $p(y)$, and thus adaptation with a batch of consecutive test samples negatively impacts the adaptation objective, such as entropy minimization~\cite{tent}.
To combat this imbalance, we propose Prediction-Balanced Reservoir Sampling (PBRS) that mimics i.i.d.\ samples from temporally correlated streams with the assistance of a small (e.g., a mini-batch size) memory. PBRS combines \textit{time-uniform} sampling and \textit{prediction-uniform} sampling to simulate i.i.d.\ samples from the non-i.i.d.\ streams. For time-uniform sampling, we adopt reservoir sampling (RS)~\cite{rs}, a proven random sampling algorithm to collect time-uniform data in a single pass on a stream without the prior knowledge of the total length of data. For prediction-uniform sampling, we first use the predicted labels to compute the majority class(es) in the memory. We then replace a random instance of the majority class(es) with a new sample. We detail the algorithm of PBRS as a pseudo-code in Algorithm~\ref{alg:pbrs}. We found that these two heuristics can effectively balance samples among both time and class axes, which mitigates the bias in temporally correlated data.
\begin{algorithm}[t]
\caption{Prediction-Balanced Reservoir Sampling}\label{alg:pbrs}
\begin{algorithmic}[1]
\Require {target stream $\mathbf{x}_t \sim P_{\mathcal{T}}(\mathbf{x}|t)$, memory bank $M$ of capacity $N$
}
\State $M[i] \leftarrow \phi$ for $i = 1, \cdots N$; and $n[c] \leftarrow 0$ for $c \in \mathcal{Y}$
\For {$t\in\{1, \cdots, T\}$}
\State $n[{\hat{y}_t}] \leftarrow n[{\hat{y}_t}] + 1$ \textcolor{gray}{\quad// increase the number of samples encountered for the class}
\State $m[c] \leftarrow |\{ (\mathbf{x}, y) \in M | y = c \}|$ for $c \in \mathcal{Y}$ \textcolor{gray}{\quad// count instances per class in memory}
\If {$|M|\ < N$} \textcolor{gray}{\quad// if memory is not full}
\State Add $(\mathbf{x}_t, \hat{y}_t)$ to $M$
\Else
\State $C^*\gets \argmax_{c \in \mathcal{Y}} m[c]$ \textcolor{gray}{\quad// get majority class(es)}
\If {$\hat{y}_t \notin C^*$} \textcolor{gray}{\quad// if the new sample is not majority} \Comment{Prediction-Balanced}
\State Randomly pick $M[i] := (\mathbf{x}_i, \hat{y_i})$ where $\hat{y_i} \in C^*$
\State $M[i] \leftarrow (\mathbf{x}_t, \hat{y}_t)$ \textcolor{gray}{\quad// replace it with a new sample}
\Else\Comment{Reservoir Sampling}
\State Sample $p\sim\mathrm{Uniform}(0,1)$
\If {$p < m[\hat{y}_t] / n[{\hat{y}_t}]$}
\State Randomly pick $M[i] := (\mathbf{x}_i, \hat{y_i})$ where $\hat{y_i}=\hat{y}_t$
\State $M[i] \leftarrow (\mathbf{x}_t, \hat{y}_t)$ \textcolor{gray}{\quad// replace it with a new sample}
\EndIf
\EndIf
\EndIf
\EndFor
\end{algorithmic}
\end{algorithm}
With the stored samples in the memory, we update the normalization statistics and affine parameters in the IABN layers. Note that IABN assumes $\tilde{{\boldsymbol{\mu}}}_{b,c}$ and $\tilde{{\boldsymbol{\sigma}}}^2_{b,c}$ follow the sampling distribution of $\mathcal{N}(\bar{{\boldsymbol{\mu}}}, \bar{{\boldsymbol{\sigma}}}^2)$ and corrects the normalization if $\tilde{{\boldsymbol{\mu}}}_{b,c}$ and $\tilde{{\boldsymbol{\sigma}}}^2_{b,c}$ are out of distribution. While IABN is resilient to distributional shifts to a certain extent, the assumption might not hold under severe distributional shifts. Therefore, we aim to find better estimates of $\bar{{\boldsymbol{\mu}}}, \bar{{\boldsymbol{\sigma}}}^2$ in IABN under distributional shifts via PBRS. Specifically, we update the normalization statistics, namely the means ${\boldsymbol{\mu}}$ and variances ${\boldsymbol{\sigma}}^2$, via exponential moving average: (a) ${\boldsymbol{\mu}}_{t}=(1-m){\boldsymbol{\mu}}_{t-1}+m\frac{N}{N-1}\hat{{\boldsymbol{\mu}}}_{t}$ and (b) ${\boldsymbol{\sigma}}^{2}_{t}=(1-m){\boldsymbol{\sigma}}^{2}_{t-1}+m\frac{N}{N-1}\hat{{\boldsymbol{\sigma}}}^{2}_{t}$ where $m$ is a momentum and $N$ is the size of the memory. We further optimize the affine parameters, scaling factor $\gamma$ and bias term $\beta$ via a single backward pass with entropy minimization, similar to a previous study~\cite{tent}. These parameters account for only around 0.02\% of the total trainable parameters in ResNet18~\cite{7780459}. The IABN layers are adapted with the $N$ samples in the memory every $N$ test samples. We set the memory size $N$ as 64 following the common batch size of existing TTA methods~\cite{bnstats2, lame, tent} to ensure a fair memory constraint.
\subsection{Inference}
NOTE{} infers each sample via a single forward pass with IABN layers.
Note that NOTE{} requires only a single instance for inference, different from the state-of-the-art TTA methods~\cite{bnstats, bnstats2, tent, lame, cotta} that require batches for every inference. Moreover, NOTE{} requires only one forwarding pass for inference, while multiple forward passes are required in other TTA methods that utilize augmentations~\cite{sun2020test, cotta}. The batch-free single-forward inference of NOTE{} is beneficial in latency-sensitive tasks such as autonomous driving and human health monitoring. After inference, NOTE{} determines whether to store the sample and predicted label to the memory via PBRS.
\section{Experiments}\label{sec:experiment}
We implemented NOTE{} and the baselines via the PyTorch framework~\cite{NEURIPS2019_9015}.\footnote{\url{https://github.com/TaesikGong/NOTE}} We ran all experiments with three random seeds and report the means and standard deviations. Additional experimental details, e.g., hyperparameters of the baselines, datasets, are specified in Appendix~\ref{app:details}.
\paragraph{Baselines.} We consider the following baselines including the state-of-the-art test-time adaptation algorithms:
\textbf{Source} evaluates the model trained from the source data directly on the target data without adaptation. Test-time normalization (\textbf{BN stats})~\cite{bnstats, bnstats2} updates the BN statistics from a batch of test data. Online Domain Adaptation (\textbf{ONDA})~\cite{onda} adapts batch normalization statistics to target domains via a batch of target data with an exponential moving average. Pseudo-Label (\textbf{PL})~\cite{pl} optimizes the trainable parameters in BN layers via hard pseudo labels. We update the BN layers only in PL, as done in previous studies~\cite{tent,cotta}. Test entropy minimization (\textbf{TENT})~\cite{tent} updates the BN parameters via entropy minimization. Laplacian Adjusted Maximum-likelihood Estimation (\textbf{LAME})~\cite{lame} takes a more conservative approach; it modifies the classifier's output probability and not the internal parameters of the model itself. By doing so, it prevents the model parameters from over-adapting to the test batch. Continual test-time adaptation (\textbf{CoTTA})~\cite{cotta} reduces the error accumulation by using weight-averaged and augmentation-averaged predictions. It avoids catastrophic forgetting by stochastically restoring a part of the neurons to the source pre-trained weights.
\paragraph{Adaptation and hyperparameters.} We assume the model pre-trained with source data is available for TTA. In NOTE{}, we replaced BN with IABN during training. We set the test batch size as 64 and the adaptation epoch as one for adaptation, which is the most common setting among the baselines~\cite{bnstats2, lame, tent}. Similarly, we set the memory size $N$ as 64 and adapt the model every 64 samples in NOTE{} to ensure a fair memory constraint.
We conduct online adaptation and evaluation, where the model is continually updated.
For the baselines, we adopt the best values for the hyperparameters reported in their papers or the official codes.
We followed the guideline to tune the hyperparameters when such a guideline was available~\cite{cotta}. We use fixed values for the hyperparameters of NOTE{}, soft-shrinkage width $\alpha=4$ and exponential moving average momentum $m=0.01$, and update the affine parameters via the Adam optimizer~\cite{kingma:adam} with a learning rate $l=0.0001$ unless specified. We detailed hyperparameter information of the baselines in Appendix~\ref{app:details:baselines}.
\subsection{Robustness under corruptions}
\begin{table}[t]
\centering
\caption{Average classification error (\%) and their corresponding standard deviations on CIFAR10-C/100-C and ImageNet-C under temporally correlated (non-i.i.d.) and uniformly distributed (i.i.d.) test data stream. \textbf{Bold} fonts indicate the lowest classification errors, while \ch{Red} fonts show performance degradation after adaptation. Values encompassed by parentheses refers to NOTE{} used directly with test batches (without using PBRS). Averaged over 3 runs.}\label{tab:cifar}
\vspace{0.1cm}
\resizebox{\textwidth}{!}{
\begin{tabular}{lcccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
& \multicolumn{4}{c}{{\textbf{Temporally correlated test stream}}} & \multicolumn{4}{c}{{\textbf{Uniformly distributed test stream}}} \\
\addlinespace[0.00cm]
\cmidrule(lr){2-5} \cmidrule(lr){6-9}
\addlinespace[0.0cm]
Method & CIFAR10-C & \multicolumn{1}{c}{CIFAR100-C} & ImageNet-C & Avg & CIFAR10-C & \multicolumn{1}{c}{CIFAR100-C} & ImageNet-C & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & 42.3 ± 1.1 & \multicolumn{1}{c}{66.6 ± 0.1} & 86.1 ± 0.0 & 65.0 & 42.3 ± 1.1 & 66.6 ± 0.1 & 86.1 ± 0.0 & 65.0 \\
BN Stats {\cite{bnstats}} & \ch{73.4 ± 1.3} & \multicolumn{1}{c}{65.0 ± 0.3} & \ch{96.9 ± 0.0} & \ch{78.5} & 21.6 ± 0.4 & 46.6 ± 0.2 & 76.0 ± 0.0 & 48.1 \\
ONDA {\cite{onda}} & \ch{63.6 ± 1.0} & \multicolumn{1}{c}{49.6 ± 0.3} & \ch{89.0 ± 0.0} & \ch{67.4} & 21.7 ± 0.4 & 46.5 ± 0.1 & 75.9 ± 0.0 & 48.0 \\
PL {\cite{pl}} & \ch{75.4 ± 1.8} & \multicolumn{1}{c}{66.4 ± 0.4} & \ch{98.9 ± 0.0} & \ch{80.2} & 21.6 ± 0.2 & 43.1 ± 0.3 & 74.4 ± 0.2 & 46.4 \\
TENT {\cite{tent}} & \ch{76.4 ± 2.7} & \multicolumn{1}{c}{\ch{66.9 ± 0.6}} & \ch{96.9 ± 0.0} & \ch{80.1} & 18.8 ± 0.2 & \textbf{40.3 ± 0.2} & 76.0 ± 0.0 & 45.0 \\
LAME {\cite{lame}} & {36.2 ± 1.3} & \multicolumn{1}{c}{63.3 ± 0.3} & \ch{82.7 ± 0.0} & 60.7 & \ch{44.1 ± 0.5} & \ch{68.8 ± 0.1} & \ch{86.3 ± 0.0} & \ch{66.4} \\
CoTTA {\cite{cotta}} & \ch{75.5 ± 0.7} & \multicolumn{1}{c}{64.2 ± 0.2} & \ch{97.0 ± 0.0} & \ch{78.9} & 17.8 ± 0.3 & 44.3 ± 0.2 & 71.5 ± 0.0 & 44.6 \\
\addlinespace[0.0cm]
\hline
\addlinespace[0.05cm]
NOTE & \textbf{21.1 ± 0.6} & \multicolumn{1}{c}{\textbf{47.0 ± 0.1}} & \textbf{80.6 ± 0.1} & \textbf{49.6} & \begin{tabular}[c]{@{}c@{}}20.1 ± 0.5\\ (\textbf{\small 17.6 ± 0.3})\end{tabular} & \multicolumn{1}{c}{\begin{tabular}[c]{@{}c@{}}46.4 ± 0.0\\ (\small 41.0 ± 0.2)\end{tabular}} & \begin{tabular}[c]{@{}c@{}}\textbf{70.3 ± 0.0}\\ ({\small 71.7 ± 0.0})\end{tabular} & \begin{tabular}[c]{@{}c@{}}45.6\\ \textbf{\small (\textbf{43.4)}}\end{tabular} \\
\addlinespace[0.0cm]
\Xhline{2\arrayrulewidth}
\end{tabular}%
}
\end{table}
\begin{figure}[t]
\centering
\begin{minipage}[t]{0.32\textwidth}
\centering
\includegraphics[width=\linewidth]{fig/dirichlet_settings.pdf}
\caption{Illustration of synthetic non-i.i.d.\ streams sampled from Dirichlet distribution varying $\delta$ on CIFAR10-C. \textit{uniform} denotes an i.i.d.\ condition. The lower the $\delta$, the more temporally correlated the distribution.}\label{fig:dirichlet_settings}
\end{minipage}
\hspace{0.1cm}
\begin{minipage}[t]{0.64\textwidth}
\centering
\begin{subfigure}[t]{0.44\linewidth}
\centering
\includegraphics[width=\linewidth]{fig/acc_var1.pdf}
\caption{Effect of Dirichlet concentration parameter $\delta$.}\label{fig:acc_var:delta}
\end{subfigure}
~
\begin{subfigure}[t]{0.44\linewidth}
\includegraphics[width=\linewidth]{fig/acc_var2.pdf}
\caption{Effect of batch size.}\label{fig:acc_var:bs}
\end{subfigure}
\caption{Average classification error (\%) under the non-i.i.d.\ setting with CIFAR10-C dataset. We vary (a) the Dirichlet concentration parameter $\delta$ to investigate the impact of the degree of temporal correlation and (b) batch size to understand the behaviors of the TTA methods. Averaged over 3 runs. Lower is better.}\label{fig:acc_var}
\end{minipage}
\end{figure}
\paragraph{Datasets.} We use CIFAR10-C, CIFAR100-C, and ImageNet-C~\cite{cifarc} datasets that are common TTA benchmarks for evaluating the robustness to corruptions~\cite{bnstats, bnstats2, tent, cotta, lame}. Both CIFAR10/CIFAR100~\cite{cifar} have 50,000/10,000 training/test data. ImageNet~\cite{imagenet} has 1,281,167/50,000 training/test data. CIFAR10/CIFAR100/ImageNet have 10/100/1,000 classes, respectively. CIFAR10-C/CIFAR100-C/ImageNet-C apply 15 types of corruption to CIFAR10/CIFAR100/ImageNet test data. Similar to previous studies~\cite{bnstats, bnstats2, tent, cotta}, we use the most severe corruption level of 5. We use ResNet18~\cite{7780459} as the backbone network and pre-trained it on the clean training data. Following prior studies~\cite{li2021federated, hsu2019measuring, wang2020tackling, wang2020federated}, we adopt Dirichlet distribution to generate synthetic non-i.i.d.\ test streams from the originally i.i.d.\ CIFAR10/100 data. The details are provided in Appendix~\ref{app:details:datasets}. We vary the Dirichlet concentration parameter $\delta$ to simulate diverse streams and visualize the resulting data in Figure~\ref{fig:dirichlet_settings}. We use $\delta=0.1$ as the default value unless specified. For ImageNet, we sort the test stream as the number of test samples per class is not enough for generating temporally correlated streams via Dirichlet distribution. We additionally provide an experiment with MNIST-C data~\cite{mu2019mnist} in the appendix, which shows similar takeaways to our experiments with CIFAR10-C/CIFAR100-C/ImageNet-C.
\paragraph{Overall result.} Tables~\ref{tab:cifar} shows the result under the temporally correlated (non-i.i.d.) data and the uniform (i.i.d.) data, respectively. We observe significant performance degradation in the baselines under the temporally correlated setting. For BN Stats, PL, TENT, and CoTTA, this degradation is particularly due to the dependence on the test batch for the re-calculation of the BN statistics. Updating the batch statistics from test data via exponential moving average (ONDA) also suffers from the temporally correlated data. This indicates relying on the test batch for re-calculating the BN statistics indeed cancels out meaningful instance-wise variations under temporal correlation. Interestingly, LAME works better in the non-i.i.d. setting than in the i.i.d. setting, which is consistent with previous reports~\cite{lame}. The primary reason is, as stated by the authors, it ``discourages deviations from the predictions of the pre-trained model,'' and thus it ``does not noticeably help in i.i.d and class-balanced scenarios.''
NOTE{} achieves on average 11.1\% improvement over the best baseline (LAME) under the non-i.i.d.\ setting. With the i.i.d.\ assumption, NOTE{} still achieves comparable performance to the baselines. When we know the target samples are i.i.d., we can simply use the test batch without using PBRS. For this variant version of NOTE{}, we update IABN with incoming test batches directly using a ten times higher learning rate of 0.001 following previous work~\cite{tent, cotta}. We report the result of the variant version of NOTE{} in the parentheses, which achieves on average 2.2\% improvement further when the i.i.d.\ assumption is known.
\paragraph{Effect of the degree of temporal correlation.} We also investigate the effect of the degree of temporal correlation for TTA algorithms. Figure~\ref{fig:acc_var:delta} shows the result. The lower $\delta$ is, the severer temporal correlation becomes. The error rates of most of the baselines deteriorate as $\delta$ decreases, which shows that the existing TTA baselines are susceptible to temporally correlated data. NOTE{} shows consistent performance among all $\delta$ values, indicating its robustness under temporal correlation.
\paragraph{Effect of batch size.} While we experiment with a widely-used value of 64 as the batch size (or memory size in NOTE{}), one might be curious about the impact of batch size under temporally correlated streams.
Figure~\ref{fig:acc_var:bs} shows the result with six different batch sizes. As shown, NOTE{} is not much affected by the batch size, while most of the baselines recovers performance degradation as the batch size increases. This is because a higher batch size has a better chance of adaptation with balanced samples under temporally correlated streams. Increasing the batch size, however, mitigates temporal correlation at the expense of inference latency and adaptation speed.
\subsection{Real-distributions with domain shift}
\paragraph{Datasets.} We evaluate NOTE{} under three real-world distribution datasets: object detection in autonomous driving (KITTI~\cite{kitti}), human activity recognition (HARTH~\cite{harth}), and user behavioral context recognition (ExtraSensory~\cite{extrasensory}). Additional dataset-specific details are in Appendix~\ref{app:details:datasets}.
KITTI is a well-known autonomous driving dataset that provides consecutive frames that contains natural temporal correlation in driving contexts. We adapt from KITTI to KITTI-Rain~\cite{kitti-rain} - a dataset that converted KITTI images to rainy images. This contains 7,481/7,800 train/test samples with nine classes. We use ResNet50~\cite{7780459} pre-trained on ImageNet~\cite{deng2009imagenet} as the backbone network.
HARTH was collected from 22 users in free-living environments for seven days. Each user was equipped with two three-axial Axivity AX3 accelerometers for recording human activities. We use 15 users collectively as the source domain and the remaining seven users as each target domain, which entails natural domain shifts from source users to target users as different physical conditions make domain shifts across users. HARTH contains 82,544/39,377 train/test samples with 12 classes. We report the averaged error over all target domains. We use four one-dimensional convolutional layers followed by one fully-connected layer as the backbone network for HARTH.
The Extrasensory dataset collected users' own smartphone sensory data (motion sensors, audio, etc.) in the wild for seven days, aiming to capture people’s authentic behaviors in their regular activities. We use 16 users as the source domain and seven users as target domains. ExtraSensory includes 17,777/4,862 train/test data with five classes. For ExtraSensory, we use two one-dimensional convolutional layers followed by one fully-connected layer as the backbone network. For both HARTH and ExtraSensory models, a single BN layer follows each convolutional layer.
\begin{table}[t]
\centering
\caption{Average classification error (\%) and their corresponding standard deviations on three real test data streams: KITTI, HARTH, and ExtraSensory. \textbf{Bold} fonts indicate the lowest classification errors, while \ch{Red} fonts show performance degradation after adaptation. Averaged over 3 runs.}\label{tab:real}
\vspace{0.1cm}
\begin{tabular}{lcccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
& \multicolumn{4}{c}{{\textbf{Real test stream}}} \\
\addlinespace[0.00cm]
\cmidrule(lr){2-5}
\addlinespace[0.00cm]
Method & KITTI & HARTH & \multicolumn{1}{c}{ExtraSensory} & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & 12.3 ± 2.3 & 62.6 ± 8.5 & \multicolumn{1}{c}{50.2 ± 2.2} & 41.7 \\
BN Stats \cite{bnstats} & \ch{35.4 ± 0.5} & \ch{68.6 ± 1.1} & \multicolumn{1}{c}{\ch{56.0 ± 0.9}} & \ch{53.4} \\
ONDA \cite{onda} & \ch{26.3 ± 0.5} & \ch{69.3 ± 1.1} & \multicolumn{1}{c}{48.2 ± 1.5} & \ch{47.9} \\
PL \cite{pl} & \ch{39.0 ± 0.3} & \ch{64.8 ± 0.6} & \multicolumn{1}{c}{\ch{56.0 ± 0.9}} & \ch{53.3} \\
TENT \cite{tent} & \ch{39.6 ± 0.2} & \ch{64.1 ± 0.7} & \multicolumn{1}{c}{\ch{56.0 ± 0.8}} & \ch{53.2} \\
LAME \cite{lame} & 11.3 ± 2.9 & 61.0 ± 10.0 & \multicolumn{1}{c}{\ch{50.7 ± 2.7}} & 41.0 \\
CoTTA \cite{cotta} & \ch{35.4 ± 0.6} & \ch{68.7 ± 1.1} & \multicolumn{1}{c}{\ch{56.0 ± 0.9}} & \ch{53.4} \\
\addlinespace[0.00cm]
\hline
\addlinespace[0.05cm]
NOTE & \textbf{10.9 ± 3.6} & \textbf{51.0 ± 5.6} & \multicolumn{1}{c}{\textbf{45.4 ± 2.6}} & \textbf{35.8} \\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}
\end{table}
\paragraph{Result.} Table~\ref{tab:real} shows the result for the real-world datasets. The overall trend is similar to the temporal correlation experiments with CIFAR10-C/CIFAR100-C datasets, which indicates that the real-world datasets are indeed temporally correlated. NOTE{} consistently reduces errors after adaptation under real-world distributions. We believe this demonstrates NOTE{} is a promising method to be utilized in various real-world ML applications with distributional shifts. We illustrate real-time classification error changes for real-world datasets in the appendix.
\subsection{Ablation study}
\begin{table}[t]
\centering
\caption{Average classification error (\%) and corresponding standard deviations of varying ablation settings on CIFAR10-C/100-C under temporally correlated (non-i.i.d.) and uniformly distributed (i.i.d.) test data stream. \textbf{Bold} type indicates those of lowest classification error. Averaged over 3 runs.}\label{tab:ablation}
\vspace{0.1cm}
\begin{tabular}{lcccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
& \multicolumn{3}{c}{{\textbf{Temporally correlated test stream}}} & \multicolumn{3}{c}{{\textbf{Uniformly distributed test stream}}} \\
\addlinespace[0.0cm]
\cmidrule(lr){2-4} \cmidrule(lr){5-7}
\addlinespace[0.0cm]
Method & CIFAR10-C & \multicolumn{1}{c}{CIFAR100-C} & Avg & CIFAR10-C & \multicolumn{1}{c}{CIFAR100-C} & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & 42.3 ± 1.1 & \multicolumn{1}{c}{66.6 ± 0.1} & 54.4 & 42.3 ± 1.1 & \multicolumn{1}{c}{66.6 ± 0.1} & 54.4 \\
IABN & 24.6 ± 0.6 & \multicolumn{1}{c}{54.5 ± 0.1} & 39.5 & 24.6 ± 0.6 & \multicolumn{1}{c}{54.5 ± 0.1} & 39.5 \\
PBRS & 27.5 ± 1.0 & \multicolumn{1}{c}{51.7 ± 0.2} & 39.6 & 25.8 ± 0.2 & \multicolumn{1}{c}{51.3 ± 0.1} & 38.5 \\
IABN+RS & \textbf{20.5 ± 1.5} & \multicolumn{1}{c}{48.2 ± 0.2} & 34.3 & 20.7 ± 0.6 & \multicolumn{1}{c}{48.3 ± 0.3} & 34.5 \\
IABN+PBRS & 21.1 ± 0.6 & \multicolumn{1}{c}{\textbf{47.0 ± 0.1}} & \textbf{34.0} & \textbf{20.1 ± 0.5} & \multicolumn{1}{c}{\textbf{46.4 ± 0.0}} & \textbf{33.2} \\
\addlinespace[0.0cm]
\Xhline{2\arrayrulewidth}
\end{tabular}
\end{table}
We conduct an ablative study to further investigate the effectiveness of the individual components. Table~\ref{tab:ablation} shows the result under both i.i.d.\ and non-i.i.d.\ settings. Using IABN alone significantly reduces error rates over Source, demonstrating the effectiveness of correcting normalization for out-of-distribution samples.
Using PBRS with BN shows comparable improvement with the IABN-only result. Note that there is only a marginal gap (around 1\%) between the non-i.i.d. and i.i.d. results in PBRS. This indicates that PBRS could effectively simulate i.i.d.\ samples from non-i.i.d.\ streams. The joint use of IABN and PBRS outperforms using either of them, meaning that PBRS provides IABN with better estimates for the normalizing operation. In addition, PBRS is better than Reservoir Sampling (RS) that has been a strong baseline in continual learning~\cite{10.5555/3504035.3504439, Chaudhry_tinyer_icml2019}. This shows storing prediction-balanced sampling in addition to time-uniform sampling leads to better adaptation in TTA. We also investigated the joint use of IN and PBRS with the combination of IABN and PBRS on CIFAR100-C, and the result shows that IABN+PBRS (47.0\%) achieves lower error rate than IN+PBRS (52.5\%) on CIFAR100-C under temporal correlation.
\section{Related work}
\section{Related work}
\paragraph{Test-time adaptation.} Test-time adaptation~(TTA) attempts to overcome distributional shifts with test data without the cost of data acquisition or labeling. TTA adapts to the target domain with only test data on the fly. Most existing TTA algorithms rely on a batch of test samples to adapt~\cite{bnstats, bnstats2, tent, cotta} to re-calibrate BN layers on the test data. Simply using the statistics of test batch in BN layers improves the robustness under distributional shifts~\cite{bnstats, bnstats2}. ONDA~\cite{onda} updates the BN statistics with test data via exponential moving average. TENT~\cite{tent} further updates the scaling and bias parameters in BN layers via entropy minimization.
Latest TTA studies consider distribution changes of test data~\cite{lame, cotta}. LAME~\cite{lame} corrects the output probabilities of a classifier rather than tweaking the model’s inner parameters. By restraining the model from over-adapting to the test batch, LAME allows the model to be more robust under non-i.i.d. scenarios. However, LAME does not have noticeable performance gains in class-balanced, standard i.i.d.\ scenarios. The primary reason is, as stated by the authors, it ``discourages deviations from the predictions of the pre-trained model,'' and thus it ``does not noticeably help in i.i.d and class-balanced scenarios.'' CoTTA~\cite{cotta} aims to adapt to continually changing target environments via weight-averaged teacher model, weight-averaged augmentations, and stochastic restoring. However, CoTTA assumes i.i.d.\ test data within each domain and updates the entire model that increases computational costs.
There also exist works~\cite{sun2020test, liu2021ttt++} utilizing domain-specific self-supervision to resolve the distribution shift with test data, but are complementary to ours, i.e., we can also optimize the self-supervised loss instead of the entropy loss, and
not applicable to our setups of real test data stream as designing good self-supervision for these domains is highly non-trivial.
\paragraph{Replay memory.} Replay memory is one of the major approaches in continual learning; it manages a buffer to replay previous data for future learning to prevent catastrophic forgetting. Reservoir sampling~\cite{rs} is a random sampling algorithm that collects time-uniform samples from unknown sample streams with a single pass, and it has been proven to be a strong baseline in continual learning~\cite{10.5555/3504035.3504439, Chaudhry_tinyer_icml2019}. GSS~\cite{10.5555/3454287.3455345} stores samples to a memory in a way that maximizes the gradient direction among those samples. A recent study modifies reservoir sampling to balance classes under imbalanced data, when the labels are given~\cite{pmlr-v119-chrysakis20a}. Our memory management scheme (PBRS) is inspired by these studies to prevent catastrophic forgetting in test-time adaptations.
\section{Discussion and conclusion}\label{sec:discussion}
This paper highlights that real-world distributions often change across the time axis, and existing test-time adaptation algorithms mostly suffer from the non-i.i.d.\ test data streams. To address this problem, we present a NOn-i.i.d.\ TEst-time adaptation algorithm, NOTE{}. Our experiments evaluated robustness under corruptions and domain adaptation on real-world distributions. The results demonstrate that NOTE{} not only outperforms the baselines under the non-i.i.d./real distribution settings, but it also shows comparable performance
under the i.i.d.\ assumption. We believe that the insights and findings from this study are a meaningful step toward the practical impact of the test-time adaptation paradigm.
\paragraph{Limitations.} {
NOTE{} and most state-of-the-art TTA algorithms~\cite{bnstats, pl, onda, bnstats2, tent, cotta} assume that the backbone networks are equipped with BN (or IABN) layers. While BN is a widely-used component in deep learning, several architectures, such as LSTMs~\cite{hochreiter1997long} and Transformers~\cite{vaswani2017attention}, do not embed BN layers. A recent study uncovered that BN is advantageous in Vision Transformers~\cite{9607565}, showing potential room to apply our idea to architectures without BN layers. However, more in-depth studies are necessary to identify the actual applicability of BN (or IABN) to those architectures. While LAME~\cite{lame} is applicable to models without BN, its limitation is the performance drop in i.i.d. scenarios, as shown in both its paper and our evaluation.
While NOTE shows its effectiveness in both non-i.i.d and i.i.d. scenarios, a remaining challenge is to design an algorithm that generalizes to any architecture. We believe the findings and contributions of our work could give valuable insights to future endeavors on this end.
}
\paragraph{Potential negative societal impacts.} As TTA relies on unlabeled test samples and changes the model accordingly, the model is exposed to potential data-driven biases after adaptation, such as fairness issues~\cite{barocas2017fairness} and adversarial attacks~\cite{szegedy2013intriguing}. In some sense, the utility of TTA comes at the expense of exposure to threats. This vulnerability is another crucial problem that both ML researchers and practitioners need to take into consideration. In addition, TTA entails additional computations for adaptation with test data, which may have negative impacts on environments, e.g., increasing electricity consumption and carbon emissions~\cite{10.1145/3381831}. Nevertheless, we believe NOTE{} would not exacerbate this issue as it is computationally efficient as mentioned in Section~\cref{sec:intro}.
\section*{Checklist}
\begin{enumerate}
\item For all authors...
\begin{enumerate}
\item Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
\answerYes{}
\item Did you describe the limitations of your work?
\answerYes{See Section~\cref{sec:discussion}.}
\item Did you discuss any potential negative societal impacts of your work?
\answerYes{See Section~\cref{sec:discussion}.}
\item Have you read the ethics review guidelines and ensured that your paper conforms to them?
\answerYes{}
\end{enumerate}
\item If you are including theoretical results...
\begin{enumerate}
\item Did you state the full set of assumptions of all theoretical results?
\answerNA{}
\item Did you include complete proofs of all theoretical results?
\answerNA{}
\end{enumerate}
\item If you ran experiments...
\begin{enumerate}
\item Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)?
\answerYes{Code is available at \url{https://github.com/TaesikGong/NOTE}.}
\item Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)?
\answerYes{See Section~\cref{sec:experiment} and Appendix~\ref{app:details}.}
\item Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)?
\answerYes{We ran the entire experiments with three different random seems (0,1,2) and reported the standard deviations.}
\item Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)?
\answerYes{See Appendix~\ref{app:details}.}
\end{enumerate}
\item If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
\begin{enumerate}
\item If your work uses existing assets, did you cite the creators?
\answerYes{}
\item Did you mention the license of the assets?
\answerYes{}
\item Did you include any new assets either in the supplemental material or as a URL?
\answerYes{}
\item Did you discuss whether and how consent was obtained from people whose data you're using/curating?
\answerYes{}
\item Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content?
\answerYes{}
\end{enumerate}
\item If you used crowdsourcing or conducted research with human subjects...
\begin{enumerate}
\item Did you include the full text of instructions given to participants and screenshots, if applicable?
\answerNA{}
\item Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable?
\answerNA{}
\item Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation?
\answerNA{}
\end{enumerate}
\end{enumerate}
\section{Experimental details}\label{app:details}
For the all experiments in the paper, we used three different random seeds (0, 1, 2) and report the average errors (and standard deviations). We ran our experiments on NVIDIA GeForce RTX 3090 GPUs.
\subsection{Baseline details}\label{app:details:baselines}
We referred to the official implementations of the baselines. We use the reported best hyperparameters from their paper or code. We further tuned hyperparameters if there exists a hyperparameter selection guideline. Here, we provide additional details of the baseline implementations including hyperparameters.
\paragraph{PL.} Following the previous studies~\cite{tent,cotta}, we update the BN layers only in PL. We set the learning rate as $LR=0.001$ as the same as~\cite{tent}.
\paragraph{ONDA.} ONDA~\cite{onda} has two hyperparameters, the update frequency $N$ and the decay of the moving average $m$. The authors set $N=10$ and $m=0.1$ as the default values throughout the experiments, and we follow this choice unless specified.
\paragraph{TENT.} TENT~\cite{tent} set the learning rate as $LR = 0.001$ for all datasets except for ImageNet, and we follow this choice. We referred to the official code\footnote{\url{https://github.com/DequanWang/tent}} for implementing TENT.
\paragraph{LAME.} LAME~\cite{lame} needs an affinity matrix and has hyperparameters related to it. We follow the authors' choice on hyperparameter selection specified in both the paper and in their official code. Namely, we use the kNN affinity matrix and with the value of k set as 5. We referred to the official code\footnote{\url{https://github.com/fiveai/LAME}} for implementing LAME.
\paragraph{CoTTA.} CoTTA~\cite{cotta} has three hyperparameters, augmentation confidence threshold $p_{th}$, restoration factor $p$, and exponential moving average (EMA) factor $m$. We follow the authors' choice for restoration factor ($p=0.01$) and EMA factor ($\alpha=0.999$). For augmentation confidence threshold, the authors provide a guideline to choose it, which is using 5\% quantile for the softmax predictions’ confidence on the source domains. We follow this guideline, which results in $p_{th}=0.92$ for MNIST-C and CIFAR10-C, $p_{th}=0.72$ for CIFAR100-C, and $p_{th}=0.55$ for KITTI. For 1D time-series datasets (HARTH and ExtraSensory), augmentations are not provided from the authors, and it is non-trivial to select appropriate augmentations for them. We thus do not use augmentations for these datasets. We referred to the official code\footnote{\url{https://github.com/qinenergy/cotta}} for implementing CoTTA.
\subsection{Dataset details}\label{app:details:datasets}
\subsubsection{Robustness to corruptions}
\paragraph{MNIST-C.} MNIST-C~\cite{mu2019mnist} applies 15 corruptions to the MNIST~\cite{lecun1998gradient} dataset. Specifically, the corruptions include Shot Noise, Impulse Noise, Glass Blur, Motion Blur, Shear, Scale, Rotate, Brightness, Translate, Stripe, Fog, Spatter, Dotted Line, Zigzag, and Canny Edges, as illustrated in Figure~\ref{fig:mnist_c_viz}. Note that the result of this dataset is included only in the supplementary material. In total, MNIST-C has 60,000 clean training data and 150,000 corrupted test data (10,000 for each corruption type). We use ResNet18~\cite{7780459} as the backbone network. We train it on the clean training data to generate source models, using stochastic gradient descent with momentum=0.9 and cosine annealing learning rate scheduling~\cite{loshchilov2016sgdr} for 100 epochs with an initial learning rate of 0.1.
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{fig/mnist_c_viz.pdf}
\caption{Illustration of the 15 corruption types in the MNIST-C dataset.}\label{fig:mnist_c_viz}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{fig/cifar_c_viz.pdf}
\caption{Illustration of the 15 corruption types in the CIFAR10-C/CIFAR100-C/ImageNet-C} dataset.\label{fig:cifar_c_viz}
\end{figure}
\paragraph{CIFAR10-C/CIFAR100-C.} CIFAR10-C/CIFAR100-C~\cite{cifarc} are common TTA benchmarks for evaluating the robustness to corruptions~\cite{bnstats, bnstats2, tent, cotta}. Both CIFAR10/CIFAR100~\cite{cifar} have 50,000/10,000 training/test data. CIFAR10/CIFAR100 have 10/100 classes respectively. CIFAR10-C/CIFAR100-C apply 15 types of corruptions to CIFAR10/CIFAR100 test data: Gaussian Noise, Shot Noise, Impulse Noise, Defocus Blur, Frosted Glass Blur, Motion Blur, Zoom Blur, Snow, Frost, Fog, Brightness, Contrast, Elastic Transformation, Pixelate, and JPEG Compression, as illustrated in Figure~\ref{fig:cifar_c_viz}. We use the most severe corruption level of 5, similar to previous studies~\cite{bnstats, bnstats2, tent, cotta}. This results in total 150,000 test data for CIFAR10-C/CIFAR100-C, respectively. We use ResNet18~\cite{7780459} as the backbone network. We train it on the clean training data to generate source models, using stochastic gradient descent with momentum=0.9 and cosine annealing learning rate scheduling~\cite{loshchilov2016sgdr} for 200 epochs with an initial learning rate of 0.1 and a batch size of 128.
\paragraph{ImageNet-C.} ImageNet-C is another common TTA benchmark for evaluating the robustness to corruptions~\cite{bnstats, bnstats2, tent, cotta, lame}. ImageNet~\cite{imagenet} has 1,281,167/50,000 training/test data. ImageNet-C applies the same 15 types of corruptions used in CIFAR10-C and CIFAR100-C. We use a pre-trained ResNet18~\cite{7780459} on ImageNet training data and fine-tune it by replacing BN layers with IABN layers on the clean ImageNet training data. For fine-tuning, we use stochastic gradient descent with momentum=0.9 for 30 epochs with a fixed learning rate of 0.001 and a batch size of 256.
\paragraph{Temporally correlated streams via Dirichlet distribution.} Note that most public vision datasets are not time-series data and existing TTA studies usually shuffled the order of these datasets resulting in i.i.d.\ streams, which might be unrealistic in real-world scenarios. To simulate non-i.i.d. streams from these ``static'' datasets, we utilize Dirichlet distribution that is widely used to simulate non-i.i.d. settings.~\cite{li2021federated, hsu2019measuring, wang2020tackling, wang2020federated} Specifically, we simulate a non-i.i.d partition for $T$ tokens on $C$ classes. For each class $c$, we draw a $T$-dimensional vector $\mathbf{q}_c\sim Dir(\delta \mathbf{p})$, where $Dir(\cdot)$ denotes the Dirichlet distribution, $\mathbf{p}$ is a prior class distribution over $T$ classes, and $\delta > 0$ is a concentration parameter. We assign data from each class to each token $t$, following proportion $\mathbf{q}_c[n]$. To simulate the nature of real-world online data where sequences are temporally correlated and data from same classes appear multiple times (e.g., walking, jogging, and then walking, see Figure~\ref{fig:supp_harth_online} and~\ref{fig:supp_extrasensory_online} for illustrations), we concatenate the generated $T$ tokens to create a synthetic non-i.i.d. sequential data. We use $\delta=0.1$ as the default value if not specified.
\subsubsection{Real-distributions with domain shift}
The following illustrates the summary and preprocessing steps of datasets collected in the real-world, or has resemblance to class distributions in the real-world.
\begin{figure}[t]
\centering
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_kitti_online.pdf}
\caption{Visualization of the class distribution in the entire KITTI dataset.}\label{fig:supp_kitti_online}
\vspace{0.3cm}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_kitti_org.pdf}
\caption{Original data with an interval of three frames.}\label{fig:supp_kitti_org}
\vspace{0.3cm}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_kitti_rain.pdf}
\caption{Rain data with an interval of three frames.}\label{fig:supp_kitti_rain}
\end{subfigure}
\caption{Illustration of the test stream of the KITTI dataset. We apply a 200mm/hr intensity of rain to the original data.}\label{fig:supp_kitti}
\end{figure}
\begin{figure}[t]
\centering
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_harth_online0.pdf}
\caption{S008.}\label{fig:supp_harth_online0}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_harth_online1.pdf}
\caption{S018.}\label{fig:supp_harth_online1}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_harth_online2.pdf}
\caption{S019.}\label{fig:supp_harth_online2}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_harth_online3.pdf}
\caption{S021.}\label{fig:supp_harth_online3}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_harth_online4.pdf}
\caption{S022.}\label{fig:supp_harth_online4}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_harth_online5.pdf}
\caption{S028.}\label{fig:supp_harth_online5}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_harth_online6.pdf}
\caption{S029.}\label{fig:supp_harth_online6}
\end{subfigure}
\caption{Illustration of the target streams of the HARTH dataset. We specify x-axis accelerometer values only.}\label{fig:supp_harth_online}
\vspace{-0.3cm}
\end{figure}
\begin{figure}[t]
\centering
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_extrasensory_online0.pdf}
\caption{4FC.}\label{fig:supp_extrasensory_online0}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_extrasensory_online1.pdf}
\caption{598.}\label{fig:supp_extrasensory_online1}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_extrasensory_online2.pdf}
\caption{619.}\label{fig:supp_extrasensory_online2}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_extrasensory_online3.pdf}
\caption{797.}\label{fig:supp_extrasensory_online3}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_extrasensory_online4.pdf}
\caption{A5D.}\label{fig:supp_extrasensory_online4}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_extrasensory_online5.pdf}
\caption{C48.}\label{fig:supp_extrasensory_online5}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_extrasensory_online6.pdf}
\caption{D7D.}\label{fig:supp_extrasensory_online6}
\end{subfigure}
\caption{Illustration of the target streams of the Extrasensory dataset. We specify x-axis accelerometer values only. Due to the length of the name of each domain, denoted here with the first three characters.}\label{fig:supp_extrasensory_online}
\end{figure}
\begin{figure}[h]
\centering
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_kitti_sot_online_tgt_rain-200_legend.pdf}
\caption{Rain-200.}\label{fig:supp_kitti_realtime_acc0}
\end{subfigure}
\caption{Illustration of the real-time cumulative classification error change of different methods on the KITTI dataset. The x-axis denotes the samples in order, whereas the y-axis denotes the error rate in percentage. Note that some lines are not clearly visible due to overlap.}\label{fig:supp_kitti_realtime_acc}
\end{figure}
\begin{figure}[h]
\centering
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_harth_online_tgt_S008_thigh_legend.pdf}
\caption{S008.}\label{fig:supp_harth_online_tgt_S008_thigh_legend}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_harth_online_tgt_S018_thigh_wolegend.pdf}
\caption{S018.}\label{fig:supp_harth_online_tgt_S018_thigh_wolegend}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_harth_online_tgt_S019_thigh_wolegend.pdf}
\caption{S019.}\label{fig:supp_harth_online_tgt_S019_thigh_wolegend}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_harth_online_tgt_S021_thigh_wolegend.pdf}
\caption{S021.}\label{fig:supp_harth_online_tgt_S021_thigh_wolegend}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_harth_online_tgt_S022_thigh_wolegend.pdf}
\caption{S022.}\label{fig:supp_harth_online_tgt_S022_thigh_wolegend}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_harth_online_tgt_S028_thigh_wolegend.pdf}
\caption{S028.}\label{fig:supp_harth_online_tgt_S028_thigh_wolegend}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_harth_online_tgt_S029_thigh_wolegend.pdf}
\caption{S029.}\label{fig:supp_harth_online_tgt_S029_thigh_wolegend}
\end{subfigure}
\caption{Illustration of the real-time cumulative classification error change of different methods on the HARTH dataset. The x-axis denotes the samples in order, whereas the y-axis denotes the error rate in percentage. Note that some lines are not clearly visible due to overlap.}\label{fig:supp_harth_online_realtime_acc}
\end{figure}
\begin{figure}[h]
\centering
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_extrasensory_online_tgt_4FC32141-E888-4BFF-8804-12559A491D8C_legend.pdf}
\caption{4FC.}\label{fig:supp_extrasensory_realtime_acc0}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_extrasensory_online_tgt_59818CD2-24D7-4D32-B133-24C2FE3801E5_wolegend.pdf}
\caption{598.}\label{fig:supp_extrasensory_realtime_acc1}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_extrasensory_online_tgt_61976C24-1C50-4355-9C49-AAE44A7D09F6_wolegend.pdf}
\caption{619.}\label{fig:supp_extrasensory_realtime_acc2}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_extrasensory_online_tgt_797D145F-3858-4A7F-A7C2-A4EB721E133C_wolegend.pdf}
\caption{797.}\label{fig:supp_extrasensory_realtime_acc3}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_extrasensory_online_tgt_A5CDF89D-02A2-4EC1-89F8-F534FDABDD96_wolegend.pdf}
\caption{A5D.}\label{fig:supp_extrasensory_realtime_acc4}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_extrasensory_online_tgt_C48CE857-A0DD-4DDB-BEA5-3A25449B2153_wolegend.pdf}
\caption{C48.}\label{fig:supp_extrasensory_realtime_acc5}
\end{subfigure}
\newline
\begin{subfigure}[t]{1\textwidth}
\centering
\includegraphics[width=1\linewidth]{fig/supp_extrasensory_online_tgt_D7D20E2E-FC78-405D-B346-DBD3FD8FC92B_wolegend.pdf}
\caption{D7D.}\label{fig:supp_extrasensory_realtime_acc6}
\end{subfigure}
\caption{Illustration of the real-time cumulative classification error change of different methods on the Extrasensory dataset. The x-axis denotes the samples in order, whereas the y-axis denotes the error rate in percentage. Note that some lines are not clearly visible due to overlap.}\label{fig:supp_extrasensory_realtime_acc}
\end{figure}
\textbf{KITTI, KITTI-Rain}. KITTI~\cite{kitti} is a well-known dataset used in numerous tasks such as object detection, object tracking, depth estimation, etc. It must be emphasized that the dataset was collected by driving around the city, in rural areas and on highways, which captures the real-world distribution. From the available tasks, we select object tracking task; to utilize its temporal correlation. In order to reduce the task to a single image classification task, we crop each frame with respect to the largest bounding box. Domain gap is introduced, through synthetic generation of corresponding ``rainy'' frames, hereby denoted as KITTI-Rain~\cite{kitti-rain}. KITTI-Rain is generated via a two-step procedure: (1) generation of a depth-map estimation of each frame, and (2) generation of rainy images from vanilla frame and its corresponding depth map, as described in \cite{kitti-rain}. For the depth map generation, we used Monodepth~\cite{monodepth}, and for rainy image generation we use the source code available in \cite{kitti-rain}. The rain intensity is set to 200mm/hr for training and testing. The final source domain consists of 7,481 samples, and each of the target domain consists of 7,800 samples. We use ResNet50~\cite{7780459} pre-trained on ImageNet~\cite{deng2009imagenet} as the backbone network. We fine-tune it on the KITTI training data to generate source models, using the Adam optimizer~\cite{kingma:adam} and cosine annealing learning rate scheduling~\cite{loshchilov2016sgdr} for 50 epochs with an initial learning rate of $0.1$ and a batch size of 64.
\textbf{HARTH}. Human Activity Recognition Trondheim dataset~\cite{harth} was collected from 22 users, with two three-axial Axivity AX3 accelerometers, each attached to the subject’s thigh and lower back. HARTH was also collected in a free-living environment, labelled through recorded video. We set the source domain as the accelerometer data collected from the back (15 users), and set the target domain as one collected from the thigh (from the remaining seven users). We deem such setting to be natural, for one of the most dominant forms of domain shift in wearable sensory data is by the positioning of sensors on the human body~\cite{sensor_positioning}. We use a window size of 50, and min-max scaled (0-1) the data, following the original paper~\cite{harth}. The final source domain consists of 82,544 samples, and each of the seven target domain consists of \{S008: 8,140, S018: 6,241, S019: 5,846, S021: 5,910, S022: 6,448, S028: 3,271, S029: 3,521\} samples. We use four one-dimensional convolutional layers followed by one fully-connected layers as the backbone network. We train it on the source data to generate source models, using stochastic gradient descent with momentum=0.9 for 100 epochs and cosine annealing learning rate scheduling~\cite{loshchilov2016sgdr} with an initial learning rate of 0.1 and a batch size of 64.
\textbf{ExtraSensory}. Extrasensory dataset~\cite{extrasensory} was collected from 60 users with the user’s own smartphones, over a seven day period, in-the-wild, i.e. data was collected from users whom engaged in their regular natural behavior. As there were no constraints on the subject’s activity, the distribution varied from user to user. We select five most frequently occurred, mutually exclusive activities (lying down, sitting, walking, standing, running), and omit other labels. We further process the data to only those consisting the following sensor modalities - accelerometer, gyroscope, magnetometer and audio. We used a window size of five, with no overlap, and standardly scale the datasets. After the pre-processing step, 23 users were left, and 16 of them were used as source domains and seven of them are used as target domains. The final source domain consists of 17,777 samples, and each of the seven target domain consists of \{4FC32141-E888-4BFF-8804-12559A491D8C: 844, 59818CD2-24D7-4D32-B133-24C2FE3801E5: 401, 61976C24-1C50-4355-9C49-AAE44A7D09F6: 776, 797D145F-3858-4A7F-A7C2-A4EB721E133C: 463, A5CDF89D-02A2-4EC1-89F8-F534FDABDD96 : 734, C48CE857-A0DD-4DDB-BEA5-3A25449B2153 : 850, D7D20E2E-FC78-405D-B346-DBD3FD8FC92B: 794\} samples. We use two one-dimensional convolutional layers followed by one fully-connected layers as the backbone network. We train it on the source data to generate source models, using stochastic gradient descent with momentum=0.9 for 100 epochs and cosine annealing learning rate scheduling~\cite{loshchilov2016sgdr} with an initial learning rate of 0.1 and a batch size of 64.
\textbf{Error on the source domain}. We also measure the domain gap between the source and the targets in the three real-distribution datasets: Table~\ref{tab:kitti:source} for KITTI, Table~\ref{tab:harth:source} for HARTH, and Table~\ref{tab:extrasensory:source} for Extrasensory. As shown, there is a clear performance degradation from the source domain to the target domain. For HARTH and ExtraSensory, the performance degradation was severe (30$\sim$40\%p increased error rates compared with Source), indicating the importance overcoming the domain shift problem in sensory applications.
\begin{table}[H]
\centering
\caption{Average classification error (\%) and their corresponding standard deviations on the KITTI dataset of the source model. Bold type indicates those of lowest classification error. Averaged over 3 runs.}\label{tab:kitti:source}
\vspace{0.1cm}
\setlength{\tabcolsep}{4pt}
{%
\begin{tabular}{lccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & Src domain & Rain & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & \textbf{7.4 ± 1.0} & 12.3 ± 2.3 & 9.9 \\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}}
\end{table}
\begin{table}[H]
\caption{Average classification error (\%) and their corresponding standard deviations on the HARTH dataset of the source model. Bold type indicates those of lowest classification error. Averaged over 3 runs.}\label{tab:harth:source}
\vspace{0.1cm}
\setlength{\tabcolsep}{4pt}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lccccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & Src domain & S008 & S018 & S019 & S021 & S022 & S028 & S029 & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & \textbf{11.7 ± 0.7} & 86.2 ± 1.3 & 44.7 ± 2.1 & 50.4 ± 9.5 & 74.8 ± 3.8 & 72.0 ± 2.6 & 53.0 ± 24.0 & 57.0 ± 16.7 & 56.2 \\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}}
\end{table}
\begin{table}[H]
\caption{Average classification error (\%) and their corresponding standard deviations on the ExtraSensory dataset of the source model. Bold type indicates those of lowest classification error. Averaged over 3 runs.}\label{tab:extrasensory:source}
\vspace{0.1cm}
\setlength{\tabcolsep}{4pt}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lccccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & Src domain & 4FC & 598 & 619 & 797 & A5C & C48 & D7D & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & \textbf{8.3 ± 0.7} & 34.6 ± 2.5 & 40.1 ± 0.7 & 63.8 ± 5.7 & 45.3 ± 2.4 & 64.6 ± 3.7 & 39.6 ± 6.8 & 63.0 ± 3.9 & 44.9 \\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}}
\end{table}
\section{Domain-wise results}
\subsection{Robustness to corruptions}
\begin{table}[H]
\centering
\caption{Average classification error (\%) and their corresponding standard deviations on CIFAR10-C with \textbf{temporally correlated test streams}, shown per corruption. \textbf{Bold} type indicates those of lowest classification error, while those with \ch{Red} font are performance degradation after adaptation. Averaged over 3 runs.}
\vspace{0.1cm}
\setlength{\tabcolsep}{3pt}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccccccccccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & \rotatebox{70}{Gaussian} & \rotatebox{70}{Shot} & \rotatebox{70}{Impulse} & \rotatebox{70}{Defocus} & \rotatebox{70}{Glass} & \rotatebox{70}{Motion} & \rotatebox{70}{Zoom} & \rotatebox{70}{Snow} & \rotatebox{70}{Frost} & \rotatebox{70}{Fog} & \rotatebox{70}{Brightness} & \rotatebox{70}{Contrast} & \rotatebox{70}{Elastic} & \rotatebox{70}{Pixelate} & \rotatebox{70}{JPEG} & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & \begin{tabular}[c]{@{}c@{}}74.0 \\ \small{± 3.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.8 \\ \small{± 3.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}75.3 \\ \small{± 4.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}43.3 \\ \small{± 2.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}48.0 \\ \small{± 2.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}32.6 \\ \small{± 1.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}35.2 \\ \small{± 2.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}22.0 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}33.0 \\ \small{± 2.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}25.9 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}8.5 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.1 \\ \small{± 1.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}23.4 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}53.6 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}26.8 \\ \small{± 0.7}\end{tabular} & 42.3 \\
BN Stats \cite{bnstats}& \begin{tabular}[c]{@{}c@{}}\ch{77.2} \\ \small{\ch{± 0.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{76.7} \\ \small{\ch{± 1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{78.9} \\ \small{\ch{± 0.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{70.0} \\ \small{\ch{± 1.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{78.6} \\ \small{\ch{± 0.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{70.5} \\ \small{\ch{± 1.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{71.1} \\ \small{\ch{± 1.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{72.5} \\ \small{\ch{± 1.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{71.9} \\ \small{\ch{± 1.1}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{70.6} \\ \small{\ch{± 1.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{68.7} \\ \small{\ch{± 1.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{69.1} \\ \small{\ch{± 1.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{75.1} \\ \small{\ch{± 1.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{73.6} \\ \small{\ch{± 1.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{76.8} \\ \small{\ch{± 1.4}}\end{tabular} & \ch{73.4} \\
ONDA \cite{onda}& \begin{tabular}[c]{@{}c@{}}69.3 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{68.5} \\ \small{\ch{± 1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}71.8 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{58.5} \\ \small{\ch{± 1.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{71.0} \\ \small{\ch{± 0.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{59.9} \\ \small{\ch{± 1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{59.5} \\ \small{\ch{± 1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{62.4} \\ \small{\ch{± 1.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{62.1} \\ \small{\ch{± 1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{59.6} \\ \small{\ch{± 1.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{55.6} \\ \small{\ch{± 1.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}58.4 \\ \small{± 1.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{65.6} \\ \small{\ch{± 1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{63.9} \\ \small{\ch{± 1.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{67.6} \\ \small{\ch{± 1.1}}\end{tabular} & \ch{63.6} \\
PL \cite{pl}& \begin{tabular}[c]{@{}c@{}}\ch{78.3} \\ \small{\ch{± 1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{78.0} \\ \small{\ch{± 1.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{80.4} \\ \small{\ch{± 1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{72.2} \\ \small{\ch{± 1.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{80.1} \\ \small{\ch{± 1.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{72.4} \\ \small{\ch{± 2.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{73.1} \\ \small{\ch{± 1.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{74.5} \\ \small{\ch{± 2.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{73.9} \\ \small{\ch{± 1.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{73.4} \\ \small{\ch{± 1.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{71.5} \\ \small{\ch{± 2.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{71.7} \\ \small{\ch{± 2.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{77.3} \\ \small{\ch{± 2.1}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{75.7} \\ \small{\ch{± 1.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{78.6} \\ \small{\ch{± 2.7}}\end{tabular} & \ch{75.4} \\
TENT \cite{tent}& \begin{tabular}[c]{@{}c@{}}\ch{79.0} \\ \small{\ch{± 2.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{78.8} \\ \small{\ch{± 2.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{80.6} \\ \small{\ch{± 2.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{73.3} \\ \small{\ch{± 1.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{80.5} \\ \small{\ch{± 2.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{74.4} \\ \small{\ch{± 2.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{74.5} \\ \small{\ch{± 3.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{74.8} \\ \small{\ch{± 2.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{75.0} \\ \small{\ch{± 2.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{74.0} \\ \small{\ch{± 2.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{72.3} \\ \small{\ch{± 3.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{74.9} \\ \small{\ch{± 3.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{78.2} \\ \small{\ch{± 2.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{76.5} \\ \small{\ch{± 2.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{79.0} \\ \small{\ch{± 2.9}}\end{tabular} & \ch{76.4} \\
LAME \cite{lame}& \begin{tabular}[c]{@{}c@{}}73.6 \\ \small{± 5.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}64.8 \\ \small{± 4.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}74.8 \\ \small{± 6.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}36.2 \\ \small{± 4.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}37.7 \\ \small{± 5.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}24.9 \\ \small{± 1.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}27.9 \\ \small{± 3.4}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}12.4 \\ \small{± 1.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}22.4 \\ \small{± 3.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}19.4 \\ \small{± 0.9}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}3.6 \\ \small{± 0.3}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}65.1 \\ \small{± 1.5}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}12.6 \\ \small{± 0.8}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}50.3 \\ \small{± 0.9}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}16.4 \\ \small{± 1.2}\end{tabular}} & 36.2 \\
CoTTA \cite{cotta}& \begin{tabular}[c]{@{}c@{}}\ch{77.0} \\ \small{\ch{± 0.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{76.8} \\ \small{\ch{± 0.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{79.0} \\ \small{\ch{± 0.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{74.1} \\ \small{\ch{± 0.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{79.6} \\ \small{\ch{± 0.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{74.3} \\ \small{\ch{± 0.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{74.0} \\ \small{\ch{± 0.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{74.8} \\ \small{\ch{± 1.1}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{73.3} \\ \small{\ch{± 0.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{72.9} \\ \small{\ch{± 0.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{72.2} \\ \small{\ch{± 0.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{76.5} \\ \small{\ch{± 0.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{76.5} \\ \small{\ch{± 0.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{75.1} \\ \small{\ch{± 0.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{76.6} \\ \small{\ch{± 0.6}}\end{tabular} & \ch{75.5} \\
\addlinespace[0.00cm]
\hline
\addlinespace[0.05cm]
NOTE{} & \textbf{\begin{tabular}[c]{@{}c@{}}34.9 \\ \small{± 1.6}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}32.3 \\ \small{± 3.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}39.6 \\ \small{± 2.5}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}13.6 \\ \small{± 0.5}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}35.8 \\ \small{± 1.9}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}11.8 \\ \small{± 0.8}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}14.5 \\ \small{± 0.5}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}14.1 \\ \small{± 0.6}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}15.2 \\ \small{± 1.3}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}14.2 \\ \small{± 0.6}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}7.7 \\ \small{± 0.3}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}7.6 \\ \small{± 0.6}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}20.8 \\ \small{± 0.7}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}27.7 \\ \small{± 2.6}\end{tabular}} & {\begin{tabular}[c]{@{}c@{}}26.4 \\ \small{± 0.5}\end{tabular}} & \textbf{21.1} \\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}%
}
\end{table}
\begin{table}[H]
\caption{Average classification error (\%) and their corresponding standard deviations on CIFAR100-C with \textbf{temporally correlated test streams}, shown per corruption. \textbf{Bold} type indicates those of lowest classification error, while those with \ch{Red} font are performance degradation after adaptation. Averaged over 3 runs. }
\vspace{0.1cm}
\centering
\setlength{\tabcolsep}{3pt}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccccccccccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & \rotatebox{70}{Gaussian} & \rotatebox{70}{Shot} & \rotatebox{70}{Impulse} & \rotatebox{70}{Defocus} & \rotatebox{70}{Glass} & \rotatebox{70}{Motion} & \rotatebox{70}{Zoom} & \rotatebox{70}{Snow} & \rotatebox{70}{Frost} & \rotatebox{70}{Fog} & \rotatebox{70}{Brightness} & \rotatebox{70}{Contrast} & \rotatebox{70}{Elastic} & \rotatebox{70}{Pixelate} & \rotatebox{70}{JPEG} & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & \begin{tabular}[c]{@{}c@{}}88.1 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}86.8 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}93.7 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}64.9 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}79.7 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}55.5 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}57.7 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}53.8 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.3 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}59.3 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}33.0 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}81.4 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}49.2 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}73.6 \\ \small{± 1.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}55.5 \\ \small{± 0.3}\end{tabular} & 66.6 \\
BN Stats \cite{bnstats}& \begin{tabular}[c]{@{}c@{}}73.9 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}73.5 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}77.2 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}56.9 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}72.3 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{58.8} \\ \small{\ch{± 0.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{57.9} \\ \small{\ch{± 0.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{65.3} \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}65.0 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{62.4} \\ \small{\ch{± 0.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{55.6} \\ \small{\ch{± 0.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}57.6 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{64.6} \\ \small{\ch{± 0.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}63.6 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{71.0} \\ \small{\ch{± 0.4}}\end{tabular} & 65.0 \\
ONDA \cite{onda}& \textbf{\begin{tabular}[c]{@{}c@{}}63.0 \\ \small{± 0.7}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}62.5 \\ \small{± 0.4}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}68.0 \\ \small{± 0.5}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}37.3 \\ \small{± 0.2}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}60.0 \\ \small{± 0.2}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}40.0 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}38.3 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}49.6 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}50.0 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}45.2 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{35.7} \\ \small{\ch{± 0.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}40.9 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}48.6 \\ \small{± 0.5}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}46.9 \\ \small{± 0.3}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}\ch{57.5} \\ \small{\ch{± 0.2}}\end{tabular} & 49.6 \\
PL \cite{pl}& \begin{tabular}[c]{@{}c@{}}71.9 \\ \small{± 1.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}72.0 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}76.3 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}59.3 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}73.8 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{61.5} \\ \small{\ch{± 0.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{59.9} \\ \small{\ch{± 0.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{67.1} \\ \small{\ch{± 0.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{66.7} \\ \small{\ch{± 1.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{63.0} \\ \small{± \ch{1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{57.9} \\ \small{\ch{± 0.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}62.2 \\ \small{± 1.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{67.6} \\ \small{\ch{± 1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}65.2 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{71.1} \\ \small{\ch{± 0.5}}\end{tabular} & 66.4 \\
TENT \cite{tent}& \begin{tabular}[c]{@{}c@{}}71.8 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}71.0 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}76.4 \\ \small{± 1.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}60.2 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}75.0 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{61.9} \\ \small{\ch{± 0.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{60.2} \\ \small{\ch{± 0.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{67.8} \\ \small{\ch{± 0.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{67.8} \\ \small{\ch{± 0.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{63.3} \\ \small{\ch{± 1.1}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{58.4} \\ \small{\ch{± 0.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}65.0 \\ \small{± 1.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{68.4} \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}65.0 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{71.8} \\ \small{\ch{± 0.1}}\end{tabular} & \ch{66.9} \\
LAME \cite{lame}& \begin{tabular}[c]{@{}c@{}}\ch{89.0} \\ \small{\ch{± 1.1}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{87.1} \\ \small{\ch{± 0.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{94.5} \\ \small{\ch{± 0.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}62.3 \\ \small{± 1.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}79.7 \\ \small{± 1.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}49.4 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}52.8 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}46.6 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}63.9 \\ \small{± 1.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}55.6 \\ \small{± 1.2}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}25.2 \\ \small{± 0.6}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}\ch{82.4} \\ \small{\ch{± 0.2}}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}40.8 \\ \small{± 0.5}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}71.9 \\ \small{± 1.4}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}47.8 \\ \small{± 0.7}\end{tabular}} & 63.3 \\
CoTTA \cite{cotta}& \begin{tabular}[c]{@{}c@{}}68.6 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}67.9 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}71.4 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}60.7 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}69.9 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{60.8} \\ \small{\ch{± 0.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{60.2} \\ \small{\ch{± 0.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{64.0} \\ \small{\ch{± 0.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}62.9 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{63.2} \\ \small{\ch{± 0.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{56.7} \\ \small{\ch{± 0.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}65.6 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{64.5} \\ \small{\ch{± 0.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}60.9 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{65.3} \\ \small{\ch{± 0.1}}\end{tabular} & 64.2 \\
\addlinespace[0.00cm]
\hline
\addlinespace[0.05cm]
NOTE{} & \begin{tabular}[c]{@{}c@{}}66.2 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}64.2 \\ \small{± 1.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}72.6 \\ \small{± 0.4}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}37.2 \\ \small{± 0.8}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}61.1 \\ \small{± 0.7}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}35.4 \\ \small{± 0.3}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}37.4 \\ \small{± 0.4}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}40.0 \\ \small{± 0.4}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}42.5 \\ \small{± 0.3}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}43.4 \\ \small{± 0.5}\end{tabular}} & {\begin{tabular}[c]{@{}c@{}}29.4 \\ \small{± 0.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}32.1 \\ \small{± 0.5}\end{tabular}} & {\begin{tabular}[c]{@{}c@{}}44.3 \\ \small{± 0.4}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}47.5 \\ \small{± 0.6}\end{tabular} & {\begin{tabular}[c]{@{}c@{}}51.3 \\ \small{± 0.3}\end{tabular}} & \textbf{47.0} \\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}%
}
\end{table}
\begin{table}[H]
\centering
\caption{Average classification error (\%) and their corresponding standard deviations on ImageNet-C with \textbf{temporally correlated test streams}, shown per corruption. \textbf{Bold} type indicates those of lowest classification error, while those with \ch{Red} font are performance degradation after adaptation. Averaged over 3 runs.}
\vspace{0.1cm}
\setlength{\tabcolsep}{3pt}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccccccccccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & \rotatebox{70}{Gaussian} & \rotatebox{70}{Shot} & \rotatebox{70}{Impulse} & \rotatebox{70}{Defocus} & \rotatebox{70}{Glass} & \rotatebox{70}{Motion} & \rotatebox{70}{Zoom} & \rotatebox{70}{Snow} & \rotatebox{70}{Frost} & \rotatebox{70}{Fog} & \rotatebox{70}{Brightness} & \rotatebox{70}{Contrast} & \rotatebox{70}{Elastic} & \rotatebox{70}{Pixelate} & \rotatebox{70}{JPEG} & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & \begin{tabular}[c]{@{}c@{}}98.4 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}97.7 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}98.4 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}90.6 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}92.5 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}89.8 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}81.8 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}89.5 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}85.0 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}86.4 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}51.1 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}97.2 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}85.3 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}76.9 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}71.7 \\ \small{± 0.0}\end{tabular} & 86.1 \\
BN Stats & \begin{tabular}[c]{@{}c@{}}98.3 \\ \small{± 0.0}\end{tabular} & \ch{\begin{tabular}[c]{@{}c@{}}98.1 \\ \small{± 0.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}98.4 \\ \small{± 0.0}\end{tabular} & \ch{\begin{tabular}[c]{@{}c@{}}98.7 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}98.8 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}97.8 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}96.6 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}96.2 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}96.0 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}95.1 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}93.1 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}98.6 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}96.3 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}95.6 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}96.1 \\ \small{± 0.0}\end{tabular}} & \ch{96.9} \\
ONDA & \begin{tabular}[c]{@{}c@{}}95.1 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}94.7 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}95.0 \\ \small{± 0.0}\end{tabular} & \ch{\begin{tabular}[c]{@{}c@{}}96.2 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}96.1 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}92.5 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}87.2 \\ \small{± 0.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}87.4 \\ \small{± 0.0}\end{tabular} & \ch{\begin{tabular}[c]{@{}c@{}}87.8 \\ \small{± 0.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}82.7 \\ \small{± 0.0}\end{tabular} & \ch{\begin{tabular}[c]{@{}c@{}}71.0 \\ \small{± 0.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}96.4 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}84.9 \\ \small{± 0.0}\end{tabular} & \ch{\begin{tabular}[c]{@{}c@{}}81.7 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}86.1 \\ \small{± 0.0}\end{tabular}} & \ch{89.0} \\
PL & \ch{\begin{tabular}[c]{@{}c@{}}99.3 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}99.3 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}99.4 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}99.5 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}99.4 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}99.5 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}98.8 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}99.1 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}99.2 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}98.1 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}97.3 \\ \small{± 0.1}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}99.8 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}98.4 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}98.5 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}98.5 \\ \small{± 0.0}\end{tabular}} & \ch{98.9} \\
TENT & \begin{tabular}[c]{@{}c@{}}98.3 \\ \small{± 0.0}\end{tabular} & \ch{\begin{tabular}[c]{@{}c@{}}98.1 \\ \small{± 0.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}98.4 \\ \small{± 0.0}\end{tabular} & \ch{\begin{tabular}[c]{@{}c@{}}98.7 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}98.8 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}97.8 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}96.6 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}96.2 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}96.0 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}95.1 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}93.1 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}98.6 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}96.3 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}95.6 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}96.1 \\ \small{± 0.0}\end{tabular}} & \ch{96.9} \\
LAME & \begin{tabular}[c]{@{}c@{}}98.1 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}97.1 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}98.0 \\ \small{± 0.0}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}87.9 \\ \small{± 0.0}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}90.9 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}87.1 \\ \small{± 0.0}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}78.3 \\ \small{± 0.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}87.1 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}80.2 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}81.5 \\ \small{± 0.0}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}39.8 \\ \small{± 0.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}96.4 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}82.5 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}70.7 \\ \small{± 0.0}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}64.9 \\ \small{± 0.0}\end{tabular}} & 82.7 \\
CoTTA & \begin{tabular}[c]{@{}c@{}}98.2 \\ \small{± 0.0}\end{tabular} & \ch{\begin{tabular}[c]{@{}c@{}}98.1 \\ \small{± 0.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}98.3 \\ \small{± 0.0}\end{tabular} & \ch{\begin{tabular}[c]{@{}c@{}}98.8 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}98.8 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}97.7 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}96.8 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}96.6 \\ \small{± 0.1}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}96.3 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}95.3 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}93.5 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}98.8 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}96.5 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}95.6 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}96.2 \\ \small{± 0.0}\end{tabular}} & \ch{97.0} \\
\addlinespace[0.00cm]
\hline
\addlinespace[0.05cm]
NOTE{} & \textbf{\begin{tabular}[c]{@{}c@{}}94.7 \\ \small{± 0.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}93.7 \\ \small{± 0.3}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}94.5 \\ \small{± 0.1}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}91.2 \\ \small{± 0.1}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}91.0 \\ \small{± 0.2}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}83.3 \\ \small{± 0.1}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}79.0 \\ \small{± 0.2}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}79.0 \\ \small{± 0.4}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}78.7 \\ \small{± 0.3}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}66.3 \\ \small{± 0.6}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}48.0 \\ \small{± 0.4}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}94.1 \\ \small{± 0.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}76.9 \\ \small{± 0.6}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}62.6 \\ \small{± 0.7}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}76.6 \\ \small{± 0.6}\end{tabular}} & \textbf{80.6} \\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}%
}
\end{table}
\begin{table}[H]
\caption{Average classification error (\%) and their corresponding standard deviations on MNIST-C with \textbf{temporally correlated test streams}, shown per corruption. \textbf{Bold} type indicates those of lowest classification error, while those with \ch{Red} font are performance degradation after adaptation. Averaged over 3 runs.}
\vspace{0.1cm}
\centering
\setlength{\tabcolsep}{2pt}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccccccccccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & \rotatebox{70}{Shot} & \rotatebox{70}{Impulse} & \rotatebox{70}{Glass} & \rotatebox{70}{Motion} & \rotatebox{70}{Shear} & \rotatebox{70}{Scale} & \rotatebox{70}{Rotate} & \rotatebox{70}{Brightness} & \rotatebox{70}{Translate} & \rotatebox{70}{Stripe} & \rotatebox{70}{Fog} & \rotatebox{70}{Spatter} & \rotatebox{70}{Dotted line} & \rotatebox{70}{Zigzag} & \rotatebox{70}{Canny edges} & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & \begin{tabular}[c]{@{}c@{}}3.7 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}27.3 \\ \small{± 5.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}20.4 \\ \small{± 6.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}4.6 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.2 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}5.1 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}6.5 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}21.1 \\ \small{± 22.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}13.8 \\ \small{± 1.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}17.4 \\ \small{± 17.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.6 \\ \small{± 14.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.8 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.7 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}18.2 \\ \small{± 3.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}26.4 \\ \small{± 11.4}\end{tabular} & 16.1 \\
BN Stats \cite{bnstats} & \begin{tabular}[c]{@{}c@{}}\ch{72.0} \\ \small{\ch{± 0.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{75.2} \\ \small{\ch{± 0.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{73.7} \\ \small{\ch{± 1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{72.1} \\ \small{\ch{± 0.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{71.2} \\ \small{\ch{± 1.1}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{71.4} \\ \small{\ch{± 0.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{71.2} \\ \small{\ch{± 0.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{71.6} \\ \small{\ch{± 0.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{78.5} \\ \small{\ch{± 0.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{72.3} \\ \small{\ch{± 1.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{70.8} \\ \small{\ch{± 1.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{71.6} \\ \small{\ch{± 0.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{73.8} \\ \small{\ch{± 0.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{74.6} \\ \small{\ch{± 0.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{72.3} \\ \small{\ch{± 0.3}}\end{tabular} & \ch{72.8} \\
ONDA \cite{onda}& \begin{tabular}[c]{@{}c@{}}\ch{53.3} \\ \small{\ch{± 3.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{59.9} \\ \small{\ch{± 3.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{59.2} \\ \small{\ch{± 3.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{54.1} \\ \small{\ch{± 3.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{51.6} \\ \small{\ch{± 2.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{53.9} \\ \small{\ch{± 2.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{54.6} \\ \small{\ch{± 2.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{50.5} \\ \small{\ch{± 2.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{65.2} \\ \small{\ch{± 2.1}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{57.5} \\ \small{\ch{± 0.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{54.8} \\ \small{\ch{± 2.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{54.2} \\ \small{\ch{± 3.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{55.4} \\ \small{\ch{± 2.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{61.0} \\ \small{\ch{± 2.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{56.7} \\ \small{\ch{± 2.1}}\end{tabular} & \ch{56.1} \\
PL \cite{pl}& \begin{tabular}[c]{@{}c@{}}\ch{73.7} \\ \small{\ch{± 1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{76.4} \\ \small{\ch{± 0.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{75.3} \\ \small\ch{{± 0.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{74.7} \\ \small{\ch{± 1.1}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{72.7} \\ \small{\ch{± 0.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{73.3} \\ \small{\ch{± 1.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{73.7} \\ \small{\ch{± 0.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{73.7} \\ \small{\ch{± 1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{78.7} \\ \small{\ch{± 0.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{74.1} \\ \small{\ch{± 1.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{75.8} \\ \small{\ch{± 2.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{72.5} \\ \small{\ch{± 0.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{75.8} \\ \small{\ch{± 0.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{76.9} \\ \small{\ch{± 1.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{74.5} \\ \small{\ch{± 0.1}}\end{tabular} & \ch{74.8} \\
TENT \cite{tent}& \begin{tabular}[c]{@{}c@{}}\ch{74.7} \\ \small{\ch{± 1.1}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{78.1} \\ \small{\ch{± 0.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{76.6} \\ \small{\ch{± 0.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{76.1} \\ \small{\ch{± 0.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{75.8} \\ \small{\ch{± 1.1}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{73.7} \\ \small{\ch{± 1.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{75.2} \\ \small{\ch{± 1.1}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{75.4} \\ \small{\ch{± 0.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{78.9} \\ \small{\ch{± 0.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{76.7} \\ \small{\ch{± 1.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{81.4} \\ \small{\ch{± 1.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{73.9} \\ \small{\ch{± 0.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{77.3} \\ \small{\ch{± 0.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{79.2} \\ \small{\ch{± 2.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{75.8} \\ \small{\ch{± 1.0}}\end{tabular} & \ch{76.6} \\
LAME \cite{lame}& \textbf{\begin{tabular}[c]{@{}c@{}}1.1 \\ \small{± 0.3}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}17.0 \\ \small{± 8.7}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}12.5 \\ \small{± 6.5}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}1.1 \\ \small{± 0.3}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}0.4 \\ \small{± 0.2}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}1.5 \\ \small{± 0.6}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}2.3 \\ \small{± 0.6}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}17.2 \\ \small{± 26.0}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}6.0 \\ \small{± 2.3}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}12.3 \\ \small{± 17.2}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}\ch{68.3} \\ \small{\ch{± 15.8}}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}0.7 \\ \small{± 0.3}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}0.7 \\ \small{± 0.4}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}13.2 \\ \small{± 3.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}22.1 \\ \small{± 12.3}\end{tabular} & 11.8 \\
CoTTA \cite{cotta}& \begin{tabular}[c]{@{}c@{}}\ch{76.9} \\ \small{\ch{± 0.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{79.4} \\ \small{\ch{± 0.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{79.1} \\ \small{\ch{± 0.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{77.6} \\ \small{\ch{± 0.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{75.4} \\ \small{\ch{± 0.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{76.2} \\ \small{\ch{± 1.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{77.6} \\ \small{\ch{± 0.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{76.0} \\ \small{\ch{± 0.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{81.6} \\ \small{\ch{± 0.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{76.8} \\ \small{\ch{± 0.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{78.0} \\ \small{\ch{± 0.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{77.6} \\ \small{\ch{± 0.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{79.3} \\ \small{\ch{± 0.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{80.6} \\ \small{\ch{± 1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{77.6} \\ \small{\ch{± 0.5}}\end{tabular} & \ch{78.0} \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
NOTE{} & \begin{tabular}[c]{@{}c@{}}\ch{3.9} \\ \small{\ch{± 1.3}}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}13.8 \\ \small{± 2.4}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}14.3 \\ \small{± 1.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.3 \\ \small{± 2.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}1.7 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.8 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}6.5 \\ \small{± 0.3}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}0.9 \\ \small{± 0.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}8.0 \\ \small{± 1.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}14.4 \\ \small{± 8.1}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}1.6 \\ \small{± 0.3}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}3.9 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}4.5 \\ \small{± 1.2}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}12.6 \\ \small{± 2.5}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}13.4 \\ \small{± 3.9}\end{tabular}} & \textbf{7.1}\\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}%
}
\end{table}
\begin{table}[H]
\caption{Average classification error (\%) and their corresponding standard deviations on CIFAR10-C with \textbf{uniformly distributed test streams}, shown per domain. \textbf{Bold} type indicates those of lowest classification error, while those with \ch{Red} font are performance degradation after adaptation. Averaged over 3 runs. NOTE{}* indicates NOTE{} used directly with test batches (without using PBRS)}
\vspace{0.1cm}
\centering
\setlength{\tabcolsep}{3pt}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccccccccccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & \rotatebox{70}{Gaussian} & \rotatebox{70}{Shot} & \rotatebox{70}{Impulse} & \rotatebox{70}{Defocus} & \rotatebox{70}{Glass} & \rotatebox{70}{Motion} & \rotatebox{70}{Zoom} & \rotatebox{70}{Snow} & \rotatebox{70}{Frost} & \rotatebox{70}{Fog} & \rotatebox{70}{Brightness} & \rotatebox{70}{Contrast} & \rotatebox{70}{Elastic} & \rotatebox{70}{Pixelate} & \rotatebox{70}{JPEG} & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & \begin{tabular}[c]{@{}c@{}}74.0 \\ \small{± 3.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.8 \\ \small{± 3.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}75.3 \\ \small{± 4.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}43.3 \\ \small{± 2.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}48.0 \\ \small{± 2.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}32.6 \\ \small{± 1.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}35.2 \\ \small{± 2.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}22.0 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}33.0 \\ \small{± 2.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}25.9 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}8.5 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.1 \\ \small{± 1.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}23.4 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}53.6 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}26.8 \\ \small{± 0.7}\end{tabular} & 42.3 \\
BN Stats \cite{bnstats}& \begin{tabular}[c]{@{}c@{}}33.1 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}31.1 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}39.8 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}12.3 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}34.8 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}13.7 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}12.6 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}18.3 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}19.9 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}14.5 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{9.3} \\ \small{\ch{± 0.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}13.0 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}23.3 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}20.8 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{28.0} \\ \small{\ch{± 0.6}}\end{tabular} & 21.6 \\
ONDA \cite{onda}& \begin{tabular}[c]{@{}c@{}}33.4 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}31.3 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}40.0 \\ \small{± 1.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}12.3 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}34.6 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}13.7 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}12.4 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}18.3 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}19.8 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}14.3 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{9.1} \\ \small{\ch{± 0.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}14.0 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}23.3 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}20.9 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{28.0} \\ \small{\ch{± 0.7}}\end{tabular} & 21.7 \\
PL \cite{pl}& \begin{tabular}[c]{@{}c@{}}29.4 \\ \small{± 1.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}26.3 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}36.8 \\ \small{± 1.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}13.7 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}36.5 \\ \small{± 1.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}14.0 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}13.5 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}19.7 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}21.2 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}15.6 \\ \small{± 1.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{10.0} \\ \small{\ch{± 0.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}14.8 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{24.5} \\ \small{\ch{± 2.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}20.1 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{27.4} \\ \small{\ch{± 1.3}}\end{tabular} & 21.6 \\
TENT \cite{tent}& \begin{tabular}[c]{@{}c@{}}25.3 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}23.1 \\ \small{± 1.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}32.1 \\ \small{± 1.2}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}11.7 \\ \small{± 0.6}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}33.1 \\ \small{± 3.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}13.2 \\ \small{± 1.1}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}11.2 \\ \small{± 0.1}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}15.9 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}18.8 \\ \small{± 0.7}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}12.9 \\ \small{± 0.8}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}\ch{8.6} \\ \small{\ch{± 0.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}14.4 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}21.7 \\ \small{± 0.9}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}16.5 \\ \small{± 0.8}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}23.6 \\ \small{± 0.7}\end{tabular} & 18.8 \\
LAME \cite{lame}& \begin{tabular}[c]{@{}c@{}}78.2 \\ \small{± 3.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{70.6} \\ \small{\ch{± 4.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{80.5} \\ \small{\ch{± 4.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{46.6} \\ \small{\ch{± 1.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}48.0 \\ \small{± 3.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{34.2} \\ \small{\ch{± 0.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{37.4} \\ \small{\ch{± 1.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}20.8 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}30.5 \\ \small{± 4.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{26.9} \\ \small{\ch{± 1.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{9.8} \\ \small{\ch{± 0.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{71.9} \\ \small{\ch{± 1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{24.2} \\ \small{\ch{± 0.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{56.4} \\ \small{\ch{± 0.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}25.8 \\ \small{± 0.9}\end{tabular} & \ch{44.1} \\
CoTTA \cite{cotta}& \textbf{\begin{tabular}[c]{@{}c@{}}23.1 \\ \small{± 0.7}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}21.5 \\ \small{± 0.6}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}28.0 \\ \small{± 0.3}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}11.7 \\ \small{± 0.5}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}29.2 \\ \small{± 0.6}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}13.3 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}12.0 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}16.6 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}16.6 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}13.8 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{8.8} \\ \small{\ch{± 0.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}14.9 \\ \small{± 0.5}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}20.6 \\ \small{± 0.7}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}17.3 \\ \small{± 0.5}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}19.9 \\ \small{± 0.4}\end{tabular}} & 17.8 \\
\addlinespace[0.00cm]
\hline
\addlinespace[0.05cm]
NOTE{} & \begin{tabular}[c]{@{}c@{}}33.5 \\ \small{± 1.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}30.0 \\ \small{± 1.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}38.2 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}12.6 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}34.4 \\ \small{± 0.8}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}11.5 \\ \small{± 0.5}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}12.9 \\ \small{± 0.6}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}14.1 \\ \small{± 0.2}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}15.2 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}14.0 \\ \small{± 0.6}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}7.4 \\ \small{± 0.2}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}7.8 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}20.7 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}24.7 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}24.2 \\ \small{± 0.4}\end{tabular} & 20.1 \\
NOTE{}* & \begin{tabular}[c]{@{}c@{}}23.8 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}23.0 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}31.1 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}11.8 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}30.9 \\ \small{± 1.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}11.8 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}11.9 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}15.3 \\ \small{± 1.3}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}14.0 \\ \small{± 0.7}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}13.3 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{8.6} \\ \small{\ch{± 0.2}}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}7.5 \\ \small{± 0.3}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}21.2 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}16.9 \\ \small{± 0.6}\end{tabular} & {\begin{tabular}[c]{@{}c@{}}23.0 \\ \small{± 1.2}\end{tabular}} & \textbf{17.6}
\\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}%
}
\end{table}
\begin{table}[H]
\caption{Average classification error (\%) and their corresponding standard deviations on CIFAR100-C with \textbf{uniformly distributed test streams}, shown per domain. \textbf{Bold} type indicates those of lowest classification error, while those with \ch{Red} font are performance degradation after adaptation. Averaged over 3 runs. NOTE{}* indicates NOTE{} used directly with test batches (without using PBRS)}
\vspace{0.1cm}
\centering
\setlength{\tabcolsep}{3pt}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccccccccccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & \rotatebox{70}{Gaussian} & \rotatebox{70}{Shot} & \rotatebox{70}{Impulse} & \rotatebox{70}{Defocus} & \rotatebox{70}{Glass} & \rotatebox{70}{Motion} & \rotatebox{70}{Zoom} & \rotatebox{70}{Snow} & \rotatebox{70}{Frost} & \rotatebox{70}{Fog} & \rotatebox{70}{Brightness} & \rotatebox{70}{Contrast} & \rotatebox{70}{Elastic} & \rotatebox{70}{Pixelate} & \rotatebox{70}{JPEG} & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & \begin{tabular}[c]{@{}c@{}}88.1 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}86.8 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}93.7 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}64.9 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}79.7 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}55.5 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}57.7 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}53.8 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.3 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}59.3 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}33.0 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}81.4 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}49.2 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}73.6 \\ \small{± 1.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}55.5 \\ \small{± 0.3}\end{tabular} & 66.6 \\
BN Stats \cite{bnstats}& \begin{tabular}[c]{@{}c@{}}60.9 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}59.9 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}65.7 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}33.7 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}57.6 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}36.5 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}35.2 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}46.7 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}46.9 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}42.8 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{32.3} \\ \small{\ch{± 0.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}35.6 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}45.8 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}43.6 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}55.5 \\ \small{± 0.2}\end{tabular} & 46.6 \\
ONDA \cite{onda}& \begin{tabular}[c]{@{}c@{}}60.8 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}60.2 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.0 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}33.9 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}57.5 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}36.3 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}34.6 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}46.5 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}47.2 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}42.1 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{32.1} \\ \small{\ch{± 0.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}36.4 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}45.5 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}43.4 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}55.1 \\ \small{± 0.1}\end{tabular} & 46.5 \\
PL \cite{pl}& \begin{tabular}[c]{@{}c@{}}52.2 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}50.3 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}59.4 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}33.5 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}54.0 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}35.7 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}33.1 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}42.8 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}44.5 \\ \small{± 1.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}39.2 \\ \small{± 1.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}30.9 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}35.5 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}45.5 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}39.9 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}50.4 \\ \small{± 1.3}\end{tabular} & 43.1 \\
TENT \cite{tent}& \textbf{\begin{tabular}[c]{@{}c@{}}48.7 \\ \small{± 0.8}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}47.2 \\ \small{± 0.6}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}55.6 \\ \small{± 0.9}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}31.5 \\ \small{± 0.2}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}50.9 \\ \small{± 0.5}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}33.5 \\ \small{± 0.4}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}31.7 \\ \small{± 0.2}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}39.6 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}41.0 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}36.8 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}29.4 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}33.6 \\ \small{± 0.4}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}42.3 \\ \small{± 0.6}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}36.8 \\ \small{± 0.5}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}46.4 \\ \small{± 0.5}\end{tabular}} & \textbf{40.3} \\
LAME \cite{lame}& \begin{tabular}[c]{@{}c@{}}\ch{91.0} \\ \small{\ch{± 1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}{89.5} \\ \small{{± 1.0}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{95.2} \\ \small{\ch{± 0.7}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{68.1} \\ \small{\ch{± 0.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{82.7} \\ \small{\ch{± 1.1}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{57.1} \\ \small{\ch{± 0.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{60.2} \\ \small{\ch{± 0.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{54.7} \\ \small{\ch{± 0.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{68.9} \\ \small{\ch{± 1.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{61.8} \\ \small{\ch{± 0.6}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{33.7} \\ \small{\ch{± 0.5}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{85.2} \\ \small{\ch{± 0.4}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{50.3} \\ \small{\ch{± 0.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{76.7} \\ \small{\ch{± 1.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{56.2} \\ \small{\ch{± 0.5}}\end{tabular} & \ch{68.8} \\
CoTTA \cite{cotta}& \begin{tabular}[c]{@{}c@{}}52.8 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}51.0 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}56.9 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}35.8 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}53.9 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}37.9 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}36.8 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}45.2 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}44.5 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}44.0 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}32.2 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}41.3 \\ \small{± 1.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}46.1 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}39.7 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}46.9 \\ \small{± 0.7}\end{tabular} & 44.3 \\
\addlinespace[0.00cm]
\hline
\addlinespace[0.05cm]
NOTE{} & \begin{tabular}[c]{@{}c@{}}65.6 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}62.6 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}72.0 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}36.8 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}60.5 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}34.9 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}36.7 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}39.6 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}41.7 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}42.3 \\ \small{± 0.3}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}28.6 \\ \small{± 0.2}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}32.3 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}43.8 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}47.7 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}50.9 \\ \small{± 0.2}\end{tabular} & 46.4 \\
NOTE{}* & \begin{tabular}[c]{@{}c@{}}51.8 \\ \small{± 1.0}\end{tabular} & {\begin{tabular}[c]{@{}c@{}}50.0 \\ \small{± 0.3}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}60.7 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}32.6 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}54.4 \\ \small{± 0.3}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}33.0 \\ \small{± 0.2}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}33.5 \\ \small{± 0.4}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}38.5 \\ \small{± 0.3}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}38.6 \\ \small{± 0.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}36.7 \\ \small{± 0.3}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}29.7 \\ \small{± 0.5}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}27.3 \\ \small{± 0.3}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}43.2 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}37.1 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}47.6 \\ \small{± 0.9}\end{tabular} & 41.0 \\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}%
}
\end{table}
\begin{table}[H]
\centering
\caption{Average classification error (\%) and their corresponding standard deviations on ImageNet-C with \textbf{temporally correlated test streams}, shown per corruption. \textbf{Bold} type indicates those of lowest classification error, while those with \ch{Red} font are performance degradation after adaptation. Averaged over 3 runs.}
\vspace{0.1cm}
\setlength{\tabcolsep}{3pt}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccccccccccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & \rotatebox{70}{Gaussian} & \rotatebox{70}{Shot} & \rotatebox{70}{Impulse} & \rotatebox{70}{Defocus} & \rotatebox{70}{Glass} & \rotatebox{70}{Motion} & \rotatebox{70}{Zoom} & \rotatebox{70}{Snow} & \rotatebox{70}{Frost} & \rotatebox{70}{Fog} & \rotatebox{70}{Brightness} & \rotatebox{70}{Contrast} & \rotatebox{70}{Elastic} & \rotatebox{70}{Pixelate} & \rotatebox{70}{JPEG} & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & \begin{tabular}[c]{@{}c@{}}98.4 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}97.7 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}98.4 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}90.6 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}92.5 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}89.8 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}81.8 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}89.5 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}85.0 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}86.4 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}51.1 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}97.2 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}85.3 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}76.9 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}71.7 \\ \small{± 0.0}\end{tabular} & 86.1 \\
BN Stats & \begin{tabular}[c]{@{}c@{}}89.4 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}88.5 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}89.2 \\ \small{± 0.2}\end{tabular} & \ch{\begin{tabular}[c]{@{}c@{}}90.8 \\ \small{± 0.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}90.0 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}81.3 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}69.8 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}72.6 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}73.8 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}62.6 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}44.3 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}92.1 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}64.5 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}60.3 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}70.7 \\ \small{± 0.0}\end{tabular} & 76.0 \\
ONDA & \begin{tabular}[c]{@{}c@{}}89.2 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}88.2 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}89.0 \\ \small{± 0.1}\end{tabular} & \ch{\begin{tabular}[c]{@{}c@{}}90.9 \\ \small{± 0.1}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}90.0 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}81.6 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}69.5 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}72.6 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}73.7 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}62.7 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}43.9 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}92.1 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}64.3 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}60.1 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}70.0 \\ \small{± 0.0}\end{tabular} & 75.9 \\
PL & \begin{tabular}[c]{@{}c@{}}89.8 \\ \small{± 1.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}86.1 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}88.5 \\ \small{± 1.6}\end{tabular} & \ch{\begin{tabular}[c]{@{}c@{}}93.0 \\ \small{± 1.1}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}92.5 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}82.2 \\ \small{± 0.0}\end{tabular} & {\begin{tabular}[c]{@{}c@{}}64.6 \\ \small{± 0.3}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}70.2 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}79.7 \\ \small{± 0.4}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}55.8 \\ \small{± 0.2}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}43.9 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}97.2 \\ \small{± 0.5}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}57.8 \\ \small{± 0.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}52.7 \\ \small{± 0.2}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}60.5 \\ \small{± 0.1}\end{tabular}} & 74.4 \\
TENT & \begin{tabular}[c]{@{}c@{}}91.1 \\ \small{± 2.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}89.7 \\ \small{± 1.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}91.0 \\ \small{± 2.5}\end{tabular} & \ch{\begin{tabular}[c]{@{}c@{}}93.1 \\ \small{± 3.2}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}92.2 \\ \small{± 3.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}84.7 \\ \small{± 4.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}72.4 \\ \small{± 3.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}73.3 \\ \small{± 1.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}78.7 \\ \small{± 6.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}59.8 \\ \small{± 4.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}44.5 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}95.2 \\ \small{± 4.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}61.6 \\ \small{± 4.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}56.4 \\ \small{± 5.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}67.4 \\ \small{± 4.7}\end{tabular} & 76.5 \\
LAME & \ch{\begin{tabular}[c]{@{}c@{}}98.6 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}97.8 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}98.6 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}90.7 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}92.6 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}89.9 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}81.9 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}89.8 \\ \small{± 0.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}85.0 \\ \small{± 0.0}\end{tabular} & \ch{\begin{tabular}[c]{@{}c@{}}86.5 \\ \small{± 0.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}51.1 \\ \small{± 0.0}\end{tabular} & \ch{\begin{tabular}[c]{@{}c@{}}97.3 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}85.6 \\ \small{± 0.0}\end{tabular}} & \ch{\begin{tabular}[c]{@{}c@{}}77.0 \\ \small{± 0.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}71.7 \\ \small{± 0.0}\end{tabular} & \ch{86.3} \\
CoTTA & \textbf{\begin{tabular}[c]{@{}c@{}}85.7 \\ \small{± 0.2}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}84.6 \\ \small{± 0.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}85.4 \\ \small{± 0.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}87.8 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}86.4 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}74.6 \\ \small{± 0.0}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}64.2 \\ \small{± 0.2}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}67.9 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}69.7 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}56.1 \\ \small{± 0.1}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}42.7 \\ \small{± 0.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}88.5 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}60.0 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}54.2 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}64.9 \\ \small{± 0.1}\end{tabular} & 71.5 \\
\addlinespace[0.00cm]
\hline
\addlinespace[0.05cm]
NOTE{} & \begin{tabular}[c]{@{}c@{}}87.6 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}85.7 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}87.2 \\ \small{± 0.2}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}83.3 \\ \small{± 0.2}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}83.2 \\ \small{± 0.2}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}73.6 \\ \small{± 0.0}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}65.4 \\ \small{± 0.2}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}65.0 \\ \small{± 0.0}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}68.6 \\ \small{± 0.1}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}57.9 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}43.5 \\ \small{± 0.1}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}75.9 \\ \small{± 0.1}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}61.2 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}54.1 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}62.8 \\ \small{± 0.1}\end{tabular} & \textbf{70.3} \\
NOTE{}* & \begin{tabular}[c]{@{}c@{}}89.5 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}87.9 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}88.9 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}84.6 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}83.7 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}74.4 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.6 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.1 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}71.2 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}58.2 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}44.7 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}78.8 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}61.2 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}54.8 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}64.8 \\ \small{± 0.1}\end{tabular} & 71.7 \\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}%
}
\end{table}
\begin{table}[H]
\caption{Average classification error (\%) and their corresponding standard deviations on MNIST-C with \textbf{uniformly distributed test streams}, shown per domain. \textbf{Bold} type indicates those of lowest classification error, while those with \ch{Red} font are performance degradation after adaptation. Averaged over 3 runs. NOTE{}* indicates NOTE{} used directly with test batches (without using PBRS)}
\vspace{0.1cm}
\centering
\setlength{\tabcolsep}{2pt}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccccccccccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & \rotatebox{70}{Shot} & \rotatebox{70}{Impulse} & \rotatebox{70}{Glass} & \rotatebox{70}{Motion} & \rotatebox{70}{Shear} & \rotatebox{70}{Scale} & \rotatebox{70}{Rotate} & \rotatebox{70}{Brightness} & \rotatebox{70}{Translate} & \rotatebox{70}{Stripe} & \rotatebox{70}{Fog} & \rotatebox{70}{Spatter} & \rotatebox{70}{Dotted line} & \rotatebox{70}{Zigzag} & \rotatebox{70}{Canny edges} & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & \begin{tabular}[c]{@{}c@{}}3.7 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}27.3 \\ \small{± 5.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}20.4 \\ \small{± 6.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}4.6 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.2 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}5.1 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}6.5 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}21.1 \\ \small{± 22.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}13.8 \\ \small{± 1.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}17.4 \\ \small{± 17.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.6 \\ \small{± 14.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.8 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.7 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}18.2 \\ \small{± 3.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}26.4 \\ \small{± 11.4}\end{tabular} & 16.1 \\
BN Stats \cite{bnstats} & \begin{tabular}[c]{@{}c@{}}2.9 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}7.0 \\ \small{± 1.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}9.1 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.0 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.0 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.8 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}6.1 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}1.1 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}12.5 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}6.5 \\ \small{± 2.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.2 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.3 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.5 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}11.4 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}6.7 \\ \small{± 0.9}\end{tabular} & 5.3 \\
ONDA \cite{onda}& \begin{tabular}[c]{@{}c@{}}2.6 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}6.5 \\ \small{± 1.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}8.6 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.8 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}1.8 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.5 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}5.7 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}1.0 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}11.7 \\ \small{± 1.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}6.1 \\ \small{± 2.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.6 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.0 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.2 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}11.0 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}6.2 \\ \small{± 0.8}\end{tabular} & 5.0 \\
PL \cite{pl} & \begin{tabular}[c]{@{}c@{}}1.6 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.5 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}4.8 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}1.7 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}1.5 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.3 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}4.9 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.8 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}6.8 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.7 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}1.0 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.2 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}1.7 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}5.3 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.9 \\ \small{± 0.9}\end{tabular} & 3.0 \\
TENT \cite{tent}& \begin{tabular}[c]{@{}c@{}}1.4 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.8 \\ \small{± 0.4}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}3.8 \\ \small{± 0.5}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}1.5 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}1.2 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}1.8 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.6 \\ \small{± 0.2}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}0.7 \\ \small{± 0.1}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}4.6 \\ \small{± 0.7}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}1.9 \\ \small{± 0.2}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}0.8 \\ \small{± 0.0}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}1.7 \\ \small{± 0.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}1.3 \\ \small{± 0.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}4.5 \\ \small{± 0.6}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}3.1 \\ \small{± 0.5}\end{tabular}} & 2.3 \\
LAME \cite{lame}& \begin{tabular}[c]{@{}c@{}}3.0 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{30.7} \\ \small{\ch{± 8.3}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}18.9 \\ \small{± 5.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.4 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}1.9 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}4.2 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}6.3 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{25.9} \\ \small{\ch{± 29.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{13.9} \\ \small{\ch{± 1.9}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{18.5} \\ \small{\ch{± 21.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{78.2} \\ \small{\ch{± 9.8}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.3 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.2 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{19.3} \\ \small{\ch{± 3.2}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{28.0} \\ \small{\ch{± 12.7}}\end{tabular} & \ch{17.2} \\
CoTTA \cite{cotta}& \begin{tabular}[c]{@{}c@{}}2.6 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}6.6 \\ \small{± 1.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}8.7 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.7 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}1.8 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.2 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}5.6 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}1.0 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\ch{14.3} \\ \small{\ch{± 1.1}}\end{tabular} & \begin{tabular}[c]{@{}c@{}}7.7 \\ \small{± 6.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}1.9 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.9 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.2 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}13.6 \\ \small{± 1.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}6.1 \\ \small{± 0.6}\end{tabular} & 5.4 \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
NOTE{} & \begin{tabular}[c]{@{}c@{}}2.5 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}10.7 \\ \small{± 1.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}10.9 \\ \small{± 2.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.0 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}1.5 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.4 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}5.5 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.9 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}5.5 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}12.1 \\ \small{± 5.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}1.2 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}2.8 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.0 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}10.9 \\ \small{± 1.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}9.1 \\ \small{± 0.4}\end{tabular} & 5.4 \\
NOTE{}* & \textbf{\begin{tabular}[c]{@{}c@{}}1.3 \\ \small{± 0.2}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}2.7 \\ \small{± 0.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}3.8 \\ \small{± 0.5}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}1.3 \\ \small{± 0.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}1.1 \\ \small{± 0.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}1.6 \\ \small{± 0.0}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}3.5 \\ \small{± 0.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}0.7 \\ \small{± 0.0}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}2.8 \\ \small{± 0.0}\end{tabular}} & {\begin{tabular}[c]{@{}c@{}}2.2 \\ \small{± 0.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}0.7 \\ \small{± 0.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}1.7 \\ \small{± 0.4}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}1.4 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}4.8 \\ \small{± 1.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}3.5 \\ \small{± 0.1}\end{tabular} & \textbf{2.2} \\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}%
}
\end{table}
\subsection{Real distributions with domain shift}
Since the adaptation is done from a single source domain to a single target domain in KITTI, no further per-domain tables are specified here.
\begin{table}[H]
\caption{Average classification error (\%) and their corresponding standard deviations on HARTH with \textbf{real test streams}, shown per domain. \textbf{Bold} type indicates those of lowest classification error, while those with \ch{Red} font are performance degradation after adaptation. Averaged over 3 runs. }
\vspace{0.1cm}
\setlength{\tabcolsep}{4pt}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & S008 & S018 & S019 & S021 & S022 & S028 & S029 & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & 86.2 ± 1.3 & 44.7 ± 2.1 & 50.4 ± 9.5 & 74.8 ± 3.8 & 72.0 ± 2.6 & 53.0 ± 24.0 & 57.0 ± 16.7 & 62.6 \\
BN Stats \cite{bnstats}& 70.3 ± 1.4 & \ch{73.8 ± 1.3} & \ch{68.1 ± 3.0} & 64.9 ± 0.9 & 68.5 ± 0.3 & \ch{65.5 ± 0.5} & \ch{69.4 ± 1.4} & \ch{68.6} \\
ONDA \cite{onda}& 75.3 ± 4.0 & \ch{60.4 ± 0.9} & \ch{63.1 ± 4.6} & 67.9 ± 0.4 & 70.0 ± 3.8 & \ch{73.6 ± 0.7} & \ch{74.5 ± 4.4} & \ch{69.3} \\
PL \cite{pl}& {60.4 ± 1.3} & \ch{71.4 ± 1.5} & \ch{62.9 ± 1.9} & 61.8 ± 1.2 & 63.1 ± 0.4 & \ch{64.5 ± 0.8} & \ch{69.4 ± 2.0} & \ch{64.8} \\
TENT \cite{tent}& \textbf{59.5 ± 0.3} & \ch{71.0 ± 1.6} & \ch{62.2 ± 1.9} & \textbf{61.1 ± 1.1} & \textbf{61.7 ± 0.4} & \ch{64.1 ± 0.5} & \ch{69.3 ± 2.1} & \ch{64.1} \\
LAME \cite{lame}& 85.5 ± 1.7 & 43.4 ± 2.0 & 48.8 ± 10.9 & 73.2 ± 3.8 & 70.7 ± 2.6 & 51.2 ± 29.4 & 54.1 ± 20.6 & 61.0 \\
CoTTA \cite{cotta}& 70.4 ± 1.4 & \ch{73.8 ± 1.3} & \ch{68.2 ± 2.9} & 64.9 ± 1.0 & 68.5 ± 0.2 & \ch{65.5 ± 0.5} & \ch{69.4 ± 1.4} & \ch{68.7} \\
\addlinespace[0.00cm]
\hline
\addlinespace[0.05cm]
NOTE{} & 84.8 ± 0.7 & \textbf{32.9 ± 1.8} & \textbf{36.3 ± 10.9} & 69.1 ± 2.4 & 67.1 ± 1.2 & \textbf{30.0 ± 13.8} & \textbf{36.6 ± 9.8} & \textbf{51.0} \\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}}
\end{table}
\begin{table}[H]
\caption{Average classification error (\%) and their corresponding standard deviations on Extrasensory with \textbf{real test streams}, shown per domain. \textbf{Bold} type indicates those of lowest classification error, while those with \ch{Red} font are performance degradation after adaptation. Due to the length of the name of each domain, denoted here with the first three characters. Averaged over 3 runs.}
\vspace{0.1cm}
\setlength{\tabcolsep}{4pt}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & 4FC & 598 & 619 & 797 & A5D & C48 & D7D & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & 34.6 ± 2.5 & {40.1 ± 0.7} & 63.8 ± 5.7 & 45.3 ± 2.4 & 64.6 ± 3.7 & 39.6 ± 6.8 & 63.0 ± 3.9 & 50.2 \\
BN Stats\cite{bnstats} & \ch{61.7 ± 4.2} & \ch{50.1 ± 5.1} & {51.6 ± 1.5} & \ch{59.4 ± 1.1} & {54.4 ± 1.0} & \ch{52.4 ± 2.8} & 62.6 ± 2.9 & \ch{56.0} \\
ONDA \cite{onda} & \ch{36.3 ± 3.5} & \ch{44.0 ± 2.2} & \textbf{50.8 ± 2.4} & \ch{56.1 ± 1.9} & 59.7 ± 2.7 & \ch{43.5 ± 5.9} & \textbf{46.7 ± 4.2} & 48.2 \\
PL \cite{pl}& \ch{62.2 ± 4.3} & \ch{50.0 ± 5.1} & 51.7 ± 1.8 & \ch{59.2 ± 1.1} & {53.9 ± 1.1} & \ch{52.3 ± 2.9} & 62.8 ± 3.0 & \ch{56.0} \\
TENT \cite{tent}& \ch{62.1 ± 4.6} & \ch{49.8 ± 5.0} & {51.6 ± 1.9} & \ch{59.4 ± 1.2} & {53.9 ± 1.0} & \ch{52.2 ± 2.9} & 62.8 ± 3.0 & \ch{56.0} \\
LAME \cite{lame}& \textbf{33.1 ± 2.4} & \textbf{37.8 ± 0.4} & \ch{68.0 ± 8.8} & \textbf{37.1 ± 6.7} & \ch{73.2 ± 2.6} & 39.0 ± 7.6 & \ch{66.4 ± 4.0} & \ch{50.7} \\
CoTTA \cite{cotta} & \ch{61.7 ± 4.2} & \ch{50.0 ± 4.9} & {51.6 ± 1.5} & \ch{59.4 ± 1.1} & 54.4 ± 1.0 & \ch{52.4 ± 2.8} & 62.6 ± 2.9 & \ch{56.0} \\
\addlinespace[0.00cm]
\hline
\addlinespace[0.05cm]
NOTE{} & \ch{41.7 ± 5.9} & \ch{40.7 ± 0.8} & 55.5 ± 10.8 & \ch{45.8 ± 4.6} & \textbf{45.8 ± 10.4} & \textbf{32.9 ± 1.1} & 55.5 ± 10.4 & \textbf{45.4} \\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}}
\end{table}
\subsection{Ablation study}
\begin{table}[H]
\caption{Average classification error (\%) and their corresponding standard deviations of varying ablation settings on CIFAR10-C with \textbf{temporally correlated test streams}, shown per domain. \textbf{Bold} type indicates those of lowest classification error. Averaged over 3 runs. }
\vspace{0.1cm}
\centering
\setlength{\tabcolsep}{3pt}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccccccccccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & \rotatebox{70}{Gaussian} & \rotatebox{70}{Shot} & \rotatebox{70}{Impulse} & \rotatebox{70}{Defocus} & \rotatebox{70}{Glass} & \rotatebox{70}{Motion} & \rotatebox{70}{Zoom} & \rotatebox{70}{Snow} & \rotatebox{70}{Frost} & \rotatebox{70}{Fog} & \rotatebox{70}{Brightness} & \rotatebox{70}{Contrast} & \rotatebox{70}{Elastic} & \rotatebox{70}{Pixelate} & \rotatebox{70}{JPEG} & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & \begin{tabular}[c]{@{}c@{}}74.0 \\ \small{± 3.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.8 \\ \small{± 3.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}75.3 \\ \small{± 4.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}43.3 \\ \small{± 2.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}48.0 \\ \small{± 2.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}32.6 \\ \small{± 1.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}35.2 \\ \small{± 2.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}22.0 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}33.0 \\ \small{± 2.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}25.9 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}8.5 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.1 \\ \small{± 1.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}23.4 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}53.6 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}26.8 \\ \small{± 0.7}\end{tabular} & 42.3 \\
IABN & \begin{tabular}[c]{@{}c@{}}44.5 \\ \small{± 2.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}41.3 \\ \small{± 2.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}48.0 \\ \small{± 1.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}16.3 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}39.9 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}13.8 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}16.1 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}14.9 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}17.8 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}16.3 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}7.6 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}8.8 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}22.5 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}34.0 \\ \small{± 1.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}26.7 \\ \small{± 0.6}\end{tabular} & 24.6 \\
PBRS & \begin{tabular}[c]{@{}c@{}}45.2 \\ \small{± 3.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}38.5 \\ \small{± 4.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}46.8 \\ \small{± 3.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}24.5 \\ \small{± 2.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}38.2 \\ \small{± 2.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}19.1 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}20.0 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}16.5 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}19.1 \\ \small{± 2.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}16.5 \\ \small{± 0.4}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}7.1 \\ \small{± 0.7}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}34.4 \\ \small{± 3.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}21.5 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}39.8 \\ \small{± 4.7}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}25.2 \\ \small{± 0.4}\end{tabular}} & 27.5 \\
IABN + RS & \textbf{\begin{tabular}[c]{@{}c@{}}33.7 \\ \small{± 6.4}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}30.0 \\ \small{± 6.7}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}37.6 \\ \small{± 2.9}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}13.6 \\ \small{± 0.3}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}34.9 \\ \small{± 1.9}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}12.4 \\ \small{± 1.2}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}14.5 \\ \small{± 1.7}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}13.9 \\ \small{± 1.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}15.0 \\ \small{± 3.1}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}14.0 \\ \small{± 1.3}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}7.2 \\ \small{± 0.0}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}7.4 \\ \small{± 0.7}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}21.1 \\ \small{± 0.9}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}26.2 \\ \small{± 4.4}\end{tabular}} & {\begin{tabular}[c]{@{}c@{}}25.9 \\ \small{± 1.1}\end{tabular}} & \textbf{20.5} \\
IABN + PBRS & \begin{tabular}[c]{@{}c@{}}34.9 \\ \small{± 1.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}32.3 \\ \small{± 3.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}39.6 \\ \small{± 2.5}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}13.6 \\ \small{± 0.5}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}35.8 \\ \small{± 1.9}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}11.8 \\ \small{± 0.8}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}14.5 \\ \small{± 0.5}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}14.1 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}15.2 \\ \small{± 1.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}14.2 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}7.7 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}7.6 \\ \small{± 0.6}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}20.8 \\ \small{± 0.7}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}27.7 \\ \small{± 2.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}26.4 \\ \small{± 0.5}\end{tabular} & 21.1 \\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}%
}
\end{table}
\begin{table}[H]
\caption{Average classification error (\%) and their corresponding standard deviations of varying ablation settings on CIFAR100-C with \textbf{temporally correlated test streams}, shown per domain. \textbf{Bold} type indicates those of lowest classification error. Averaged over 3 runs. }
\vspace{0.1cm}
\centering
\setlength{\tabcolsep}{3pt}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccccccccccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & \rotatebox{70}{Gaussian} & \rotatebox{70}{Shot} & \rotatebox{70}{Impulse} & \rotatebox{70}{Defocus} & \rotatebox{70}{Glass} & \rotatebox{70}{Motion} & \rotatebox{70}{Zoom} & \rotatebox{70}{Snow} & \rotatebox{70}{Frost} & \rotatebox{70}{Fog} & \rotatebox{70}{Brightness} & \rotatebox{70}{Contrast} & \rotatebox{70}{Elastic} & \rotatebox{70}{Pixelate} & \rotatebox{70}{JPEG} & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & \begin{tabular}[c]{@{}c@{}}88.1 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}86.8 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}93.7 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}64.9 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}79.7 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}55.5 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}57.7 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}53.8 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.3 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}59.3 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}33.0 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}81.4 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}49.2 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}73.6 \\ \small{± 1.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}55.5 \\ \small{± 0.3}\end{tabular} & 66.6 \\
IABN & \begin{tabular}[c]{@{}c@{}}79.3 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}77.2 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}84.2 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}45.0 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}69.6 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}40.9 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}43.1 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}42.5 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}48.6 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}52.5 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}30.4 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}40.5 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}47.6 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}59.8 \\ \small{± 1.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}56.2 \\ \small{± 0.4}\end{tabular} & 54.5 \\
PBRS & \begin{tabular}[c]{@{}c@{}}68.8 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.2 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}73.3 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}46.2 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}64.9 \\ \small{± 1.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}41.8 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}41.7 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}44.2 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}48.5 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}44.7 \\ \small{± 0.2}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}28.3 \\ \small{± 0.2}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}60.1 \\ \small{± 0.4}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}44.2 \\ \small{± 0.4}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}51.9 \\ \small{± 0.8}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}50.5 \\ \small{± 0.5}\end{tabular}} & 51.7 \\
IABN + RS & \begin{tabular}[c]{@{}c@{}}66.8 \\ \small{± 2.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}65.2 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}73.1 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}38.7 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}63.0 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}36.6 \\ \small{± 0.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}38.0 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}41.9 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}43.9 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}44.6 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}29.5 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}33.5 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}46.0 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}49.9 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}52.4 \\ \small{± 0.4}\end{tabular} & 48.2 \\
IABN + PBRS & \textbf{\begin{tabular}[c]{@{}c@{}}66.2 \\ \small{± 0.8}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}64.2 \\ \small{± 1.6}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}72.6 \\ \small{± 0.4}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}37.2 \\ \small{± 0.8}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}61.1 \\ \small{± 0.7}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}35.4 \\ \small{± 0.3}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}37.4 \\ \small{± 0.4}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}40.0 \\ \small{± 0.4}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}42.5 \\ \small{± 0.3}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}43.4 \\ \small{± 0.5}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}29.4 \\ \small{± 0.1}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}32.1 \\ \small{± 0.5}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}44.3 \\ \small{± 0.4}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}47.5 \\ \small{± 0.6}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}51.3 \\ \small{± 0.3}\end{tabular} & \textbf{47.0} \\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}%
}
\end{table}
\begin{table}[H]
\caption{Average classification error (\%) and their corresponding standard deviations of varying ablation settings on CIFAR10-C with \textbf{uniformly distributed test streams}, shown per domain. \textbf{Bold} type indicates those of lowest classification error. Averaged over 3 runs. }
\vspace{0.1cm}
\centering
\setlength{\tabcolsep}{3pt}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccccccccccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & \rotatebox{70}{Gaussian} & \rotatebox{70}{Shot} & \rotatebox{70}{Impulse} & \rotatebox{70}{Defocus} & \rotatebox{70}{Glass} & \rotatebox{70}{Motion} & \rotatebox{70}{Zoom} & \rotatebox{70}{Snow} & \rotatebox{70}{Frost} & \rotatebox{70}{Fog} & \rotatebox{70}{Brightness} & \rotatebox{70}{Contrast} & \rotatebox{70}{Elastic} & \rotatebox{70}{Pixelate} & \rotatebox{70}{JPEG} & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & \begin{tabular}[c]{@{}c@{}}74.0 \\ \small{± 3.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.8 \\ \small{± 3.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}75.3 \\ \small{± 4.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}43.3 \\ \small{± 2.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}48.0 \\ \small{± 2.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}32.6 \\ \small{± 1.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}35.2 \\ \small{± 2.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}22.0 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}33.0 \\ \small{± 2.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}25.9 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}8.5 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.1 \\ \small{± 1.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}23.4 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}53.6 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}26.8 \\ \small{± 0.7}\end{tabular} & 42.3 \\
IABN & \begin{tabular}[c]{@{}c@{}}44.5 \\ \small{± 2.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}41.4 \\ \small{± 2.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}48.1 \\ \small{± 1.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}16.3 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}39.9 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}13.9 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}16.2 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}14.9 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}17.9 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}16.4 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}7.6 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}8.8 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}22.5 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}34.1 \\ \small{± 1.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}26.7 \\ \small{± 0.6}\end{tabular} & 24.6 \\
PBRS & \begin{tabular}[c]{@{}c@{}}43.4 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}37.9 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}46.2 \\ \small{± 1.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}21.8 \\ \small{± 2.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}36.8 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}18.1 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}17.6 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}16.1 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}19.3 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}15.2 \\ \small{± 0.3}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}7.1 \\ \small{± 0.4}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}32.5 \\ \small{± 1.5}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}20.0 \\ \small{± 0.2}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}30.7 \\ \small{± 0.7}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}23.8 \\ \small{± 0.1}\end{tabular}} & 25.8 \\
IABN + RS & \begin{tabular}[c]{@{}c@{}}33.8 \\ \small{± 1.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}31.1 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}40.4 \\ \small{± 1.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}13.3 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}35.6 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}11.8 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}13.2 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}14.6 \\ \small{± 0.3}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}14.9 \\ \small{± 0.6}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}14.7 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}7.7 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}8.1 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}22.3 \\ \small{± 0.5}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}24.6 \\ \small{± 1.9}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}25.1 \\ \small{± 1.2}\end{tabular} & 20.7 \\
IABN + PBRS & \textbf{\begin{tabular}[c]{@{}c@{}}33.5 \\ \small{± 1.7}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}30.0 \\ \small{± 1.6}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}38.2 \\ \small{± 0.9}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}12.6 \\ \small{± 0.8}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}34.4 \\ \small{± 0.8}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}11.5 \\ \small{± 0.5}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}12.9 \\ \small{± 0.6}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}14.1 \\ \small{± 0.2}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}15.2 \\ \small{± 0.8}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}14.0 \\ \small{± 0.6}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}7.4 \\ \small{± 0.2}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}7.8 \\ \small{± 0.2}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}20.7 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}24.7 \\ \small{± 0.7}\end{tabular} & {\begin{tabular}[c]{@{}c@{}}24.2 \\ \small{± 0.4}\end{tabular}} & \textbf{20.1} \\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}%
}
\end{table}
\begin{table}[H]
\caption{Average classification error (\%) and their corresponding standard deviations of varying ablation settings on CIFAR100-C with \textbf{uniformly distributed test streams}, shown per domain. \textbf{Bold} type indicates those of lowest classification error. Averaged over 3 runs. }
\vspace{0.1cm}
\centering
\setlength{\tabcolsep}{3pt}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccccccccccccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
Method & \rotatebox{70}{Gaussian} & \rotatebox{70}{Shot} & \rotatebox{70}{Impulse} & \rotatebox{70}{Defocus} & \rotatebox{70}{Glass} & \rotatebox{70}{Motion} & \rotatebox{70}{Zoom} & \rotatebox{70}{Snow} & \rotatebox{70}{Frost} & \rotatebox{70}{Fog} & \rotatebox{70}{Brightness} & \rotatebox{70}{Contrast} & \rotatebox{70}{Elastic} & \rotatebox{70}{Pixelate} & \rotatebox{70}{JPEG} & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & \begin{tabular}[c]{@{}c@{}}88.1 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}86.8 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}93.7 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}64.9 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}79.7 \\ \small{± 0.9}\end{tabular} & \begin{tabular}[c]{@{}c@{}}55.5 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}57.7 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}53.8 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.3 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}59.3 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}33.0 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}81.4 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}49.2 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}73.6 \\ \small{± 1.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}55.5 \\ \small{± 0.3}\end{tabular} & 66.6 \\
IABN & \begin{tabular}[c]{@{}c@{}}79.3 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}77.2 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}84.3 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}45.0 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}69.6 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}40.9 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}43.1 \\ \small{± 0.6}\end{tabular} & \begin{tabular}[c]{@{}c@{}}42.5 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}48.6 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}52.5 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}30.5 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}40.5 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}47.6 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}59.8 \\ \small{± 1.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}56.2 \\ \small{± 0.4}\end{tabular} & 54.5 \\
PBRS & \begin{tabular}[c]{@{}c@{}}68.6 \\ \small{± 1.0}\end{tabular} & \begin{tabular}[c]{@{}c@{}}66.0 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}72.9 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}45.3 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}64.1 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}40.9 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}41.6 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}43.7 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}47.9 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}44.2 \\ \small{± 0.3}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}28.3 \\ \small{± 0.3}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}59.9 \\ \small{± 0.7}\end{tabular} & \begin{tabular}[c]{@{}c@{}}44.2 \\ \small{± 0.5}\end{tabular} & \begin{tabular}[c]{@{}c@{}}51.1 \\ \small{± 1.6}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}50.4 \\ \small{± 0.6}\end{tabular}} & 51.3 \\
IABN + RS & \begin{tabular}[c]{@{}c@{}}67.1 \\ \small{± 1.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}65.6 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}74.0 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}39.0 \\ \small{± 0.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}61.4 \\ \small{± 1.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}36.5 \\ \small{± 0.1}\end{tabular} & \begin{tabular}[c]{@{}c@{}}38.7 \\ \small{± 0.8}\end{tabular} & \begin{tabular}[c]{@{}c@{}}41.4 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}44.0 \\ \small{± 0.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}45.0 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}30.0 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}34.0 \\ \small{± 0.2}\end{tabular} & \begin{tabular}[c]{@{}c@{}}46.0 \\ \small{± 1.4}\end{tabular} & \begin{tabular}[c]{@{}c@{}}48.8 \\ \small{± 1.3}\end{tabular} & \begin{tabular}[c]{@{}c@{}}52.5 \\ \small{± 0.5}\end{tabular} & 48.3 \\
IABN + PBRS & \textbf{\begin{tabular}[c]{@{}c@{}}65.6 \\ \small{± 1.0}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}62.6 \\ \small{± 0.7}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}72.0 \\ \small{± 0.2}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}36.8 \\ \small{± 0.7}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}60.5 \\ \small{± 0.7}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}34.9 \\ \small{± 0.5}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}36.7 \\ \small{± 0.2}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}39.6 \\ \small{± 0.2}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}41.7 \\ \small{± 0.6}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}42.3 \\ \small{± 0.3}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}28.6 \\ \small{± 0.2}\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}32.3 \\ \small{± 0.9}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}43.8 \\ \small{± 0.2}\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}47.7 \\ \small{± 0.4}\end{tabular}} & \begin{tabular}[c]{@{}c@{}}50.9 \\ \small{± 0.2}\end{tabular} & \textbf{46.4} \\
\addlinespace[0.00cm]
\Xhline{2\arrayrulewidth}
\end{tabular}%
}
\end{table}
\section{Replacing BN with IABN during test time}
\begin{table}[H]
\centering
\caption{Average classification error (\%) and corresponding standard deviations of varying ablation settings on CIFAR10-C/100-C under temporally correlated (non-i.i.d.) and uniformly distributed (i.i.d.) test data stream. IABN* refers to replacing BN with IABN during test time (no pre-training with IABN layers). \textbf{Bold} type indicates those of lowest classification error. Averaged over 3 runs.}\label{tab:ablation_wofinetune}
\vspace{0.1cm}
\begin{tabular}{lcccccc}
\Xhline{2\arrayrulewidth}
\addlinespace[0.08cm]
\textbf{} & \multicolumn{3}{c}{\textbf{Temporally correlated test stream}} & \multicolumn{3}{c}{\textbf{Uniformly distributed test stream}} \\
\addlinespace[0.0cm]
\cmidrule(lr){2-4} \cmidrule(lr){5-7}
\addlinespace[0.0cm]
Method & CIFAR10-C & CIFAR100-C & Avg & CIFAR10-C & CIFAR100-C & Avg \\
\addlinespace[0.02cm]
\hline
\addlinespace[0.05cm]
Source & 42.3 ± 1.1 & 66.6 ± 0.1 & 54.4 & 42.3 ± 1.1 & 66.6 ± 0.1 & 54.4 \\
\textbf{IABN*} & 27.1 ± 0.4 & 60.8 ± 0.1 & 44.0 & 27.1 ± 0.4 & 60.8 ± 0.2 & 44.0 \\
IABN & 24.6 ± 0.6 & 54.5 ± 0.1 & 39.5 & 24.6 ± 0.6 & 54.5 ± 0.1 & 39.5 \\
\textbf{IABN*+PBRS} & 24.9 ± 0.2 & 55.9 ± 0.2 & 40.4 & 23.2 ± 0.4 & 55.3 ± 0.1 & 39.3 \\
IABN+PBRS & \textbf{21.1 ± 0.6} & \textbf{47.0 ± 0.1} & \textbf{34.0} & \textbf{20.1 ± 0.5} & \textbf{46.4 ± 0.0} & \textbf{33.2} \\
\addlinespace[0.0cm]
\Xhline{2\arrayrulewidth}
\end{tabular}
\end{table}
For pre-trained models with BN layers such as ResNet~\cite{7780459}, NOTE{} needs to re-train the model by replacing BN layers with IABN layers in order to utilize the effectiveness of IABN. This requires additional computational cost of re-training, which might make it inconvenient to utilize off-the-shelf models. We further investigate whether simply switching BN to IABN without re-training still leads to performance gain.
Table~\ref{tab:ablation_wofinetune} shows the result of this experiment, where IABN* refers to replacing BN with IABN during test time. We note that IABN* still shows significant reduction of errors under CIFAR10-C and CIFAR100-C datasets compared with BN (Source). We interpret this as the normalization correction in IABN is valid to some extent without re-training of the model. We notice that IABN* outperforms the baselines in CIFAR10-C with 27.1\% error, while the second best (LAME) shows 36.2\% error~\ref{tab:cifar}. In addition, IABN* also shows improvement combined with PBRS. This implies that IABN can be used without re-training the model, which aligns with the fully test-time adaptation paradigm introduced in a recent study~\cite{tent}.
\section{License of assets}\label{app:license}
\paragraph{Datasets} KITTI dataset (CC-BY-NC-SA 3.0), KITTI-rain dataset (CC-BY-NC-SA 3.0), CIFAR10, 100 (MIT License), ImageNet-C (Apache 2.0), MNIST-C (CC-BY-NC-SA 4.0), HARTH datset (MIT License), and the Extrasensory dataset (CC-BY-NC-SA 4.0)
\paragraph{Codes} Code for rain augmentation on KITTI dataset (Apache 2.0), torch-vision for ResNet18 and ResNet50 (Apache 2.0), code for depth estimation used in rain augmentation on KITTI dataset (UCLB ACP-A License), code for generating Dirichlet distributions (Apache 2.0), official repository of CoTTA (MIT License), official repository of TENT (MIT License), and the official repository of LAME (CC BY-NC-SA 4.0).
\clearpage
|
\section{Introduction}
Throughout this paper we denote by $(X, p, g, J, \omega, \Omega)$ a pointed complete Calabi-Yau manifold of complex dimension $n$. This means that $p\in X$, $(g, J, \omega)$ is a K\"ahler metric, $\Omega$ is a holomorphic volume form and the following complex Monge-Amp\`ere equation is satisfied
\begin{equation*}\label{e:CMA}\omega^n=(\sqrt{-1})^{n^2} \Omega\wedge\overline\Omega.
\end{equation*}
We will always impose the following two conditions
\begin{itemize}
\item Euclidean volume growth: there exists a $\kappa>0$ such that $\Vol(B_g(p, R))\geqslant \kappa R^{2n}$ for all $R>0$.
\item Quadratic curvature decay: there exists a $C>0$ such that $|Rm_g(q)|\leqslant Cd_g(p, q)^{-2}$ for all $q\neq p$.
\end{itemize}
We also denote by $(\mathcal C, O, g_\mathcal C, J_\mathcal C, \omega_\mathcal C, \Omega_{\mathcal C})$ a Calabi-Yau cone. This means that $O$ is a
distinguished point called the \emph{vertex}, $\mathcal C\setminus \{O\}$ is a smooth manifold diffeomorphic to $\mathbb{R}_{>0}\times L$ for some compact manifold $L$, $(g_\mathcal C, J_{\mathcal C}, \omega_{\mathcal C}, \Omega_{\mathcal C})$ is a Calabi-Yau metric on $\mathcal C\setminus \{O\}$ and $g_\mathcal C=dr^2+r^2g_L$ is a Riemannian cone. A Calabi-Yau cone is naturally a normal affine algebraic cone.
It is known by \cite{DS2, Liu} that any $X$ as above is asymptotic to a unique Calabi-Yau cone $\mathcal C$ at infinity. Moreover, $X$ is naturally a quasi-projective variety and there is an algebro-gometric 2-step degeneration from $X$ to $\mathcal C$, via a possible intermediate K-semistable cone $W$. The main result of this paper is
\begin{theorem}[No semistability at infinity]\label{t:main0}
$W$ is isomorphic to $\mathcal C$ as normal affine algebraic cones.
\end{theorem}
This comes out of our expectation, since the 2-step degeneration theory of \cite{DS2} works symmetrically with respect to asymptotic cones at infinity and local tangent cones at a singularity and it is known that $W$ \emph{can not} always be eliminated in the local setting. Theorem \ref{t:main0} reveals a sharp contrast between the two cases. Our proof makes a novel use of the Tian-Yau construction of complete Calabi-Yau metrics and the Bishop-Gromov volume monotonicity in Riemannian geometry.
As a byproduct, we also show that $X$ is polynomially asymptotic to $\mathcal C$ at infinity:
\begin{theorem}[Polynomial rate to the asymptotic cone]
\label{t:main2}
There exists a diffeomorphism $\Psi$ from the open subset $\{r>1\}\subset \mathcal C$ to the complement of a compact set in $X$ such that for some $\delta>0$ and all $k\geqslant0$,
\begin{equation}\label{e:polynomial rate}
|\nabla_{g_\mathcal C}^k(\Psi^*g-g_{\mathcal C})|_{g_{\mathcal C}}+|\nabla_{g_\mathcal C}^k(\Psi^*J-J_{\mathcal C})|_{g_{\mathcal C}}
+|\nabla_{g_\mathcal C}^k(\Psi^*\Omega-\Omega_{\mathcal C})|_{g_{\mathcal C}}=O(r^{-\delta-k}), \ \ \ r\rightarrow\infty.
\end{equation}
\end{theorem}
Theorem \ref{t:main2} bridges the gap between two previous results. On the one hand, on the level of Riemannian metrics the result of Colding-Minicozzi \cite{CM}, which uses the \L{}ojasiewicz-Simon theory, implies that $X$ is asymptotic to a unique cone at a \emph{logarithmic} rate. The latter was expected to be optimal in general. On the other hand, Conlon-Hein \cite{CH3} classified complete Calabi-Yau manifolds asymptotic to a Calabi-Yau cone, however under the stronger extra assumption of \emph{polynomial} rate (as given exactly by \eqref{e:polynomial rate}).
\subsection*{Acknowledgements} The first author would like to thank Xuemiao Chen for many discussions on related questions in the context of Hermitian-Yang-Mills connections. The second author thanks Ronan Conlon for his lectures at UC Berkeley on noncompact Calabi-Yau manifolds. We thank Lorenzo Foscolo for helpful discussions, and Yuji Odaka for comments. Both authors are partially supported by NSF Grant DMS-1708420. The first author is partially supported by the Simons Collaboration Grant in special holonomy.
\section{Outline of the proof}\label{two step degeneration}
\subsection{2-step degeneration theory}
Fix $\lambda\equiv1/\sqrt{2}$. Denote by $(X_j, J_j, g_j, \Omega_j, p_j)$ the rescaling of $X$ by a factor $\lambda^{j}$. This means that we take $J_j=J, p_j=p, g_j=\lambda^{2j}g, \Omega_j=\lambda^{nj}\Omega$. It follows from the Bishop-Gromov volume monotonicity that after passing to subsequences $(X_j, p_j)$ converges in the pointed Gromov-Hausdorff sense to a Calabi-Yau cone. We summarize below the theory developed in \cite{DS2} and \cite{Liu}. In \cite{DS2} a more general singular setting is considered but under a technical assumption that $X$ is a rescaled Gromov-Hausdorff limit of \emph{polarized} K\"ahler-Einstein manifolds. This assumption was later removed in \cite{Liu} for complete Calabi-Yau manifolds with Euclidean volume growth and quadratic curvature decay, so the results of \cite{DS2} apply in our setting.
One conclusion of \cite{DS2, Liu} is that the asymptotic cone is unique as a Calabi-Yau cone, i.e., it does not depend on the choice of subsequences involved in the definition. This may also follow from the more general result of Colding-Minicozzi \cite{CM} (which does not concern the complex strucutre), but the complex geometric proof in \cite{DS2, Liu} more importantly gives an algebro-geometric description of the asymptotic cone. This is crucial for us so below we review the relevant statements.
Let $(\mathcal C, g_{\mathcal C}, J_{\mathcal C}, \omega_{\mathcal C}, \Omega_{\mathcal C}, O)$ be the unique asymptotic Calabi-Yau cone of $X$. Since $\mathcal C$ is a K\"ahler cone, there is a Reeb vector field $\xi=J_{\mathcal C}(r\partial_r)$, which is holomorphic and Killing. It generates a holomorphic isometric action of a compact torus $\mathbb T$ on $\mathcal C$.
The K\"ahler form has a simple expression given by $\omega_{\mathcal C}=\frac{1}{2}\partial_{J_\mathcal C}\bar\partial_{J_\mathcal C} r^2$. Furthermore, $\mathcal C$ is naturally a normal affine variety. The coordinate ring $R(\mathcal C)$ can be intrinsically characterized as the space of holomorphic functions on $\mathcal C\setminus \{O\}$ that have polynomial growth at infinity. More explicitly,
there is a holomorphic embedding $\Phi_\infty: (\mathcal C, O)\rightarrow(\mathbb{C}^N, 0)$ such that the Reeb vector field $\xi$ extends to a linear vector field on $\mathbb{C}^N$ of the form $Re(\sqrt{-1}\sum_j d_j z_j\partial_{z_j})$, where $d_j\in \mathbb R_{>0}$ for all $j$.
Let $\Lambda: \mathbb{C}^N\rightarrow\mathbb{C}^N$ be the diagonal linear transformation of $\mathbb{C}^N$ given by
$$\Lambda(z_1,\cdots, z_N)=(\lambda^{d_1}z_1, \cdots, \lambda^{d_N}z_N), $$
and let $T_\xi$ be the one parameter group of linear transformations generated by $\xi$.
Denote by $G_\xi$ the subgroup of $GL(N;\mathbb{C})$ consisting of elements that commute with $T_\xi$ and denote $K_\xi=G_\xi\cap U(N)$. Notice $T_\xi$ naturally acts on $\mathcal C$.
Denote by $B_j$ the unit ball around $p_j$ in $X_j$, which we can identify with the ball of radius $\lambda^{-j}$ around $p$ in $X$. Denote by $B_\infty$ the unit ball around the vertex $O$ in $\mathcal C$. Then we know that $B_\infty$ is the Gromov-Hausdorff limit of $B_j$ as $j\rightarrow\infty$. Fix a distance function on $B_j\sqcup B_\infty$ that realizes the Gromov-Hausdorff convergence. We have the following
\begin{enumerate}[(a).]
\item The ring $R(X)$ of holomorphic functions on $X$ with polynomial growth is finitely generated and $\text{Spec}(R(X))$ defines a normal affine algebraic variety $X'$ with isolated singularities. There is a natural map $\pi:(X,p)\rightarrow (X',p')$ which is a crepant resolution of singularities. In particular, $\pi^{-1}(p')$ is a compact analytic set in $X$.
\item For all $j\geqslant 1$, there are holomorphic embeddings $\Phi_j: (X',p')\rightarrow (\mathbb{C}^N,0) $ and $F_j\in G_\xi$ satisfies $F_j=\Lambda+\tau_j$ for linear maps $\tau_j\rightarrow 0$, such that
\begin{enumerate}[(1).]
\item $\Phi_{j+1}=F_{j+1}\circ\Phi_{j}$.
\item For any subsequence of $\{j\}$ tends to infinity, passing to a further subsequence $\Phi_j(B_j)$ converges $h.\Phi_\infty(B_\infty)$ in the Hausdorff sense in $\mathbb{C}^N$ for some $h\in K_\xi$. This convergence is compatible with the Gromov-Hausdorff convergence in the sense that given any sequence $x_j\in B_j$ converging to $x_\infty\in B_\infty$, we have $\Phi_j(x_j)$ converges to $h.\Phi_\infty(x_\infty)$.
\item We will always identify $(X', p')$ with $(\Phi_1(X'),0)$. Denote $Y_j\equiv\Lambda^j.X'$. Then the weighted asymptotic cone $W=\lim_{j\rightarrow\infty}Y_j$ is a normal affine algebraic variety in $\mathbb{C}^N$. In particular, $W$ is also invariant under the $T_\xi$ action. This limit is to be understood in a strong sense: there exists a generating set $\{P_1, \cdots, P_l\}$ of the ideal $\mathcal I_{X'}$ of polynomials on $\mathbb{C}^N$ vanishing on $X'$, such that the leading terms $\{\tilde P_1, \cdots, \tilde P_l\}$ generate the ideal $\mathcal I_{W}$. Here the leading term of a polynomial is defined via the weighted degree associated to $\Lambda$.
\item The coordinate ring $R(W)$ can be intrinsically described as the graded ring associated to a filtration on $R(X)$ determined by a \emph{degree} function associated to the Calabi-Yau metric $g$.
\item
$\mathcal C$ is in the closure of the $G_\xi$ orbit of $W$ in a suitable multi-graded Hilbert scheme. In particular, $W$ has only one isolated singularity and the canonical bundle of $W$ is trivial.
\end{enumerate}
\end{enumerate}
The motivation in \cite{DS2} was to relate the metric scaling on $X$ to the algebro-geometric rescaling $\Lambda$. The main discovery there is that they are \emph{almost but not exactly} the same, and one
may realize the asymptotic cone $\mathcal C$ as a 2-step degeneration from $X'$ through the intermediate cone $W$. From the technical aspect the central issue is that as $j\rightarrow\infty$, even though $F_j$ converges to $\Lambda$ the accumulated error in the composition $F_j\circ F_{j-1}\circ \cdots \circ F_1$ may still diverge away from $\Lambda^j$. Notice $W$ does not inherit any canonical metric from $X$ or $\mathcal C$. In
\cite{DS2} a parallel result is also proved for tangent cones at local singularities of K\"ahler-Einstein metrics. However there are important distinctions between the two settings.
One such distinction was already pointed out in \cite{DS2}, which shows that the local setting is more \emph{rigid} and the asymptotic setting is more \emph{flexible}. In the case of local tangent cones, it is conjectured in \cite{DS2} that there is a local notion of (K-)stability which characterizes both $(W, \xi)$ and $\mathcal C$ as \emph{invariants} of the underlying algebraic singularity (by analogy with the Harder-Narasimhan-Seshadri filtration for sheaves, one should think of $W$ as a \emph{semistable} object and $\mathcal C$ as a \emph{polystable} object). Later Li \cite{ChiLi} related this picture to a generalized volume minimization in Sasaki geometry and reformulated this conjecture using more algebro-geometric terminologies.
There has been active research on this topic. The uniqueness part of the conjecture in \cite{DS2} was confirmed by Li-Wang-Xu \cite{LWX}. By way of contrast as suggested at the end of \cite{DS2}, for asymptotic cones of complete Calabi-Yau metrics one should not expect either $W$ or $\mathcal C$ is a canonical object associated to the underlying algebraic variety $X$. Explicit examples have been constructed on $\mathbb{C}^n$ if we allow $W$ and $\mathcal C$ to have singular cross sections (see \cite{HN, YLi, Gabor1, CR}).
The main result of this paper shows that the asymptotic setting is however more \emph{rigid} from a different perspective.
In the case of tangent cones at singularities of K\"ahler-Einstein metrics, it is known that both of the 2 steps are necessary in order to degenerate the singularity to the metric tangent cone $\mathcal C$. In the case of asymptotic cones at infinity we will show that $W$ is always isomorphic to $\mathcal C$, in other words, one can degenerate the variety to $\mathcal C$ in one step. Theorem \ref{t:main0} can be reformulated as
\
\noindent {\bf Theorem \ref{t:main0}} (No semistability at infinity).
\emph{ $W$ is isomorphic to $\mathcal C$ as affine varieties with $T_\xi$ action. }
\subsection{Outline of the proof}
Now we explain the key ideas in the proof of Theorem \ref{t:main0}. Theorem \ref{t:main2} will be a consequence.
Our strategy is to construct a Calabi-Yau metric ${\omega_W}$ on $W\setminus\{0\}$ so that
\begin{enumerate}
\item the asymptotic cone of ${\omega_W}$ at infinity is given by $\mathcal C$;
\item the metric completion of ${\omega_W}$ is homeomorphic to $W$ and is a pointed Gromov-Hausdorff limit of complete Calabi-Yau metrics. Furthermore the result of \cite{DS2} can be extended to apply to the singular point $0\in W$.
\end{enumerate}
Now the Bishop-Gromov inequality applies on $W$, yielding that the volume ratio $\nu(R)=\text{Vol}(B(0, R))R^{-2n}$ is a non-increasing function in $R$. On a Calabi-Yau cone $\mathcal C$ we denote by $\kappa(\mathcal C)=\text{Vol}(B(O, 1))$ the volume density of $\mathcal C$. Now as $R\rightarrow\infty$, $\nu(R)$ converges to $\kappa(\mathcal C)$, while as $R\rightarrow0$ it converges to the volume density of the metric tangent cone at $0$. By \cite{DS2} and the generalized volume minimization in \cite{LX} we know the latter is not bigger than $\kappa(\mathcal C)$. So $\nu(R)$ is constant in $R$, which implies that $(W, {\omega_W})$ must be a metric cone hence is isometric to $\mathcal C$. With more work one can show that $W$ and $\mathcal C$ are $T_\xi$-equivariantly isomorphic.
The existence of ${\omega_W}$ is therefore the crux of the matter. For this purpose we will make use of the Tian-Yau construction \cite{TY2} which produces complete Calabi-Yau metrics out of lower dimensional compact K\"ahler-Einstein metrics. In \cite{TY2} one starts with a Fano manifold $X$ and a smooth divisor $D$ which is $\mathbb Q$-linearly equivalent to $K_X^{-1}$, then assuming that $D$ admits a K\"ahler-Einstein metric with positive Ricci curvature, one can find a complete Calabi-Yau metric on $X\setminus D$. The rough idea is to first use the Calabi ansatz to write down a background complete K\"ahler metric which is approximately Ricci-flat and then solve the complex Monge-Amp\`ere equation on the noncompact manifold $X\setminus D$. The work of \cite{TY2} has been extended widely to produce more examples of Calabi-Yau metrics.
In the proof of Theorem \ref{t:main0} we will apply the Tian-Yau construction in a non-traditional manner. That is, we will use the unknown Calabi-Yau metric $\omega$ on $X$ as a model at infinity to construct a Calabi-Yau metric ${\omega_W}$ on $W$. This is achieved by constructing a diffeomorphism between the ends of $X$ and $W$, then grafting $\omega$ to a K\"ahler form on $W$ that is slowly asymptotic to the cone $\mathcal C$, but satisfies the complex Monge-Amp\`ere equation with polynomially decaying error. Most of the work in this paper is devoted to making this strategy rigorous. There is an extra difficulty caused by the fact that $W$ has an isolated singularity. To get around the issue, we notice that $W$ is the limit of $Y_j$ in $\mathbb{C}^N$, and $Y_j$ admits a crepant resolution $\hat Y_j$ (which is isomorphic to $X$). We will first construct a family of Calabi-Yau metrics $\hat\omega_{j, \epsilon}$ on $\hat Y_j$ with uniform estimates, where $\epsilon$ measures the size of the exceptional set, and then take a double limit by first letting $\epsilon\rightarrow 0$ and then letting $j\rightarrow\infty$. In the course of the proof we will bring in more robust techniques, such as the H\"ormander $L^2$ estimate, which is of independent interest.
In the rest of this paper we will present a detailed proof of the main results.
In Section \ref{S3} we construct a family of background K\"ahler metrics on $\hat Y_j$ which are asymptotic to the cone $\mathcal C$ and are approximate solutions to the complex Monge-Amp\`ere equation at infinity. In Section \ref{s5} we solve the complex Monge-Amp\`ere equation with respect to these background metrics and derive uniform estimates. In Section \ref{s6} we complete the proof of Theorem \ref{t:main0} and \ref{t:main2}.
We make a convention on the notations throughout this paper: $K$ will denote a compact set in some space; $C$ will denote a positive constant; $C_{\sharp}$ will denote a constant that depends only on $\sharp$; in particular when $j$ does not appear in the subscript it means that the constant is uniform in $j$. The precise meaning of these objects may vary between lines.
\section{Construction of background metrics} \label{S3}
\subsection{Weak conical approximation at infinity} Denote by $r=d(O, \cdot)$ the radial function on the asymptotic Calabi-Yau cone $\mathcal C$, by $\mathcal Q_R$ the subset $\{r\geqslant R\}$ and by $A_{R, 2R}$ the annulus $\{R\leqslant r\leqslant 2R\}$. The following lemma is standard so we only give a sketch of proof.
\begin{lemma}\label{l:asymptotic decay}
There exist an $R_0>0$, a compact set $K\subset X$ and a diffeomorphism $\Phi: \mathcal Q_{R_0}\rightarrow X\setminus K$ such that for all $l\geqslant 0$, $\lim_{s\rightarrow\infty}\sup_{\partial \mathcal Q_s}s^l\left (\left|\nabla^l_{g_\mathcal C}(\Phi^*g-g_{\mathcal C})\right|_{g_\mathcal C}+\left |\nabla^l_{g_\mathcal C}(\Phi^*J-J_{\mathcal C})\right |_{g_\mathcal C}\right)=0$.
\end{lemma}
\begin{proof}
Since $\mathcal C$ is the unique asymptotic cone of $X$, for $i$ large we may find a diffeomorphism $\Phi_i$ from $A_{4^{i}, 8\cdot 4^{i}}$ onto $B(p, , 8\cdot 4^i(1+\epsilon_i))\setminus B(p, 4^i(1+\epsilon_i))$ such that $4^{il}\sup_{A_{4^i, 8\cdot 4^i}}(|\nabla^l_{g_\mathcal C}(\Phi_i^*g-g_{\mathcal C})|_{g_\mathcal C}+|\nabla^l_{g_\mathcal C}(\Phi_i^*J-J_{\mathcal C})|_{g_\mathcal C})\leqslant\epsilon_{i,l}$, where $\epsilon_i\rightarrow0$ and for each $l$ fixed, $\epsilon_{i, l}\rightarrow 0$ as $i\rightarrow\infty$. Moreover, the rescaled maps $\Phi_i\circ 4^i$ realize the Gromov-Hausdorff convergence to the cone $\mathcal C$. It suffices to glue together $\Phi_i$'s to get the desired $\Phi$. This can done in a straightforward way. By a contradiction argument one can prove that for $i$ large, in the intersection $\text {Im}(\Phi_i)\cap\text{Im} (\Phi_{i+1})$, the composition $4^{-i-1}\circ \Phi_{i+1}^{-1}\circ \Phi_i\circ 4^{i+1}$ is close to a holomorphic isometric embedding $\Psi_i$ of the annulus $A_{1,2}$ into itself with $\Psi_i^*r=r$. It follows that $\Psi_i$ preserves the Reeb vector field on $\mathcal C$, hence it extends to a global holomorphic isometry of $\mathcal C$ which fixes the vertex. Then we replace $\Phi_{i+1}$ by $\Phi_{i+1}\circ \Psi_i$ and glue it with $\Phi_i$. Then one can proceed inductively on $i$.
\end{proof}
Notice that at this point we only know the metric $\Phi^*g$ is asymptotic to $g_{\mathcal C}$ at infinity, without any quantitative \emph{rate} of decay. Nevertheless, the difference of the Levi-Civita connections does decay at the order $o(r^{-1})$. The general result of Colding-Minicozzi \cite{CM} yields a logarithmic decay of the Riemannian metric but we will not need it in this paper. Using the diffeomorphism constructed in Lemma \ref{l:asymptotic decay}, we may naturally view $r$ as a function on the end of $X$, and we have
$$C^{-1} d_g(p, \cdot)\leqslant r(\cdot)\leqslant C d_g(p, \cdot).$$
Notice $r$ is a smooth function so is better suited for the purpose of doing analysis.
\subsection{Weighted elliptic estimates}
Here we collect some facts on weighted elliptic theory for ends of Riemannian manifolds asymptotic to cones. These are not sharp results but suffice for our purpose in this paper.
For any $k\in \mathbb Z_{\geqslant 0}$, $\alpha\in (0, 1)$, $\delta\in \mathbb R$, $R>0$, we define the Banach space $C^{k, \alpha}_{\delta}(\mathcal Q_R)$ using the following weighted norm on a function $f$ defined on $\mathcal Q_R\subset\mathcal C$:
\begin{align*}
\|f\|_{C_{\delta}^{k, \alpha}(\mathcal \mathcal Q_R)}& \equiv \sup_{s\geqslant R}\Big\{ \sum\limits_{m=0}^ks^{-\delta+m} \|\nabla_{g_{\mathcal C}}^m f\|_{C^0_{g_\mathcal C}(A_{s,2s})} + s^{-\delta+k+\alpha}[f]_{C^{k,\alpha}_{g_{\mathcal C}}(A_{s,2s})}\Big\},\end{align*}
where
\begin{align*}
[f]_{C^{k,\alpha}(A_{s,2s})} \equiv \sup\limits_{y_1, y_2\in A_{s,2s}, d_{g_\mathcal C}(y_1,y_2)<\text{inj}_{g_\mathcal C}(y_1)}\bigg\{\frac{|\nabla^k_{g_\mathcal C}f(y_1)-\nabla^k_{g_\mathcal C}f(y_2)|}{d_{g_\mathcal C}(y_1,y_2)^\alpha} \bigg\}.
\end{align*}
Fix $k\geqslant 2n+1$ and $\alpha\in (0, 1)$.
The following can be proved in exactly the same way as Proposition 6.7 in \cite{SZ}, so we omit the proof here.
\begin{lemma}\label{p:weighted-esimate-quotient-space}
There exists a finite set $\Gamma\subset(0, 1)$ such that for all $\delta\in (0, 1)\setminus \Gamma$ and for all $R\geqslant 1$, one can find a bounded linear map $\mathcal S_R: C^{k, \alpha}_{-\delta}(\mathcal Q_R)\to C^{k+2, \alpha}_{-\delta+2}(\mathcal Q_R)$ with $\Delta_{g_\mathcal C}\circ \mathcal S_R=\Id$ and with the operator norm $\|\mathcal S_R\|\leqslant C$ for some $C$ depending on $\mathcal C$, $\delta$, $k$, $\alpha$ (but not on $R$).
\end{lemma}
Now fix $R_1>0$. Suppose there is another Riemannian metric $g$ on $\mathcal Q_{R_1}$ such that for all $l\geqslant 0$ and $s\geqslant R_1$,
\begin{equation}\label{new riemannian metric}
s^l\sup_{\partial \mathcal Q_s}|\nabla^l_{g_{\mathcal C}} (g-g_{\mathcal C})|\leqslant e_l(s)
\end{equation}
for some function $e_l$ defined on $[R_1, \infty)$ with $\lim_{s\rightarrow \infty}e_l(s)=0$. Then using similar expressions as above, we can define $C^{k, \alpha}_{\delta}(\mathcal Q_R)$ with respect to the Riemannian metric $g$.
\begin{proposition} \label{p:Poisson equation general}
Given $\delta\in (0,1)\setminus \Gamma$, there exist $R\geqslant R_1$ and $C>0$ depending only on $\mathcal C$, $R_1$, $k, \delta, \alpha$ and the functions $e_1,\cdots, e_{k+2}$, such that there is a bounded linear map $\mathcal T_R: C^{k,\alpha}_{-\delta}(\mathcal Q_R)\to C^{k+2, \alpha}_{-\delta+2}(\mathcal Q_R)$ with the properties that $\Delta_{g}\circ \mathcal T_R=\Id$ and $\|\mathcal T_R\|\leqslant C$. \end{proposition}
\begin{proof}
One can write $\Delta_{g}=\Delta_{g_\mathcal C}+(g-g_{\mathcal C})*\nabla^2_{g_\mathcal C}+\nabla_{g_{\mathcal C}}g*\nabla_{g_\mathcal C}$. Then we have
$$\|\Delta_g-\Delta_{g_\mathcal C}\|\leqslant C \sup_{s\geqslant R, \ 0\leqslant l \leqslant k+2}e_l(s).$$
The right hand side is small when $R\gg1$. Then the conclusion follows from standard functional analysis. \end{proof}
The following is an application of the standard elliptic regularity by a straightforward rescaling argument. We omit the proof.
\begin{proposition}\label{c:elliptic regularity} Let $g$ be a Riemannian metric on $\mathcal{Q}_{R_1}$ satisfying \eqref{new riemannian metric}.
Suppose $u$ is a function defined on $\mathcal Q_{R_1}$ and for some $s, s'\in \mathbb R$ and for all $l\geqslant0$, $|\nabla^l_g(\Delta_gu)|_g\leqslant C_l r^{-l-2+s}$, and $\int_{\mathcal{Q}_{R_1}}u^2r^{-2s'-2n}d\operatorname{Vol}_g\leqslant \hat C$. Then for any $R>R_1$ and $l>0$, there exists a constant $A_l>0$ depending only on $\mathcal C$, $R_1$, $R$, $C_0, C_1, \cdots, C_l, \hat C$ and the functions $e_1,\cdots, e_{l}$, such that $|\nabla_g^{l}u|_g\leqslant A_l r^{\gamma-l}$ on $\mathcal{Q}_{R}$ for all $\gamma>\min(s, s')$.
\end{proposition}
\subsection{Rough algebraic approximation of the Calabi-Yau metric}
We continue to use the notations in Section 2.1. Since the coordinate functions on $\mathbb{C}^N$ have positive weight with respect to $\xi$, we can find a K\"ahler cone metric $\omega_\xi$ on $\mathbb{C}^N$ with Reeb vector field $\xi$. Fix a choice of such $\omega_\xi$. Then we have $\Lambda^*\omega_\xi=\lambda^{2}\omega_\xi$. Denote by $r_\xi=d_{\omega_\xi}(0, \cdot)$ the radial function on $\mathbb{C}^N$ defined by $\omega_\xi$.
\begin{proposition}\label{rough estimate}
There exists a compact set $K\subset X$ such that for all $\delta>0$ and $k\geqslant 1$, on $X\setminus K$ we have \begin{equation}\label{e:rough comparison} C_\delta^{-1} r^{-\delta}\omega\leqslant \Phi_1^*\omega_\xi\leqslant C_\delta r^{\delta}\omega,
\end{equation}
\begin{equation}\label{eqn4.3}|\nabla^k_{\omega}\Phi_1^*\omega_{\xi}|_{\omega}\leqslant C_{\delta,k} r^{\delta-k},
\end{equation}
\begin{equation}\label{e:rough comparison of distance}
C_\delta^{-1} r^{1-\delta}\leqslant \Phi_1^*r_\xi\leqslant C_\delta r^{1+\delta}.
\end{equation}
\end{proposition}
\begin{proof}
By definition $\Phi_j=F_j\circ \cdots F_2\circ \Phi_1$. By Item (b) in Section \ref{two step degeneration} we know that for all $j\geqslant 1$ and $k\geqslant 0$, the following hold on $B_{j+1}\setminus B_{j}$:
\begin{equation}\label{e:compare omega and omegaxi}C^{-1}\lambda^{2j}\omega\leqslant \Phi_j^*\omega_\xi\leqslant C \lambda^{2j}\omega, \ \ \ \ |\nabla^k_{\omega}(\lambda^{-2j}\Phi_j^*\omega_{\xi})|_{\omega}\leqslant C_{k} \lambda^{kj}.
\end{equation}
By construction we have $F_j^{-1}=\Lambda^{-1}\circ \sigma_j$ for a linear map $\sigma_j\in G_\xi$ with $\|\sigma_j-\Id\|\rightarrow0$, where $\|\cdot\|$ is a fixed norm on the space of $N\times N$ matrices. Since $\sigma_j$ commutes with $\Lambda$ we may write
$$F_2^{-1}\circ \cdots\circ F_j^{-1}=\Lambda^{1-j}\sigma_2\circ \cdots\circ \sigma_j. $$
For any $\sigma\in G_\xi$ we have $\mathcal L_{r_\xi\partial_{r_\xi}}\sigma^*\omega_\xi=2\sigma^*\omega_\xi$. It follows that $\sigma^*\omega_\xi$ is homogeneous of degree 2 on $\mathbb{C}^N$. So one can find a small neighborhood $U$ of $\text{Id}$ such that for all $\sigma\in U$ and all $k\geqslant 0$, the following holds on $\mathbb{C}^N$:
\begin{equation}\label{e:pull back metric estimate}|\nabla^k_{\omega_\xi}(\sigma^*\omega_\xi-\omega_\xi)|_{\omega_\xi} \leqslant C_k r_\xi^{-k}\|\sigma-\Id\|.\end{equation}
In particular we can also assume
\begin{equation}\label{e:connection difference estimate}|\nabla^k_{\omega_\xi}(\nabla_{\sigma^*\omega_\xi}-\nabla_{\omega_\xi})|_{\omega_\xi}\leqslant C_k r_\xi^{-k-1}\|\sigma-\Id\|,
\end{equation}
where as usual we view the difference of the two connections as a tensor.
Since $\|\sigma_j-\Id\|\rightarrow 0$, for $j\gg1$ we have $\sigma_j\in U$.
We notice the elementary fact that for all $\delta>0$, $$\prod_{l=2}^j\big(1+C\|\sigma_l-\Id\|\big)\leqslant C_\delta \lambda^{-j\delta}.$$ From this, iterating \eqref{e:pull back metric estimate} and using \eqref{e:compare omega and omegaxi} yield \eqref{e:rough comparison}. The latter also implies \eqref{e:rough comparison of distance}.
We now prove the case $k=1$ in \eqref{eqn4.3}. The case $k>1$ can be proved similarly by induction.
Notice for any $\sigma\in U$ and 2-form $\gamma$, we have
\begin{eqnarray*}
\nabla_{\omega_\xi}\sigma^*\gamma&=&\nabla_{\sigma^*\omega_\xi}\sigma^*\gamma+(\nabla_{\omega_\xi}-\nabla_{\sigma^*\omega_\xi})\sigma^*\gamma\\&=&\sigma^*(\nabla_{\omega_\xi}\gamma)+(\nabla_{\omega_\xi}-\nabla_{\sigma^*\omega_\xi})\sigma^*\gamma.
\end{eqnarray*}
Therefore,
$$|\nabla_{\omega_\xi}(\sigma_j^*\cdots \sigma_2^*\omega_\xi)|_{\omega_\xi}=|\sum_{l=2}^{j} \sigma_j^*\cdots \sigma_{l+1}^*(\nabla_{\omega_\xi}-\nabla_{\sigma_l^*\omega_\xi})\sigma_{l-1}^*\cdots\sigma_2^*\omega_\xi|_{\omega_\xi}.$$
Using \eqref{e:pull back metric estimate} and \eqref{e:connection difference estimate} we obtain that on the annulus $\{\lambda\leqslant r_\xi\leqslant \lambda^{-1}\}\subset\mathbb{C}^N$ for all $\delta>0$,
\begin{equation}\label{eqn4.7}|\nabla_{\omega_\xi}(F_2^{-1}\circ\cdots\circ F_j^{-1})^*\omega_\xi|_{\omega_\xi}\leqslant C \lambda^{-2j}\prod_{l=2}^j (1+C\|\sigma_l-\Id\|)\leqslant C_\delta \lambda^{-(2+\delta)j}.
\end{equation}
By definition $\Phi_1^*\omega_\xi=\Phi_j^*\left((F_2^{-1}\circ\cdots\circ F_j^{-1})^*\omega_\xi\right)$, so we have
$$|\nabla_{\omega}\Phi_1^*\omega_\xi|_\omega=|\Phi_j^*(\nabla_{\omega_\xi}(F_2^{-1}\circ\cdots\circ F_j^{-1})^*\omega_\xi)|_\omega+|(\nabla_{\omega}-\nabla_{\Phi_j^*\omega_\xi})\Phi_j^*(F_2^{-1}\circ\cdots\circ F_j^{-1})^*\omega_\xi|_\omega.$$
The first term of the right hand side can be handled using \eqref{eqn4.7} and \eqref{e:compare omega and omegaxi}, and the second term can be dealt with using \eqref{e:compare omega and omegaxi} and \eqref{e:rough comparison}. Together we get \eqref{eqn4.3} when $k=1$.
\end{proof}
We may view $\omega_\xi$ as a rough approximation of the unknown Calabi-Yau metric $\omega$. Although not needed in this paper, we remark that the above $\omega_\xi$ is not unique and we can find one with explicit algebraic formula. We emphasize that from the proof we can not get rid of the $\delta$ in the estimates. In particular, we can \emph{not} claim the uniform equivalence between $\omega$ and $\omega_\xi$. Nonetheless we know the error is smaller than any polynomial order. This suffices for our applications. Moreover, the differences between the covariant derivatives do decay at infinity. Namely,
\begin{corollary}\label{estimates for different covariant derivatives}
There exists a compact set $K\subset X$ such that on $X\setminus K$, \text{for all $\delta>0$ and $k\geqslant 0$} we have
\begin{equation}\label{estimate for covariant derivatives}
|\nabla^k_{\omega}(\nabla_\omega-\nabla_{\Phi_1^*\omega_{\xi}})|_{\omega}\leqslant C_{\delta,k} r^{\delta-k-1}.
\end{equation}
\end{corollary}
\begin{proof}
Two Riemannian metrics $g_1$ and $g_2$ on a manifold determine a section $T$ of $\operatorname{End}(TM)$ via
\begin{equation*}
g_2(\cdot,\cdot)=g_1(T\cdot,\cdot),
\end{equation*}
which is positive and self-adjoint with respect to both $g_1$ and $g_2$. Using Koszul formula, one can check directly that
\begin{equation*}
\nabla_{g_2}-\nabla_{g_1}=T^{-1}*\nabla_{g_1}T
\end{equation*} where we view both sides as (1,2)-tensors and $*$ denotes some algebraic operations. Therefore to estimate $|\nabla_{g_1}^k(\nabla_{g_2}-\nabla_{g_1})|_{g_1}$ it suffices to estimate $T^{-1}$ and $\nabla^k_{g_1}T$. Then \eqref{estimate for covariant derivatives} follows from Proposition \ref{rough estimate}.
\end{proof}
\subsection{Plurisubharmonic weight functions}
We construct here some plurisubharmonic functions that will be used later. As mentioned in Section \ref{two step degeneration} we always identify $(X',p')$ with $(\Phi_1(X'),0)$ in $\mathbb C^N$. Then we can naturally restrict functions and forms defined on $\mathbb{C}^N$ to $X'$ and then pull-back to $X$ via the crepant resolution map $\pi$. When there is no confusion we will omit the various pull-back notations.
\begin{lemma}\label{psh function on a kahler cone}The following hold
\begin{itemize}
\item[(1)] Let $(\mathcal K, \omega_{\mathcal K})$ be a K\"ahler cone and let $\rho$ denote the distance function from the vertex of the cone. Then the function $\varphi =\log(\rho^2+1)-\frac{1}{100}(\log(\rho^2+3))^{\frac{1}{2}}$ is a strictly plurisubharmonic function and there exists a compact set $K\subset \mathcal K$ such that on $\mathcal K\setminus K$,\begin{equation*}
\ensuremath{\sqrt{-1}}\partial\bar\partial \varphi\geqslant C \rho^{-2}(\log \rho)^{-\frac{3}{2}}\omega_{\mathcal K} .
\end{equation*}
\item[(2)]The function on $X$ defined by $\varphi=\log(r_{\xi}^2+1)-\frac{1}{100}(\log(r_{\xi}^2+3))^{\frac{1}{2}}$ satisfies $\sqrt{-1}\partial\bar\partial\varphi>0$ away from the exceptional set of $\pi$. Moreover outside a compact set $K$ it satisfies the following for all $\epsilon>0$: $$(1-\epsilon)\log (r^2+1)-C_{\epsilon}\leqslant \varphi\leqslant (1+\epsilon)\log (r^2+1)+C_{\epsilon},$$ $$\ensuremath{\sqrt{-1}}\partial\bar\partial \varphi\geqslant {C_{\epsilon}^{-1}}{r^{-2-\epsilon}}\omega,$$
$$|\nabla_\omega\varphi|_{\omega}\leqslant C_\epsilon r^{-1+\epsilon}. $$
\end{itemize}
\end{lemma}
\begin{proof}
(1) follows from a straightforward computation: the first term is strictly plurisubharmonic but is decaying at the rate $\rho^{-4}$ along the radial direction and the second term is used to strengthen the positivity along the radial direction. Applying (1) to the K\"ahler cone $(\mathbb{C}^N, \omega_\xi)$ and then using Proposition \ref{rough estimate}, we obtain (2).
\end{proof}
\subsection{Decomposition of the K\"ahler form}
We will use the following version of H\"ormander $L^2$ estimate.
\begin{theorem}\label{Hormander}
Suppose $M$ is an $n$ dimensional complex manifold with trivial canonical bundle which admits a complete K\"ahler metric. Let $\omega_M$ be any K\"ahler metric on $M$ and $\varphi_M$ be a smooth function with $\sqrt{-1}\partial\bar\partial\varphi_M\geqslant \Upsilon\omega_M$ for a continuous non-negative function $\Upsilon$. Let $q$ be a positive integer. Then for any $(n,q)$ form (resp. $(0,q)$ form) $\eta$ on $M$ with $\bar\partial \eta=0$ and $\int_{M}\Upsilon^{-1} |\eta|_{\omega_M}^2e^{-\varphi_M}\omega_M^n<\infty$, there exists a $(n, q-1)$ form (reps. $(0, q-1)$ form ) $\zeta$ on $M$ satisfying $\bar\partial\zeta=\eta$ and with estimate
\begin{equation*}
\int_{M} |\zeta|_{\omega_M}^2e^{-\varphi_M}\omega_M^n\leqslant \int_{M}q^{-1}\Upsilon^{-1}|\eta|_{\omega_M}^2e^{-\varphi_M}\omega_M^n.
\end{equation*}
\end{theorem}
This follows from for example \cite[Chapter VIII, Theorem 6.1]{Demailly}: one simply takes $L$ to be the trivial bundle and use a trivialization of $K_M$ to identify a $(0,q)$ form with a $(n,q)$ form.
The main result of this subsection is
\begin{theorem}\label{decomposition of kahler forms}
For every $\epsilon>0$, there exists a K\"ahler form $\beta$ and a real-valued smooth function $\psi$ on $X$ such that
\begin{itemize}
\item [(1)]$\omega=\beta+\ensuremath{\sqrt{-1}}\partial\bar\partial \psi$;
\item [(2)]$|\nabla^k_{\omega}\beta|_{\omega}=O(r^{-2+\epsilon-k})$ for all $k\geqslant 0$;
\item [(3)]$|\nabla_{\omega}^k\psi|_{\omega}=O(r^{2+\epsilon-k})$ for all $k\geqslant 0$.
\end{itemize}
\end{theorem}
\begin{proof}
Let $\varphi$ be the plurisubharmonic function obtained in Lemma \ref{psh function on a kahler cone}-(2). Fix $\epsilon>0$.
\
\textbf{Step 1.} By Lemma \ref{l:asymptotic decay} we can identify $X\setminus K$ (for $K$ compact) with $\mathcal Q_{R_0}\subset\mathcal C$ using the diffeomorphism $\Phi$. Then one can write $\omega=\tilde\beta+d\tilde\eta$ with $|\nabla^k_{\omega} \tilde\beta|_{\omega}=O(r^{-2-k})$ for all $k\geqslant 0$.
\
\textbf{Step 2.}
We decompose $\tilde\beta$ according to types:
$
\tilde\beta=\tilde \beta^{2,0}+\tilde \beta^{1,1}+\tilde \beta^{0,2}.
$
Since $d\tilde \beta=0$, we have $\bar\partial\tilde\beta^{0, 2}=0$. Notice by \textbf{Step 1}, we know that for all $\tau>0$, $$\int_{X} |\tilde\beta^{0,2}|_\omega^2 e^{-(n-2+\tau)\varphi}\omega^n<\infty.$$
So we may apply Theorem \ref{Hormander} (with $\varphi_M=(n-1+\tau)\varphi$) to obtain that for every $\tau>0$, there exists a (0,1)-form $\zeta$ on $X$ satisfying $\bar\partial\zeta=\tilde\beta^{0,2}$ and $\int_{X}|\zeta|_{\omega}^2 e^{-(n-1+\tau)\varphi}\omega^n<\infty$. It follows from Proposition \ref{c:elliptic regularity} that $|\nabla_{\omega}^k\zeta|_{\omega}=O(r^{-1+2\tau-k})$ for all $k\geqslant 0$.
Choose $\tau=\frac{\epsilon}{4}$.
Then we obtain $\omega=\beta_2+d\eta_2$, where \begin{equation*}
\beta_2=\tilde\beta^{1,1}-\partial\zeta-\bar\partial\olsi{\zeta}
, \ \ \ \ \eta_2=\zeta+\olsi{\zeta}+\tilde\eta,
\end{equation*} and $|\nabla_{\omega}^k\beta_2|_{\omega}=O(r^{-2+\frac{\epsilon}{2}-k})$ for all $k\geqslant 0$. Notice $\beta_2$ and $d\eta_2$ are both real-valued $(1,1)$ forms.
\
\textbf{Step 3.} Again we use the map $\Phi$ in Lemma \ref{l:asymptotic decay} to identify $X\setminus K$ with $\mathcal Q_{R_0}\subset \mathcal C$. Let $\chi$ be a smooth compactly supported function on $X$ which equals 1 in a neighborhood of $K$. Then $\eta_2=\chi\eta_2+\eta_2'$, where $\eta_2'$ vanishes in a neighborhood of $K$. Writing
$d\eta_2'=dr\wedge\alpha_1+\alpha_2$, where $\partial_r\lrcorner \alpha_1=\partial_r\lrcorner \alpha_2=0$, we define $\lambda=\int_{R_0}^{r}\alpha_1 dr$. Then $\lambda$ vanishes near $\partial K$ so extends to a 1-form on $X$ which vanishes on $K$. Moreover, one can check that $d\eta_2'=d\lambda$. The decay of $\beta_2$ implies $|d\eta_2|_{\omega}\leqslant C$ on $X$, so we have $|\lambda|=O(r^{1+\delta})$ for all $\delta>0$. It follows that $d\eta_2=d\gamma$, where $\gamma=\chi\eta_2+\lambda$, with $|\gamma|=O(r^{1+\delta})$ for all $\delta>0$.
\
\textbf{Step 4.} Since $d\gamma=\omega-\beta_2$ is of type $(1,1)$, we have $\bar\partial\gamma^{0, 1}=0$. We now apply Theorem \ref{Hormander} (with $\varphi_M=(n+2+\frac \epsilon 2)\varphi$) to obtain a function $u$ satisfying $\bar\partial u=\gamma^{0,1}$ and
\begin{equation*}
\int_X |u|^2e^{-(n+2+\frac{\epsilon}{2})\varphi}\omega^n<\infty.
\end{equation*}
Then $d\gamma=\sqrt{-1}\partial\bar\partial\psi_1$, where $\psi_1=\sqrt{-1}(\bar u-u)$.
Since $\Delta_{\omega}\psi_1=\operatorname{tr}_{\omega}(\sqrt{-1}\partial\bar\partial \psi_1)=\operatorname{tr}_{\omega}(d\eta_2)$, by Proposition \ref{c:elliptic regularity} we see
$|\nabla_\omega^k\psi_1|_{\omega}=O(r^{2+\epsilon-k})$ for all $k\geqslant 0$.
\
\textbf{Step 5.} We have achieved that $\omega=\beta_2+\sqrt{-1}\partial\bar\partial\psi_1$ with estimates on both $\beta_2$ and $\psi_1$. It remains to modify $\beta_2$ to be a K\"ahler form. Recall that using the diffeomorphism constructed in Lemma \ref{l:asymptotic decay}, we view $r$ as a smooth function on the end of $X$. Moreover by the asymptotics for the complex structure and the Riemannian metric obtained in Lemma \ref{l:asymptotic decay}, we know that there exists a large constant $\olsi{R}>R_0$ such that on the region $r>\olsi{R}$ we have $\ensuremath{\sqrt{-1}}\partial\bar\partial r^{\epsilon}\geqslant c r^{-2+\epsilon}\omega,$ for some positive number $c>0$.
Choose a smooth non-decreasing convex function $u:\mathbb R\rightarrow \mathbb R$ such that $u(t)=1$ for $t\leqslant 2\olsi{R}^{\epsilon}$ and $u(t)=t$ for $t\gg1$ and let $h=u\circ r^{\epsilon}$. Then $h$ can be naturally viewed as a smooth plurisubharmonic on $X$ and outside a compact set $K'$ we have
\begin{equation*}
\ensuremath{\sqrt{-1}}\partial\bar\partial h\geqslant c r^{-2+\epsilon}\omega,
\end{equation*}
for some positive number $c>0$. Choose a smooth compactly supported function $\chi$ on $X$ which equals 1 on a neighborhood of $K'$. Then
we define $$\beta=\beta_2+\sqrt{-1}(\chi\psi_1)+\sqrt{-1}\partial\bar\partial(Ah),$$
$$\psi=(1-\chi)\psi_1-Ah, $$
where $A$ is a constant. One can choose $A$ large so that $\beta$ is K\"ahler.
Then $\omega=\beta+\sqrt{-1}\partial\bar\partial\psi$. It follows from the above construction that we have the desired estimates.
\end{proof}
\subsection{Construction of background K\"ahler forms}\label{section:defining projection map} Recall that $W$ is the weighted asymptotic cone of $X'$ with respect to $\Lambda$ and $Y_j=\Lambda^j.X'$, so in particular away from $0$ we have $Y_j$ converges to $W$ as smooth submanifolds in $\mathbb C^N$. For notational convenience we also write $Y_\infty=W$. We denote by $\text{Sing}(Y_j)$ the singular set of $Y_j$, which consists of finitely many points, and denote $Y_j^\circ=Y_j\setminus \text{Sing}(Y_j)$. Define $\mathfrak S\equiv\cup_{1\leqslant j\leqslant \infty}\text{Sing}(Y_j)$.
\begin{proposition}\label{construction of projection maps}
There exists $\delta_0>0$, a compact set $K\subset \mathbb C^N$ which contains $\mathfrak S$, and for each $j\in \mathbb Z_{\geqslant 1}\cup\{\infty\}$ a diffeomorphism $p_j:X'\setminus K\rightarrow Y_j\setminus K$ such that for all $k\geqslant 0$
\begin{equation*}
|\nabla_{\omega_\xi}^k (p_j-\Id)|_{\omega_\xi}\leqslant C_kr_\xi^{-\delta_0-k},
\end{equation*}
where we view both $p_j$ and $\Id$ as maps into $\mathbb{C}^N$, and \begin{equation*} \label{e: complex strucutre close}
\left|\nabla^k_{\omega_\xi}\left(p_j^*(J_{Y_j})-J_{X'}\right)\right|_{\omega_\xi}\leqslant C_kr_\xi^{-\delta_0-k}. \end{equation*}
\end{proposition}
\begin{proof}
By definition we can find finitely many holomorphic polynomials $F_1, \cdots, F_m$ on $\mathbb{C}^N$ which generate the ideal of $X'$, such that $G_1, \cdots, G_m$ generate the ideal of $W$. Here $G_\alpha$ is the leading term of $F_\alpha$ with respect to the $\Lambda$ action. Then we know that for all $l\in \mathbb N$ we have $F_\alpha(\Lambda^{-l}. z)=\lambda^{-l c_\alpha} (G_\alpha(z)+\lambda^{l \delta_\alpha}E_{\alpha l}(z))$ for some $\delta_{\alpha}, c_\alpha>0$ and some polynomial $E_{\alpha l}$ whose coefficients are uniformly bounded independent of $l$. Denote $F_{\alpha, l}= (\Lambda^{-l})^*F_\alpha$, then $F_{\alpha, 1}\cdots, F_{\alpha, l}$ generate the ideal of $Y_l$. Let $\delta_0=\min_\alpha \delta_\alpha. $
We define $p_\infty$ to be the normal projection map from $X'$ to $W$, i.e., for $x\in X'$, we let $p_\infty(x)$ to the unique point in $W$ that is closest to $X'$ with respect to the metric $\omega_\xi$. We need to show this is well-defined outside a compact set in $X$ and satisfies the desired properties.
For $l\geqslant 1$ we denote by $A_l$ the annulus in $\mathbb{C}^N$ defined by $\lambda^{l+1}\leqslant r_\xi\leqslant \lambda^{l-1}$. Then $\Lambda^l$ maps $X'\cap A_l$ to $Y_l\cap A_1$. By the conical nature of $\omega_\xi$, it suffices to consider the normal projection from $Y_l\cap A_1$ to $W$ for $l\gg1$.
\begin{claim}
there is a covering of $W\cap A_1$ by finitely many open sets of the form $U_\gamma\cap W$ where $U_\gamma\subset\mathbb{C}^N$ such that for $l$ large there are holomorphic functions $h_{\gamma,l}: U_\gamma\cap W\rightarrow\mathbb{C}^N$ with $|\nabla^k_{\omega_\xi}(h_{\gamma, l}-\Id)|\leqslant C_k \lambda^{l\delta_0}$ for all $k\geqslant 0$ and $Y_l\cap A_1\subset \cup_\gamma \text{Im}(h_{\gamma,l})$.
\end{claim}
To prove the \textbf{Claim}, we fix a point $w\in W\cap A_1$. Then we can find a neighborhood $U\subset \mathbb{C}^N$, such that $W\cap U$ is given by the zero set of $N-n$ number of $G_\alpha$'s (in the scheme-theoretic sense). For simplicity of notation we may assume these are $G_1, \cdots G_{N-n}$. Shrinking $U$ if necessary we may find local holomorphic coordinates $\{\zeta_1, \cdots, \zeta_N\}$ such that $\zeta_\alpha=G_\alpha$ for $\alpha=1, \cdots N-n$. Now we have $F_{\alpha, l}=\lambda^{-lc_\alpha}(G_\alpha(z)+\lambda^{l\delta_\alpha} E_{\alpha l}(z))$. It follows that for $l$ large the common zero set of $F_{\alpha, l}(\alpha=1, \cdots, N-n)$ is a smooth complex submanifold in $U$, so in particular it agrees with $Y_l\cap U$. Using the local coordinates $\{\zeta_1, \cdots, \zeta_N\}$ it is easy to see that $Y_l\cap U$ is contained in the image of a holomorphic function $h_l:U\rightarrow \mathbb{C}^N$ such that $|\nabla^k_{\omega_\xi}(h_l-\Id)|_{\omega_\xi}\leqslant C_k\lambda^{l\delta_0}$ for all $k\geqslant 0$. Since $A_1$ is compact the \textbf{Claim} follows.
\
Since the normal injectivity radius of $W\cap A_1$ is uniformly bounded, there is a tubular neighborhood $\mathcal N$ of $W\cap A_1$ such that the normal projection map $\Pi: \mathcal N\rightarrow W$ is smooth. It follows from the \textbf{Claim} that for $l$ large, $Y_l\cap A_1\subset \mathcal N$. So $p_\infty$ is well-defined and smooth outside a compact set in $X'$. It is straightforward to check that it satisfies the desired derivative bounds. The estimates on the complex structures follow from the fact that $h_{\gamma, l}$ is holomorphic.
Now for each $j<\infty$ we can perform the above arguments uniformly (the closeness between $Y_j$ and $W$ gets improved as $j$ gets larger) to construct the corresponding projection maps $q_j: Y_j\setminus K\rightarrow W$ with uniform estimates. Then we let $p_j=q_j^{-1}\circ p_\infty$.
\end{proof}
Now we fix $\epsilon_0\ll \delta_0$ and apply Theorem \ref{decomposition of kahler forms} with $\epsilon=\epsilon_0$ to get a decomposition
\begin{equation}\label{e:decomposition}
\omega=\beta+\sqrt{-1}\partial\bar\partial\psi.
\end{equation} Let $\psi_j'=\psi\circ p_j^{-1}$ and $r_j=r\circ p_j^{-1}.$ Denote $\gamma_j\equiv\sqrt{-1}\partial\bar\partial\psi_j'$.
By Lemma \ref{l:asymptotic decay}, Proposition \ref{rough estimate}, Corollary \ref{estimates for different covariant derivatives} and Proposition \ref{construction of projection maps}, we have
\begin{corollary}\label{c:metric close}
There exists a compact set $K\subset \mathbb{C}^N$ such that for all $j\in \mathbb Z_{\geqslant1}\cup\{\infty\}$, $\gamma_j$ is a K\"ahler form on $Y_j\setminus K$. Moreover we have
\begin{itemize}
\item for all $k\geqslant 0$ and $\delta>0$,
\begin{equation*}C_\delta^{-1}r_j^{1-\delta}\leqslant r_\xi\leqslant C_\delta r_j^{1+\delta} \text{ and } |\nabla_{\gamma_j}^k(\gamma_j-\omega_\xi)|_{\gamma_j}\leqslant C_{\delta,k} r_j^{\delta-k},
\end{equation*}
\item for all $k\geqslant 0$ and $\delta<\delta_0-\epsilon_0$, \begin{equation} \label{e:pull back metric close}
|\nabla_\omega^k(p_j^*\gamma_j-\omega)|_{\omega}\leqslant C_{\delta, k} r_j^{-\delta-k}.
\end{equation}
\end{itemize}
\end{corollary}
Next we extend $\gamma_j$ to a global K\"ahler form on $Y_j^\circ$, in a uniformly controlled way.
By Proposition \ref{rough estimate} and Proposition \ref{construction of projection maps}, we know that there exist constants $c, T_1>0$ such that
$\ensuremath{\sqrt{-1}}\partial\bar\partial \psi_j'\geqslant cr_\xi^{-\frac{1}{2}}\omega_{\xi}$
on $\{r_{\xi}\geqslant T_1/2\}\cap Y_j$.
Using a cut-off function on $\mathbb{C}^N$ we may extend $\psi_j'$ smoothly to $Y_j^\circ$, which vanishes when $r_{\xi}\leqslant T_1/2$, equals $\psi\circ p_j^{-1}$ when $r_{\xi}\geqslant T_1$ and satisfies
$ \ensuremath{\sqrt{-1}}\partial\bar\partial \psi_j'\geqslant -\frac{A}{2}\omega_{0}
$ on $\{r_{\xi}\leqslant T_1\}\cap Y_j$ for some constant $A>0$ independent of $j$.
\begin{lemma}\label{construction of exact background kahler forms}
There exists a compact set $K\subset \mathbb C^N$ containing a neighborhood of $\mathfrak S$ and
a compactly supported smooth function $\tilde\psi$ on $\mathbb{C}^N$, such that for all $j\in \mathbb Z_{\geqslant1}\cup\{\infty\}$, the form $\omega_j=\sqrt{-1}\partial\bar\partial\psi_j$, where $\psi_j\equiv\psi_j'+\tilde\psi|_{Y_j}$, is a K\"ahler form on $Y_j^{\circ}$ and satisfies $\omega_j=A_0\omega_0$ on $K\cap Y_j$, where $\omega_0$ denotes the K\"ahler form for the standard Euclidean metric on $\mathbb{C}^N$ and $A_0$ is a constant independent of $j$.
\end{lemma}
\begin{proof}
Choose a smooth increasing and convex function $u: \mathbb{R} \rightarrow \mathbb{R}$ such that $u(t)=2 T_1^2 / 3$ for $t \leqslant T_1^2 / 2$ and $u(t)=t$ for $t \geqslant T_1^2$. Define the function $r_{1}^2= u \circ r^2_{\xi}$ on $\mathbb{C}^N$. Then $\ensuremath{\sqrt{-1}}\partial\bar\partial r_1^2 \geqslant 0$ and equals $\omega_{\xi}$ on $\{r_{\xi}\geqslant T_1\}$. Let $r_0$ denote the radius function for the standard Euclidean metric $\omega_0$. Choose a cutoff function $\rho:\mathbb R\rightarrow\mathbb R$ which equals 1 when ${t\leqslant 2T_1^2}$ and equals $0$ when ${t\geqslant 4T_1^2}$. Then it is straightforward to check that there exists a small positive constant $\epsilon$ such that $\phi=\epsilon(\rho \circ r_{\xi}^2)\cdot r_0^2+r_1^2$ satisfies that $\sqrt{-1}\partial\bar\partial\phi$ is a K\"ahler form on $\mathbb{C}^N$. By construction there exists a constant $T>0$ such that $\{r_{\xi}\leqslant 2T_1\}\subseteq \{\phi\leqslant T\}$.
It is easy to find a smooth function $\chi:[0,\infty)\rightarrow\mathbb R$ such that the following hold
\begin{itemize}
\item $\chi(t)=\frac{At}{\epsilon}$ if $t\leqslant T$;
\item $\chi(t)=0$ for $t\gg1$;
\item for all $t>0$ we have $\chi''\geqslant-\frac{c}{8}t^{-\frac{3}{2}}$ and $\chi'\geqslant -\frac{c}{8}t^{-\frac{1}{2}}$.
\end{itemize}Then $\tilde\psi=\chi\circ \phi$ satisfies the desired properties.
\end{proof}
Recall we have the crepant resolution map $\pi: X\rightarrow X'$, thus for $j\in \mathbb Z_{\geqslant 1}$ the composition $\Lambda^j\circ \pi$ is a crepant resolution of $Y_j$. We denote the latter by $\pi_j:\hat Y_j=X\rightarrow Y_j$. For all $j\in \mathbb Z_{\geqslant 1}$ and $\epsilon\in (0,\lambda^{4j})$, we have the following family of K\"ahler forms on $\hat Y_j$
$$\widetilde{\omega}_{j,\epsilon}\equiv \epsilon \beta +\pi_j^*\omega_j=\epsilon\beta+\ensuremath{\sqrt{-1}}\partial\bar\partial \pi_j^*\psi_j,$$ where $\beta$ is given in \eqref{e:decomposition}. Notice $\widetilde{\omega}_{j,\epsilon}$ can also be naturally viewed as a form on $Y_j^\circ$ via $(\pi_j^{-1})^*$. The following lemma shows that they are uniformly asymptotic to the cone $\mathcal C$.
\begin{lemma}\label{decay of beta}
There exists a compact set $K\subset \mathbb C^N$ such that for all $j\in \mathbb Z_{\geqslant 1}$ and $k\geqslant0$ on $Y_j\setminus K$ we have
\begin{equation*}
\lambda^{4j}\left|\nabla^k_{\omega_j}(\pi_j^{-1})^*\beta\right|_{\omega_j}\leqslant C_k r_j^{-\frac{1}{2}-k}.
\end{equation*}
\end{lemma}
\begin{proof}
Outside a compact set, we identify $X$ with $X'$ and therefore we may just view $\pi_j$ as $\Lambda^j$. Let us first prove the case $k=0$. By Corollary \ref{c:metric close}, we know that for any $\epsilon_1,\epsilon_2>0$, there exists constants $C_{\epsilon_1}$ and $C_{\epsilon_2}$ such that
\begin{equation}\label{replacement of omega and omega_xi}
\begin{aligned}
\left|(\Lambda^{-j})^*\beta\right|_{\omega_j}&\leqslant C_{\epsilon_1}\left|(\Lambda^{-j})^*\beta\right|_{\omega_{\xi}} r_{\xi}^{\epsilon_1}\leqslant C_{\epsilon_1}(\Lambda^{-j})^*(\left|\beta\right|_{(\Lambda^j)^*\omega_{\xi}}) r_{\xi}^{\epsilon_1}\\
&\leqslant C_{\epsilon_1}\lambda^{-2j} (\Lambda^{-j})^*(\left|\beta\right|_{\omega_{\xi}})r_{\xi}^{\epsilon_1}\leqslant C_{\epsilon_1}C_{\epsilon_2}\lambda^{-2j} (\Lambda^{-j})^*(\left|\beta\right|_{\omega}r_{\xi}^{\epsilon_2})r_{\xi}^{\epsilon_1} \\
&\leqslant C_{\epsilon_1}C_{\epsilon_2}\lambda^{-2j(1+\epsilon_2)}(\Lambda^{-j})^*(r^{-1}) r_{\xi}^{\epsilon_1+\epsilon_2},
\end{aligned}
\end{equation}where for the last inequality we used the estimate of $\beta$ with respect to the metric $\omega$ obtained in Theorem \ref{decomposition of kahler forms}.
By choosing $\epsilon_1$ and $\epsilon_2$ small enough, one can directly show that there exists a positive constant $C_0$ such that
\begin{equation*}
\lambda^{4j}\left|(\Lambda^{-j})^*\beta\right|_{\omega_j}\leqslant C_0 r_j^{-\frac{1}{2}}.
\end{equation*}
This proves the case $k=0$. Note that suppose we have two Riemannian metrics $g_1$ and $g_2$ then for any tensor $\eta$ and another fixed background metric $g_0$, we have
\begin{equation}\label{covariant derivative with respect to different metrics}
\left|\nabla_{g_1}^k\eta\right|_{g_0}\leqslant C_k \sum_{\substack{k_1+k_2+k_3=k\\k_i\geqslant0, k_1\leqslant
k-1}} \left|\nabla_{g_2}^{k_1}(\nabla_{g_1}-\nabla_{g_2})\right|_{g_0}\left|\nabla_{g_1}-\nabla_{g_2}\right|_{g_0}^{k_2}\left|\nabla_{g_2}^{k_3}\eta\right|_{g_0}.
\end{equation}
Then the proof for $k\geqslant 1$ follows from the same argument as we did in \eqref{replacement of omega and omega_xi} by using \eqref{covariant derivative with respect to different metrics} together with Corollary \ref{estimates for different covariant derivatives} and Theorem \ref{decomposition of kahler forms}.
\end{proof}
Recall that $\Omega_X$ is the given holomorphic volume form on $X$. We can view it naturally as a holomorphic volume form on $X'=Y_1$ (notice the canonical bundle of each $Y_j$ is trivial).
\begin{lemma}\label{construction of holomorphic volume form}
For all $j\in \mathbb Z_{\geqslant 1}\cup\{\infty\}$, there exists a holomorphic volume form $\Omega_j$ on $Y_j$ satisfying the following two properties:
\begin{itemize}
\item [(1)]there exists a constant $\delta_1>0$ and a compact set $K\subset \mathbb{C}^N$ such that on $Y_j\setminus K$ the function $f_j=\log\left(\frac{p_j^*(\Omega_j\wedge\ols{\Omega}_j)}{\Omega_{X}\wedge\ols{\Omega}_{X}}\right)$ satisfies the following for all $k\geqslant0$ and all $j$,
\begin{equation*}
|\nabla^k_{\omega_j}f_j|\leqslant C_k r^{-\delta_1-k}.
\end{equation*}
\item[(2)] for every compact set $K_1\subset \mathbb C^N$ we have
\begin{equation*}
\sup_{j\in \mathbb Z_{\geqslant1}\cup\{\infty\}}\int_{K_1\cap Y_j}(\ensuremath{\sqrt{-1}})^{n^2}\Omega_j\wedge\ols{\Omega}_j<\infty.
\end{equation*}
\end{itemize}
\end{lemma}
\begin{remark}
It follows from the proof that for $j<\infty$, $\Omega_j$ coincides with $(\Lambda^{-j})^*\Omega_X$ up to a nonzero multiplicative constant depending on $j$. For our purpose we need the precise asymptotics of $\Omega_j$ and the stated properties of $f_j$ essentially determine this constant.
\end{remark}
\begin{proof}
Let $\varphi =\log(r_{\xi}^2+1)-\frac{1}{100}(\log(r_{\xi}^2+3))^{\frac{1}{2}}$ and denote $\varphi_j=\varphi|_{Y_j}$. Then by Lemma \ref{psh function on a kahler cone}-(1) and Corollary \ref{c:metric close}, we know that outside a compact set $K$ for every $\epsilon>0$, there is a positive constant $c_{\epsilon}$ such that
$
\ensuremath{\sqrt{-1}}\partial\bar\partial \varphi_j\geqslant c_{\epsilon}r_j^{-2-\epsilon}\omega_j.
$
Let $p_j$ be given in Proposition \ref{construction of projection maps}. For all $\delta\in (0, \delta_0-\epsilon_0)$ we have
\begin{equation} \label{e:almost constant}
\left|\left((p_j^{-1})^*\Omega_X\right)^{n,0}\right|_{\omega_j}-1=O(r_j^{-\delta}),
\end{equation}
and for all $k\geqslant 1$,
$
\left|\nabla^k_{\omega_j}\left((p_j^{-1})^*\Omega_X\right)^{n,0}\right|_{\omega_j}\leqslant C_{\delta, k} r_j^{-\delta-k}.
$
Fix $T$ large so that $K\subset \{r_\xi\leqslant T\}$, and fix a cut off function $\chi$ on $\mathbb{C}^N$ which equals 0 when $r_\xi\leqslant T$, equals 1 when $r_\xi>2T$. Since $Y_j$ is affine and $ \text{Sing}(Y_j)$ consists of finitely many points, we know the smooth locus $Y_j\setminus \text{Sing}(Y_j)$ admits a complete K\"ahler metric. Then for any $\tau>0$ small, we can apply Theorem \ref{Hormander} on $M=Y_j\setminus \text{Sing}(Y_j)$, with $\omega_M=\omega_j$ and $\varphi_M=(n-1+\tau)\varphi_j$, to obtain an $(n,0)$ form $v_j$ satisfying
$\label{partial bar equation 2}
\bar\partial v_j=\bar\partial \left(\chi\cdot ((p_j^{-1})^*\Omega_X)^{n,0}\right),
$ such that
$$\int_{Y_j\setminus \text{Sing}(Y_j)}e^{-(n-1+\tau)\varphi_j}(\ensuremath{\sqrt{-1}})^{n^2}v_j\wedge \ols{v}_j\leqslant C. $$
By Proposition \ref{c:elliptic regularity}, when $r_\xi\geqslant 2T$, we have $|\nabla^k_{\omega_j}v_j|\leqslant C_k r_j^{-\delta_1-k}$ for some constant $\delta_1>0$ and all $k\geqslant0$.
Now let $\Omega_j=\chi ((p_j^{-1})^*\Omega_X)^{n,0}-v_j$. It is holomorphic on $Y_j\setminus \text{Sing}(Y_j)$ and is non-zero by \eqref{e:almost constant} and the decay of $v_j$. Since $Y_j$ is normal it defines a holomorphic section of $K_{Y_j}$. Moreover the function $\frac{(\Lambda^{j})^*\Omega_j}{\Omega_X}$ is a holomorphic function on $X$ which is $O(r^{\epsilon})$ for all $\epsilon>0$, so must be a nonzero constant. In particular, $\Omega_j$ is a holomorphic volume form on $Y_j$. The desired estimates follow from straightforward computation.
\end{proof}
\begin{proposition}\label{preconditioning}
There exist a $\delta_2>0$, a compact set $K\subset \mathbb{C}^N$ containing a neighborhood of $\mathfrak S$ and for all $j\in\mathbb Z_{\geqslant 1}$ and $\epsilon\in (0, \lambda^{4j})$ a smooth function $u_{j,\epsilon}$ on $Y_j\setminus \text{Sing}(Y_j)$ which is constant on $K$ and satisfies $\widetilde\omega_{j,\epsilon}+\sqrt{-1}\partial\bar\partial u_{j,\epsilon}>0$ on $Y_j^\circ$, such that on $Y_j^\circ$,
\begin{equation*}
\left(\widetilde\omega_{j,\epsilon}+\ensuremath{\sqrt{-1}}\partial\bar\partial u_{j,\epsilon}\right)^n=e^{-f_{j,\epsilon}}(\ensuremath{\sqrt{-1}})^{n^2}\Omega_j\wedge\ols{\Omega}_j,
\end{equation*}
and on $Y_j\setminus K$ we have for all $k\geq0$, $$
\left|\nabla^k_{\widetilde\omega_{j,\epsilon}}u_{j,\epsilon}\right|_{\widetilde\omega_{j,\epsilon}}\leqslant C_k r_j^{-\delta_2+2-k}, \ \ \ \ \ \ \left |\nabla^k_{\widetilde\omega_{j,\epsilon}}f_{j,\epsilon}\right|_{\widetilde\omega_{j,\epsilon}}\leqslant C_k r_j^{-\delta_2-2-k}.$$
\end{proposition}
\begin{proof}
We first identify the end of $X'$ differentiably with $\mathcal Q_{R_0}$ via Lemma \ref{l:asymptotic decay}. Then using the maps $p_j$ constructed in Proposition \ref{construction of projection maps} we may identify the end of all $Y_j$ with $\mathcal Q_{R_2}$ for some $R_2>R_1+R_0$, where $R_1$ is given in Proposition \ref{p:Poisson equation general}. In particular we may view the metrics $\widetilde g_{j,\epsilon}$ (associated to $\widetilde\omega_{j,\epsilon}$) on $Y_j^\circ$ as a family of Riemannian metrics on $\mathcal Q_{R_2}$. By Lemma \ref{l:asymptotic decay}, Corollary \ref{c:metric close} and Lemma \ref{decay of beta} we know that for all $j$, $s^l\sup_{\partial\mathcal Q_s}|\nabla^l_{g_\mathcal C}(\widetilde g_{j,\epsilon}-g_\mathcal C)|_{g_\mathcal C}\leqslant e_l(s)$ for some $e_l(s)$ with $\lim_{s\rightarrow\infty} e_l(s)=0$. Then we can apply Proposition \ref{p:Poisson equation general} to solve the Poisson equation with uniform estimates for the metric $\widetilde g_{j, \epsilon}$ on $\mathcal Q_{R_2}$. More precisely, we solve the equation in a fixed H\"older space for $k$ large and then apply elliptic regularity as Proposition \ref{c:elliptic regularity} to gain the decay for all derivatives. Now the conclusion follows from the argument of Conlon-Hein in \cite[Section 2.4]{CH1}.
\end{proof}
For all $j\in \mathbb Z_{\geqslant 1}$ and $\epsilon\in (0,\lambda^{4j})$, we define $$\omega_{j, \epsilon}\equiv\epsilon(\pi_j^{-1})^*\beta+\ensuremath{\sqrt{-1}}\partial\bar\partial (\psi_j+u_{j, \epsilon}),$$ and we denote by $g_{j, \epsilon}$ and $J_{j, \epsilon}$ the corresponding Riemannian metric and complex structure. It follows from Lemma \ref{l:asymptotic decay}, Proposition \ref{construction of projection maps}, Corollary \ref{c:metric close}, Lemma \ref{decay of beta} and Proposition \ref{preconditioning} that
\begin{proposition}\label{close to asymptotic cone C}
There exist a compact set $K\subset \mathbb C^N$ and a large positive number $R_2$ such that for all $j\in \mathbb Z_{\geqslant 1}$, the diffeomorphism $P_j=p_j\circ \Phi:\mathcal Q_{R_2}\rightarrow Y_j\setminus K$ satisfies that for all $l\geqslant 0$,
\begin{equation*}
\lim_{s\rightarrow\infty}\sup_{j\geqslant 1}\sup_{\partial \mathcal Q_s}s^l\left (|\nabla^l_{g_\mathcal C}(P_j^*g_{j, \epsilon}-g_{\mathcal C})|_{g_\mathcal C}+|\nabla^l_{g_\mathcal C}(P_j^*J_{j, \epsilon}-J_{\mathcal C})|_{g_\mathcal C}\right)=0.
\end{equation*}
\end{proposition}
\section{Uniform estimates for the complex Monge-Amp\`ere equation}\label{s5}
From now on for $j\in \mathbb Z_{\geqslant 1}$ we will abuse notations and identify $\Omega_j$ and $r_j$ with their pull-backs under $\pi_j:\hat Y_j\rightarrow Y_j$. Denote $\mu_j=(\ensuremath{\sqrt{-1}})^{n^2}\Omega_j\wedge\ols{\Omega}_j$. We now look for Calabi-Yau metrics on $\hat Y_j$ which are asymptotic to $\omega_{j, \epsilon}$. To do this we need to solve on $\hat Y_j$ the equation
\begin{equation}\label{complex monge-ampere equation}
\begin{aligned}
\left(\omega_{j, \epsilon} +\ensuremath{\sqrt{-1}}\partial\bar\partial \varphi_{j, \epsilon}\right)^n=\mu_j=e^{f_{j, \epsilon}}\omega_{j, \epsilon}^n.
\end{aligned}
\end{equation}
By Proposition \ref{preconditioning},
there exists a compact set $K\subset \mathbb C^N$ such that on $\hat Y_j\setminus \pi_j^{-1}(K\cap Y_j)$ for all $k\geqslant0$,
\begin{equation}\label{decaying of Ricci potential}
|\nabla^k_{\omega_{j, \epsilon}}f_{j, \epsilon}|\leqslant C_k r_{j}^{-\delta_2-2-k}.
\end{equation}
The following theorem can be derived from Tian-Yau \cite[Section 4]{TY2}. See also Hein \cite[Section 4]{hein1}.
\begin{theorem}[{\cite{TY2,hein1}}]\label{ existence result of Tian-Yau}
For fixed $j$ and $\epsilon$, (\ref{complex monge-ampere equation}) admits a smooth solution $\varphi_{j, \epsilon}$ on $ Y_j$ such that
\begin{itemize}
\item [(1)] there exists a constant $p_0$ depending only on $\delta_2$ such that $\Vert\varphi_{j, \epsilon}\Vert_{L^{p_0}(\mu_j)}<\infty$ and $\varphi_{j, \epsilon}(x)\rightarrow 0$ as $ r_j(x)\rightarrow \infty$.
\item [(2)] $\left\Vert\nabla^k_{\omega_{j, \epsilon}}\varphi_{j, \epsilon}\right\Vert_{L^{\infty}} <\infty$ for all $k\geqslant0$.
\end{itemize}
\end{theorem}
\begin{remark}
The proof of the $L^{\infty}$ bound of $\varphi_{j, \epsilon}$ is via the Moser iteration, thus indeed we know that $\left \Vert\varphi_{j, \epsilon}\right\Vert_{L^p(\mu_j)}<\infty$ for all $p\gg1$. In the following, we fix a large $p_0$ such that $q_0=\frac{n(p_0+1)}{n+p_0}$ satisfies
\begin{equation}\label{assumption on index q_0}
(2+\delta_2)q_0>2n.
\end{equation}
\end{remark}
Since the background metrics $\omega_{j,\epsilon}$ are asymptotic to the cone metric on $\mathcal{C}$, by the method in \cite[Section 5]{TY2} and using ${r_j}^{-2\delta}$ as a barrier to apply the maximum principle, we can prove the following decay estimates for $\varphi_{j,\epsilon}$. We remark that alternatively, one can also use the method of \cite{CH1, CGT} to solve the complex Monge-Amp\`ere equation in a suitable weighted space (since we have Lemma \ref{p:weighted-esimate-quotient-space}), then the decay property will be a direct by-product.
\begin{lemma}\label{non uniform estimate}
There exist compact sets $K_{j,\epsilon}\subset \mathbb C^N$ depending only on $\left\Vert\varphi_{j,\epsilon}\right\Vert_{L^{\infty}}$ and constants $Q_{k, j, \epsilon}$ depending only on $k$ and $\left\Vert\varphi_{j,\epsilon}\right\Vert_{L^{\infty}}$ such that on $Y_j\setminus K_{j,\epsilon}$ for some $\delta_3>0$ and all $k\geqslant0$,
\begin{equation*}
|\nabla^k_{\omega_{j, \epsilon}}\varphi_{j, \epsilon}|\leqslant Q_{k, j, \epsilon} { r_j}^{-\delta_3-k}.
\end{equation*}
\end{lemma}
\begin{proof}
We focus on the $C^0$ decay property since the higher order estimates follow from standard rescaling argument and elliptic estimates (see \cite[Proposition 5.1]{TY2}). Fix a positive constant $\delta< \min\left\{\frac{\delta_2}{2},\frac{1}{2}\right\}$.
\begin{claim}
there is a large constant $R_*$ such that for every $R\geqslant R_*$, on the region $\{r_j\geqslant R\}$ we have
\begin{equation*}
\left(\omega_{j, \epsilon}+R\ensuremath{\sqrt{-1}}\partial\bar\partial r_{j}^{-2\delta}\right)^n< e^{f_{j, \epsilon}}\omega_{j, \epsilon}^n\text{ and } \left(\omega_{j, \epsilon}-R\ensuremath{\sqrt{-1}}\partial\bar\partial r_{j}^{-2\delta}\right)^n> e^{f_{j, \epsilon}}\omega_{j, \epsilon}^n.
\end{equation*}
\end{claim}
The decay of $\varphi_{j,\epsilon}$ follows from this claim by applying maximal principle to the region $\{r_j\geqslant R_{j,\epsilon}\}$ with
$
R_{j,\epsilon}= \max\{R_*, \left\Vert\varphi_{j,\epsilon}\right\Vert_{L^{\infty}}^{\frac{1}{1-2\delta}}\}.
$
Let us prove the \textbf{Claim}. We just prove the first inequality since the second inequality follows from the same argument.
By \eqref{decaying of Ricci potential} and Proposition \ref{close to asymptotic cone C}, we know that outside a compact set $K_0$, there exists a constant $C_0>0$ such that
\begin{equation}\label{supersubharmonic}
\begin{aligned}
1-C_0r^{-2-\delta_2}\leqslant e^{f_{j,\epsilon}}&\leqslant 1+C_0r^{-2-\delta_2},\\
\left|\ensuremath{\sqrt{-1}}\partial\bar\partial r_j^{-2\delta}\right|_{\omega_{j,\epsilon}}&\leqslant C_0r_j^{-2(\delta+1)},\\
\Delta_{\omega_{j,\epsilon}}r_j^{-2\delta}&\leqslant -\delta r_j^{-2(\delta+1)}.
\end{aligned}
\end{equation} The last inequality here is due to the fact that on the cone $\mathcal C$, for the radial function $r$ we have $$\Delta_{\omega_{\mathcal C}}r^{-2\delta}=4\delta(-n+\delta+1)r^{-2(\delta+1)}.$$
Choose $R_*$ large such that $K_0\subseteq \{r_j\leqslant R_*\}$ and by the second inequality in \eqref{supersubharmonic} and the choice of $\delta$ we may also assume that for all $R\geqslant R_*$, on $\{r_j\geqslant R\}$ we have
\begin{equation}\label{Laplacian dominates}
\frac{\left(\omega_{j, \epsilon}+R\ensuremath{\sqrt{-1}}\partial\bar\partial r_{j}^{-2\delta}\right)^n}{\omega_{j,\epsilon}^n}\leqslant 1+R\Delta_{\omega_{j,\epsilon}}r_j^{-2\delta}+\frac{\delta}{2}r_j^{-2(\delta+1)},
\end{equation}
and
\begin{equation}\label{choice of delta}
(-R\delta+\frac{\delta}{2})r_j^{-2(\delta+1)}\leqslant -2C_0r_j^{-2-\delta_2}.
\end{equation}
Then the claim follows from \eqref{supersubharmonic}, \eqref{Laplacian dominates} and \eqref{choice of delta}.
\end{proof}
Denote $$\hat\omega_{j, \epsilon}=\omega_{j, \epsilon}+\ensuremath{\sqrt{-1}}\partial\bar\partial \varphi_{j, \epsilon}.$$
Below we will derive uniform estimates on $\varphi_{j, \epsilon}$. Then we will take a double limit, first as $\epsilon\rightarrow0$ and then as $j\rightarrow\infty$, to obtain a Calabi-Yau metric on $W\setminus\{0\}$. The following arguments are essentially standard, given the large body of literature in this field; see for example the recent work by Collins-Guo-Tong \cite{CGT} which deals with the case when the background metric is asymptotically conical in the stronger sense (with polynomial decay rate). For our purpose we will not need uniform estimates for all $j$ and $\epsilon\in (0, \lambda^{4j})$. Instead we first fix $j$ and derive uniform estimates as $\epsilon\rightarrow 0$, which allows us to take a limit $\varphi_j$. Then we will derive uniform estimates for $\varphi_j$ as $j\rightarrow\infty$. Notice we are free to pass to subsequences when taking limits and we do not need uniqueness of the limits.
Let $$A_{j, \epsilon}=\sup_{q\geqslant q_0} \Vert e^{-f_{j, \epsilon}}-1\Vert_{L^q(\mu_j)}, \ \ \ \ A_{j}=\sup_{q\geqslant q_0} \Vert e^{-f_j}-1\Vert_{L^q(\mu_j)}. $$ By \eqref{decaying of Ricci potential} and \eqref{assumption on index q_0} we know for each $j,\epsilon$, $A_{j, \epsilon}<\infty$, and for each fixed $j$, as $\epsilon\rightarrow0$ we have $A_{j, \epsilon}\rightarrow A_j$.
\begin{lemma}\label{l:Aj uniform bound}
There exists a constant $C>0$ such that $A_j\leqslant C$ for all $j$.
\end{lemma}
\begin{proof}
By (\ref{decaying of Ricci potential}) and Lemma \ref{construction of holomorphic volume form}, it suffices to prove a uniform $C^0$ bound on $e^{-f_j}$ over $K\cap Y_j$ for a fixed large compact set $K\subset \mathbb{C}^N$. By Lemma \ref{construction of exact background kahler forms} and Lemma \ref{preconditioning}, there is a constant $C$ independent of $j$ such that
$\omega_j\leqslant C \omega_0$ on $Y_j\cap K$, where $\omega_0$ denotes the K\"ahler form for the Euclidean metric on $\mathbb C^N$. By the Gauss-Codazzi equation we know the Ricci curvature of $\omega_0|_{Y_j}$ is non-positive, it follows that the function $ \frac{\omega_0^n}{(\ensuremath{\sqrt{-1}})^{n^2}{\Omega_j}\wedge\ols{\Omega}_j}$ is plurisubharmonic on $Y_j\setminus \text{Sing}(Y_j)$. Pulling back to $\hat Y_j$ it is a smooth plurisubharmonic function. Then the conclusion follows from the maximal principle and Lemma \ref{construction of holomorphic volume form}.
\end{proof}
As observed by \cite{CGT} (following an earlier idea of Tosatti) it is more convenient to use the Calabi-Yau metrics $\hat\omega_{j, \epsilon}$ as background metrics when we do Moser iteration. By Lemma \ref{non uniform estimate} each $\hat\omega_{j, \epsilon}$ is weakly asymptotic to the cone $\mathcal C$, so by Bishop-Gromov inequality we know the volume of a metric ball of radius $R $ in $(\hat Y_j, \hat \omega_{j, \epsilon})$ is bounded below by $\kappa(\mathcal C) R^{2n}$, where $\kappa(\mathcal C)$ is the volume density of $\mathcal C$. So we have a uniform Sobolev inequality, namely, there exists $S>0$ independent of $j$ and $\epsilon$ such that for all $v\in C^\infty_0(\hat Y_j)$,
$$\|v\|_{L^{\frac{2n}{n-1}}(\hat Y_j, \hat \omega_{j, \epsilon})}\leqslant S \|\nabla_{\hat\omega_{j, \epsilon}}v\|_{L^2(\hat Y_j, \hat\omega_{j, \epsilon})}.$$
\begin{lemma}\label{C^0 bound}
There exists a continuous function $\mathcal P: (0, \infty)\rightarrow (0, \infty)$ such that for all $j\in \mathbb Z_{\geqslant 1}$ and $\epsilon\in (0,\lambda^{4j})$,
\begin{itemize}
\item [(1).] $\left\Vert\varphi_{j, \epsilon}\right\Vert_{L^{\infty}}\leqslant \mathcal P(A_{j, \epsilon}).$
\item [(2).] $
\mathcal P(A_{j, \epsilon})^{-1}\omega_j\leqslant \hat\omega_{j, \epsilon}\leqslant \mathcal P(A_{j, \epsilon})\frac{\mu_j}{\omega_j^n}\omega_j.$
\end{itemize}
\end{lemma}
\begin{proof} The proof is standard, see for example \cite[Section 4]{CGT}. For completeness, we include some details.
(1). We have
$\hat\omega^{n}_{j, \epsilon}-\omega_{j, \epsilon}^{n}= d d^{c} \varphi_{j, \epsilon} \wedge \sum_{k=0}^{n-1}\left(\hat\omega^{k}_{j, \epsilon} \wedge \omega_{j, \epsilon}^{n-1-k}\right).$
For any exponent $p\geqslant p_0$, multiplying both sides by $|\varphi_{j, \epsilon}|^{p-1}\varphi_{j, \epsilon}$ and integrating by parts (the estimates obtained in Theorem \ref{ existence result of Tian-Yau} justify this), we get
\begin{equation*}
\int_{X}|\varphi_{j, \epsilon}|^{p-1}\varphi_{j, \epsilon}(e^{-f_{j, \epsilon}}-1)\mu_j= \frac{4p}{(p+1)^2}\int_{X} \ensuremath{\sqrt{-1}}\partial |\varphi_{j, \epsilon}|^{\frac{p+1}{2}}\wedge \bar\partial |\varphi_{j, \epsilon}|^{\frac{p+1}{2}}\wedge \sum_{k=0}^{n-1}\left(\hat\omega^{k}_{j, \epsilon} \wedge \omega_{j, \epsilon}^{n-1-k}\right).
\end{equation*}
In the following $C$ always denote a constant depending only on $n$ and $p_0$. By Sobolev inequality, we obtain
\begin{equation*}
\int_{X}|\varphi_{j, \epsilon}|^{p-1}\varphi_{j, \epsilon}(e^{-f_{j, \epsilon}}-1)\mu_j\geqslant C\frac{S p}{(p+1)^2}\left(\int_{X}|\varphi_{j, \epsilon}|^{(p+1)\frac{n}{n-1}}\mu_j\right)^{\frac{n-1}{n}}.
\end{equation*}
Applying the H\"older's inequality on the left hand side, we have
\begin{itemize}
\item []\begin{equation*}
\Vert\varphi_{j, \epsilon}\Vert_{L^{\frac{n}{n-1}(p_0+1)}(\mu_j)}\leqslant C S \Vert e^{-f_{j, \epsilon}}-1\Vert_{L^{q_0}(\mu_j)},
\end{equation*}
\item[]\begin{equation*}
\Vert\varphi_{j, \epsilon}\Vert^{p+1}_{L^{\frac{n}{n-1}(p+1)}(\mu_j)}\leqslant C \frac{p}{(p+1)^2}
S \Vert e^{-f_{j, \epsilon}}-1\Vert_{L^{p+1}(\mu_j)}\Vert\varphi_{j, \epsilon}\Vert^p_{L^{p+1}(\mu_j)} \text{ for all $p\geqslant p_0$}.
\end{equation*}
\end{itemize}Then a standard Moser iteration argument completes the proof.
(2). Note that the second inequality follows from the first one and the elementary inequality that $\operatorname{tr}_{\beta}(\alpha)\leqslant \frac{\alpha^n}{\beta^n}\left(\operatorname{tr}_{\alpha}(\beta)\right)^{n-1}$ for any two positive (1,1)-forms $\alpha$ and $\beta$. The first inequality here is a consequence of Chern-Lu's inequality. More precisely, by construction the bisectional curvature of $\omega_j$ has a uniform upper bound $B$ independent of $j$. Then Chern-Lu's inequality implies that
\begin{equation*}
\Delta_{\hat\omega_{j, \epsilon}}\log(\operatorname{tr}_{\hat\omega_{j, \epsilon}}\omega_j)\geqslant -B \operatorname{tr}_{\hat\omega_{j, \epsilon}}\omega_j.
\end{equation*}
By the definition of $\hat\omega_{j, \epsilon}$, we have
$
\operatorname{tr}_{\hat\omega_{j, \epsilon}}(\ensuremath{\sqrt{-1}}\partial\bar\partial \varphi_{j, \epsilon})\leqslant n-\operatorname{tr}_{\hat\omega_{j, \epsilon}}\omega_j.
$
Therefore,
\begin{equation*}
\Delta_{\hat\omega_{j, \epsilon}}\left(\log(\operatorname{tr}_{\hat\omega_{j, \epsilon}}\omega_j)-(B+1)\varphi_{j, \epsilon}\right) \geqslant \operatorname{tr}_{\hat\omega_{j, \epsilon}}\omega_j-(B+1)n.
\end{equation*} Now we apply the maximum principle to the function $F=\log(\operatorname{tr}_{\hat\omega_{j, \epsilon}}\omega_j)-(B+1)\varphi_{j, \epsilon}$. Since $\varphi_{j, \epsilon}$ tends to zero at infinity, $F$ converges to $\log n$ at infinity. Notice $\pi_j^*\omega_j$ is degenerate along the exceptional set, so $F$ tends to $-\infty$ on $\pi_j^{-1}(\text{Sing}(Y_j))$. If the supremum of $F$ is attained at infinity, then we automatically have a uniform upper bound for $\log (\operatorname{tr}_{\hat\omega_{j, \epsilon}}(\omega_j))$. If the maximum is achieved on $\hat Y_j$, Then $\log (\operatorname{tr}_{\hat\omega_{j, \epsilon}}(\omega_j))\leqslant C$, where $C$ depends only on $\left\Vert\varphi_{j, \epsilon}\right\Vert_{L^{\infty}}$.
\end{proof}
Using this we can also improve Lemma \ref{non uniform estimate}, namely, we can make the dependence of $K_{j,\epsilon}$ and $Q_{k, j, \epsilon}$ on $j$ and $\epsilon$ to be through $A_{j, \epsilon}.$
For fixed $j$ since as $\epsilon\rightarrow 0$ we have $A_{j, \epsilon}\rightarrow A_j$, passing to a subsequence $\epsilon_{j,l}$ we can make $\varphi_{j, \epsilon_{j, l}}$ converge smoothly to a limit $\varphi_j$ (as $l\rightarrow\infty$) on $Y_j^\circ$. The metric $\hat\omega_j=\omega_j+\sqrt{-1}\partial\bar\partial\varphi_j$ defines a Calabi-Yau metric on $Y_j\setminus \text{Sing}(Y_j).$ Moreover, we have
\begin{itemize}
\item On $Y_j^\circ$, \begin{equation}\label{e:phij uniform bound}|\varphi_{j}|\leqslant \mathcal P(A_j).
\end{equation}This implies $\hat\omega_j$ extends as a closed positive $(1,1)$ current on $Y_j$.
\item On $Y_j^{\circ}$, \begin{equation}\label{e:lower bound}\hat\omega_{j}\geqslant \mathcal P(A_j)^{-1}\omega_j.
\end{equation}
\item On $Y_j\setminus K$, for all $k\geqslant0$, $|\nabla^k_{\omega_{j}}\varphi_{j}|\leqslant \mathcal Q(A_j) r_j^{-\delta_3-k}.
$
\end{itemize}
Moreover, standard argument using the Bishop-Gromov volume comparison (c.f. \cite[Lemma 4.14]{CGT}) implies that for any compact set $K\subset\mathbb{C}^N$, there is constant $C_K>0$ independent of $j$ and $\epsilon$ such that \begin{equation*}
\diam (\pi_j^{-1}(K\cap Y_j),\hat\omega_{j, \epsilon})\leqslant C_K.
\end{equation*}
So the diameter of $K\cap Y_j^\circ$ with respect to $\hat\omega_j$ is also uniformly bounded independent of $j$.
By construction we know as $j\rightarrow\infty$, $\omega_j$ converges smoothly to $\omega_\infty=\sqrt{-1}\partial\bar\partial (\psi_\infty+u_\infty)$ on $W\setminus \{0\}$, and $\psi_\infty$ and $u_\infty$ are continuous across $0$. Using the above estimates by Lemma \ref{l:Aj uniform bound} by passing to a subsequence we may assume that $\varphi_j$ converges smoothly to a smooth function $\varphi_\infty$ on $W\setminus \{0\}$ which is uniformly bounded, such that $\hat\omega_\infty=\omega_\infty+\sqrt{-1}\partial\bar\partial\varphi_\infty$ is a Calabi-Yau metric on $W\setminus\{0\}$ with
$\hat\omega_\infty\geqslant C^{-1}\omega_\infty$ and $|\nabla_{\omega_\infty}^k\varphi_\infty|\leqslant C r_\infty^{-\delta_3-k}$ on $W\setminus K$, where we recall $r_\infty=r\circ p_\infty^{-1}$. Moreover the metric has finite diameter on $K\cap W$, and it is asymptotic to the cone $\mathcal C$ at infinity.
\section{Proof of the main results}\label{s6}
We now investigate the singular behavior of $\hat\omega_\infty$ at $0$. For this purpose we need to study the singular behavior of $\hat\omega_j$, so we first fix $j$ and consider the family $\hat\omega_{j, l}\equiv \hat\omega_{j, \epsilon_{j,l}}$ as $l\rightarrow\infty$.
Passing to a further subsequence we let $({Z_j},d_{{Z_j}})$ be a pointed Gromov-Hausdorff limit of $(\hat Y_j, \hat\omega_{j, l}, p_j)$ (where $p_j$ is a point on the exceptional set of $\pi_j)$ as $l\rightarrow\infty$. As mentioned before, the metrics $\hat\omega_{j, \epsilon}$ are uniformly volume non-collapsing, so the general convergence theory leads to a decomposition ${Z_j}=\mathcal R\cup \mathcal S$, where the regular set $\mathcal R$ is a smooth open manifold with $d_{Z_j}|_{\mathcal R}$ induced by a Calabi-Yau metric $(\omega_{\mathcal R}, J_{\mathcal R})$, and the singular set $\mathcal S$ is closed and has Hausdorff codimension at least 4.
\begin{proposition}\label{p: metric completion}
${Z_j}$ agrees with the metric completion of $(Y_j^\circ, \hat\omega_j)$ and is naturally homeomorphic to $Y_j$.
\end{proposition}
This would follow from the arguments of J. Song \cite{Song} (and its generalization in \cite{CGT}). We outline a different and more direct proof here, which may be of independent interest. It is not hard to see that we can naturally view $ (Y_j^\circ,\hat\omega_j)$ as embedded in $(\mathcal R, \omega_{\mathcal R})$ as an open K\"ahler manifold. Moreover, since $\omega_j\leqslant \mathcal{P}(A_j)\hat\omega_{j,l}$, by passing to a further subsequence, one can take the limit of the holomorphic maps $\pi_j: (\hat Y_j, \hat\omega_{j, l})\rightarrow Y_j\subset \mathbb{C}^N$ and obtain a surjective continuous map $\pi_\infty: {Z_j}\rightarrow Y_j\subset \mathbb{C}^N$, which is the identity on $Y_j^\circ$. It then follows that $\mathcal R\setminus Y_j^\circ$ is a complex analytic set in $\mathcal R$, given by $\pi_\infty^{-1}(\text{Sing}(Y_j))\cap \mathcal{R}$.
On $Y_j^\circ$ we may write $\hat\omega_j=\sqrt{-1}\partial\bar\partial\hat\varphi_j$, where $\hat\varphi_j=\varphi_j+\psi_j$ is locally bounded across $\text{Sing}(Y_j)$. A key point is the following
\begin{lemma}\label{l:potential gradient bound}
The function $\hat\varphi_j$ extends smoothly to $\mathcal R$. Moreover, for any compact set $K\subset {Z_j}$ we have
$$\sup_{K\cap \mathcal R} |\nabla_{\omega_{\mathcal R}}\hat\varphi_j|_{\omega_{\mathcal R}}<\infty.$$
\end{lemma}
\begin{proof}
On $Y_j^{\circ}$, the form $\sqrt{-1}\partial\bar\partial\hat\varphi_j=\hat\omega_j=\omega_{\mathcal R}$ is smooth. Since $\mathcal R\setminus Y_j^{\circ}$ is a closed pluripolar set and $\hat\varphi_j$ is locally bounded across $\text{Sing}(Y_j)$, by well-known removable singularity theorem for plurisubharmonic functions (see \cite[Chapter I,Theorem 5.24]{Demailly}), we see that the same equation holds on $\mathcal R$ in the sense of currents, so elliptic regularity gives a smooth extension of $\hat\varphi_j$ to $\mathcal R$.
To prove the gradient bound we fix an open subset $U\subset Y_j$ containing $\text{Sing}(Y_j)$ and with $\partial U$ smooth. Denote $\hat U=\pi_j^{-1}(U)$. Its boundary $\partial\hat U$ is naturally identified with $\partial U$. We can solve the Poisson equation $\Delta_{\hat\omega_{j, l}}\eta_{l}=n$ on $\hat U$ with boundary condition given by $\hat\varphi_j|_{\partial U}$. Then by the Cheng-Yau gradient estimate we have $|\nabla_{\hat\omega_{j, l}} \eta_l|\leqslant C$ for all $l$. So passing to a subsequence $\eta_l$ converges uniformly to a limit $\eta$ on $\pi_\infty^{-1}(U)\subset {Z_j}$, with $|\nabla_{\hat\omega_j}\eta|\leqslant C$, $\Delta_{\hat\omega_j}\eta=n$ on $\pi_\infty^{-1}(U)\cap \mathcal R$ and $\eta|_{\partial U}=\hat\varphi_j|_{\partial U}$.
Now consider the function $f=\hat\varphi_j-\eta$ on $\pi_\infty^{-1}(U)\cap\mathcal R$. It is harmonic and $f|_{\partial U}=0$. One can conclude that $f$ is identically zero from standard removable singularity theorems for harmonic functions on Ricci limit spaces. For example, using the fact that the Hausdorff codimension of $\mathcal S$ is at least 4, for $\epsilon>0$ small, one can construct (c.f. \cite[Proposition 3.5]{DS1}) a cut-off function $\chi_\epsilon: {Z_j}\rightarrow \mathbb R$ which is equal to 1 when $d_{Z_j}(x, \mathcal S)\geqslant \epsilon$ and vanishes in a neighborhood of $\mathcal S$, such that $\int_{Z_j} d_{Z_j}(\cdot, \mathcal S)^{-1} |\nabla_{\omega_\mathcal R} \chi_\epsilon|\omega_{\mathcal R}^n<\epsilon$. By the local Cheng-Yau gradient estimate we have $|\nabla_{\omega_{\mathcal R}}\hat\varphi_j|\leqslant Cd(\cdot, \mathcal S)^{-1}$. Then by a straightforward integration by part argument one sees that $f\equiv0.$
\end{proof}
Given this Lemma, one can follow the strategy of \cite{DS1, DS2} to construct holomorphic functions by the H\"ormander $L^2$ technique (see a related observation in \cite{Gabor2} for studying the asymptotic cone of complete $\sqrt{-1}\partial\bar\partial$-exact Calabi-Yau manifolds with Euclidean volume growth). Here one can work on the trivial holomorphic line bundle $L$ over $Y_j^{\circ}$ endowed with the hermitian metric $|\cdot|_{\hat\varphi_j}\equiv e^{-\hat\varphi_j}$. Notice $Y_j^{\circ}$ admits a complete K\"ahler metric, so one can apply Theorem \ref{Hormander} here. Let $\nabla_L$ be the corresponding Chern connection on $L$. For $k\geqslant 1$, denote by $\mathcal V_k$ the space of holomorphic sections $f$ of $L^k$ over $Y_j^{\circ}$ with $$\|f\|_k^2\equiv \int_{Y_j^{\circ}}|f|^2_{k\hat\varphi_j}(k\hat{\omega}_j)^n<\infty.$$
Since $Y_j$ is normal it is clear that any holomorphic function $f$ on $Y_j^{\circ}$ extends to a holomorphic function on $Y_j$. In particular, for any compact set $K\subset \mathbb{C}^N$, $|f|$ and $|df|_{\omega_j}$ are uniformly bounded over $Y_j^{\circ}\cap K$. By \eqref{e:phij uniform bound}, \eqref{e:lower bound} and Lemma \ref{l:potential gradient bound}, we know that $|f|_{k\hat\varphi_j}$ and $|\nabla_L f|_{k\hat\varphi_j}$ are also uniformly bounded over $Y_j^{\circ}\cap K$. Viewed in ${Z_j}$, it follows that $f$ extends to a holomorphic function on $\mathcal R$ and the above norms are locally bounded near $\mathcal S$.
To apply the idea in \cite{DS1} we need to show that for any compact set $K\subset Y_j$, there exist constants $K_0, K_1>0$ depending on $K$ such that for all $k\geqslant 1$ and any $f\in \mathcal V_k$, we have
\begin{equation}
\sup_K|f|_{k\hat\varphi_j}\leqslant K_0 \|f\|_k
,\ \ \ \ \sup_K|\nabla_L f|_{k\hat\varphi_j}\leqslant K_1\|f\|_k.
\end{equation}
These were proved by applying the Moser iteration to the corresponding differential inequalities $$\Delta_{k\hat\omega_j}|f|_{k\hat\varphi_j}\geqslant -C|f|_{k\hat\varphi_j}, \ \ \ \ \Delta_{k\hat\omega_j}|\nabla_L f|_{k\hat\varphi_j}\geqslant -C|\nabla_L f|_{k\hat\varphi_j}, $$
which a priori only hold on $\mathcal R$, but again using the existence of a good cut-off function as in \cite{DS1}, Proposition 3.5 and noticing that there is a Sobolev inequality on ${Z_j}$ (since $Z_j$ is a volume non-collapsing Ricci limit space), one can make the usual arguments go through.
Using the existence of tangent cones at points in ${Z_j}$, it is straightforward to adapt the arguments in \cite{DS1, DS2} to construct holomorphic functions in $\mathcal V_k$ that separate points. Namely, given any $x_1\neq x_2\in {Z_j}$, one can find $k\geqslant 1$ and $f_1, f_2\in \mathcal V_k$ such that
$ |f_\alpha(x_\alpha)|_{k\hat\varphi_j}\geqslant 1, $ and $|f_\alpha(x_\beta)|_{k\hat\varphi_j}\leqslant 1/10$ for $\alpha\neq\beta$.
\begin{proof}[Proof of Proposition \ref{p: metric completion}]
We first prove $\pi_\infty$ is injective (hence is a homeomorphism). Suppose $x_1\neq x_2\in {Z_j}$. Then we can construct the functions $f_1, f_2$ as above. But both can be viewed as holomorphic functions on $Y_j$, so it follows that $\pi_\infty(x_1)\neq \pi_\infty(x_2)$.
Since $\text{Sing}(Y_j)$ consists of finitely many points, we see ${Z_j}\setminus Y_j^{\circ}=\pi_\infty^{-1}(\text{Sing}(Y_j))$ is also a finite set. Clearly this implies that ${Z_j}$ is indeed the metric completion of $(Y_j^{\circ}, \hat\omega_j)$.
\end{proof}
Now we further let $j\rightarrow\infty$. Passing to a subsequence we may take a pointed Gromov-Hausdorff limit $(Z_j, d_{Z_j})\rightarrow (Z_\infty, d_\infty)$. The same argument as above combined with the uniform estimates in Section \ref{s5} give
\begin{proposition}
$Z_\infty$ agrees with the metric completion of $(W\setminus \{0\}, \hat\omega_\infty)$ and is naturally homeomorphic to $W$.
\end{proposition}
In particular, we may naturally identify $Z_\infty$ with $W$.
\begin{proof}[Proof of Theorem \ref{t:main0}] As is proved in Section \ref{s5}, $Z_\infty$ has a unique asymptotic cone given by $\mathcal C$.
From the above discussion it is straightforward to extend the arguments of \cite{DS2} to show that there is a unique metric tangent cone $\mathcal C'$ at $0$. Furthermore, by the minimizing property of K-semistable valuations proved by Li-Xu \cite{LX}, we know the volume densities satisfy $\kappa(\mathcal C')\leqslant \kappa(\mathcal C)$, where the right hand side can be interpreted as the normalized volume associated to the natural valuation of $W$ at $0$.
Notice $Z_\infty$ is a pointed Gromov-Hausdorff limit of complete Ricci-flat manifolds, so the Bishop-Gromov inequality applies to $Z_\infty$, which gives that $\kappa(\mathcal C')\geqslant \kappa(\mathcal C)$. So the equality holds and then $Z_\infty$ must be a K\"ahler cone itself, hence is isomorphic to $\mathcal C$ as a K\"ahler cone. Denote by $\xi'$ the Reeb vector field on $Z_\infty$ and by $\tilde r$ the radial function on $Z_\infty$.
It remains to show that $\xi'=\xi$. For this purpose we need to use some results in \cite{DS2}. The algebraic structure of $Z_\infty$ can be intrinsically characterized in terms of the ring $R(Z_\infty)$ holomorphic functions on $Z_\infty$ of polynomial growth (measured using the K\"ahler cone metric). The cone structure yields a decomposition \begin{equation}\label{e:coordinate ring decomposition}R(Z_\infty)=\bigoplus_{\mu\geqslant0} R_\mu(Z_\infty),
\end{equation}where $R_\mu(Z_\infty)$ consists of $f$ which are of homogeneous of degree $\mu$, i.e., $\mathcal L_{J\xi'}f=-\mu f$. Conversely, $\xi'$ can be recovered from this decomposition.
The above decomposition is determined by a filtration associated with a \emph{degree} function.
For any nonzero $f\in R(Z_\infty)$, the following is well-defined (c.f. \cite{DS2})
$$d_{Z_\infty}(f)\equiv\lim_{R\rightarrow\infty}\frac{\sup_{\tilde r(x)=R}\log |f(x)|}{\log R}<\infty.$$
Then the decomposition in \eqref{e:coordinate ring decomposition} is determined by the filtration of $R(W)$ with respect to $d_{Z_\infty}$.
On the other hand, the coordinate ring $R(W)$ of $W$ consists of restrictions of polynomial functions on $\mathbb{C}^N$. We define
$$d_{W}(f)\equiv\lim_{R\rightarrow\infty}\frac{\sup_{r_\xi(x)=R}\log |f(x)|}{\log R}.$$
There is a decomposition $R(W)=\bigoplus_{\mu\geqslant 0} R_\mu(W)$, where $R_\mu(W)$ is the space of polynomials $f$ with $\mathcal L_{J\xi}f=-\mu f$. It can be checked that this decomposition is also determined by the filtration of $R(W)$
with respect to $d_W$.
Notice that by the construction of $\hat\omega_\infty$ we know for all $\epsilon>0$ ,$C_\epsilon^{-1}r_\xi^{1-\epsilon}\leqslant \tilde r\leqslant C_\epsilon r_\xi^{1+\epsilon}$ on $W\setminus K$ for a fixed compact set $K$. It follows that $R(Z_\infty)=R(W)$ and $d_W(f)=d_{Z_\infty}(f)$ for any $f\in R(W)$. So we must have $R_\mu(Z_\infty)=R_\mu(W)$ for all $\mu\geqslant 0$.
In particular, $\xi'=\xi$. Hence we have constructed a $T_\xi$ equivariant isomorphism between $W$ and $\mathcal C$. This completes the proof of Theorem \ref{t:main0}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{t:main2}] We simply notice that by the estimate of $\varphi_\infty$ the background metric $\omega_\infty$ is polynomially asymptotic to the metric $\hat\omega_\infty$, but the latter is exactly the Calabi-Yau cone $Z_\infty$ (which is the same as $\mathcal C$). Then $p_\infty$ exactly provides the diffeomorphism that shows $\omega$ and $\Omega$ are asymptotic to $\hat\omega_\infty$ and $\Omega_\infty$ in a polynomial rate.
\end{proof}
\begin{remark}Using the same proof one can weaken the Calabi-Yau condition in Theorem \ref{t:main0} and \ref{t:main2} to be Ricci-flat K\"ahler. Alternatively, one can show this by working on the universal cover since it is well known (cf. \cite{Anderson, PLi}) that Riemannian manifolds with non-negative Ricci curvature and Euclidean volume growth have finite fundamental groups.
\end{remark}
\section{Discussions}
We point out some further related directions that one can explore. The guiding problem is
\begin{problem}[Algebraization for complete Calabi-Yau metrics]\label{q6.1}
Give an algebro-geometric characterization of all complete Calabi-Yau metrics with Euclidean volume growth.
\end{problem}
\begin{remark}
Without the volume growth condition, the situation is more complicated since in general the Calabi-Yau metric does not have to be conical at infinity. Extra assumptions are needed in order to make connections with algebraic geometry. In complex dimension 2 under a natural finite energy condition, the models at infinity are completely classified by Sun-Zhang \cite{SZ}, and there is a complete classification of the Calabi-Yau metrics in terms of algebro-geometric data by the work of many people; see \cite{SZ} for references.
\end{remark}
Given Theorem \ref{t:main2}, the results of Conlon-Hein \cite{CH3} provide an answer to Problem \ref{q6.1} under the extra quadractic curvature decay condition. Namely, such Calabi-Yau metrics are always constructed as follows (ignoring uniqueness issues for the moment), where each step is essentially algebro-geometric.
\begin{enumerate}
\item Choose a Calabi-Yau cone $(\mathcal C, \xi)$. This is equivalent to choosing a K-polystable algebraic cone \cite{CS}.
\item Choose a nomal affine variety $X'$ with $(\mathcal C, \xi)$ as a weighted asymptotic cone at infinity.
\item Choose a crepant resolution $\pi:X\rightarrow X'$.
\item Choose a K\"ahler class on $X$. This is a numerical condition according to \cite{CoTo,CH3}.
\end{enumerate}
We want to ask how much of the above picture holds in general, if we allow singularities in the relevant objects. There are several technical obstacles to realize this program, mostly related to the possible appearance of singularities on the asymptotic cone.
First one needs to extend the 2-step degeneration theory of \cite{DS2} to this setting. In particular, we would like to construct abundant holomorphic functions with polynomial growth on $X$. This is related to a special case of Yau's compactification conjecture, which we reiterate as follows
\begin{conjecture}[Yau's Compactification conjecture] \label{Yau conjecture}
A complete Calabi-Yau manifold with Euclidean volume growth is naturally a quasi-projective variety.
\end{conjecture}
There has been partial progress towards this. For example, it is known that any asymptotic cone is naturally a normal affine variety \cite{LS}. It is also observed in \cite{Gabor2} that the results of \cite{DS2} automatically extend to the setting when the Calabi-Yau metric is $\partial\bar\partial$-exact so Conjecture \ref{Yau conjecture} holds in this case. Now suppose the 2-step degeneration theory works. The following is natural to expect
\begin{conjecture}[No semistability at infinity, the general version]\label{conj6.2}
Theorem \ref{t:main0} holds without the quadratic curvature decay assumption.
\end{conjecture}
\begin{remark}
Theorem \ref{t:main2} also has a general version, where one uses the Gromov-Hausdorff distance to measure the rate of convergence to the asymptotic cone.
\end{remark}
One difficulty in proving Conjecture \ref{conj6.2} using the strategy of this paper is related to the generalization of the Tian-Yau construction, which is itself an interesting question. A first step would be
\begin{question}[Prescribing asymptotic cone]\label{q6.3}
Given a singular Calabi-Yau cone $(\mathcal C, \xi)$ and a normal affine variety $X'$ with $(\mathcal C, \xi)$ as a weighted asymptotic cone at infinity, when does $X'$ admit a (possibly singular) Calabi-Yau metric asymptotic to $(\mathcal C, \xi)$ in the Gromov-Hausdorff sense?
\end{question}
There are some recent extensions of the Tian-Yau construction with singular asymptotic cones for special examples; see \cite{HN, YLi, Gabor1, CR, Chiu}.
\
As mentioned in Section \ref{two step degeneration} the classification of complete Calabi-Yau metrics with Euclidean volume growth on a fixed underlying algebraic variety is a subtle problem. Even on $\mathbb{C}^n(n\geqslant 3)$ the situation can be complicated; the recent works by Sz\'ekelyhidi \cite{Gabor2} and Chiu \cite{Chiu} make progress in this direction. We make an attempt here to formulate some general questions. To minimize technical issues we restrict to the case of $\partial\bar\partial$-exact metrics on smooth affine varieties. As mentioned above the results of \cite{DS2} apply here. Fix a smooth affine variety $X$ with the coordinate ring $R(X)$. A complete $\partial\bar\partial$-exact Calabi-Yau metric $\omega$ on $X$ with Euclidean volume growth defines an affine structure on $X$ with coordinate ring $R_\omega(X)$ given by the space of holomorphic functions on $X$ with polynomial growth with respect to $\omega$. We say that $\omega$ is a \emph{compatible} metric on $X$ if $R_\omega(X)=R(X)$.
A compatible Calabi-Yau metric $\omega$ defines a degree function $d_\omega$ on $R(X)$ satisfying
\begin{itemize}
\item $d_\omega(f)=0$ if and only if $f$ is a nonzero constant;
\item $d_\omega(fg)=d_\omega(f)+d_\omega(g)$;
\item $d_\omega(f+g)\leqslant \max(d_\omega(f), d_\omega(g))$.
\end{itemize}
Furthermore, $d_\omega$ gives rise to a filtration of $R(X)$ whose associated graded ring is the coordinate ring $R(W)$ of the intermediate K-semistable cone $W$, and $W$ degenerates to the unique asymptotic cone $\mathcal C$ as affine algebraic cones. The result of Li-Wang-Xu \cite{LWX} implies that $(\mathcal C, \xi)$ is uniquely determined by $d_\omega$.
Clearly, if $\omega$ is compatible, then for any $\lambda>0$ and any algebraic automorphism $F$ of $X$, $\lambda F^*\omega$ is also compatible. Furthermore, $d_\omega=d_{\lambda\omega}$ and $d_{F^*\omega}=F^*d_{\omega}$.
In terms of algebro-geometric language one notices that $v_\omega\equiv-d_\omega$ is a valuation on $R(X)$ with values in $\mathbb R_{\leqslant0}$. We call such valuations a \emph{negative valuation}. We say a negative valuation $v$ is \emph{semistable} (resp. \emph{polystable}) if it defines a filtration whose associated graded ring is finitely generated and defines a K-semistable (resp. K-polystable) algebraic cone $\mathcal C_v$. Geometrically, the cone is a weighted asymptotic cone at infinity of $X$ under some affine embedding. Conjecture \ref{conj6.2} implies that the negative valuation induced by a compatible Calabi-Yau metric on $X$ is always polystable.
Given a semistable valuation $v$ on $X$, we define $G_v$ as the group of algebraic automorphisms of $X$ that preserve $v$ and define $\tilde G_v$ as the group of algebraic automorphisms of $\mathcal C_v$ that fixes the vertex and preserves the cone structure. Both groups are finite-dimensional. Notice that $\tilde G_v$ contains the algebraic torus $\mathbb T_v$ generated by the cone vector field. There is a natural homomorphism $\varphi: G_v\rightarrow \tilde G_v$.
\
\
To classify the space of all compatible Calabi-Yau metrics on $X$, assuming Conjecture \ref{conj6.2}, one can proceed as follows:
\
\textbf{Step 1}. Determine the space $\mathcal V$ of all possible polystable negative valuations on $X$ up to algebraic automorphism. This is essentially an algebro-geometric question. One can also study the related question with polystability replaced by semistability. In general there may be infinitely many polystable negative valuations on a given $X$. Indeed, given any polystable cone $V$ in $\mathbb{C}^n$ which is a hypersurface, one can find a polystable negative valuation $v$ on $X=\mathbb{C}^n$ with $\mathcal C_v=\mathbb{C}\times V$. Indeed, the result of Sz\'ekelyhidi \cite{Gabor1} implies that if $V$ has only isolated singularities, then the associated negative valuation actually arise from a compatible Calabi-Yau metric on $\mathbb{C}^n$. Still, it seems interesting to investigate the structure of the space $\mathcal V$.
\
\textbf{Step 2}. Now we fix a polystable negative valuation $v$ on $X$. Denote by $\mathcal M_v$ the space of all compatible Calabi-Yau metrics $\omega$ on $X$ with $v=v_\omega$. Is $\mathcal M_v$ nonempty?
This is related to Question \ref{q6.3}.
\
\textbf{Step 3}. Suppose $\mathcal M_v\neq \emptyset$. We know all $\omega\in \mathcal M_v$ have the same asymptotic cone at infinity given by $\mathcal C_v$. But we should expect more and one would like to have
\begin{conjecture}[Uniform equivalence]\label{conj6.5}
There exists a constant $C>0$ such that for all $\omega_1, \omega_2\in \mathcal M_v$, we have $C^{-1}\omega_1\leqslant \omega_2\leqslant C\omega_1$.
\end{conjecture}
It is easy to see the converse is true: if two compatible Calabi-Yau metrics $\omega_1$ and $\omega_2$ are uniformly equivalent, then $v_{\omega_1}=v_{\omega_2}$. We remark that one can also ask an analogous question for local singularities of K\"ahler-Einstein metrics (on algebraic varieties with klt singularities), that is, whether any two such metrics are indeed locally uniformly equivalent. The answer to this question, currently not known, is also related to Conjecture \ref{conj6.5}. The recent work by Chiu-Sz\'ekelyhidi \cite{ChSz} made progress in understanding local singular behavior of K\"ahler-Einstein metrics.
\
\textbf{Step 4}. Denote by $\mathcal N_v$ the space of Calabi-Yau cone metrics on $\mathcal C_v$ (with the given Reeb vector field determined by $v$). We have a natural map $\mathfrak C: \mathcal M_v\rightarrow \mathcal N_v$ given by taking the asymptotic cone of a compatible Calabi-Yau metric. Then we make
\begin{conjecture}[Generalized uniqueness]\label{conj6.7}
The map $\mathfrak C$ is bijective.
\end{conjecture}
We explain this holds if we assume $\mathcal C_v$ has only an isolated singularity at the vertex. In this case any $\mathcal M_v$ has quadratic curvature decay so we may apply the arguments of this paper, and one can show that Conjecture \ref{conj6.5} and \ref{conj6.7} indeed hold in this case. That is, since we have fixed the negative valuation $v$, this fixes the algebro-geometric degeneration from $X$ to $\mathcal C_v$, in terms of a weighted asymptotic cone. Then by Section \ref{section:defining projection map} we can fix the diffeomorphism between the end of $X$ and the end of $\mathcal C_v$. For every $\omega_{\infty}\in \mathcal N_v$, by \cite{CH3} one gets a compatible Calabi-Yau metric on $X$ which is asymptotic to $\omega_\infty$ at a polynomial rate. Conversely, given a compatible Calabi-Yau metric $\omega\in \mathcal M_v$, using the arguments in this paper, we can get a Calabi-Yau cone metric on $\mathcal C_v$ which is the asymptotic cone of $\omega$ at a polynomial rate. Then the uniqueness theorem proved by Conlon-Hein (c.f. \cite[Theorem 3.1]{CH1}) implies that $\mathfrak C$ is bijective. Notice even in this case it is not clear that the action of $G_v$ is transitive on $\mathcal M_v$ (modulo scaling): we know that the action of $\widetilde G_v$ on $\mathcal{N}_v$ is transitive by the generalized Bando-Mabuchi theorem (see for example, \cite[Proposition 4.8] {DS2}). Indeed, $\mathcal N_v$ can be identified with the homogeneous space $\widetilde G_v/K_v$ for a maximal compact subgroup $K_v$ whose complexification is $\widetilde G_v$. But the map $\varphi: G_v\rightarrow\tilde G_v$ could be complicated in general: neither the injectivity nor the surjectivity seems obvious. It seems to us that Conjecture \ref{conj6.7} is a more natural formulation than a naive uniqueness statement. We mention that the above conjecture is compatible with the result and conjecture in \cite{Gabor2, Chiu} for complete Calabi-Yau metrics on $\mathbb{C}^3$ with specific asymptotic cones.
\
The above four steps altogether would lead to a satisfactory classification of compatible Calabi-Yau metrics on a fixed smooth affine variety. More generally, when the metrics are not necessarily $\partial\bar\partial$-exact, one needs to further involve K\"ahler classes in the discussion.
\
From another perspective, one can also try to classify complete Calabi-Yau metrics asymptotic to a given Calabi-Yau cone, without fixing the underlying algebraic variety.
Notice again this should contrast with the local situation. As is well known (cf. \cite{DS2, HN, Gabor1}), there is no boundedness of local singularities of Calabi-Yau metrics with a given tangent cone if the latter has nonisolated singularities. One example is given by the $A_k$ singularities in dimension $n\geqslant 3$ for $k\geqslant 4$, whose local tangent cones are all given by the product of an $n-1$ dimensional $A_1$ singularity with $\mathbb C$. However, for asymptotic cones, there are more constraints. For example, the Bishop-Gromov inequality ensures a strong uniform non-collapsing condition, which gives a preliminary compactness in this situation. In particular, one would expect a well-behaved moduli space when the volume density of the asymptotic cone is bounded away from zero.
\begin{problem}[Moduli problem]
Given a positive number $\lambda>0$, study the structure of the moduli space of (possibly singular) $\partial\bar\partial$-exact Calabi-Yau metrics on affine varieties whose asymptotic cones have volume density at least $\lambda$.
\end{problem}
It is straightforward to formulate an algebro-geometric counterpart of the above problem. That is, one can study the moduli space $\mathfrak M_{\geq \lambda}$ of pairs $(X, v)$, where $X$ is an affine variety and $v$ is a polystable negative valuation on $X$ with $\mathcal C_v$ having volume density at least $\lambda$. It seems to be an approachable question given recent advances in birational algebraic geometry. However, as pointed out by Yuji Odaka there are some complications. For simplicity we fix a K-polystable cone $\mathcal C$, then the deformation space $\operatorname{Def}(\mathcal C)$ of $\mathcal C$ as affine varieties is in general infinite dimensional, but it is natural to expect that the subspace $\operatorname{Def}_{-}(\mathcal C)$ which consists of those deformations to the above pairs $(X, v)$ with $\mathcal C_v=\mathcal C$ is finite dimensional, and there is an action of $\widetilde G_v$ on $\text{Def}_{-}(\mathcal C)$. Now since $\widetilde G_v$ is noncompact the quotient is in general not Hausdorff, so any meaningful construction of such moduli space will have to address this issue. On the other hand, if we accept the above picture, then the space of $\partial\bar\partial$-exact Calabi-Yau metrics asymptotic to $\mathcal C$ (up to isomorphism) is given by the quotient $\text{Def}_{-}(\mathcal C)/K_v$. It seems an interesting question to us to explore the structure here in more detail.
\
Finally, we point out that while Theorem \ref{t:main2} is a Riemannian geometric statement, our proof in this paper hinges on complex geometry.
One can ask whether the polynomial convergence rate
is a more general phenomenon in geometric analysis, or is it only special in the complex geometric world. Notice this is a question involving \emph{global} geometry, not only the end at infinity. Indeed, we expect that the technique of Adams-Simon \cite{AS} can be used to show that given a Calabi-Yau cone $\mathcal C$ with obstructed deformations in an appropriate sense, one can produce an end of Calabi-Yau metric which is asymptotic to $\mathcal C$ at only a logarithmic rate; see \cite{GChen} for a related work on isolated singularities of $G_2$ metrics.
\begin{question}[Polynomial rate to the asymptotic cone]\label{conj6.9}
Is a complete Ricci-flat Riemannian manifold with Euclidean volume growth always asymptotic to a unique metric cone at a polynomial rate (in the Gromov-Hausdorff distance)?
\end{question}
Even the uniqueness of asymptotic cones is not known to date without the quadratic curvature decay condition. An answer to Question \ref{conj6.9}, either in the positive or the negative, would be interesting. A special setting is given by manifolds with special holonomy, i.e., $G_2$ and $Spin(7)$ manifolds. One can also explore similar questions for global solutions of other geometric PDEs, such as minimal submanifolds and Yang-Mills connections. These will be important in classifying singularity models.
\printbibliography
\end{document}
|
\section{Introduction}
In order to solve general convex optimization, many versions of Black-Box first-order algorithms have been proposed. However, most of those commonly used algorithms require a Euclidean ($L_2$) projection onto the domain set at each iteration, such as the famous Projected sub-Gradient Descent (PGD) algorithm.\footnote{We use the more common acronym PGD for projected sub-gradient descent instead of PSD.} Unless there is special structure, it is not always computationally easy to perform the projection steps. This can make PGD less attractive for large-scale constrained optimization problems. This paper develops an optimization algorithm that replaces projections with (computationally easier) Linear Optimization steps.
Specifically, let $\mathcal{X}\in\mathbb{R}^n$ be a closed and convex set. The problem of this paper is to minimize a general convex function $f:\mathbb{R}^n\rightarrow\mathbb{R}$ over all $x\in\mathcal{X}$. The Euclidean projection, i.e. $L_2$ projection, is defined as follows. Let $\mathcal{X}$ be a closed convex set. Define
\begin{displaymath}
\pi_{\mathcal{X}}\left(y\right)
\overset{def}{=}
\argmin _{x \in \mathcal{X}}\left\|x-y\right\|
,\quad \forall y \in \mathbb{R}^{n}
\end{displaymath}
We call $\pi_{\mathcal{X}}\left(y\right)$ the Euclidean projection of the point $y$ onto the set $\mathcal{X}$ (as $\mathcal{X}$ is closed and convex, the projection exists and is unique). As our algorithm is similar to PGD, let us review PGD:
\begin{algorithm}[H]
\caption{Projected sub-Gradient Descent (PGD)}
\label{alg: PGD}
\begin{algorithmic}[1]
\STATE{Require $T\in\{1,2,\ldots\}, \beta>0, x_0\in\mathcal{X}$}
\FOR{$k=0\cdots T-1$}
\STATE \label{eq:PGD2SGD} Calculate $g_k\in\partial f(x_k)$
\STATE $x_{k+1} = \pi_\mathcal{X}(x_k-\beta g_k)$
\ENDFOR
\RETURN $\Bar{x} = \frac{1}{T+1} \sum_{k=0}^{T} x_k$
\end{algorithmic}
\end{algorithm}
It can be shown that the PGD's update rule is equivalent to solve the following quadratic optimization problem:
\begin{displaymath}
x_{k+1}
= \pi_\mathcal{X}(x_k-\beta g_k)
=\argmin_{x\in\mathcal{X}}
\left\{\langle g_k,x\rangle
+\frac{1}{2 \beta}\|x-x_k\|^2\right\}
\end{displaymath}
where $\langle x,y\rangle$ is the inner product of $x$ and $y$ in $\mathbb{R}^n$.
For a general convex set, the projection step may involve a numerical procedure that can be computationally expensive. Different algorithms to compute a projection of a point onto a convex set have been suggested \cite{lin2003class,lin2003projection,han1988successive,iusem1994row}. However, for many convex sets, a linear optimization, defined by the following equation:
\begin{equation}
\label{eq:linear opt}
\min_{x\in\mathcal{X}}\left\{\langle g, x\rangle\right\}
,\quad
\text{for a given }g\in\mathbb{R}^n
\end{equation}
can be carried out easier than a projection. Let us call sets $\mathcal{X}$ for which problems of the type Equation \cref{eq:linear opt} can be solved more easily than a projection as Appropriate sets. Our algorithm works for any domain set that is convex. However, it shows its potential when the domain set $\mathcal{X}$ is an Appropriate set.
Many examples of Appropriate sets are known and studied. The work \cite{jaggi2013revisiting} unifies many Appropriate sets under the umbrella of \emph{Atomic sets}. One famous example is the set of matrices with bounded \textit{trace norm}. The trace norm of a matrix $A$, shown as $\|A\|_*$, is the sum of its singular values. For a fixed $\tau>0$, define set $\mathcal{X}\overset{def}{=}\{A\in\mathbb{R}^{m\times n} : \|A\|_*\leq \tau$\}. The projection of matrix $X\in\mathbb{R}^{m\times n}$ on the set $\mathcal{X}$ requires a Singular Value Decomposition (SVD) of $X$ which costs $O\left(m n \min\{m,n\}\right)$ time. Meanwhile, a linear optimization: $\min_{X\in\mathcal{X}}\langle P,X\rangle$, requires the calculation of the largest singular value and its respective singular vectors of the matrix $P$. This calculation costs linear time in the number of non-zero entries in that matrix.
\subsection{Main Result}
Similar to Frank-wolf type algorithms that are projection-free but only work for smooth objective functions, this paper offers a projection-free algorithm for general convex programming. This is an alternative for PGD when the Equation \cref{eq:linear opt} is easier then a projection. The algorithm is also compatible with stochastic (noisy) sub-gradients, which makes it an alternative for SGD.
\subsection{History and Related Work}
The original Frank-Wolfe algorithm \cite{https://doi.org/10.1002/nav.3800030109} was introduced and analyzed for polyhedral domains in $\mathbb{R}^n$, relying on line-search on a quadratic upper bound on $f$. A general framework to analyze Frank-wolf type algorithms is achieved in \cite{jaggi2013revisiting} which also provides a comprehensive overview and comparison between many existing algorithms of that type. Using the idea of Variance-Reduced, \cite{pmlr-v48-hazana16} gets better bounds for both smooth convex functions and smooth strongly-convex functions while presenting another comparison of different variants of Frank-Wolfe algorithms.
Using randomized smoothing, \cite{hazan2012projection} proposed a variant of Frank-Wolfe for online optimization of non-smooth and smooth convex functions but not achieving the optimal convergence rates. This result has been improved in \cite{hazan2020faster} using the so-called blocking technique.
In the case when the domain only consists of convex functional constrains ($\mathcal{X}=\set{x | h_i(x)\leq0 , \forall i\in\{1,\cdots,I\}}$), \cite{NIPS2012_c52f1bd6} develops a unique algorithm with only one projection which (with high probability) achieves an $O(1/\sqrt{T})$ rate for general convex optimization, and an $O(\log(T)/T)$ rate for strongly convex optimization. Using similar assumptions (and adding an extra assumption of smoothness of constraint functions $h_i$), \cite{pmlr-v89-levy19a} achieves a one projection algorithm with $O(1/\sqrt{T})$ regret for general online convex optimization, and an $O(\log(T)/T)$ regret for online strongly convex optimization. Similar functional constraints are treated in \cite{7798542} with $O(1/T)$ convergence for general composite smooth convex and/or separable convex programs using knowledge of a Lagrange Multiplier bound, and \cite{https://doi.org/10.48550/arxiv.1708.00322} removes the need to know the Lagrange Multiplier bound.
In another line of work, \cite{NEURIPS2018_36d75342} proposed a generalization of the conditional gradient algorithm achieving rates similar to the standard stochastic gradient algorithm using only zeroth-order information.
\section{Preliminaries}
$\mathbb{R}$ denotes the set of real numbers, and $\mathbb{R}^n$ is the usual vector space of real $n$-tuples $x=(x^{(1)},\cdots,x^{(n)})$. The inner product of $x$ and $y$ in $\mathbb{R}^n$ is expressed by
\begin{displaymath}
\langle x , y\rangle \overset{def}{=} \sum_{i=1}^n x^{(i)} y^{(i)}
\end{displaymath}
In this paper the norm is standard Euclidean which is defined as:
\begin{displaymath}
\|x\|
\overset{def}{=}
\sqrt{\sum_{i=1}^n \left(x^{(i)}\right)^2}
=
\sqrt{\langle x , x\rangle}
\end{displaymath}
Consider an arbitrary function $f$ defined on the set $S$. Fix the point $x\in S$. A vector $g$ is called a \emph{subgradient} of the function $f$ at the point $x$ if
\begin{equation}
\label{eq: subgrad}
f(y)\geq f(x) + \langle g , y-x\rangle
,\quad
\forall y\in S
\end{equation}
The set of all subgradients of $f$ at $x$, $\partial f(x)$, is called the \emph{subdifferential} of the function $f$ at the point $x$.
Function $f$ is called $G$-Lipschitz continuous on the set $S$ if $\forall x,y\in S $
\begin{equation}
\label{eq: lip def}
|f(x)-f(y)|\leq G \|x-y\|
\end{equation}
A subset $C$ of $\mathbb{R}^n$ is said to be \emph{convex} if $(1-\lambda)x+\lambda y \in C$ whenever $x\in C$, $y\in C$ and $0\leq\lambda\leq1$. A function $f$ is called \emph{convex} on set $C$ if $C$ is a convex set and for all $x, y \in C$ and $\alpha\in[0, 1]$ the following inequality holds:
\begin{displaymath}
f(\alpha x + (1-\alpha)y)
\leq
\alpha f(x) + (1-\alpha)f(y).
\end{displaymath}
\subsection{Extension of Convex Function}
An important assumption used in this paper is that $f$ is convex and $G$-Lipschitz continuous on the \textit{whole $\mathbb{R}^n$}. While this is a valid assumption in many practical cases, there are cases which can violate it.
The following theorem collects important parts from a larger \emph{McShane-Whitney extension theorem} of \cite{lmcs:6105}. It shows that if $f$ is convex and $G$-Lipschitz continuous on the set $\mathcal{X}$, then $f$ can be extended to a convex and $G$-Lipschitz continuous function on the whole of $\mathbb{R}^n$.
\begin{theorem}
\label{thm:Extend}
Let $\mathcal{X} $ be a convex set in $\mathbb{R}^n$, $f:\mathcal{X}\rightarrow\mathbb{R}$ a convex $G$-Lipschitz continuous function defined on $\mathcal{X} $. Then the function $\Tilde{f}$ defined by:
\begin{displaymath}
\Tilde{f}(w)
\overset{def}{=}
\inf_{x\in\mathcal{X} }
\{f(x)+G\|x-w\|\}
\end{displaymath}
satisfies the following:
\begin{enumerate}[(i)]
\item $\Tilde{f}$ is a real valued function.
\item $\Tilde{f}$ is a $G$-Lipschitz continuous function on $\mathbb{R}^n$ .
\item For any $x\in\mathbb{R}^n$, any sub-gradient $g\in\partial\Tilde{f}(x)$ is bounded by G: $\|g\|\leq G$.
\item $\Tilde{f}$ is a convex function on $\mathbb{R}^n$.
\item $f(w)=\Tilde{f}(w)$, $\forall w\in\mathcal{X} $
\end{enumerate}
\end{theorem}
\begin{proof}
This theorem is a special case of the \emph{McShane-Whitney extension theorem} and it is fully borrowed from different parts of \cite{lmcs:6105}. However, for the sake of convenience, the proof is collected and presented at \cref{APP:Proof Extend}.
\end{proof}
\subsection{Problem setup}
We want to provide a projection-free algorithm to solve the following optimization:
\begin{displaymath}
\min_{x\in\mathcal{X}} f(x)
\end{displaymath}
We assume that the set $\mathcal{X}$ is contained in an Euclidean ball centered at $x_1\in\mathcal{X}$ and of radius $R$. This has two consequences:
\begin{equation}
\label{eq: 2R}
\|x-y\|\leq 2 R, \quad \forall x,y\in\mathcal{X}
\end{equation}
\begin{equation}
\label{eq: R}
\|x-x_1\|\leq R, \quad \forall x\in \mathcal{X}
\end{equation}
Furthermore we assume that $f$ is convex and $G$-Lipschitz on $\mathbb{R}^n$ (if not we can use its extension $\Tilde{f}$ according to the \cref{thm:Extend}).
\section{Projection-free Algorithm for Deterministic Optimization}
We assume that the function $f$ is represented by a Black-Box first-order oracle. This means that we can only get information by requesting the oracle to return a sub-gradient of the requested point $x$ which we call $g(x)\in\partial f(x)$. We provide the upper bound the on number of requests from the oracle to achieve a desired accuracy.
\begin{algorithm}[H]
\caption{Projection-free Algorithm}
\label{alg:Deterministic}
\begin{algorithmic}[1]
\STATE{Require integer $T\geq1$}
\STATE{Choose constants $\alpha>0$ and $\eta>0$ and the initial point $x_1\in\mathcal{X} $}
\STATE\label{eq: y1} Choose $y_1=x_1$
\STATE\label{eq: Q0} Choose $Q_{0}=\mathbf{0}$
\FOR{$1\leq k\leq T-1$}
\STATE\label{eq: Q update} Choose $Q_{k}=Q_{k-1}+y_k-x_k$
\STATE\label{eq: subgrad update} Choose $g_k=g(y_k)\in \partial f(y_k)$
\STATE\label{eq: x update} Choose $x_{k+1}\in\argmin_{x\in\mathcal{X} } \langle -Q_k, x\rangle$
\STATE\label{eq: y update} Choose $y_{k+1} =\frac{1}{\alpha+\eta}\left(\alpha y_k + \eta x_{k+1} -\eta Q_k - g_k\right)$
\ENDFOR
\RETURN \label{eq: xbar}$\Bar{x} = \frac{1}{T}\sum_{k=1}^{T}x_k$
\end{algorithmic}
\end{algorithm}
\begin{theorem}[Convergence of \cref{alg:Deterministic}]
\label{Main result}
Define $x^*\in\argmin_{x\in\mathcal{X}}\{f(x)\}$. Then with the following parameters,
\begin{displaymath}
\alpha = \frac{G\sqrt{T}}{R}
\end{displaymath}
\begin{displaymath}
\eta = \frac{G}{2 R\sqrt{T}}
\end{displaymath}
the \cref{alg:Deterministic} ensures
\begin{displaymath}
f(\Bar{x})-f(x^*)
\leq
\frac{3 R G}{\sqrt{T}}
\end{displaymath}
for any integer $T\geq1$.
\end{theorem}
\begin{proof}
See \cref{Proof of the Main result}.
\end{proof}
Let us remember the \cref{alg: PGD} convergence too:
\begin{theorem}[PGD convergence rate]
Define $x^*\in\argmin_{x\in\mathcal{X}}\{f(x)\}$. With the following parameter,
\begin{displaymath}
\beta = \frac{R}{ G\sqrt{T}}
\end{displaymath}
the \cref{alg: PGD} ensures
\begin{displaymath}
f(\Bar{x})-f(x^*)
\leq
\frac{R G }{\sqrt{T}}
\end{displaymath}
for any integer $T\geq1$.
\end{theorem}
\begin{proof}
See Theorem 3.2. from \cite{bubeck2015convex}.
\end{proof}
Upper bound on the error in both algorithms are of the same order (up to a constant) regarding the problem's parameters. However, this does not mean that they will always perform similarly as these are only upper bounds. Simulations, provided on the \cref{sec:simul}, compare these two algorithms numerically.
\section{Projection-free Algorithm for Stochastic Optimization}
In many applications, function $f$ is represented by a first-order \emph{stochastic} oracle. The oracle takes $Y$ as input and returns $\hat{g}(Y)$ such that $\mathbb{E}\{\hat{g}(Y)|Y\}\in\partial f(Y)$, meaning we only have access to a noisy but unbiased version of a sub-gradient. We also need to add an extra condition:
\begin{displaymath}
\mathbb{E}\{\|\hat{g}(Y) \|^2| Y\}\leq B^2
\end{displaymath}
for some known fixed $B\geq G$ (this simply means that the noise must have a bounded variance). Surprisingly the algorithm works for this case too, only the $g_k=g(y_k)$ in \cref{eq: subgrad update} of \cref{alg:Deterministic} is replaced with $g_k = \hat{g}(Y_k)$.
\begin{theorem}[Convergence of \cref{alg:Deterministic} with stochastic sub-gradient]
\label{main stochastic}
Define $x^*\in\argmin_{x\in\mathcal{X}}\{f(x)\}$.
\begin{itemize}
\item
Then with the following parameters,
\begin{displaymath}
\alpha = \frac{ B\sqrt{T}}{R}
\end{displaymath}
\begin{displaymath}
\eta = \frac{G}{2 R\sqrt{T}}
\end{displaymath}
the \cref{alg:Deterministic} with stochastic sub-gradient ensures
\begin{displaymath}
\mathbb{E}\left\{f(\Bar{X})\right\}-f(x^*)
\leq
\frac{B R + 2 G R}{\sqrt{T}}
\end{displaymath}
for any integer $T\geq1$.
\item
Then with following parameters,
\begin{displaymath}
\alpha = \frac{B\sqrt{T}}{R}
\end{displaymath}
\begin{displaymath}
\eta = \frac{2 B}{R\sqrt{T}}
\end{displaymath}
the \cref{alg:Deterministic} with stochastic sub-gradient ensures
\begin{displaymath}
\mathbb{E}\left\{f(\Bar{X})\right\}-f(x^*)
\leq
\frac{3 G D}{2\sqrt{T}}
\end{displaymath}
for any integer $T\geq1$.
\end{itemize}
\end{theorem}
\begin{proof}
See \cref{APP:Proof of the stochastic}.
\end{proof}
SGD algorithm is also very similar to PGD (\cref{alg: PGD}). The $g_k=g(x_k)$ in \cref{eq:PGD2SGD} of \cref{alg: PGD} is replaced with $g_k = \hat{g}(X_k)$.
\begin{theorem}[SGD convergence rate]
Define $x^*\in\argmin_{x\in\mathcal{X}}\{f(x)\}$. With the following parameter,
\begin{displaymath}
\beta = \frac{R}{B\sqrt{T}}
\end{displaymath}
the SGD algorithm ensures
\begin{displaymath}
\mathbb{E}\{f(\Bar{X})\}-f(x^*)
\leq
\frac{B R}{\sqrt{T}}
\end{displaymath}
for any integer $T\geq1$.
\end{theorem}
\begin{proof}
See Theorem 6.3. from \cite{bubeck2015convex}.
\end{proof}
Upper bound of the average error in both algorithms are of the same order (up to a constant) regarding the problem's parameters. However, if in addition to parameter $B$ the parameter $G$ is provided, unlike PGD/SGD, our algorithm is able to take advantage of it. Simulations, provided on the \cref{sec:simul}, compare these two algorithms numerically.
\section{Numerical results}
\label{sec:simul}
In this section we provide two examples where our algorithm is compared with PGD and SGD.
\subsection{Hypercube domain with \texorpdfstring{$L1$}{Lg}-norm objective function }
This example is a good case study to see the numerical accuracy of our algorithm as the exact minima is analytically calculable. However, this is not an example to illustrate the computational gains. Our domain set is a $n$-dimensional hypercube:
\begin{displaymath}
\mathcal{X}=\left\{x\in\mathbb{R}^n : |x^{(i)}|\leq 1, \, \forall i\in\{1,\cdots,n\}\right\}
\end{displaymath}
The set $\mathcal{X}$ is inside the Euclidean ball with radius $R=2\sqrt{n}$ centered at $x_1=\mathbf{0}$. Define the $L_1$-norm as: $\|x\|_1=\sum_{i=1}^n |x^{(i)}|$. Let us choose the objective function $f_\omega(x)=\|x-\omega\|_1$. It is easy to see that this function is convex and Lipschitz continuous over $\mathbb{R}^n$ with coefficient
\begin{displaymath}
G = \sqrt{n}
\end{displaymath}
When the exact sub-gradient is given to us without any error, we choose the parameters as follows:
$\beta = \frac{R}{G\sqrt{T}}$,
$\alpha=\frac{G\sqrt{T}}{R}$,
$\eta = \frac{G}{2 R\sqrt{T}}$.
In order to simulate an inexact sub-gradient, we added a mean-zero normal distribution with covariance matrix $\sigma^2 I_n$ to the exact sub-gradient, i.e., $\hat{g}(y_k) =g(y_k) + N_k$ where $\{N\}_{k=1}^T$ are i.i.d, $n$ dimensional vector, samples of $\mathcal{N}(0,\sigma^{2} I_n)$. Thus,
\begin{displaymath}
\mathbb{E}\{\|\hat{g}(x)\|^2\} = G^2+n\sigma^2
\end{displaymath}
so
\begin{displaymath}
B = \sqrt{ G^2+n\sigma^2}
\end{displaymath}
We choose the parameters as following:
$\beta = \frac{R}{B\sqrt{T}}$,
$\alpha=\frac{B\sqrt{T}}{R}$,
$\eta = \frac{G}{2 R\sqrt{T}}$.
The experiments were done for different noise levels, $\sigma$, different dimensions, $n$, and for different values of $T$. Two cases are considered in which the parameter $\omega$ is inside or outside of the domain set, $\mathcal{X}$. Clearly, in all of the figures, by increasing $T$, the error decreases monotonically for both algorithms.
Four first figure show the results of the simulations for $\omega \notin \mathcal{X}$. First observation from \cref{fig:Ex1_out_n=10} is that both algorithms are sensitive to the noise levels in low dimension (i.e. parameter $n$). Another observation is that our algorithm gives a slightly better result in high noise levels while the PGD/SGD works better in the low noises. In the case of no-noise, they have the same performance but our algorithm shows a more smooth convergence curve. By gradually increasing the dimension from $10$ in \cref{fig:Ex1_out_n=10} to $500$ in the next three figures: \cref{fig:Ex1_out_n=100}, \cref{fig:Ex1_out_n=250}, and \cref{fig:Ex1_out_n=500}, both algorithms start to show less sensitivity
to the noise levels. Still our algorithm gives a slightly better result in high noise levels while the PGD/SGD works better in the low noise levels.
The next four first figure show the results of the simulations for $\omega \in \mathcal{X}$. First observation from \cref{fig:Ex1_in_n=10} is that both algorithms are sensitive to the noise levels in low dimension (i.e. parameter $n$). While both algorithms perform similarly in the presence of the noise, PGD/SGD algorithm shows a very non-smooth convergence curve in the absence of the noise. By gradually increasing the dimension from $10$ in \cref{fig:Ex1_in_n=10} to $500$ in the next three figures: \cref{fig:Ex1_in_n=100}, \cref{fig:Ex1_in_n=250}, and \cref{fig:Ex1_in_n=500}, both algorithms perform similarly in the presence of the noise. PGD/SGD algorithm shows a non-smooth and inferior convergence curve in the absence of the noise.
\newpage
\begin{figure}[H]
\label{fig:Ex1_out_n=10}
\centering
\includegraphics[width=0.76\linewidth]{image_out/W=out_dim=10.eps}
\caption{Dimension $n=10$ and $\omega$ is outside the set $\mathcal{X}$. Both algorithms show a lot of sensitivity to the noise levels in low dimension. Another observation is that our algorithm gives a slightly better result in high noise levels while the PGD/SGD works better in the low noises. In the case of no-noise, they have the same performance but our algorithm shows a more smooth convergence curve.}
\end{figure}
\begin{figure}[H]
\label{fig:Ex1_out_n=100}
\centering
\includegraphics[width=0.76\linewidth]{image_out/W=out_dim=100.eps}
\caption{Dimension $n=100$ and $\omega$ is outside the set $\mathcal{X}$. Both algorithms show less sensitivity to the noise levels in a higher dimension. Again, our algorithm gives a slightly better result in high noise levels while the PGD/SGD works better in the low noises.}
\end{figure}
\begin{figure}[H]
\label{fig:Ex1_out_n=250}
\centering
\includegraphics[width=0.8\linewidth]{image_out/W=out_dim=250.eps}
\caption{Dimension $n=250$ and $\omega$ is outside the set $\mathcal{X}$. Both algorithms show less sensitivity to the noise levels in a higher dimension. Again, our algorithm gives a slightly better result in high noise levels while the PGD/SGD works better in the low noises.}
\end{figure}
\begin{figure}[H]
\label{fig:Ex1_out_n=500}
\centering
\includegraphics[width=0.8\linewidth]{image_out/W=out_dim=500.eps}
\caption{Dimension $n=500$ and $\omega$ is outside the set $\mathcal{X}$. Both algorithms show less sensitivity to the noise levels in a higher dimension. Again, our algorithm gives a slightly better result in high noise levels while the PGD/SGD works better in the low noises.}
\end{figure}
\begin{figure}[H]
\label{fig:Ex1_in_n=10}
\centering
\includegraphics[width=0.78\linewidth]{image_in/W=in_dim=10.eps}
\caption{Dimension $n=10$ and $\omega$ is inside the set $\mathcal{X}$. Both algorithms show a lot of sensitivity to the noise levels in low dimension. While both algorithms perform similarly in the presence of the noise, PGD/SGD algorithm shows a very non-smooth convergence curve in the absence of the noise.}
\end{figure}
\begin{figure}[H]
\label{fig:Ex1_in_n=100}
\centering
\includegraphics[width=0.78\linewidth]{image_in/W=in_dim=100.eps}
\caption{Dimension $n=100$ and $\omega$ is inside the set $\mathcal{X}$. Again, both algorithms perform similarly in the presence of the noise, PGD/SGD algorithm shows a non-smooth and inferior convergence curve in the absence of the noise.}
\end{figure}
\begin{figure}[H]
\label{fig:Ex1_in_n=250}
\centering
\includegraphics[width=0.8\linewidth]{image_in/W=in_dim=250.eps}
\caption{Dimension $n=250$ and $\omega$ is inside the set $\mathcal{X}$. Again, both algorithms perform similarly in the presence of the noise, PGD/SGD algorithm shows a non-smooth and inferior convergence curve in the absence of the noise.}
\end{figure}
\begin{figure}[H]
\label{fig:Ex1_in_n=500}
\centering
\includegraphics[width=0.8\linewidth]{image_in/W=in_dim=500.eps}
\caption{Dimension $n=500$ and $\omega$ is inside the set $\mathcal{X}$.Again, both algorithms perform similarly in the presence of the noise, PGD/SGD algorithm shows a non-smooth and inferior convergence curve in the absence of the noise.}
\end{figure}
\newpage
\subsection{Bounded nuclear norm domain with \texorpdfstring{$L1$}{Lg}-norm objective function}
The bounded nuclear norm domain, is a well known example in which the linear optimization has significant computational advantage over projection.
\begin{definition}[Singular Value Decomposition]
SVD of a real valued $m\times n$ matrix $A$ is a factorization of the form $A=U S V^\top$, where U is an $m\times m$ orthogonal matrix, $S$ is an diagonal matrix with non-negative real numbers on the diagonal, and $V$ is an $n\times n$ orthogonal matrix. The diagonal entries $s_i=S_{i,i}$ are uniquely determined by $A$ and are known as the singular values of $A$. Also, let us call the $i$-th column of matrix $U$ by $u_i$ and the $i$-th column of matrix $V$ by $v_i$.
\end{definition}
\begin{definition}[Nuclear norm] Nuclear norm or Trace norm of a $m\times n$ matrix $A$ with SVD $A=US V^\top$ is defined as:
\begin{displaymath}
\|A\|_*=\sum_{i=1}^{\min(m,n)} s_i
\end{displaymath}
\end{definition}
Now let us define the set $\mathcal{X}$ of all real valued $m\times n$ matrices with nuclear norm smaller that $\tau$. Fix $\tau>0$,
\begin{displaymath}
\mathcal{X}\overset{def}{=}\{A\in\mathbb{R}^{m\times n} : \|A\|_*\leq \tau\}
\end{displaymath}
Until this section, algorithms were defined in the vector form. However, they can work for matrices just by rearranging the matrices into vector forms. Frobenius norm of a matrix is equal to its vector form $L_2$-norm.
\begin{definition}[Frobenius norm] Frobenius norm can be defined as:
\begin{displaymath}
\|A\|_F = \sqrt{\sum_{i,j=1}^{m,n}|a_{i,j}|^2}
\end{displaymath}
\end{definition}
\begin{lemma}
Frobenius norm can be upper bounded by Nuclear norm:
\begin{displaymath}
\|A\|_F\leq\|A\|_*
\end{displaymath}
\end{lemma}
\begin{proof}
See \cite{golub2013matrix}.
\end{proof}
Now we are ready to come up with a enclosing Euclidean ball for the set $\mathcal{X}$. If we choose $ X_1=\mathbf{0}$ then
\begin{displaymath}
\|X_1-Z\|_F = \|Z\|_F\leq \|Z\|_* \leq \tau, \quad\forall Z\in \mathcal{X}
\end{displaymath}
Let us choose the objective function
\begin{displaymath}
f_W(X)=\sum_{i=1}^m\sum_{j=1}^n |X_{i,j}-W_{i,j}|
\end{displaymath}
It is easy to see that this function is convex and Lipschitz continuous with coefficient
\begin{displaymath}
G = \sqrt{n m}
\end{displaymath}
When the exact sub-gradient is given to us without any error, we choose the parameters as before:
$\beta = \frac{R}{G\sqrt{T}}$,
$\alpha=\frac{G\sqrt{T}}{R}$,
$\eta = \frac{G}{2 R\sqrt{T}}$.
In order to simulate inexact sub-gradients, we added a mean-zero normal distribution with covariance matrix $\sigma^2 I_{n m}$ to the exact sub-gradients, i.e., $\hat{g}(y_k) =g(y_k) + N_k$ where $\{N_k\}$ are i.i.d, $nm$ vectors, samples of $\mathcal{N}(0,\sigma^2 I_{n m})$. Thus,
\begin{displaymath}
\mathbb{E}\{\|\hat{g}(x)\|^2\} = G^2+n m \sigma^2
\end{displaymath}
so
\begin{displaymath}
B = \sqrt{ G^2+n m \sigma^2}
\end{displaymath}
Thus the parameters are:
$\beta = \frac{R}{B\sqrt{T}}$,
$\alpha=\frac{B\sqrt{T}}{R}$,
$\eta = \frac{G}{2 R\sqrt{T}}$.
We used the two following lemmas to implement our algorithm and PGD/SGD:
\begin{lemma}[Linear optimization on nuclear-norm ball]
Fix a parameter $\tau>0$. Let $A$ be a $m\times n$ matrix and consider its singular-value decomposition $A=U S V^\top$. Then
\begin{displaymath}
\argmin_{\|X\|_*\leq\tau} \langle A,X\rangle_F
=
\tau u_1 v_1^\top
\end{displaymath}
\end{lemma}
\begin{proof}
See \cite{jaggi2013revisiting}.
\end{proof}
\begin{lemma}[Projection onto the nuclear-norm ball]
Fix a parameter $\tau>0$. Let $A$ be a $m\times n$ matrix and consider its singular-value decomposition $A=U S V^\top$.
If $\|A\|_*\geq \tau$, then the Euclidean projection of $A$ onto the nuclear-norm ball is given by
\begin{displaymath}
\pi_{\|X\|_*\leq\tau} (A) = \sum_{i=1}^{\min(m,n)}
\max(0,s_i-\lambda) u_i v_i^\top
\end{displaymath}
where $\lambda\geq0$ is the solution to the equation
\begin{displaymath}
\sum_{i=1}^{\min(m,n)}
\max(0,s_i-\lambda) = \tau
\end{displaymath}
\end{lemma}
\begin{proof}
See \cite{beck2017first}, \cite{garber2021convergence}.
\end{proof}
The experiments were done for different noise levels, $\sigma$, different dimensions, and for different values of $T$. Two cases are considered in which the parameter $W$ is inside or outside of the domain set, $\mathcal{X}$. When $W$ is inside, it is the optimal point. However, when the $W$ is outside, then the optimal point is not calculable analytically (unlike the first experiment) so we only demonstrate the value of the function $f(\Bar{x})$ instead of the error $f(\Bar{x})-f(x^*)$. Clearly, in all of the figures, by increasing $T$, the error decreases monotonically for both algorithms.
Four first figure show the results of the simulations for $W \notin \mathcal{X}$. First observation from \cref{fig:Ex2_out_N=5_M=5} is that PGD/SGD shows a superior result over our algorithm in the low noise levels. This superiority vanishes in the higher noise levels. By gradually increasing the dimensions in the next three figures: \cref{fig:Ex2_out_N=5_M=10}, \cref{fig:Ex2_out_N=10_M=10}, and \cref{fig:Ex2_out_N=10_M=20}, PGD/SGD, still, shows a superior result over our algorithm in the low noise levels. This superiority vanishes in the higher noise levels and our algorithm starts to do a better job. Clearly, PGD/SGD is more sensitive to the noise level.
The next four first figure show the results of the simulations for $W \in \mathcal{X}$. First observation from \cref{fig:Ex2_in_N=5_M=5} is that in the absence of the noise, PGD/SDG does an inferior job. However, introducing small amount of noise improves the SGD/PGD performance dramatically. In other words, our algorithm is showing a superior result over PGD/SDG in the low noise levels, this superiority vanishes in the higher noise levels. By gradually increasing the dimension in the next three figures: \cref{fig:Ex2_in_N=5_M=10}, \cref{fig:Ex2_in_N=10_M=10}, and \cref{fig:Ex2_in_N=10_M=20}, still, in the absence of the noise, PGD/SDG does an inferior job. Again, this superiority vanishes in the higher noise levels. One more time, introducing small amount of noise improve the SGD/PGD performance dramatically.
\newpage
\begin{figure}[H]
\label{fig:Ex2_out_N=5_M=5}
\centering
\includegraphics[width=0.9\linewidth]{image_2_out/W=out_N=5_M=5.eps}
\caption{Parameters are $n=5$, $m=5$. $W$ is outside of the set $\mathcal{X}$. PGD/SGD is showing a superior result over our algorithm in the low noise levels. This superiority vanishes in the higher noise levels.}
\end{figure}
\begin{figure}[H]
\label{fig:Ex2_out_N=5_M=10}
\centering
\includegraphics[width=0.9\linewidth]{image_2_out/W=out_N=5_M=10.eps}
\caption{Parameters are $n=5$, $m=10$. $W$ is outside of the set $\mathcal{X}$. PGD/SGD is showing a superior result over our algorithm in the low noise levels. This superiority vanishes in the higher noise levels.}
\end{figure}
\begin{figure}[H]
\label{fig:Ex2_out_N=10_M=10}
\centering
\includegraphics[width=0.9\linewidth]{image_2_out/W=out_N=10_M=10.eps}
\caption{Parameters are $n=10$, $m=10$. $W$ is outside of the set $\mathcal{X}$. PGD/SGD is showing a superior result over our algorithm in the low noise levels. This superiority vanishes in the higher noise levels and our algorithm does a better job. Clearly, PGD/SGD is more sensitive to the noise level.}
\end{figure}
\begin{figure}[H]
\label{fig:Ex2_out_N=10_M=20}
\centering
\includegraphics[width=0.9\linewidth]{image_2_out/W=out_N=10_M=20.eps}
\caption{Parameters are $n=10$, $m=20$. $W$ is outside of the set $\mathcal{X}$. PGD/SGD is showing a superior result over our algorithm in the low noise levels. This superiority vanishes in the higher noise levels and our algorithm does a better job. Clearly, PGD/SGD is more sensitive to the noise level.}
\end{figure}
\begin{figure}[H]
\label{fig:Ex2_in_N=5_M=5}
\centering
\includegraphics[width=\linewidth]{image_2_in/W=in_N=5_M=5.eps}
\caption{Parameters are $n=5$, $m=5$. $W$ is inside of the set $\mathcal{X}$. In the absence of the noise, PGD/SDG does an inferior job. However, introducing small amount of noise improves the SGD/PGD performance dramatically. In other words, our algorithm is showing a superior result over PGD/SDG in the low noise levels, this superiority vanishes in the higher noise levels.}
\end{figure}
\begin{figure}[H]
\label{fig:Ex2_in_N=5_M=10}
\centering
\includegraphics[width=\linewidth]{image_2_in/W=in_N=5_M=10.eps}
\caption{Parameters are $n=5$, $m=10$. $W$ is inside of the set $\mathcal{X}$. In the absence of the noise, PGD/SDG does an inferior job. However, introducing small amount of noise improve the SGD/PGD performance dramatically. In other words, our algorithm is showing a superior result over PGD/SDG in the low noise levels, this superiority vanishes in the higher noise levels.}
\end{figure}
\begin{figure}[H]
\label{fig:Ex2_in_N=10_M=10}
\centering
\includegraphics[width=\linewidth]{image_2_in/W=in_N=10_M=10.eps}
\caption{Parameters are $n=10$, $m=10$. $W$ is inside of the set $\mathcal{X}$. In the absence of the noise, PGD/SDG does an inferior job. However, introducing small amount of noise improve the SGD/PGD performance dramatically. In other words, our algorithm is showing a superior result over PGD/SDG in the low noise levels, this superiority vanishes in the higher noise levels.}
\end{figure}
\begin{figure}[H]
\label{fig:Ex2_in_N=10_M=20}
\centering
\includegraphics[width=\linewidth]{image_2_in/W=in_N=10_M=20.eps}
\caption{Parameters are $n=10$, $m=20$. $W$ is inside of the set $\mathcal{X}$. In the absence of the noise, PGD/SDG does an inferior job. However, introducing small amount of noise improve the SGD/PGD performance dramatically. In other words, our algorithm is showing a superior result over PGD/SDG in the low noise levels, this superiority vanishes in the higher noise levels.}
\end{figure}
\newpage
\section{Conclusions and Open Problems}
This paper addresses the problem of solving general constrained convex optimization without using projection. Prior works on projection-free algorithms focus mainly on smooth problems and/or problems with special constraint structure. The simulations and the convergence theorems show that our algorithm's performance is comparable with the PGD/SGD algorithm, thus it can be a substitute for the PGD/SGD algorithm if projection-free algorithms are favoured.
An open problem is whether our algorithm can be modified to exploit the gains of mirror descent \cite{doi:10.1137/1027074,beck2003mirror}. A well known algorithm that replaces the Euclidean norm in PGD/SGD with a Bregman divergence which results in improved performance with respect to system dimension in certain cases (such as when the domain set is a probability simplex). Another question arises considering that PGD/SGD with time varying step-size reaches the optimal rate, $O(1/T)$, for non-smooth strongly convex optimization. It is unclear how to extend our algorithm in this scenario.
|
\section{Introduction} \label{introduction}
\begin{center}
\begin{figure*}
\includegraphics[width=0.65\textwidth]{Porter-Temple22_f1.pdf}
\caption{Galaxy Zoo 4 GAMA-KiDS decision tree. The decision tree can be viewed at \url{https://data.galaxyzoo.org/gz_trees/gz_trees.html} under the 'GZ GAMA-KiDS' section.}
\label{fig:GZ4}
\end{figure*}
\end{center}
Though they are visually distinctive, the exact properties that affect the formation of arms in spiral galaxies are not yet fully explored.
The potential and observed links between spiral galaxy structure and pitch angle, arm strength \citep{Seigar98,Yu20,Kendall14} are now being examined with increasing samples sizes and sophistication in analysis \citep{Hart17a,Yu18a,Lingard21}. The motivation for increased interest in spiral morphology in representative samples is to constrain the dominant formation mechanism of spiral arm structure \citep{Masters21}. For example, \cite{Pringle19a} find a constant distribution of pitch angles of arms, consistent with the density wave theory origin of spiral structure. \cite{Hart18} found 40 per cent of arm formation in massive spirals can be by ``swing amplification"; the number of arms is consistent with the prediction from this mechanism with the remainder originating from other mechanisms.
\cite{Diaz-Garcia19c} do not find observational evidence that spiral arms are driven by stellar bars \citep[as do][]{Hart17a} or through ``manifolds", pathways of infalling material, which would show as a dependence of arm strength and pitch angle \citep{Athanassoula09}. They found that bar and arm strength are correlated, while bar strength and pitch angle are not. In multi-wavelength data, \cite{Yu18b} found younger stars to reside in tighter arms and \cite{Miller19a} found that these stars then trailed out from the arms.
\cite{Seigar98} found no correlation between pitch angle and Hubble type, which was reiterated by \cite{Kendall14, Yu18}; however, they note that not finding a correlation is unsurprising given the small range of pitch angles they examined. \cite{Yu20} later found a loose correlation between pitch angle and spiral arm strength, with an overall tendency for pitch angle to decrease with weaker arm strength, while \cite{Savchenko20} find no strong difference (except for number of arms) between grand design, multi-armed, and flocculent spirals in pitch angle, arm width or strength. The link between arm strength, pitch angle and formation mechanisms remains complex.
Instead of focusing on the formation of spiral arm structure, one can examine the correlations with global properties of the galaxies such as star formation, stellar mass or specific star formation.
\cite{Hart17} investigated spiral structure using the Sloan Digital Sky Survey (SDSS) main galaxy sample, with morphological data from the public release of Galaxy Zoo 2 \citep{Willett13}, stellar mass from \cite{Chang15}, and star formation from GALEX fluxes \citep{Martin05}. Using these data, they determined no significant dependence of spiral arm number on specific star formation rate (sSFR).
In this paper, we explore the connection of spiral arm number with stellar mass, star formation rate (SFR), and specific star formation rate (sSFR), using similar methods to \cite{Hart17}. We make use of the improved star formation and stellar mass estimates by the Galaxy And Mass Assembly \citep[GAMA,][]{Driver09,Liske15} survey using self-consistent {\sc magphys} SED fits to the full uv to sub-mm SED \citep{Driver16c,Wright16} and GalaxyZoo voting base on deeper and higher-resolution KiDS data \cite[][Kelvin et al. \textit{in prep.}]{Holwerda19b}. With this improved quality data, we investigate the trends with spiral arm numbers that the results from \cite{Hart17} suggested. We compare spiral arm number subsamples of stellar mass, SFR, and sSFR to the whole set of galaxies to determine any notable differences. The sets defined by a spiral number (m = 1, 2, 3, 4, 5+) are from the visual classification from the GAMA-KiDS Galaxy Zoo project, detailed in Section \ref{galaxy zoo}. This paper is organized as follows: section \ref{data} describes the data used in the paper and how subsamples are defined, section \ref{results} presents the results for star formation, stellar mass, and specific star formation as a function of the number of spiral arms,
section \ref{discussion} discusses these results and section \ref{conclusions} lists our conclusions.
\begin{table*}
\large
\centering
\renewcommand{\arraystretch}{1.2}
\begin{tabular}{llcccccccccccc} \toprule
& & \multicolumn{2}{c}{2-Sample K-S Test} & \multicolumn{2}{c}{k-Sample A-D Test} & \multicolumn{7}{c}{k-Sample A-D Test Critical Values} \\ \cmidrule(lr){3-4} \cmidrule(lr){5-6} \cmidrule(lr){7-13}
& & Statistic & Significance & Statistic & Significance & 25\% & 10\% & 5\% & 2.5\% & 1\% & 0.5\% & 0.1\%\\\midrule
\multirow{5}{*}{ \rotatebox{90}{Stellar Mass} }
& m=1 & 0.138 & \textbf{0.038} & 3.977 & 0.008 & 0.33 & 1.23 & 1.96 & 2.72 & \textbf{3.75} & 4.59 & 6.55\\
& m=2 & 0.034 & 0.203 & 0.763 & 0.159 & \textbf{0.33} & 1.23 & 1.96 & 2.72 & 3.75 & 4.59 & 6.55\\
& m=3 & 0.152 & \textbf{0.001} & 6.642 & 0.001 & 0.33 & 1.23 & 1.96 & 2.72 & 3.75 & 4.59 & \textbf{6.55}\\
& m=4 & 0.252 & 0.081 & -0.071 & 0.250 & \textbf{0.33} & 1.23 & 1.96 & 2.72 & 3.75 & 4.59 & 6.55\\
& m=5+ & 0.264 & \textbf{0.000} & 10.216 & 0.001 & 0.33 & 1.23 & 1.96 & 2.72 & 3.75 & 4.59 & \textbf{6.55}\\ \midrule
\multirow{5}{*}{ \rotatebox{90}{SFR} }
& m=1 & 0.099 & 0.256 & 0.278 & 0.250 & \textbf{0.33} & 1.23 & 1.96 & 2.72 & 3.75 & 4.59 & 6.55\\
& m=2 & 0.110 & \textbf{0.000} & 36.514 & 0.001 & 0.33 & 1.23 & 1.96 & 2.72 & 3.75 & 4.59 & \textbf{6.55}\\
& m=3 & 0.281 & \textbf{0.000} & 39.523 & 0.001 & 0.33 & 1.23 & 1.96 & 2.72 & 3.75 & 4.59 & \textbf{6.55}\\
& m=4 & 0.291 & \textbf{0.028} & 4.342 & 0.006 & 0.33 & 1.23 & 1.96 & 2.72 & \textbf{3.75} & 4.59 & 6.55\\
& m=5+ & 0.217 & \textbf{0.004} & 7.721 & 0.001 & 0.33 & 1.23 & 1.96 & 2.72 & 3.75 & 4.59 & \textbf{6.55}\\ \midrule
\multirow{5}{*}{ \rotatebox{90}{sSFR} }
& m=1 & 0.234 & \textbf{0.000} & 9.270 & 0.001 & 0.33 & 1.23 & 1.96 & 2.72 & 3.75 & 4.59 & \textbf{6.55}\\
& m=2 & 0.072 & \textbf{0.000} & 19.745 & 0.001 & 0.33 & 1.23 & 1.96 & 2.72 & 3.75 & 4.59 & \textbf{6.55}\\
& m=3 & 0.187 & \textbf{0.000} & 10.309 & 0.001 & 0.33 & 1.23 & 1.96 & 2.72 & 3.75 & 4.59 & \textbf{6.55}\\
& m=4 & 0.261 & 0.064 & 0.901 & 0.139 & \textbf{0.33} & 1.23 & 1.96 & 2.72 & 3.75 & 4.59 & 6.55\\
& m=5+ & 0.143 & 0.126 & 1.976 & 0.050 & 0.33 & 1.23 & \textbf{1.96} & 2.72 & 3.75 & 4.59 & 6.55\\\bottomrule
\end{tabular}
\caption{\label{tab:statstable} Spiral arm number (m), Kolmogorov-Smirnov test statistic and significance for stellar mass, star formation rate, and specific star formation rate are shown under the header 2-sample K-S Test. Bold values for the K-S test significance are the statistically significant values discussed in section \ref{discussion}. The Anderson Darling test statistic and estimated significance level for Stellar Mass, SFR, and sSFR are shown under the header k-sample A-D Test. The critical values for different levels of significance are listed, with the critical value that each subsample meets in bold. The Anderson Darling test significance estimates are floored at 0.1\% and capped at 25\%.}
\end{table*}
\section{Data} \label{data}
The data used comes from the Galaxy and Mass Assembly (GAMA) survey \citep{Driver09, Liske15}. We use the GAMA DR3 \citep{Baldry18} and the Kilo Degree Survey \citep[KiDS,][]{de-Jong13,de-Jong15,de-Jong17,Kuijken19} imaging. Additionally, we use the MAGPHYS table described in the GAMA DR3. MAGPHYS computes the stellar mass and specific star formation rate used and is fully described in \cite{da-Cunha08}.
\subsection{GAMA} \label{gama}
GAMA is a combined spectroscopic and multi-wavelength imaging survey designed to study spatial structure in the nearby ($z < 0.25$) Universe on kpc to Mpc scales \citep[see][for an overview]{Driver09, Driver11}. The survey, after completion of phase 2 \citep{Liske15}, consists of three equatorial regions each spanning 5 deg in Dec and 12 deg in RA, centered in RA at approximately 9h (G09), 12h (G12) and 14.5h (G15) and two Southern fields, at 05h (G05) and 23h (G23). The three equatorial regions, amounting to a total sky area of 180 deg$^2$, were selected for this study. For the purpose of visual classification, 49,851 galaxies were selected from the equatorial fields with redshifts $z<0.15$ (see below). The GAMA survey is $>$98\% redshift complete to r $<$ 19.8 mag in all three equatorial regions. We use the {\sc magphys} SED fits data-products \citep{Driver18} from the third GAMA data-release \citep[DR3,][]{Baldry18}.
\subsection{KiDS} \label{kids}
The Kilo Degree Survey \citep[KiDS,][]{de-Jong13,de-Jong15,de-Jong17,Kuijken19} is an ongoing optical wide-field imaging survey with the OmegaCAM camera at the VLT Survey Telescope. It aims to image 1350 deg$^2$ in four filters (u g r i). The core science driver is mapping the large-scale matter distribution in the Universe, using weak lensing shear and photometric redshift measurements. Further science cases include galaxy evolution, Milky Way structure, detection of high-redshift clusters, and finding rare sources such as strong lenses and quasars.
KiDS image quality is typically 0\farcs6 resolution (for sdss-r) and depths of 23.5, 25, 25.2, 24.2 magnitude for i, r, g and u respectively. This imaging was the input for the GalaxyZoo citizen science classifications.
\subsection{Galaxy Zoo} \label{galaxy zoo}
Information on galaxy morphology is based on the GAMA-KiDS Galaxy Zoo classification \citep[][Kelvin et al., \textit{in prep.}]{Lintott08}. The GAMA-KiDS Galaxy Zoo project is described in Kelvin et al., \textit{in prep}. RGB cutouts were constructed from KiDS g-band and r-band imaging with the green channel as the mean of these. KiDS cutouts were introduced to the classification pool and mixed in with the ongoing classification efforts.
For the Galaxy Zoo classification, 49,851 galaxies were selected from the equatorial fields with redshifts $z < 0.15$. The Galaxy Zoo provided a monumental effort with almost 2 million classifications received from over 20,000 unique users over the course of the first 12 months.
This classification has been used by the GAMA team to identify dust lanes in edge-on galaxies \citep{Holwerda19}, searches for strong lensing galaxy pairs \citep{Knabel20}, and the morphology of green valley galaxies (Smith et al {\em in prep}.).
In this paper we use the visual classifications of spiral galaxies from the Galaxy Zoo project; the full decision tree for the GAMA-KiDS Galaxy Zoo project is shown in Figure \ref{fig:GZ4}.
\subsection{MAGPHYS SED} \label{magphys}
In addition to the GAMA-KiDS Galaxy Zoo classifications, we use the {\sc magphys} \citep{da-Cunha08}, spectral energy distribution fits to the GAMA multi-wavelength photometry \citep{Wright17}, presented in \cite{Driver18}. {\sc magphys} computes stellar mass, star formation rate, and specific star formation rates which will serve as comparison data for the Galaxy Zoo arm classifications.
\subsection{Sample Selection} \label{sample selection}
\begin{figure}
\centering
\includegraphics[scale=0.49]{Porter-Temple22_f2.png}
\caption{Stellar Mass vs. Redshift for the GAMA-KiDS Galaxy Zoo project data. The limited sample, which includes only those galaxies with $z < 0.08$ and $M_* > 10^9$, is indicated by the red box. Only galaxies with 30\% or more votes in favor of being a spiral galaxy are included.}
\label{fig:redshiftLIM}
\end{figure}
To be included in the subset of the GAMA-KiDS Galaxy Zoo project used (herein after referred to as 'the limited sample'), a galaxy must meet three criteria. First, the galaxy must have a stellar mass \(M_* > 10^9\). Any galaxies below that limit are excluded. Second, the galaxy must have received at least 30\% of votes in favor of it being a spiral galaxy. This is represented by question T03 in the Galaxy Zoo decision tree shown in Figure \ref{fig:GZ4}. This avoids galaxies that were misclassified as spiral galaxies due to a low number of votes. Third, included galaxies must have a redshift less than 0.08, meaning any galaxies with $z \geq 0.08$ are not included in the limited sample. Doing this excludes those galaxies whose spiral arms are not correctly represented by Galaxy Zoo votes because of unclear imaging or lack of distinction between, for example, two-armed and four-armed spirals at $z\geq 0.08$.
The limits on the limited sample from the full GAMA-KiDS Galaxy Zoo project is shown in Figure \ref{fig:redshiftLIM}.
\subsection{Defining Subsamples} \label{subsample}
Each subsample of spiral galaxies is defined by their spiral arm number as voted by Galaxy Zoo participants. This is represented by question T06 in Figure \ref{fig:GZ4}, with answers A0, A1, ..., A4 being classified in this paper as m=1, m=2, ..., m=5+.
In addition to fulfilling all the criteria described in section \ref{sample selection}, to fall into any given subsample m=x, a galaxy must meet two additional criteria. First, it must have received at least 50\% of votes in favor of having x spiral arms; that is, a galaxy is in the m=x subsample if the fraction of votes in favor of x arms is \( > 0.5\). The cutoff at 50\% means that the majority of votes dictates what subsample the galaxy falls into, so no galaxy falls into more than one subsample. Second, the galaxy must have less than 100\% of votes in favor of it having x spiral arms. This eliminates some galaxies that have a very low number of votes. So, a galaxy that with a fraction of votes $f_m$ in the range $(0.5 < f_m < 1)$ for answer A0 in Table \ref{fig:GZ4} would be included in the m=1 subsample, and similar for m=2, 3, 4, and 5+ spiral arms.
\section{Results} \label{results}
The limited sample of galaxies is compared with each subsample as determined in section \ref{subsample}, with respect to stellar mass, SFR, and sSFR. The number of galaxies N given in each subsample is shown in Table \ref{tab:statstable}, along with the Kolmogorov–Smirnov test (K-S test) statistic and p-value.
The K-S test statistic indicated how similar the subsample is to the parent sample, with smaller values being more similar and larger values being less similar, where a statistic of 0.0 indicates two identical distributions. The p-value associated with each K-S statistic dictates the significance in the K-S statistic, and we consider a p-value of .05 or lower to be significant.
For an additional test of sample similarity, we perform the Anderson-Darling test on the above samples with the resulting statistic and p-values also listed in Table \ref{tab:statstable}. The critical values for each A-D test are returned for different levels of confidence and we bold the value that is exceeded by the A-D statistic in each case. The benefit of the A-D test over the K-S test is that it identifies confidence levels independently from the reported p-value. The A-D test is much more sensitive to the tails of any distribution and the K-S test is more dependent of the center of distribution. As our distributions are all non-Gaussian, this makes the A-D test better suited for the comparison.
Broadly the K-S and A-D tests agree on which populations differ but they disagree on the level of significance. For example Stellar Mass and one arm (m=1) or SFR (m=4), the A-D test assigns higher significance to the difference. We note that the K-S test reports a small, but low significance difference for the m=5+ sSFR distribution but the A-D identified a (just) significant result (5\% critical value exceeded, Table \ref{tab:statstable}).
\begin{figure*}
\centering
\includegraphics[width=.32\textwidth]{Porter-Temple22_f3a.png}\quad
\includegraphics[width=.32\textwidth]{Porter-Temple22_f3b.png}
\smallskip
\includegraphics[width=.32\textwidth]{Porter-Temple22_f3c.png}\quad
\includegraphics[width=.32\textwidth]{Porter-Temple22_f3d.png}
\smallskip
\includegraphics[width=.32\textwidth]{Porter-Temple22_f3e.png}
\caption{Stellar mass histograms for each of the subsamples selected from the limited GAMA-KiDS Galaxy Zoo sample. The gray filled histogram shows the distributions of the entire limited data set, while the colored outlines show the distribution for the individual spiral arm number subsamples.}
\label{pics:stellarmassfig}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{Porter-Temple22_f4.png}
\caption{Stellar mass distribution densities for each of the spiral arm subsamples. The shaded gray region indicates $\pm$1 standard deviation of the whole sample. The dotted gray line indicates the mean for the whole sample. Each spiral arm distribution shows the high range, mean, and low range, indicated by horizontal dash marks. The number of galaxies in each subsample is shown above each distribution.}
\label{fig:stmsummary}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=.32\textwidth]{Porter-Temple22_f5a.png}\quad
\includegraphics[width=.32\textwidth]{Porter-Temple22_f5b.png}
\smallskip
\includegraphics[width=.32\textwidth]{Porter-Temple22_f5c.png}\quad
\includegraphics[width=.32\textwidth]{Porter-Temple22_f5d.png}
\smallskip
\includegraphics[width=.32\textwidth]{Porter-Temple22_f5e.png}
\caption{SFR histograms for each of the subsamples selected from the limited GAMA-KiDS Galaxy Zoo sample. The gray filled histogram shows the distributions of the entire limited data set, while the colored outlines show the distribution for the individual spiral arm number subsamples.}
\label{pics:sfrfig}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{Porter-Temple22_f6.png}
\caption{SFR distribution densities for each of the spiral arm subsamples. The shaded regions are equivalent to the definitions in Figure \ref{fig:stmsummary}.}
\label{fig:sfrsummary}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=.32\textwidth]{Porter-Temple22_f7a.png}\quad
\includegraphics[width=.32\textwidth]{Porter-Temple22_f7b.png}
\smallskip
\includegraphics[width=.32\textwidth]{Porter-Temple22_f7c.png}\quad
\includegraphics[width=.32\textwidth]{Porter-Temple22_f7d.png}
\smallskip
\includegraphics[width=.32\textwidth]{Porter-Temple22_f7e.png}
\caption{sSFR histograms for each of the subsamples selected from the limited GAMA-KiDS Galaxy Zoo sample. The gray filled histogram shows the distributions of the entire limited data set, while the colored outlines show the distribution for the individual spiral arm number subsamples.}
\label{pics:specificfig}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{Porter-Temple22_f8.png}
\caption{sSFR distribution densities for each of the spiral arm subsamples. The shaded regions are equivalent to the definitions in Figure \ref{fig:stmsummary}.}
\label{fig:SSFRsummary}
\end{figure}
\subsection{Stellar Mass} \label{results: stellar mass}
Figure \ref{pics:stellarmassfig} shows the histograms resulting from the process described in sections \ref{sample selection} and \ref{subsample}, and from the K-S test described above.
The m=1, m=3, and m=5 subsamples show visual differences in their stellar mass distributions. The m=5 subsample, though limited by a small number of galaxies, shows a notable shift towards higher stellar masses. Likewise the m=3 subsample tends towards higher masses as well, with the peak falling just below 10.0 for stellar mass, versus the peak at 9.5 for the limited sample. The m=1 subsample shows a tendency to lower stellar masses.
These are reflected in the K-S statistic in Table \ref{tab:statstable}, with the m=5 subsample having the greatest difference from the limited sample. The m=4 sample has the second highest statistic value, but with the lowest number of galaxies and p-value of .08 on that statistic, we do not consider it as significant. The m=1, 3, and 5 values are significant in their A-D test as well, with slightly higher significance for m=1.
Figure \ref{fig:stmsummary} shows the distributions in the colored violin plots, with the gray band indicating the median and $\pm$ 1 standard deviation for the limited sample. This also reflects the shift in distribution, with m=1 having both a lower median than the limited sample and showing a greater quantity of galaxies at lower stellar masses. Likewise, the m=3 subsample is shifted to slightly higher stellar masses, and m=5 visibly higher than the median from the limited sample.
\subsection{Star Formation Rate} \label{results: sfr}
As for stellar mass, Figure \ref{pics:sfrfig} shows the histograms resulting from the above process. The m=3, 4, and 5 distributions show a notable difference in their star formation rate distributions, with each of them having a higher SFR distribution than the limited sample. Visually, the m=3 subsample appears to have the highest distribution for SFR, and the K-S statistic reflects a greater difference from the limited sample than most other subsamples. We find that these three subsamples have a notable difference in their distributions from the limited sample, as reflected in Table \ref{tab:statstable} with their K-S statistics being much higher than the m=1 or m=2 samples. This difference is reflected again in their A-D statistics with high significance for the m=2, 3, 4, and 5+ samples. The agreement between K-S and A-D statistics is due to the mostly Gaussian shape of the distributions in SFR with only a weak tail to lower SFR.
Figure \ref{fig:sfrsummary}, as above, shows the summarized distributions for SFR. This reflects a higher average distribution for m=3, 4, and 5 subsamples. Though the m=1 and 2 subsamples appear to have higher SFR distributions than the limited sample in Figures \ref{pics:sfrfig} and \ref{fig:sfrsummary}, they also have relatively low K-S statistics compared to the other subsamples, showing a higher similarity to the limited sample than m=3, 4, or 5.
\subsection{Specific Star Formation Rate} \label{results:SSFR}
As in sections \ref{results: stellar mass} and \ref{results: sfr}, Figure \ref{pics:specificfig} shows the histograms for the sSFR values. The distribution for subsamples m=4 and m=5 are shifted to lower sSFRs. The K-S statistics for m=4,and 5 reflect these distribution shifts, but we consider the p-values of their K-S statistics to be less significant.
Conversely, we see a significant shift in the m=1 population towards higher sSFRs, and the m=3 population's distribution weighted heavily towards -10. Again this is reflected by the K-S statistics and high significance p-values in table \ref{tab:statstable}, where the m=1 subsample has a greater difference in the limited sample, and m=3 fitting the limited sample quite well. The A-D tests confirm the significance of different distributions depending on the number of spiral arms and add a significant difference for the m=5+ distribution (5\% critical value exceeded by the A-D test). This lends high confidence to the conclusion that sSFR and arm number are strongly correlated.
As above, Figure \ref{fig:SSFRsummary} shows the summarized distributions for sSFR. We see that m=1 has a higher than average sSFR compared to the other samples, and that the distribution of m=3 is more concentrated into one peak region.
\section{Discussion} \label{discussion}
In section \ref{sample selection}, we detail how the spiral arm number subsamples are defined. In categorizing them based on the Galaxy Zoo votes (question T06, shown in Figure \ref{fig:GZ4}) we are treating spiral arms as integers. However, this does not take into account whether a spiral galaxy has well-defined arms; flocculent spiral galaxies with poorly-defined or discontinuous arms can not be well classified with an integer number of arms. The voting pattern does reflect this somewhat, as the classifications came from real people voting, and so any galaxies with poorly-defined arms would be best categorized through the majority vote. So, a galaxy is classified as accurately as it can be into an integer number of spiral arms. From question T03 (Figure \ref{fig:GZ4}), we do know that these are all spiral galaxies.
In Figures \ref{pics:stellarmassfig}, \ref{pics:sfrfig}, and \ref{pics:specificfig} we can see that the low number of galaxies in each subsample does leave the m=4 and m=5+ distribution lacking in statistical weight for the K-S test results compared to, for example, the m=2 subsample. \cite{Hart17} used the optical-WISE SED inferred stellar masses and separately estimated star formation rates from either FUV flux or 22$\mu$m flux. The improvement in our data is the use of a self-consistent SED to determine both from 21 filters spanning ultraviolet through sub-mm \citep{Wright16,Driver16c}. Additionally, the A-D test results lend more statistical significance to the m=5+ distribution in particular, while giving a more significant result (to 0.1\%) for sSFR for m=1, 2, and 3. Because of the small sample size of the m=4 subsample, the A-D test (which is more sensitive to the tails of the distribution) does not show a higher significance than the m=4 sSFR result.
Overall, we find that spiral galaxies are less efficient at forming stars if they have more spiral arms. The m=1 subsample has a much lower stellar mass on average, but a higher than average distribution for sSFR (see Figures \ref{pics:stellarmassfig} and \ref{pics:specificfig}). This is supported by the findings of \cite{Hart17}, who noted that two armed spiral galaxies are more gas deficient than other galaxies, and so are more efficient at converting gas to stars.
Galaxies with stronger bars have fewer but stronger arms \citep{Yu20}, and arm strength has been found to correlate well with SFR as a function of stellar mass \citep{Yu21}. Given our results, it is unclear if the causation is more arms leads to weaker arms which in turn leads to lower sSFR. Alternatively, it is possible that the perceived change in sSFR is caused by a subtle bias in the MAGPHYS SED results (Section \ref{magphys}), because the arm patterns rearrange the dusty interstellar medium (ISM) in the disc, skewing SED measurements of star formation. Arms are more opaque than the disc \citep{kw00b,Holwerda05b}, and therefore better at hiding directly measured star formation. The many-armed spirals with low sSFR might simply be hiding their directly measurable star formation instead of having lower rates overall. However, \cite{Hart17} found that two-armed spirals have more mid-infrared (MIR) dust emission, indicating that a greater proportion of new stars in two-armed spirals are in heavily obscured region and the MAGPHYS SED result is based on balancing the missing ultraviolet light with the observed heated dust emission. Given this, it seems unlikely that low sSFRs in many-armed spirals are caused by a higher obscuration fraction of new stars.
Higher star formation for a given mass will likely highlight the spiral structure in these disks as the site of recent star formation. \cite{Hart17} note that the mean of their distribution shifts with only 0.05 dex with each additional spiral arm. We point to Figure \ref{fig:SSFRsummary} to show that the mode of the distribution is a better indication of the change with the number of arms. Between the shift in the distribution of sSFR values and the much improved star formation and stellar mass accuracy thanks to a consistent SED treatment rather than single-flux based estimates, we find the trend in lowering sSFR with number of spiral arms convincing.
\section{Conclusions} \label{conclusions}
In this paper, we examined the connection of spiral arm number with stellar mass, star formation rate (SFR), and specific star formation rate (sSFR). Using the data from GAMA DR3 and the morphological classifications from Galaxy Zoo GAMA-KiDS, we compared subsamples consisting of galaxies with 1, 2, 3, 4, or 5+ spiral arms. Overall, we find the following:
\begin{enumerate}
\item Galaxies with more spiral arms tend towards higher stellar masses (Figure \ref{fig:stmsummary}) and higher star formation rates (Figure \ref{fig:sfrsummary}).
\item Galaxies with more spiral arms tend towards lower specific star formation rates (Figures \ref{pics:specificfig} and \ref{fig:SSFRsummary}, Table \ref{tab:statstable}).
\item The single arm (m=1) subsample tends to have lower stellar mass and higher specific star formation than both the full sample and any other subsample.
\end{enumerate}
A different, non-integer classification of the number of spiral arms, allowing for the voting tally to assign fractions of spiral arms to galaxies may reflect the reality of these galaxies better. Additionally, changing the limited sample to include only galaxies with a sufficient number of votes to ensure accuracy in arm classification (as opposed to percentages of votes in favor of spiral arm pattern) may yield a higher sample size with stronger statistical significance.
The Rubin Observatory and future iterations of the Galaxy Zoo are expected to improve the statistics of spiral arm numbers on galaxies in the nearby Universe. Equally important however, are good stellar mass and star-formation estimates from SED models for similar comparisons as this work and in \cite{Hart17a}.
The Euclid and Roman space telescopes will a wealth of morphological data on higher-redshift spiral galaxies. These will allow for a direct comparison of the evolution of spiral structure.
\section*{Acknowledgements}
The material is supported by NASA Kentucky award No: 80NSSC20M0047 (NASA-REU to L. Haberzettl and R. Porter-Temple).
BWH is supported by an Enhanced Mini-Grant (EMG). The material is based upon work supported by NASA Kentucky under NASA award No: 80NSSC20M0047.
This research made use of Astropy, a community-developed core Python package for Astronomy \citep{Astropy-Collaboration13,Astropy-Collaboration18}.
\section{Data Availability}
The data for this project is available from the GAMA DR3 website \url{http://www.gama-survey.org/dr3/schema/table.php?id=82}.
|
\section*{Outline}
\begin{abstract}
Virtual reality (VR) is an emerging technology of great societal potential. Some of its most exciting and promising use cases include remote scene content and untethered lifelike navigation. This article first highlights the relevance of such future societal applications and the challenges ahead towards enabling them.
It then provides a broad and contextual high-level perspective of several emerging technologies and unconventional techniques and argues that only by their synergistic integration can the fundamental performance bottlenecks of hyper-intensive computation, ultra-high data rate, and ultra-low latency be overcome to enable untethered and lifelike VR-based remote scene immersion. A novel future system concept is introduced that embodies this holistic integration, unified with a rigorous analysis, to capture the fundamental synergies and interplay between communications, computation, and signal scalability that arise in this context, and advance its performance at the same time. Several representative results highlighting these trade-offs and the benefits of the envisioned system are presented at the end.
\end{abstract}
\setcounter{footnote}{1}
\section{Introduction and Motivation}
\label{sec:intro}
Virtual reality (VR) holds tremendous potential to advance our society. It enables visual immersion in virtual worlds created by means of computer graphics on a head-mounted display worn by a user and has found applications so far in training, education, entertainment, and gaming. An even broader set of use cases is anticipated ahead.
Looking forward, VR is expected to make impact on quality of life, energy conservation, and the economy \cite{DigiCapital:16a,ApostolopoulosCCKTW:12}, and reach a \$62B market by 2027 \cite{GrandViewResearch:20}. As the Internet-of-Things (IoT) is becoming a reality, modern technologists envision transferring remote contextual and environmental immersion experiences as part of an online VR session. In particular, together with another emerging technology, known as 360$^\circ$ video, VR can suspend our disbelief of being at a remote location, akin to {\em virtual human teleportation} -- a truly momentous advance for our society \cite{Chakareski:18a}. The present state of the world (online classes, work from home, telemedicine, and so on) due to the COVID-19 pandemic aptly illustrates the importance of remote 360$^\circ$ video VR immersion and communication, enabled in a seamless, untethered, and lifelike manner across the spectrum of our society, as illustrated in Figure~\ref{fig:VR_360_Applications}.
\begin{figure}[htb]
\centering
\includegraphics[width=1.5\figwidth]{Figures/VR_360_Applications_v2.eps}
\caption{Seamless untethered virtual reality and lifelike remote scene \tsd video for virtual traveling/teleportation, to advance first responders, environmental monitoring, remote robot navigation and teleoperation, education and training, collaborative work, healthcare and rehabilitation, urban planning, and large-scale infrastructure inspection.}
\vspace{-0.2cm}
\label{fig:VR_360_Applications}
\end{figure}
However, two main highly-intertwined communication system challenges stand in the way of realizing this vision: VR requires {\bf (1) hyper-intensive computation and (2) ultra-low latency gigabit-per-second wireless networking}. Neither of these challenges can be met by current and upcoming conventional network systems \cite{Forbes_VR_2016,EdwardKnightlyKeynoteINFOCOM2017}, as the content to be delivered is too voluminous and the VR headsets' computing and storage capabilities are insufficient within an acceptable and wearable form factor. For instance, MPEG recommends a minimum of 12K high-pixel-quality spatial resolution and 100 frames per second temporal display rate, for a 360$^\circ$ video panorama experienced by a VR user \cite{ChampelSFTK:16}. These requirements would map to a data rate of several Gbps, even after applying state-of-the-art High Efficiency Video Coding (HEVC) compression \cite{SullivanOHW:12}. Similarly, mobile GPUs lag their desktop counterparts in computing power by a factor of ten and will not have the required TFlops to provide the necessary VR decoding and rendering computation, at the minimum resolutions and frame rates indicated above, in the foreseeable future, given the current semiconductor technology trends \cite{CuervoCK:18}.
Emerging 360$^\circ$ video practices compound these challenges, as they are highly inefficient and reuse traditional video/networking technologies in virtual reality contexts without integrating their specifics \cite{Petrangeli2017}. This considerably degrades the quality of experience and application utility. The under performing of these practices is even more dramatic in mobile settings, due to the much lower wireless bandwidth and computing capability of such devices. Regrettably, this is the context where such applications are expected to have the highest societal impact, {\em advancing disaster relief, the environmental sciences, public safety, transportation, search and rescue, and urban planning}, among others.
Therefore, these considerable challenges and shortcomings limit present VR applications to off-line operation, low-fidelity graphics content, tethered high-end computing equipment, and predominantly gaming and entertainment settings.
\section{Objectives of the Feature Article}
This article has multiple objectives of educating the broader Signal Processing Magazine readership. Its first aim is to educate about the importance of enabling next generation virtual reality applications comprising high-fidelity remote scene immersion and seamless untethered lifelike navigation of the reconstructed remote environment. Its second aim is to emphasize that traditional technology upgrade cycles alone would not suffice to bridge the performance gap to make such applications possible and that a holistic integration of unconventional techniques and emerging technologies would be required instead. Its third aim is to provide a tutorial of these methods and their synergistic interplay towards enabling the envisioned next generation applications. Finally, its fourth aim is to illustrate a case study of rigorous high-level integration of these strategies and systematic end-to-end modeling and analysis into an embodiment of a future mobile multi-user virtual reality system for six degrees of freedom (6DOF) immersion. Besides its educational objectives, a fifth aim of the feature article is to identify a research framework and stimulate novel research and community building, given the emerging nature of virtual reality and 360$^\circ$ video and their prospective broad societal impact.
Another benefit of the fifth aim is that the technical advances it can lead to can facilitate fundamental research in other application areas of high-volume high-speed/low-latency data transfer in emerging cyber-physical systems and IoT settings, where for the first time the spatiotemporal aspects of the data need to be closely explored and tightly integrated with the user navigation actions, to maintain the desired quality of experience for the end user, given the limited available system resources.
\section{Existing Tutorials and Distinctions}
There have been tutorials appearing before at conferences such as IEEE VR and 3D User Interfaces, and ACM SIGGRAPH. However, they focused on unrelated aspects such as eye tracking in 360-degree video, human perception in virtual environments, and virtual reality content creation \cite{LeMeurJ:19,GabbardSE:12,Isdale:03}, which are traditionally associated with the fields of computer vision and human perception. Similarly, another recent brief article from the IEEE Communications Magazine \cite{BastugBMD:17} narrowly focuses on 5G (a technology upgrade cycle) and an abstract wireless VR application as one prospective use case, using assumptions that do not relate well to practice and thus do not lead to insightful observations. These include, for instance, traditional models of user arrival processes from communications/information theory, immersive experience measures expressed in percentages, and unrealistic compression rate characteristics of 4K \tsd video content.
The scope of this feature article is very different and broader, as it aims to highlight and educate about the present challenges of enabling future virtual reality applications, deployed in an untethered manner and with high-fidelity remote scene content, focusing on fundamental problems and trade-offs between signal processing, communications, and computation that arise in this context. Another objective of the article is to provide a tutorial like coverage of a non-conventional framework of research that can help bridge the present performance gap to enable such future applications and harness their expected societal benefit.
\section{Progress to date, synergistic advances, and broader implications}
We outline related work in \tsd video VR streaming using traditional approaches and systems, and synergistic advances in other technological domains that can help overcome the fundamental bottlenecks of the former.
Relative to traditional video streaming \cite{ChakareskiF:06,ChakareskiAWTG:04a,ReisCKS:10,ChouM:02}, 360$^\circ$ video streaming to VR headsets introduces the additional challenging requirements of ultra-high data rate, hyper-intensive computation, and ultra-low latency, as introduced earlier. Though some advances have been made in \tsd video streaming using traditional network systems, by investigating intelligent resource allocation and content representation \cite{Petrangeli2017,ChakareskiACSS:17,CorbillonDSC:17,HosseiniS:16}, the delivered immersion is still limited to low to moderate quality and 4K spatial panorama resolution, encoded at a temporal rate of 30 frames per second. This outcome stems from the fundamental limits in data rate and latency of such systems and their use of traditional server-client architectures. Moreover, a shared key shortcoming of the majority of emerging studies that is important to note is the pursuit of heuristic design choices and the lack of analysis of the fundamental performance trade-offs among the delivered immersion quality, user navigation patterns, signal representation, and system resources.
Free-space optics (FSO) and millimeter wave (mmWave) are emerging wireless technologies that are presently investigated and developed to help overcome the bottlenecks of traditional wireless systems. Both have the potential to enable multi-Gbps data transmission rates. FSO exploits the light intensity of a light emitting diode (LED) or a laser diode (LD) to modulate a message signal. FSO technologies using the former approach are known as visible light communication or VLC, as they provide illumination at the same time. After propagating through the optical wireless channel, the (infrared or visible) light message is detected by a photo-diode (PD) \cite{dimitrov2015principles}.
Unlike the radio frequency spectrum, plentiful unlicensed spectrum is available for light communications (300GHz–800THz), which has put FSO on the road-map towards sixth generation (6G) networks \cite{8792135}. While being a novel technology, a few studies of design concepts and experimental testbeds have already appeared \cite{rahman2018fso,8998135}. In the radio frequency spectrum, mmWave wireless communication is considered the enabling technology of next-generation wireless systems, as in the range of 10-100 GHz, more than 20 GHz of spectrum is available for use by cellular or wireless LAN applications. mmWave has seen its first commercial products operating in the 60~GHz band appear in the early 2010s. More complex transmission schemes to increase even further the achievable data rate are currently being investigated \cite{blandino2018multi}.
A few disparate preliminary studies emerged so far examining the potential of FSO and mmWave to advance mobile VR. A mmWave-based VR system was proposed in \cite{LiuZZLZZG:18} that uses WiGig modules for wireless connectivity between two laptop computers, one acting as a server and the other as a client, with the VR headset attached to it. Similarly, a mmWave reflector was developed in \cite{AbariBDK:17} to aid in connectivity maintenance with a mobile VR headset in the event of blocking of the direct (line of sight) wireless link. In each case, only synthetic computer graphics content was considered for transmission to and rendering on a VR headset. The company HTC released recently a wireless adapter operating at 60 GHz to enable untethered delivery of computer graphics content to VR headsets. However, the quality of experience is limited, due to the low quality and low resolution of the delivered content, and the highly ineffective real-time compression applied at the transmitting server. A design concept for using narrow beam FSO transceivers was proposed in \cite{rahman2018fso}, highlighting the challenge of link maintenance in such settings. The study in \cite{khan2019visible} explored the design of an FSO-enabled VR headset featuring hemispherically organized layers of highly directional PDs to facilitate connectivity maintenance under challenging head navigation movements of a VR user. Finally, Microsoft patented last year a design concept for a free space optics enabled VR headset and transmitting system \cite{CuervoGCK:19}, indicating vision and interest in the tech sector for actual products in the future.
In earlier related developments highlighting the prospective benefits of raw video data transmission, delivery of traditional uncompressed HD video over a short range 60 GHz wireless link was studied for home settings targeting stationary consumer electronics that do not integrate video compression, e.g., gaming consoles \cite{SinghQSNKK:08}. More recently, due to its ability to reduce network latency and help mobile devices offload part of their computation, edge computing and caching have started to be investigated for delivering \tsd content in wireless cellular systems \cite{Chakareski:18}.
The studies and endeavors highlighted heretofore either address specific technology aspects or lack a systematic end-to-end analysis. The envisioned next generation VR system
can help broadly advance the above efforts and the state-of-the-art, and make simultaneous impact on other emerging application areas of similar characteristics, as noted earlier.
Our preliminary advances highlight the substantial benefits and potential of the envisioned system \cite{ChakareskiK:21,ChakareskiG:20}.
As complementary to the above discussion, we examine if technologies employed in online multi-player VR gaming could be leveraged to enable further benefits. Foveated rendering exploits the very narrow field of view (the central $1.5-2^\circ$ of the entire field of view of the human eye) of the fovea, responsible for sharp central vision, to reduce the rendering workload of generating an image to be displayed on a VR headset, by greatly reducing its quality in the peripheral vision (outsize the zone gazed by the fovea) \cite{FoveatedRenderingWikipedia:21}. Essentially, much fewer pixels and at lower fidelity are rendered outside the fovea area. This considerably accelerates the frame rate of display of successive image frames displayed on a headset, which is helpful in dynamic gaming environments. Foveated rendering can be integrated into the envisioned VR system towards the same goal. Still, two aspects need to be considered carefully. Gaming content can be degraded considerably without noticing artifacts up close, as the computer graphics content is not overly complex and is evolving rapidly. On the other hand, high fidelity remote immersion via \tsd video may be susceptible to visibly noticeable degradation in quality under the same setting, as actual scene content is much more complex (to render) and thus sensitive to pixel resolution or fidelity reductions across the entire field of view. Even more importantly, the most challenging computing task faced by a headset in \tsd video VR applications is decoding the massive content at the target frame rate, not rendering the user's viewpoint, as explained later.
Online gaming applications also facilitate server-client architectures featuring multiple distributed servers to handle the massive client load \cite{BharambePS:06}. Each server maintains and updates a copy of the shared gaming environment, in response to the players' actions controlling their respective avatars in the game. An updated state of the environment is then sent back to all players every 100ms for first person shooter games (most dynamic) or at a lower rate for other types of games. Each player updates its own view of the environment using its gaming device based on the received updated state. Though such traditional computer network systems are suitable for this type of applications, the latencies, data rates, and weak state inconsistencies they exhibit would not be conducive to enable lifelike remote scene immersion on untethered VR headsets.
\section{A Virtual Reality and \tsd Video Primer}
A high level system illustration of a VR application is included in Figure~\ref{fig:VR_Primer}, left and middle. A user wearing a VR headset is linked to a powerful (typically gaming) computer, equipped with a Graphics Processing Unit (GPU), computer graphics software, and a 3D scene model. The link comprises a thick long integrated cable featuring as its main component a high data rate multimedia cable such as HDMI that transports in the direction of the user her present 3D viewpoint in the VR simulation rendered by the server computer based on the direction of viewing of the user communicated over the cable in the opposite direction. The navigation actions of the user comprise three head rotation angles denoted as yaw, pitch, and roll, and are illustrated in Figure~\ref{fig:VR_Primer} right. They are measured with respect to three coordinate axes centered on the user's head, using gyroscopes built into the VR headset, and determine the direction of viewing of the user in the synthesized 3D scene.
\begin{figure}[htb]
\centering
\includegraphics[width=\columnwidth]{Figures/VR_Primer.eps}
\caption{VR 101: {\bf (Left/Middle)} A user is linked to a powerful (gaming) desktop computer, equipped with a GPU and computer graphics software, to experience a VR simulation on her headset. The rendered 3D viewpoint of the user in the simulation and the direction of viewing of the user are exchanged over a long high data rate multimedia cable. {\bf (Right)} The navigation actions of the user comprise rotation angles yaw, pitch, and roll around three coordinate axes centered on the user's head and determine the direction of viewing.}
\vspace{-0.2cm}
\label{fig:VR_Primer}
\end{figure}
The virtual scene may be static or dynamically evolving. The simulation may also include spatial audio content that is reproduced in parallel on stereo headphones worn by the user. More recent VR application systems include the possibility for limited spatial movement of the user in the virtual scene. The spatial coordinates of the user headset are then tracked externally using infrared base stations mounted in the room where the system resides. The present 3D viewpoint of the user is rendered based on her spatial coordinates and direction of viewing in this case.
The major computing load of the application is reconstructing the 3D viewpoint of the user dynamically in response to her navigation actions and can be quite intensive. It is handled by the server computer and its GPU that execute demanding computer vision algorithms on the voluminous geometric representation of the 3D virtual scene. The collocated server computer and VR headset, and the cabled high data rate connection between them help to minimize the interactive latency of the application and to deliver high volume 3D computer graphics at the display frame rate required to avoid motion sickness \cite{Moss2011}.
Still, having a tethered VR headset can represent a tripping hazard and reduces the quality of experience of the user and the utility/scope of the application. Thus, most recent VR application systems feature wireless headsets, with stand alone display and computing capabilities or a slot in which a mobile phone is inserted to provide them. However, the capabilities of such systems in terms of delivered content quality, frame rate, and interactive latency are not on par with their tethered counterparts, as described earlier.
A \tsd video VR application system replaces the computer graphics 3D scene model from Figure~\ref{fig:VR_Primer}, left and middle, with actual remote scene 3D content. In particular, 360$^\circ$ video is a recent video format that is recorded by an omnidirectional camera that captures incident light waves from every direction (see Figure~\ref{fig:360VideoBackground}, top left). Such cameras comprise two or more wide-field lenses and rely on computer vision image stitching to produce a full 360$^\circ$ horizontal plane field of view and 180$^\circ$ vertical plane field of view \cite{Omni360CameraWikipedia:20}. Thereby, the constructed scene content would appear captured on the interior surface of a sphere centered on the camera location, akin to how we perceive the world around us.
\begin{figure}[htb]
\centering
\includegraphics[width=\figwidth]{Figures/360VideoVRStreamingBackground_v2.eps}
\caption{360$^\circ$ video capture and streaming, and user viewport $V_c$.}
\vspace{-0.2cm}
\label{fig:360VideoBackground}
\end{figure}
Concretely, \tsd video enables a 3D 360-degree look-around of the surrounding scene for a remote user, virtually placed at the camera location, on his VR headset, as illustrated in Figure~\ref{fig:360VideoBackground} right. After capture, the raw spherical or 360$^\circ$ video frames are first mapped to a wide equirectangular panorama (illustrated in Figure~\ref{fig:360VideoBackground}, bottom left) and then compressed using state-of-the-art (planar) video compression such as HEVC. The former intermediate step is introduced, as compression techniques operating directly on spherical data are much less mature and performing relative to traditional video compression operating on 2D video frames. Beyond the equirectangular mapping, which is most widely adopted, cube, pyramid, and dodecahedron planar projections have also been studied \cite{CorbillonDSC:17}.
The computing workload of a \tsd video VR application system is even more intensive and comprises decoding the compressed \tsd content and rendering the current 3D viewpoint of the user dynamically in response to his navigation actions. The latter is informally known as viewport and corresponds to only a small portion of the 360$^\circ$ view sphere denoted as $V_c$ in Figure~\ref{fig:360VideoBackground} right. The task of reconstructing the viewport on the VR headset requires remapping the decoded 360$^\circ$ panorama to the original spherical format and then projecting pixels from $V_c$ to their planar equivalents on the display of the headset.
For remote service, when the user and the stored 360$^\circ$ data are not collocated, the entire monolithic 360$^\circ$ panorama is commonly streamed to the user using traditional state-of-the-art video streaming (MPEG-Dynamic Adaptive Streaming over HTTP (DASH) \cite{Sodagar:11}). This considerably penalizes the quality of experience, due to the overwhelming volume of 360$^\circ$ data that needs to be delivered and that exceeds by orders of magnitude the available network bandwidth $C$ and the computing capabilities of the receiving device. These shortcomings are compounded by the reliance on ineffective network protocols such as HTTP and TCP to deliver latency-sensitive multimedia data of this nature, which have been adopted to lower the cost of an intelligent streaming system at the penalty of not having good control of the data delivery process. Thus, only lower quality, frame rate, and resolution 360$^\circ$ videos can presently be delivered online over the Internet. Yet, the streaming also lacks the ultra-low latency interactivity needed for truly immersive experiences, as traditional server-client Internet architectures are used in this case. The quality of experience and application utility are even lower in traditional wireless settings, due to the even lower data rates and computing/storage capabilities available therein, as noted earlier.
It should be noted that the recent advances in spatially adaptive streaming of broad or omnidirectional video panoramas \cite{CorbillonDSC:17,Petrangeli2017,ChakareskiACSS:17,HosseiniS:16} have contributed to further activities within MPEG and the follow-up DASH-SRD and OMAF standards that integrate them \cite{MPEG-DASH-SRD:15,MPEG-DASH-OMAF:18}. However, most of the bottlenecks highlighted above of such standards-based traditional approaches and their fundamental limitations in terms of enabled data rate, delay, and computing capabilities still remain.
\section{Bridging the Present Performance Gap}
In the following, we provide a high-level tutorial-style description of several unconventional techniques and emerging technologies and their synergistic interplay towards enabling next generation virtual reality applications comprising high-fidelity remote scene immersion and seamless untethered lifelike navigation of the reconstructed remote environment. To set the discussion, we illustrate in Figure~\ref{fig:NextGenerationVRSystem} a novel 6DOF VR system concept that embodies their holistic integration to help bridge the present performance gap.
\begin{figure}[htb]
\centering
\includegraphics[width=1.2\figwidth]{Figures/NextGenerationVRSystem_v2.eps}
\caption{Future 6DOF virtual reality system to bridge the present performance gap. It integrates synergistic edge computing, viewport-adaptive scalable \tsd video representation, and dual-connectivity transmission via traditional/emerging wireless technologies.}
\vspace{-0.2cm}
\label{fig:NextGenerationVRSystem}
\end{figure}
Its key components are an {\em edge server equipped with storage and computing capabilities} that will help bring the content and required computing closest to the user without overwhelming her device, to minimize the interactive latency; a {\em viewport-adaptive scalable \tsd video representation} that will facilitate dynamic spatiotemporal adaptation of the 6DOF content in response to the user's actions, to maximize the transmission efficiency; and {\em dual connectivity transmission via traditional and emerging wireless technologies} that will simultaneously provide reliability and robustness, and high-fidelity immersion. The dual-connectivity transmission will stream in parallel two synergistic content layers (baseline and enhancement) over the two respective wireless technologies. The enhancement layer will comprise raw and compressed content data to enable further performance synergies.
To enable its effective use, the transmission over the emerging wireless technology will integrate transmitter steering or switching that will facilitate user tracking feedback capturing the navigation actions of the user and her spatial position. The user will be equipped with a dual-connectivity VR device that will integrate the two content layers.
Given the nature of the application, it will be an effective system design to employ dual-connectivity transmission only for the down-link communication from the edge server to the mobile VR users. The up-link communication in the opposite direction, i.e., from the users to the edge server, can be effectively carried out by employing solely single-connectivity transmission via reliable lower data rate traditional wireless technology links, as it will only carry miscellaneous low-rate control information.
We implicitly have an indoor setting in mind for the envisioned system, due to the nature of the target application. Moreover, some of the emerging technologies and unconventional techniques the system integrates have advanced further and would be easier to deploy in such a setting. Still, all its key components can either directly apply to an outdoor scenario or have outdoor counterparts that have been developed in parallel. Thus, an outdoor deployment of the envisioned system can potentially be pursued as well.
\subsection{Raw data transmission of ultra-high resolution and frame rate content}
Video compression has made the present Internet possible and has fueled its growth for many years \cite{CiscoVNI_BarnettJAK:18}. It enables a reduction of the required network transmission data rate by a few orders of magnitude. However, it induces a decoding delay at the receiving client, at the same time, proportional to the compressed data volume. This can be penalizing for ultra-low latency applications such as online remote immersion via virtual reality, especially in mobile settings where client devices have limited computing capabilities and the delivered compressed data is massive. Intelligent transmission of the required 360$^\circ$ video content as raw data in this setting can help overcome this challenge, as explained next.
\subsection{High-frequency directional wireless transmission}
It may seem contradictory at first to send raw video data as that would scale back up the transmission delay, by a few orders of magnitude. However, the emerging wireless technologies of free-space optics and millimeter wave transmit data at much higher spectrum frequencies (relative to traditional sub-6 GHz wireless technologies), to enable several orders of magnitude higher data rates. Using such transmission can make sending raw video data appealing, as it would cancel out the increase in transmission delay and still provide the benefit of lower computing delay at the client device.
However, these emerging technologies exhibit different transmission characteristics relative to their traditional counterparts that need to be addressed to enable their efficient utilization. Concretely, their transmission beams are very narrow and directed, as reflected waves from obstacles and the environment feature very poor signal quality, due to the high career frequency that is used for transmission \cite{Popovski:20}. Therefore, the transmitter and receiver need to be positioned in a direct line-of-sight of each other and be actively aligned to maintain that property, in case of a mobile receiver. This challenge can be addressed in the envisioned system by benefiting from the available user tracking information and the limited spatial mobility of the users in an indoor setting.
In particular, high-frequency free-space optical transmitters can be mounted on steerable platforms and actively directed towards the users using servo motors, based on the tracking information. Two servo motors can control the horizontal and vertical rotations, or azimuth and elevation angles, of a transmitter tracking its user over the spatial area of the system. Such motors can operate at 50 Hz control signal frequency and can cover a \tsd rotation angle per second \cite{TowerProMG995ServoMotor:21}, which is much faster than a typical mobile user maximum speed of one meter per second indoor \cite{TavakkolniaSAGASH:19}. Similarly, high-frequency millimeter wave transmitters adequate for mobile communication can be realized with multi-element phased-antenna arrays (PAAs) to enable dynamic beam forming in both azimuth and elevation in response to the navigation/mobility actions of the respective receiving users \cite{blandino2018multi}.
VR headsets can be equipped with respective receivers for such high-frequency transmission.
In the case of free-space optics, the upper portion of a headset can be covered with an array of photo-diode detectors, which can receive and decode in parallel the optical beam signal incident on their surface. The decoded incident signals at every photo-detector can be combined using diversity combining techniques for more effective performance. The dimensions and arrangement of the photo-detector array can be configured to match the accuracy of the steering motors of the respective transmitters, the width of the optical beam, and the head movement navigation actions of a user. Similarly, a small form factor millimeter wave multi-element PAA receiver\footnote{A prototype of size $5 \times 20$ millimeters, featuring a 16-element PAA, was recently demonstrated in \cite{blandino2018multi}.}
can be placed on the top portion of a VR headset.
As the respective transmitters for each technology would be typically placed high above the users, blockage by human movement would not be a challenge and the placement of the receivers on top of the VR headsets would be a natural choice.
\begin{figure}[htb]
\centering
\begin{subfigure}{.5\columnwidth}
\centering
\includegraphics[width=0.9\linewidth]{Figures/LD_beam.eps}
\end{subfigure
\begin{subfigure}{.5\columnwidth}
\centering
\includegraphics[width=0.9\linewidth]{Figures/headset-magazine.eps}
\end{subfigure}
\vspace{-0.3cm}
\caption{RGB-LD transmitter (left) and VR headset equipped with photo-detectors (right).}
\label{fig:RGB_LD_VLC_Transmitter_VR_HeadsetPhotoDiodes}
\vspace{-0.3cm}
\end{figure}
We discuss in slightly more detail now an embodiment of the FSO transmitters and receivers in the envisioned system. LDs can enable higher transmission rates relative to LEDs, and thus may be the preferred choice. As their transmission beam is very narrow (typically $< 1^\circ$), a combination of red, green, and blue LDs, together with a diffuser, as illustrated in Figure~\ref{fig:RGB_LD_VLC_Transmitter_VR_HeadsetPhotoDiodes} (left), can be used to produce a slightly broader transmission beam that will increase the transmission coverage and reliability, while limiting the reduction in data rate. Simultaneously, this will lead to coherent white light emitted by the transmitter that can prospectively serve as lighting. Early prototypes of such transmitters have demonstrated data rates of up to 10 Gbps \cite{zafar2017laser}. Beam splitters and detectors can be used to monitor and control the different laser powers, for safety/performance considerations.
To mitigate the effect of narrow transmission beams and maintain high receive data rate, an angle-diversity-receiver can be implemented, featuring multiple small-area PDs installed on a platform at different inclination angles\footnote{The receive data rate of a PD rapidly drops with its surface area. However, a large surface area receiver increases the reliability of the optical link. The envisioned receiver design integrates the best of both approaches.}. Such a multi-PD design can be particularly helpful in maintaining the highly directional optical link during dynamic 360$^{\circ}$ navigation head movements of a VR user. We recently investigated this approach towards the design of a helmet-shaped VR headset, comprising multiple PDs placed along different upper-hemispherical layers on the headset, as illustrated in Figure~\ref{fig:RGB_LD_VLC_Transmitter_VR_HeadsetPhotoDiodes} (right). One PD is placed on top, and the rest are distributed along different layers on the hemispherical headset surface. The design optimization aims to minimize the number of PDs used, while ensuring uninterrupted connectivity with an FSO transmitter. It selects the number of layers, their placement along the headset hemisphere, and the number of PDs per layer, in pursuit of this objective, and integrates VR head movement navigation data for robust connectivity \cite{khan2019visible}. Power dissipation of the PDs is important in this setting and can be integrated into the design optimization, to impact the number of PDs placed on the headset.
\subsection{Low-high-frequency dual connectivity wireless streaming}
High-frequency wireless transmission induces brittle pencil-beam like directed communication links. Streaming 360$^\circ$ video over them alone would lower the reliability, as such links can be fragile and sensitive to line-of-sight misalignment between the transmitter and the receiver. Integrating low-frequency (traditional sub-6 GHz) wireless transmission as a synergistic supplement can help maintain the robustness of the system, as such radio waves do not need a sender-receiver alignment and provide multi-path and reflected signal benefits. Concretely, the traditional wireless connectivity can be employed to stream a baseline layer or representation of the 360$^\circ$ content that alone will ensure an uninterrupted service and reliable application quality, if a transient high-frequency link loss occurs prospectively. The high-frequency wireless connectivity can then be used to stream an enhancement content layer or representation that will build upon the baseline layer to enable high fidelity immersion and augment the quality of experience of the user. Due to the plentiful network transmission bandwidth enabled by the high-frequency connectivity, the enhancement content layer can be streamed at least in part as raw data, to enable further system performance advances. An effective operation of such dual-connectivity 360$^\circ$ video streaming towards reliable high-fidelity immersion can be advanced via a synergistic and efficient design of the data representation that will enable the construction of the two content layers, as explained next.
\subsection{Scalable 360$^\circ$ video tiling and viewport-driven adaptation}
\label{sec:Scal360TilingViewportAdaptation}
Rather than streaming the entire 360$^\circ$ panorama, as conventionally done, one can construct a scalable tiling-based representation of the 360$^\circ$ panorama to enable effective spatiotemporal adaptation of the transmitted data stream to the dynamic viewport of the user and the available network bandwidth. This will lead to efficient use of resources and higher quality of experience for the user. Moreover, the scalable 360$^\circ$ tiling can facilitate rigorous mathematical analysis of performance aspects of streaming systems, as we show later. Tiling of traditional wide-panorama video has been introduced as an option in HEVC, to facilitate parallel compression of the content in multi-core processor systems. More recently, its benefits have started to be recognized for 360$^\circ$ video streaming \cite{ChakareskiACSS:17,HosseiniS:16}.
A sample scalable tiling-based 360$^\circ$ video representation is illustrated in the right portion of Figure~\ref{fig:360VideoBackgroundTilingStreaming}. Concretely, each 360$^\circ$ video frame comprising a Group of Pictures or GOP\footnote{This is a block of consecutive frames that are compressed together with no reference to other frames.} is spatially partitioned into a set of tile sectors or simply tiles $(i,j)$ along its longitude and latitude dimensions. The collection of tiles across the video frames of a GOP at the same spatial location is then compressed into multiple embedded layers of progressively increasing signal fidelity. The first (bottom) layer of a compressed 360$^{\circ}$ GOP-tile is commonly known as a base layer, and the remaining layers are identified as enhancement layers. The reconstruction fidelity of a GOP-tile improves incrementally as more layers are being decoded progressively starting from the first layer.
\begin{figure}[htb]
\centering
\includegraphics[width=1\linewidth]{Figures/ScalableMulti-Layer360Tiling_BaselineEnhancementContentLayers_v2.eps}
\caption{GOP-level scalable signal tiling of an equirectangular 360$^{\circ}$ panorama.}
\vspace{-0.2cm}
\label{fig:360VideoBackgroundTilingStreaming}
\end{figure}
The baseline content layer can be constructed from the scalable tiling-based 360$^\circ$ representation such that it comprises the first $L_b$ layers for a broader set of tiles encompassing the user viewport, to account for (i) a prospective mismatch between viewport knowledge at the server/sender, used to construct a complementary enhancement content layer, and the actual user viewport at the receiving client, induced by rapid user head movements, or (ii) a prospective transient high-frequency link loss. Accounting for these two possibilities thereby will ensure that the viewport can be reconstructed continuously and will augment the application reliability considerably. The enhancement content layer can then be constructed such that it comprises the subsequent $L_e$ layers for a narrower set of tiles focusing closely on the viewport, to maximize its expected quality, as illustrated in the right portion of Figure~\ref{fig:360VideoBackgroundTilingStreaming}. The choice of number of layers and tile selection from the scalable 360$^\circ$ tiling to construct the baseline and enhancement content layers can be made rigorous, inclusive of the integrated selection of a subset of enhancement content layer tile to be transmitted as raw data, as introduced earlier.
We illustrate this as part of an end-to-end example analysis of the envisioned VR system concept that we present later.
We note that cheap storage and easier implementation have made non-scalable video compression preferred in practice. With minor adjustments, our system design can directly apply to the case of constructing both content layers independently using non-scalable compression. Moreover, scalable video coding has been successfully deployed in cutting-edge technologies for diverse low-latency multi-party telepresence settings \cite{Vidyo:22}, providing therein considerable benefits in terms of lower server complexity and higher client quality of experience over state-of-the-art non-scalable video compression based solutions. These benefits and related system context arise in the setting investigated here as well. Finally, scalable video coding is consistently enhanced through research and every subsequent generation of video codecs \cite{BoyceYCR:17}. The next generation VR system we explore in this feature article can provide avenues for advancing further such efforts.
\subsection{6DOF Virtual Reality and \tsd Video}
A single \tsd video (of actual or synthetic content) enables three degrees of navigation freedom (3DOF) to a VR user, in the form of rotational head movements around three orthogonal axes (as noted earlier, see Figure~\ref{fig:VR_Primer} right), in experiencing a remote scene immersively from a single location. The streaming strategy described in Section~\ref{sec:Scal360TilingViewportAdaptation} can be extended naturally to the case when the application will also allow for spatial movement of the user in the remote scene, to enable a 6DOF immersion experience over the spatial area of the VR arena, where the content is navigated.
Here, the user will have the ability to select not only her direction of viewing but also the spatial location of the \tsd viewpoint in the scene to be explored, which will augment the quality of experience. The content for each such spatial \tsd video viewpoint available to be navigated can be represented using the scalable \tsd tiling approach highlighted above and the edge server can apply in this case dynamic viewpoint adaptation and viewport-driven content adaptation jointly, in response to the 6DOF navigation actions of the user, as illustrated in Figure~\ref{fig:6DOF_360VideoStreaming}.
\begin{figure}[htb]
\centering
\includegraphics[width=0.85\linewidth]{Figures/6DOF_360VideoStreaming_v2.eps}
\caption{6DOF remote scene immersion, and dynamic spatial \tsd viewpoint and viewport adaptation for a mobile VR user navigating the scene.}
\vspace{-0.2cm}
\label{fig:6DOF_360VideoStreaming}
\end{figure}
Complementary signal processing explorations can be pursued in such a 6DOF system, for further enhancements. For instance, additional {\em virtual} \tsd video location viewpoints can be dynamically synthesized by the edge server for a navigating VR user using geometric signal processing techniques from multi-view imaging such as depth-based view interpolation \cite{ChakareskiVS:12}. Similarly, analysis can be carried out to establish the smallest number of captured \tsd video viewpoints and their spatial distribution over the scene to enable a required quality level for such intermediate interpolated viewpoints.
\subsection{Edge-based operation and mobile edge computing}
Streaming the content from an edge server (bringing the content closest to the user) will minimize the delivery latency compared to traditional server-client network architectures. Moreover, the edge server's computing capabilities can be leveraged to relieve the computing requirements induced upon a mobile VR headset. The former's much more powerful multiple GPUs can be effectively used to this end. In addition to the content decoding computation noted earlier, another computational requirement induced upon a client device running a virtual reality application represents the dynamic rendering of the present field of view of the user (aka viewport). This introduces an additional delay component into the entire end-to-end operational latency chain. Sharing the conventional headset decoding and rendering computation with a nearby edge server can be explored to optimize the delivered immersion fidelity, while meeting the strict system latency of the application. Raw video transmission and scalable 360$^\circ$ tiling can synergistically integrate with this objective, as well.
\subsection{System Design and Integration Embodiment}
\label{sec:SystemDesignArenaIntegrationEmbodiment}
We highlight here in greater detail how a system level integration of the techniques and technologies outlined heretofore can be pursued towards realizing the envisioned VR system concept in practice. To facilitate the discussion, we have included a high level illustration of an arena embodiment of this future application system in Figure~\ref{fig:SystemIntegrationDesign}. There can be multiple traditional sub-6 GHz wireless access points mounted on the walls of the arena and a higher number of emerging high-frequency wireless transmitters mounted on the ceiling of the arena. Both types of transmitters can be linked via fiber optical links to the edge server installed on site. The server can control the dynamic transmission scheduling of the traditional access points and next generation (xGen) transmitters. The user headsets will be equipped with respective sub-6-GHz/xGen dual connectivity receivers. Moreover, the edge server will be equipped with high-end storage and computing capabilities to execute the envisioned dual-connectivity 360$^\circ$ video streaming. In the case of on-demand remote immersion, the compressed 6DOF VR or \tsd video content can already reside at the edge server and be deployed there ahead of time. In the case of live remote immersion, the content can be deployed in real time over a high-speed fiber optic Internet link to the edge server, from the remote location where it would be captured and compressed using scalable \tsd tiling.
\begin{figure}[htb]
\centering
\includegraphics[width=0.65\linewidth]{Figures/SystemIntegrationDesign_v2.eps}
\caption{An indoor arena integration of the envisioned virtual reality system. AP: access point; Tx: transmitter}
\vspace{-0.2cm}
\label{fig:SystemIntegrationDesign}
\end{figure}
To track the spatial location and head orientation of the users, IR tracking base stations can be mounted on the walls or ceiling of the arena. Present commercially available products of this nature provide accurate positional tracking of less than 1-mm error at temporal resolutions of 250 samples per second or higher. The tracking information is relayed to the edge server in real-time over high-speed links to enable its operation. The spatial area of the arena can be conceptually split into multiple small sectors or cells, and one xGen transmitter can be mounted above each such sector, as illustrated in Figure~\ref{fig:SystemIntegrationDesign}. At the onset of a streaming session for remote immersion, a user can be assigned to a given sector and its xGen transmitter.
During the session, the user will dynamically navigate the 6DOF VR or \tsd video content and may move across the arena. If this movement is limited within the original sector, the same xGen transmitter can be dynamically steered towards the user, as explained earlier.
On the other hand, as such high-frequency links experience strong signal degradation over short distances,
dynamically (re)assigning another xGen transmitter to the user can be carried out, every time she will exit her present sector as the session evolves. This can help maximize the received signal quality and thus data rate on the high frequency link of the user, and minimize the interference to other high frequency links in the arena serving other users, at the same time. Thereby, the quality of experience of all simultaneous users in the arena can be augmented. Finally, if more than one user happen to transiently pass through or reside in the same sector at the same time, during the course of the session, the resources of the high-frequency xGen transmitter of that sector can be split across these users during that time, uniformly or preferentially, for some alternative transmitter assignment and steering settings/methods for the system\footnote{We recall that the system design to exclusively serve each user by one xGen transmitter via its narrow transmission beam that is dynamically directed towards the user, as outlined in the Section "High-Frequency Directional Wireless Transmission" earlier, is the default choice that we focus on as the most adequate and performing for the next generation VR system we investigate.}
In particular, we highlight briefly here the prospective options for xGen transmitter steering and (re)assignment in our system, to provide further clarity and instruction to the reader. In the case of FSO wireless technology, electronic, mechanical, or hybrid electro-mechanical transmitter steering can be carried out. (i) The first approach is also known as electronic transmitter switching or assignment, as it effectively activates an FSO transmitter associated with a spatial cell into which a user enters, and deactivates the FSO transmitter associated with the cell that the user exited. This approach comprises the same number of FSO transmitters as spatial cells and their beam width is selected to ensure coverage of the entire spatial area of their respective cells, while minimizing the overlap with transmitter beams of adjacent cell, to minimize the inter-cell interference. Here, the FSO transmitters are fixed (static) in orientation and point directly downwards to the spatial area of the cells. Lastly, when multiple users are present in the same cell, the transmission resources of the respective FSO transmitter can be shared across them using multi-access techniques, e.g., time-division multiple access (TDMA). (ii) Employing mechanical steering requires having at least the same number of transmitters as users in the arena, if there are more users than spatial sectors/cells in the arena\footnote{In the experimental evaluation, we address this choice simply by setting the number of xGen transmitters per cell to be equal to $\lceil N_u/N_c \rceil$, where $N_u$ and $N_c$ denote respectively the number of spatial cells and users in the arena system, and $\lceil \cdot \rceil$ denotes the ceiling operator.}. In this case, there can be multiple transmitters mounted above the spatial area of a single cell and their orientation is dynamically steered towards their respective users to whom they have been assigned using servo motors, as introduced earlier. Periodic reassignment of users to transmitters over the course of a session can be carried out here as well, as the users move across the spatial area of the arena during the course of a media session, to enhance the system performance, as noted a little earlier. (iii) Finally, the hybrid electro-mechanical steering aims to integrate both prior approaches, where a user can be tracked by his or her (re)assigned transmitter using mechanical steering until the user enters a cell where another user is already present. In that case, both users will start to be served using a single static transmitter mounted above the center of that cell using a multi-access technique. Once a user becomes alone again in his or her present cell, the system will switch back to mechanical steering of the respective transmitter assigned to the user, towards the user.
In the case of mm-wave wireless technology, electronic beam formation and steering can be employed to ensure transmitter-receiver alignment as the users move across the arena, as noted earlier in the Section "High-Frequency Directional Wireless Transmission". Here, one option can be to have at least as many transmitters as users in the arena, if there are more users than spatial cells in the arena. Thus, potentially there can be multiple transmitters mounted above each cell in the arena system. Alternatively, the system can comprise the same number of transmitters as cells in the arena, mounted above the center of each respective cell. In this latter case, multiple users present in the same cell can be served by the same transmitter of that cell using multi-beam formation and steering, to share its transmission resources. Multi-beam mm-wave transmitters or access points are prevalently encountered in practice. Finally, in each case, dynamic reassignment of users to transmitters can be carried out during the course of the session as the users move across the spatial area of the arena, to enhance the system performance.
The benefits and shortcomings of the transmitter steering and (re)assignment methods available for each xGen wireless technology and outlined here, in the context of our system, have been explored recently in \cite{ChakareskiKRB:20,ChakareskiKRB:21}.
\section{Example End-to-end Analysis}
We guide the reader here through several high-level analysis examples that build upon each other to highlight the prospective benefits of the envisioned VR system and motivate new studies that can help advance it further. In particular, we first illustrate how the \tsd video tiling and the ability to collect user navigation data can facilitate the development of statistical models of user navigation. In conjunction, we illustrate how effective rate-distortion modeling of the spatiotemporal encoding characteristics of the \tsd panorama across its tiles can be pursued. We then highlight how these modeling advances can be integrated to pursue further analysis that captures the fundamental trade-offs between computing, communication, and signal representation in the context of the end-to-end application performance enabled by the system. Finally, we illustrate how rigorous system level optimization can be built upon these integrated modeling and analysis advances to select key resource allocation decisions.
{\bf Navigation modeling.} The head movement navigation actions of a user can be collected over time at the edge server and can be used to characterize the probability that a \tsd tile will appear in the user viewport during a time period. This will inform the understanding of how important each tile and its content are for the quality of experience of the user. One way to capture this information is to integrate the number of instances a tile will appear in the viewport during navigation of the content over the time interval of interest. This approach can be made rigorous by identifying the spatial portion $S^{nm,V_c}_j$ of each \tsd tile $(n,m)$ present in the user viewport $V_c$ at the temporal display instance $t_j$ of the $j$-th panoramic video frame comprising the \tsd content. Care must be taken here to account for the widely varying viewport size depending on its latitude on the \tsd panorama, to integrate this information properly then over time. This phenomenon is induced by the equirectangular mapping of the native spherical data that is employed at compression, as explained earlier. It is illustrated in Figure~\ref{fig:ViewportSizeLatitudeImpact} that highlights the shape and size of a viewport on the \tsd panorama for two different yaw and pitch rotation angles, denoted therein as azimuth and polar angles $(\varphi,\theta)$ in spherical coordinates.
\begin{figure}[htb]
\centering
\begin{subfigure}{0.45\columnwidth}
\centering
\includegraphics[width=\linewidth]{Figures/viewport0-0.eps}
$(\varphi,\theta) = (0^{\circ},0^{\circ})$
\end{subfigure}
\begin{subfigure}{0.45\columnwidth}
\centering
\includegraphics[width=\linewidth]{Figures/viewport120-60.eps}
$(\varphi,\theta) = (120^{\circ},-60^{\circ})$.
\end{subfigure}
\caption{The viewport's latitude coordinate on the \tsd view sphere (polar angle $\theta$ = head rotation angle pitch) impacts its shape and size on the \tsd panorama.}
\label{fig:ViewportSizeLatitudeImpact}
\end{figure}
Formally, the quantity $S^{nm,V_c}_j$ represents the overlap or intersection between the spatial areas of the tile and the viewport in the \tsd panorama at that time instance and can be captured as $S^{nm,V_c}_j = S^{V_c}_j \cap S^{nm}_j$, where $S^{nm}_j$ and $S^{V_c}_j$ denote the sets of pixels representing the two denoted spatial areas. To account for the unequal viewport size on the equirectangular plane across time, in developing a statistical model of user navigation, one can normalize the fractions of the spatial areas of every tile present in the user viewport $V_c$ at $t_j$, using $s^{nm}_j = \frac{|S^{nm,V_c}_j|}{\sum_{n,m}|S^{nm,V_c}_j|}$,
\noindent where $|S|$ denotes the size of a set $S$, in this case in number of pixels. Thus, $\{s^{nm}_j\}$ represents the normalized distribution of the spatial area of the user viewport across every tile in the $360^\circ$ panorama, at time instance $t_j$.
Finally, given $\{s^{nm}_j\}$, one can formulate the probability (likelihood) of the user navigating tile $(n,m)$ over a time interval spanned by the time instances $[t_i,t_j]$, as $P_{nm}^{(t_i,t_j)} = \frac{\sum_{k=i}^{j}s^{nm}_k}{j-i+1}$.
In other words, $P_{nm}^{(t_i,t_j)}$ indicates how likely tile $(n,m)$ will appear (at least in part) in the user viewport during navigation of the $360^\circ$ video from its temporal instance $t_i$ to its temporal instance $t_j$, or the popularity of the $360^\circ$ scene content captured by the tile for these user and time interval. For instance, if $t_i$ and $t_j$ correspond to the first and last video frame of the $360^\circ$ video, then, $P_{nm}^{(t_i,t_j)}$ captures the navigation probability or popularity of tile $(n,m)$ across the entire video.
To highlight the nature of these quantities, we illustrate in Figure~\ref{fig:RollerCoasterWingsuitNavigationProbabilities} their expected values (averages across a large user set) for two popular \tsd videos used in our analysis, where we considered a popular $6 \times 4$ tiling\footnote{Denser tiling layouts increase the processing complexity and reduce the compression efficiency, but enable more precise delineation of the user viewport and thus more accurate analysis and resource allocation.}. Concretely, one can observe that in the case of Roller Coaster, video tiles on the fringes of the $360^\circ$ panorama are rarely navigated by a user, i.e., they scarcely appear in the user's viewport, during a $360^\circ$ video navigation session. Conversely, video tiles indexed as $(n,m) =$ (3,2), (4,2), (3,3), and (4,3) are quite often navigated by the user, as noted by their much higher navigation likelihoods shown in the histogram.
\begin{figure}[htb]
\centering
\begin{subfigure}{.5\columnwidth}
\centering
\includegraphics[width=1\linewidth]{Figures/NavigationProbabilityRollerCoaster_v2.eps}
\end{subfigure
\begin{subfigure}{.5\columnwidth}
\centering
\includegraphics[width=1\linewidth]{Figures/NavigationProbabilityWingsuit_v2.eps}
\end{subfigure}
\vspace{-0.3cm}
\caption{Expected $\{P_{nm}\}$ for $360^\circ$ tiles for Roller Coaster (left) and Wingsuit (right).}
\label{fig:RollerCoasterWingsuitNavigationProbabilities}
\vspace{-0.3cm}
\end{figure}
The navigation probabilities of $360^\circ$ video tiles for the video Wingsuit, induced by users experiencing the content, look fairly different. This is due to the nature of this content and the induced specific interests of the users, expressed when navigating it. In particular, one can see here that the typical user is predominantly interested in navigating the southern hemisphere of the $360^\circ$ panorama, as seen from the respective tile navigation probabilities indicated in the histogram. The more dynamic and interesting content of Wingsuit resides spatially in the southern hemisphere of its $360^\circ$ view sphere.
To address the prospect of imperfect or lack of knowledge of near future navigation actions and introduce further robustness into the system, e.g., in contexts such as live remote immersion, predicting the upcoming head movement (or jointly head and body movement) navigation actions or 360-degree tile navigation likelihoods for a user can be integrated. This can be carried out accurately via regression or machine learning techniques, based on the user’s navigation history and prospectively including features extracted from the already navigated content \cite{QianHJG:16,HouDZB:21}. It has been shown that the navigation actions of a user exhibit a strong low-frequency component and short-term correlation over time that can benefit such prediction methods. Moreover, it has been observed that the dynamics of the navigation actions of a user generally correlate well with the temporal characteristics of the content that is navigated \cite{BlandinoRCCKG:21}.
{\bf Rate-distortion modeling.} One can also accurately model the rate-distortion dependency of compressed GOP tiles to facilitate analysis and optimization of the edge server's operation. Here, we examine two models, exponential ($D = c_1 e^{-d_1 R}$) and power law ($D = c_2 R^{d_2}$), to capture this dependency. Concretely, we vary the encoding rate $R$ of a tile and record its respective Mean Squared Error (MSE) reconstruction distortion $D$. We graph the obtained pairs $(R,D)$ in Figure~\ref{fig:TileRateDistortionDependency} with markers, for three representative tiles of the popular 360$^\circ$ video Roller Coaster. The two analytical models are also graphed in Figure~\ref{fig:TileRateDistortionDependency}. We can see that Tiles 3 and 16 have steeper rate-distortion dependency, due to their relatively static content. Tile 11 requires higher encoding rate to achieve the same reconstruction error, due to its more dynamic content. Figure~\ref{fig:TileRateDistortionDependency} shows that the power law model provides a more accurate characterization across all three tiles. This motivates its use in our later analysis.
\begin{figure}[htb]
\centering
\begin{subfigure}{.5\columnwidth}
\centering
\includegraphics[width=1\linewidth]{Figures/r-d-pow_v2.eps}
\end{subfigure
\begin{subfigure}{.5\columnwidth}
\centering
\includegraphics[width=1\linewidth]{Figures/r-d-exp_v2.eps}
\end{subfigure}
\vspace{-0.3cm}
\caption{Rate-distortion dependency $D(R)$: (Left) Power law model and (Right) Exponential model. Actual data points shown as markers.}
\label{fig:TileRateDistortionDependency}
\vspace{-0.3cm}
\end{figure}
We note that in video compression the raw pixel data is stored in the three-dimensional luminance-chrominance color space denoted as YUV, to enable data volume reduction. The Y (luminance) component captures exclusively the signal strength of each pixel, i.e., its monochromatic intensity or brightness. The two U and V chrominance components capture exclusively the color information associated with each pixel. Video signal distortion or quality is commonly measured only on the Y component due to its nature.
{\bf Motivating resource allocation example.}
Building upon the two modeling advances and scalable \tsd tiling, we illustrate here how data rate resources can be effectively allocated over the \tsd panorama and time. When we examine and formulate the end-to-end performance of the envisioned VR system concept later, we will integrate this motivating example analysis therein.
When the \tsd content is compressed using scalable tiling, the data rate and reconstruction error of a GOP tile $(n,m)$ can be directly related to the number of embedded layers $l_{nm}$ selected to represent the tile. These relationships can be made precise using the rate-distortion modeling described earlier. Similarly, given the statical navigation profile $\{P_{nm}\}$ of a user for this GOP, one can formulate the expected viewport distortion experienced by the user during the GOP as $\sum_{nm}P_{nm}D_{nm}(l_{nm})$. Finally, given a maximum streaming data rate of $C$, one can then aim to optimally select the number of layers $l_{nm}$ sent for every tile $(n,m)$ during the GOP time interval, such that the expected user viewport distortion over that interval is minimized. This can be formally captured via the optimization:
\beq
\min_{\{l_{nm}\}}\sum_{nm}P_{nm}D_{nm}(l_{nm}), \, \text{subject to:} \sum_{nm}R_{nm}(l_{nm})\le C. \nonumber
\eeq
To highlight its benefits and potential, we implemented the optimization above, varied $C$, and computed the optimal solution and the respective minimum expected viewport distortion in each case (the value of the objective function at the optimal solution). Simultaneously, we recorded in the same context the corresponding performance of the state-of-the-art MPEG-DASH streaming standard, implemented with spatial viewport-adaptation \cite{Petrangeli2017}. We examined two popular 4K $360^\circ$ videos Wingsuit and Angel Falls to carry out this motivational assessment. We show these outcomes in Figure~\ref{fig:TransmissionEfficiencyVsMPEG-DASH} where the y-axis indicates the delivered viewport quality or immersion fidelity via a base-10 logarithmic inverse of the expected distortion, formally known as the {\em peak-signal-to-noise ratio} (PSNR). We note that throughout this paper, the reported PSNR results are measured with respect to the luminance (Y) component of the user viewport video signal, as explained earlier. On rare occasions, these results are denoted as Y-PSNR to indicate this aspect.
\begin{figure}[htb]
\centering
\includegraphics[width=1.1\figwidth]{Figures/GraphPSNRvsDataRateTwoSequences_v3.eps}
\vspace{-0.4cm}
\caption{$360^\circ$ video transmission efficiency of the motivating resource allocation example vs. MPEG-DASH.}
\label{fig:TransmissionEfficiencyVsMPEG-DASH}
\end{figure}
We can see that the approach highlighted above enables considerable benefits over MPEG-DASH by integrating the user navigation actions and the rate-distortion trade-offs across the $360^\circ$ panorama, in deciding how transmission resources should be allocated. Approximately 6-7 dB of immersion fidelity improvement have been enabled across both $360^\circ$ videos and all network bandwidth values considered in Figure~\ref{fig:TransmissionEfficiencyVsMPEG-DASH}. These advances can in turn enable much higher operational efficiency for a streaming system for $360^\circ$ video delivery that integrates this methodology, as our subsequent analysis will highlight.
The problem formulation can be easily extended to integrate tile-level weights $w_{nm}$ that will capture the amount of deformation (stretching) that the content of each tile $(n,m)$ undergoes on the \tsd sphere when mapped back for viewing on the VR headset. This deformation is solely dependent on the tile's latitude and emphasizes equatorial tiles as more sensitive to content distortion because of this. By this integration, the formulation will capture as its objective the expected tile-level weighted to spherically-uniform viewport distortion, whose pixel-level deterministic counterpart denoted as WS-MSE has recently been introduced as a more adequate quality metric for omnidirectional content \cite{SunLY:17}.
{\bf Per-user end-to-end performance analysis.} The edge server in the envisioned VR system can benefit from a navigation profile $\{P_{nm}\}$ for a GOP to preferentially treat the tiles comprising it when streaming the content to the respective user, as motivated by previous resource allocation example. Concretely, the server can first identify as $M$ the set of GOP tiles with non-zero navigation likelihoods. To construct the baseline content layer, the server can assign streaming data rates $R_{nm,w}$ to lower indexed embedded layers from the scalable \tsd tiling exclusively for tiles $(n,m)$ from $M$, to maximize the resource utilization.
If $\Delta T$ is the temporal duration of a GOP, then the latency of streaming the baseline content layer to a user over her traditional wireless connectivity link can be formulated as $\tau^w = \frac{\sum_{(n,m)\in M}R_{nm,w}\Delta T}{C^w}$, where $C^w$ indicates the transmission capacity of the link.
Similarly, by facilitating $\{P_{nm}\}$, the server can identify a subset $M^{\text{r}}$ of tiles in $M$ that are most relevant for the quality of experience of the user. The construction of $M^{\text{r}}$ can be made rigorous as explained a little later. These tiles can be streamed as raw data as part of the enhancement content layer, to maximize their impact. They can be reconstructed first as such at the edge server at the highest quality, from the compressed scalable \tsd tiling that will reside there, by decoding each compressed tile from $M^{\text{r}}$ from all its embedded layers. If the encoding data rate of the entire set of layers for tile $(n,m)$ is $R_{nm,\max}$, then the decoding latency of reconstructing tiles from $M^{\text{r}}$ as raw data at the server can be formulated as $\tau^Z = \frac{\sum_{(n,m)\in M^{\text{r}}}R_{nm,\max}\Delta T}{Z}$, where $Z$ is the decoding speed of the server.
The construction of the enhancement content layer by the server can be completed by selecting an adequate set of higher indexed embedded layers from the scalable \tsd tiling for the remaining (compressed) tiles in $M$ (the set $M^{\text{e}} = M \setminus M^{\text{r}}$). Let $R_{nm,x}$ be the selected streaming data rate for each such tile in the enhancement content layer.
This content layer will be streamed to the user over her xGen wireless connectivity link characterized with transmission capacity $C^x$. The associated transmission delay can be characterized as $\tau^x = \frac{|M^{\text{r}}|E_r + \sum_{(n,m)\in M^{\text{e}}}R_{nm,x}\Delta T}{C^x}$, where $E_r$ denotes the data size of a raw GOP tile.
Finally, the latency on the client device involves decoding compressed tiles and rendering the viewport. These tasks will be carried out twice for the baseline and enhancement content layers streamed to the user, to enable respectively application reliability and high quality immersion, as introduced earlier. The decoding capability (speed) $z$ of the headset will need to be partitioned between decoding the baseline content layer and decoding the subset of compressed tiles from the enhancement content layer $M^{\text{e}}$. Let these two decoding speed allocations be denoted as $z^w$ and $z^x$, respectively. Thus, the required time to carry out each of these two decoding tasks can be formulated as $\tau^{z,w} = \frac{\sum_{(n,m)\in M}R_{nm,w}\Delta T}{z^w}$ and $\tau^{z,x} = \frac{\sum_{(n,m)\in M^{\text{e}}}R_{nm,x}\Delta T}{z^x}$. Similarly, the processing capability (power) $r$ of the headset will need to be partitioned between rendering the viewport solely from the baseline content layer and rendering the viewport at enhanced quality jointly from the baseline and enhancement content layers. Let these two processing power allocations be denoted as $r^w$ and $r^x$, respectively. The time delay of carrying out each of these two rendering tasks can be formulated as $\tau^{r,w} = \frac{E_v}{r^w b_h}$ and $\tau^{r,x} = \frac{E_v}{r^x b_h}$, where $E_v$ is the data size of the viewport after decoding and $b_h$ is the computed data volume per CPU cycle on the headset.
Benefiting from the earlier modeling, the quality of immersion delivered to a VR user by the envisioned dual-connectivity system can be accurately captured by formulating the expected viewport distortion experienced by the user as:
$$D\left(\{R_{nm,w},R_{nm,x}\}\right) = \sum_{(n,m)\in M^{\text{r}}} P_{nm} a_{nm} R_{nm,\max}^{b_{nm}} + \sum_{(n,m)\in M^{\text{e}}} P_{nm} a_{nm} \left(R_{nm,x} + R_{nm,w}\right)^{b_{nm}},$$
where $a_{nm}$ and $b_{nm}$ are the rate-distortion power law model parameters for tile $(n,m)$. Additionally, each summation term above can be weighted by a tile-level WS-MSE coefficient $w_{nm}$ that captures the unequal impact of each tile $(n,m)$ when the panoramic content is re-projected back on the \tsd sphere for viewing, as introduced earlier.
The analysis of latency and quality of immersion can enable pursuing end-to-end optimization of the system. One way of formally capturing this is via the problem formulation highlighted below that aims to maximize the quality of immersion delivered to a user given various system latency and resource constraints. To make the notation more compact, we henceforth replace the symbols $(n,m)$ with $(i,j)$.
\begin{align}
& \mathop{\min_{\{R_{ij,w}\},\{R_{ij,x}\},}}_{M^{\text{r}},\{z^w,z^x\},\{r^w,r^x\}} D\left(\{R_{ij,w},R_{ij,x}\}\right), \label{eq:pf1}\\
\hspace{-0.3cm} \text{subject to:} \quad & \tau^w + \tau^{z,w} + \tau^{r,w} \le \Delta T, \label{eq:c1}\\
& \tau^Z + \tau^x + \tau^{z,x} + \tau^{r,x} \le \Delta T, \label{eq:c2}\\
& R_{ij,w} \in [R_{ij,\min},R_{ij,\max}], \; R_{ij,x} \le R_{ij,\max} - R_{ij,w}, \label{eq:c3}\\
& r^w + r^x \le r, \quad z^w + z^x \le z. \label{eq:c5}
\end{align}
The inequalities \eqref{eq:c1} and \eqref{eq:c2} capture the application latency requirements associated with the transmissions on the two wireless connectivities. The inequalities in \eqref{eq:c3} capture the data rate limits enabled by the scalable \tsd tiling, where $R_{ij,\min}$ is the smallest possible data rate for tile $(i,j)$ (the data rate of its first embedded layer). We note that the transmission latency constraints \eqref{eq:c2} and \eqref{eq:c3} are stricter than and imply the respective transmission capacity constraints on the two wireless connectivity links. Thus, the latter two constraints do not need to be included in the optimization (\ref{eq:pf1})-\eqref{eq:c5}. Finally, the inequalities in \eqref{eq:c5} capture the limited computational resources of the VR headset of the user.
The problem (\ref{eq:pf1})-\eqref{eq:c5} is mixed-integer programming, which is hard to solve optimally in practice. Still, with a clever selection of $M^{\text{r}}$, one can pursue the optimal solution at lower complexity. In particular, for a given set of tiles to be sent as raw data, the problem above can be transformed to geometric programming and solved exactly via iterative approaches that converge rapidly \cite{xu2014global}. Moreover, though the number of tiles in $M$ is not excessive, instead of pursuing a combinatorial approach to identify the optimal $M^{\text{r}}$ as part of the overall solution to (\ref{eq:pf1})-\eqref{eq:c5}, one can accomplish that more effectively by sorting the tiles in $M$ according to an adequate criterion, related to the objective function, and constructing $M^{\text{r}}$ to comprise the first $k$ tiles from the sorted $M$. By sweeping $k$ across a range of values, solving (\ref{eq:pf1})-\eqref{eq:c5} for each such thereby constructed $M^{\text{r}}$ using geometric programming, and identifying the optimal value of the objective function in each case, one can then identify the overall solution.
Moreover, though $k$ can technically range up to $|M|$, the maximum number of tiles that can be sent as raw data will typically be smaller, as otherwise the system latency constraints will be violated. This is a result of the large data size of a raw GOP tile and the present transmission capacities of xGen technologies. Yet, when the latter are selected rather too conservatively in (\ref{eq:pf1})-\eqref{eq:c5}, even sending one raw tile can become unfeasible, and the choice of $k = 0$ lends itself as optimal. Finally, we observed that selecting the navigation likelihood weighted derivative of the rate-distortion dependency for a GOP tile as the criterion for sorting the tiles in $M$ in decreasing order leads to the same overall solution produced when instead a combinatorial search is used to identify $M^{\text{r}}$.
The optimization described in this section would be applied at the edge server to every subsequent GOP of the content transmitted to a given user. In a subsequent section, we will build upon the analysis and optimization developed here, to formulate an overall multi-user optimization strategy for the envisioned system.
{\bf Dynamic xGEn transmitter assignment.} Earlier, as part of the arena embodiment of the envisioned system outlined in Section~\ref{sec:SystemDesignArenaIntegrationEmbodiment}, we highlighted a prospective dynamic reassignment of xGen transmitters to mobile VR users during the course of a session for further performance enhancement, to complement the multi-user system level resource allocation optimization described next. In particular, during a session, as the mobile VR users navigate the content across the spatial area of the arena, such reassignment can be carried periodically such that, for instance, the smallest signal-to-noise ratio (SNR) across all transmitter and user pairs in the arena, is maximized. As SNR is inversely proportional to the distance from a user and an xGen transmitter, an equivalent approach can be pursued aiming to minimize the longest distance between a user and a transmitter, in an assignment. Either of these approaches will lead to a combinatorial optimization problem that may be complex to solve, depending on the number of users and transmitters in the system.
An alternative lower-complexity strategy would be to formalize the assignment as a graph-bottleneck matching on a bipartite graph comprising the users and xGen transmitters, as the two vertex set partitions, and where the graph edge weights would correspond to the distances between the users and transmitters. Concretely, in this case one will seek to find the maximum size vertex matching on the graph, whose biggest edge weight is smallest across all the matchings of that size. An efficient algorithmic solution can then be formulated that will integrate and benefit from the Dulmage–Mendelsohn decomposition of such a graph, to construct the optimal assignment $\pi$ incrementally. Our recent advances demonstrate the benefits of this strategy \cite{ChakareskiK:21}. Moreover, we have empirically observed that for the diverse settings considered in \cite{ChakareskiK:21} using the {\em max min SNR} as an objective metric for the transmitter assignment enables virtually the same performance, at much lower computational cost for the optimization, over the case of using the {\em max min SINR} (Signal-to-Interference and Noise Ratio). The performance evaluation that we carry out subsequently integrates the impact of interference via the SINR.
{\bf Multi-user system performance analysis and optimization.} The overall objective of the system is to maximize the delivered immersion quality across all users in the arena. Once users are assigned to xGen transmitters, as described above, this objective can be pursued by formulating and solving a respective optimization problem of interest. Building upon the notation and formalism introduced earlier in Section "Per-user end-to-end performance analysis", this multi-user problem can be formally described as:
\begin{align}
\text{(MU-OPT):} \quad & \min_{\left\{ \substack{\{R_{ij,w}^u\},\{R_{ij,x}^u\}, C_u^w, C_u^x, \\ M^{\text{r}}_u, \{z_u^w,z_u^x\},\{r_u^w,r_u^x\}, Z_u } \right\}_{\forall u}} \sum_u D_u\left(\{R_{ij,w}^u,R_{ij,x}^u\}\right), \label{eq:multi-user-optim}\\
\hspace{-0.3cm} \text{subject to:} \quad & \eqref{eq:c1} - \eqref{eq:c5}, \nonumber\\
& \sum_u Z_u \le Z, \label{eq:AggregateComputingCapacity}\\
& \sum_u C_u^w \le C^w, \label{eq:AggregateTransmissionCapacityLowFrequency}\\
& \sum_{u \in U_t} C_u^x \le C^{x,t}, \forall t,
\label{eq:AggregateTransmissionCapacityHighFrequencyPerTransmitter}
\end{align}
\noindent where we have introduced a subscript or superscript $u$ to all symbols used in (\ref{eq:pf1})-\eqref{eq:c5} earlier, to indicate a specific user $u$ here. Moreover, $Z$ denotes the aggregate decoding capacity of the edge server in \eqref{eq:AggregateComputingCapacity}, $C^w$ denotes the aggregate (down-link) transmission capacity of the traditional wireless technology in \eqref{eq:AggregateTransmissionCapacityLowFrequency}, and $C^{x,t}$ denotes the aggregate (down-link) transmission capacity of the xGen transmitter $t$ in \eqref{eq:AggregateTransmissionCapacityHighFrequencyPerTransmitter}. Finally, alike in \eqref{eq:AggregateTransmissionCapacityHighFrequencyPerTransmitter}, $U_t$ denotes the set of users assigned to xGen transmitter $t$.
Solving the optimization problem above (MU-OPT) in its full generality can be pursued via iterative techniques and the Lagrange multiplier method that would build upon the respective solution of the per-user optimization (\ref{eq:pf1})-\eqref{eq:c5} as a key inner step of such approaches. Moreover, solving for the respective Lagrange multipliers associated with the constraints \eqref{eq:AggregateComputingCapacity}-\eqref{eq:AggregateTransmissionCapacityHighFrequencyPerTransmitter} as part of the overall optimization can be carried out using for instance integrated fast sub-gradient techniques. Next, we outline the impact of parameter selection, user-transmitter assignment setting, and the indoor nature of our system on (MU-OPT).
With judicious analysis and choice of key system resource parameters captured by the constraints \eqref{eq:AggregateComputingCapacity}-\eqref{eq:AggregateTransmissionCapacityHighFrequencyPerTransmitter}, the general optimization outlined above can be simplified. Concretely, it is expected that in our system setting the edge server will be equipped with computing capabilities that exceed those of the mobile clients by orders of magnitude and that it will be well provisioned to assist abundantly the expected number of users in the arena system with such capabilities. Thus, one simple approach of dispensing with \eqref{eq:AggregateComputingCapacity} without preventing the integrated per-user optimization (\ref{eq:pf1})-\eqref{eq:c5} from exploring the full range of performance trade-offs and benefits is to set $Z_u = Z/N_u, \forall u$, where $N_u$ is the number of users in the arena. We have empirically verified this assessment in our extensive evaluations. It is rational to follow the same approach to dispense with \eqref{eq:AggregateComputingCapacity} by selecting $C_u^w = C^w/N_u$, with virtually no impact on the optimal performance delivered to each user for the following reasons. First, it is similarly expected that in our system setting the traditional wireless connectivity will be well provisioned to provide plentiful down-link transmission capacity to the expected number of mobile VR users to be served. Second, the envisioned arena system concept has an indoor nature and a relatively small spatial footprint. Third, the traditional wireless connectivity is integrated solely to augment the system/application reliability. Optimizing the choice of $C_u^w$ will not make an impact on this objective.
Finally, dispensing with \eqref{eq:AggregateTransmissionCapacityHighFrequencyPerTransmitter} and addressing the choice of $C_u^x, \forall u,$ is more subtle. On the one hand, in an embodiment setting of the envisioned system that integrates prospective multi-user assignment of its xGen transmitters, as discussed earlier in the last segment of Section "System design and integration embodiment", it is alike expected that each xGen transmitter $t$ will be equipped with ample additional transmission capacity to support the prospective serving of multiple users in parallel and provisioned adequately for the expected number of users to be assigned to a single xGen transmitter during a session. Then, dispensing with the constraint \eqref{eq:AggregateTransmissionCapacityHighFrequencyPerTransmitter} by splitting the transmission capacity of each transmitter $t$ uniformly across its assigned users, i.e., $C_u^x = C^{x,t}/|U_t|$, can be a sensible choice in this context as well, which may not impair notably the truly optimal high-quality immersion that can be enabled for every user in the arena via the assigned xGen transmitter. On the other hand, the rate-distortion dependency $D_u\left(\{\cdot,R_{ij,x}^u\}\right)$ that captures the delivered viewport quality may exhibit different performance trade-offs across different users $u$ over time, depending on the navigation characteristics of the users and the 6DOF content that is explored. This factor may lead to some performance loss depending on the magnitude of the difference between the uniformly assigned capacity $C_u^x$ and its rigorously selected value as an integral part of the optimization with the constraint \eqref{eq:AggregateTransmissionCapacityHighFrequencyPerTransmitter} kept in place. Still, it should be noted that for the present values of the transmission capacity $C_u^x$ that can be assigned to a user and rate-distortion dependencies of existing 3DOF/6DOF \tsd video/VR content, the operational points of a streaming system typically lie in the (far right) saturation portion of the rate-distortion performance dependency. This means that marginal differences in the value of $C_u^x$ assigned to user $u$ will only make a negligible difference in the overall system streaming performance and quality of experience delivered to the user. All the above considerations outlined here related to the choice of $C_u^x, \forall u,$ with the aim to prospectively omit \eqref{eq:AggregateTransmissionCapacityHighFrequencyPerTransmitter} from the optimization need to be carefully considered.
Yet, we have empirically observed in our extensive evaluations that the system setting comprising dynamic steering of the xGen transmitters and an exclusive assignment of one user per transmitter, i.e., the number of users and xGen transmitters in the arena is the same, provides the best performance over other prospective settings of user-transmitter assignment and transmitter steering, outlined earlier in the last segment of Section "System design and integration embodiment". For this characteristic setting of our system, the constraint \eqref{eq:AggregateTransmissionCapacityHighFrequencyPerTransmitter} will be met by default, by simply setting $C_u^x = C^{x,t}$ for the unique user $u$ assigned to that xGen transmitter $t$. This will dispense with the need for \eqref{eq:AggregateTransmissionCapacityHighFrequencyPerTransmitter} and will simultaneously enable delivering the highest possible immersion quality to user $u$. In turn, the latter will also signify that the best overall performance of the system is achieved, since the multi-user optimization (MU-OPT) formulated herein will decouple into multiple independent instances of the per-user optimization (\ref{eq:pf1})-\eqref{eq:c5} applied to every user-transmitter pair $(u,t)$, given the additive nature of the objective function in \eqref{eq:multi-user-optim} and the rationale for dispensing with the other constraints \eqref{eq:AggregateComputingCapacity}-\eqref{eq:AggregateTransmissionCapacityLowFrequency} described earlier. Lastly, beyond the argument carried out regarding the system-level constraints of (MU-OPT), as it applies to our specific context, we note that computational complexity reduction for an optimization method and user fairness are two commonly invoked reasons to justify uniform resource sharing across multiple entities in resource allocation problems, which that can be applicable here too.
Let $\{\text{GOP}_{u,k}\}_{\forall u}$ denote the collection of GOP content that needs to be transmitted to every user $u$ in the system at the onset of the GOP temporal interval $k$ of the 6-DOF remote immersion session. The optimization strategy described here would be applied at the edge server to every subsequent $\{\text{GOP}_{u,k}\}_{\forall u}$, i.e., for $k=1,2,\dots$
{\bf Performance analysis examples.} We highlight some of the key performance trade-offs arising in the arena embodiment setting of the envisioned system by implementing in simulation its most critical components and the optimization (\ref{eq:pf1})-\eqref{eq:c5}, together with the graph matching based strategy for dynamic user to xGen transmitter assignment outlined above. We considered that the spatial area of the arena is 6m $\times$ 4m and is split into six sectors. We further considered that there are six users in the arena and six xGen transmitters mounted on the ceiling above the center point of each cell, as illustrated in Figure~\ref{fig:SystemIntegrationDesign}. The beams of the transmitters are steered towards their assigned mobile users either electronically, in the case of mm-wave wireless technology, or mechanically, in the case of FSO wireless technology, as introduced earlier in the Section "High-frequency directional wireless transmission". We integrated tile-level spherical distortion coefficients $w_{nm}$ into the objective function in (\ref{eq:pf1}), when solving the optimization, and computed the PSNR of the optimal value of the objective function, to examine in our analysis. Henceforth, we denote this performance metric as WS-PSNR to distinguish it from the traditional PSNR, computed from the optimal value of the objective function without the integration of $\{w_{nm}\}$ into the optimization.
To carry out the evaluation, we used the popular 3-DOF 8K 60-fps \tsd video Runner and 6DOF VR content Museum, as well as actual 3-/6-DOF user navigation traces that we collected in our wireless VR arena lab \cite{ChakareskiK:21,ChakareskiASZ:21}.
Key system performance parameters were selected in the evaluation to correspond to the computational and transmission capabilities, respectively, of present state-of-the-art GPUs, VR headsets, and traditional/emerging wireless technologies
\cite{CuervoCK:18}. The considered values for these parameters are captured by the ranges of values representing the axes associated with the independent variables in the performance graphs included here.
The first two sets of performance analysis examples consider the \tsd video content. The third analysis example set considers the 6DOF VR content represented as a collection of spatial \tsd video viewpoints that users can dynamically select to navigate as they move across the arena (see Figure~\ref{fig:6DOF_360VideoStreaming}). Each such omnidirectional viewpoint is compressed into scalable \tsd tiling, at 120fps frame rate and 12K spatial resolution panorama.
\begin{figure}[htb]
\centering
\begin{subfigure}{.5\columnwidth}
\centering
\includegraphics[width=1\linewidth]{Figures/xGenTransCapacVsServerDecodingSpeed.eps}
\end{subfigure
\begin{subfigure}{.5\columnwidth}
\centering
\includegraphics[width=1\linewidth]{Figures/HeadsetDecodingSpeedVsHeadsetRenderingSpeed.eps}
\end{subfigure}
\vspace{-0.3cm}
\caption{Performance trade-offs: xGen transmission capability vs. edge server computing capability (left) and headset rendering vs. computing capabilities (right).}
\label{fig:PerformanceAnalysisTradeOffsCase1Case2}
\vspace{-0.3cm}
\end{figure}
{\em (i) 3-DOF \tsd video.} The performance results presented here are obtained by running the multi-user system optimization as described in the respective section, for many different placement configurations of the (static) users across the spatial area of the arena, and then computing the expected per-user performance. We recall that the users are not moving spatially across the arena during the session given the 3-DOF nature of the content. In Figure~\ref{fig:PerformanceAnalysisTradeOffsCase1Case2} (left), we examine the trade-offs between the transmission capacity of the xGen link to a user and the edge server's computing capability, and the resulting viewport quality experienced by the user. We can see that for lower xGen link capacities, increasing the server's decoding speed does not improve the delivered quality of immersion. That is because in such cases, almost all GOP tiles comprising the enhancement content layer are streamed as compressed data. Thus, the server's computing capability does not have much impact in such settings. On the other hand, for higher xGen link capacity values, the number of GOP tiles selected to be decoded at the server and streamed as raw data increases. Simultaneously, with the increasing link capacity, there is more room now for the remaining compressed tiles of the enhancement content layer to be transmitted at higher data rates. Both of these aspects contribute to higher quality of immersion for the VR user under such settings.
In Figure~\ref{fig:PerformanceAnalysisTradeOffsCase1Case2} (right), we examine the trade-offs between the rendering capability and decoding capability of the user's VR headset, and their interrelated impact on the delivered viewport quality. One can observe that the viewport quality moderately improves for around 0.2 dB as the rendering speed increases from 1.88 Gpixels/s to 9.4 Gpixels/s. Further increasing the rendering speed does not impact the delivered immersion fidelity. On the other hand, one can observe that increasing the headset decoding speed impacts the WS-PSNR more significantly. Concretely, for any given rendering speed examined in the figure, an improvement of around 2.5 dB is achieved as the device's decoding speed is increased from 100 Mbps to 500 Mbps. That is because, as the decoding speed increases, the device can decode GOP tiles compressed and transmitted at higher data rates, without violating the end-to-end latency constraints. Thereby, an improvement in the delivered viewport quality is achieved.
\begin{figure}[htb]
\centering
\begin{subfigure}{.5\columnwidth}
\centering
\includegraphics[width=1\linewidth]{Figures/xGenTransCapacVsHeadsetDecodingSpeed.eps}
\end{subfigure
\begin{subfigure}{.5\columnwidth}
\centering
\includegraphics[width=1\linewidth,trim={3cm 0 3cm 1cm},clip=true]{Figures/PerformanceComparisonDASH.eps}
\end{subfigure}
\vspace{-0.3cm}
\caption{Performance trade-offs: xGen transmission capability vs. client computing capability (left), and performance advances over a traditional state-of-the-art method (right).}
\label{fig:PerformanceAnalysisTradeOffsCase3Case4}
\vspace{-0.3cm}
\end{figure}
Next, in Figure~\ref{fig:PerformanceAnalysisTradeOffsCase3Case4} (left), we examine the trade-offs between the transmission capacity of the xGen link and the decoding speed of the headset, and their interrelated impact on the delivered quality of immersion. One can observe that the experienced WS-PSNR improves significantly as either the link capacity or the decoding speed increases. Concretely, the WS-PSNR increases for more than 1.5 dB when the link capacity increases from 700 Mbps to 1.2 Gbps. Similarly, when the link capacity is 700 Mbps, the WS-PSNR increases from 74.8 dB to 77.2 dB as the headset's decoding speed increases from 100 Mbps to 500 Mbps. Moreover, a WS-PSNR gain of around 4 dB is achieved when the link capacity is increased from 700 Mbps to 1.2 Gbps and the headset's decoding speed is increased from 100 Mbps to 500 Mbps. Increasing the value of either of these key capabilities enables GOP tiles compressed at higher data rates to be streamed from the server over the xGen link and decoded in time on the client device, thus improving the WS-PSNR significantly.
In Figure~\ref{fig:PerformanceAnalysisTradeOffsCase3Case4} (right), we examine the performance benefits enabled by the envisioned system over the traditional state-of-the-art. We implemented the current streaming standard MPEG-DASH following \cite{Petrangeli2017}, to stream the \tsd content, compressed using HEVC, over the traditional wireless connectivity link of a user. One can observe that significant performance gains in expected viewport quality and its variation are enabled. Concretely, MPEG-DASH can only achieve low average viewport quality of 63.68 dB, with a standard deviation of around 1.5 dB. These outcomes are in line with performance capabilities of emerging \tsd practices that can only stream lower resolution and lower frame rate \tsd videos at low to moderate quality at best, as outlined in the introduction of this article. On the other hand, the envisioned system enables high-quality viewport with expected WS-PSNR of 74 - 78 dB and standard deviation of 1.23 dB, for high frame rate and high resolution \tsd content, thus providing gains of 10-14 dB in immersion fidelity. These advances are motivating and are enabled by the dual connectivity streaming, scalable \tsd tiling, and edge computing that are synergistically integrated by the envisioned system together with rigorous end-to-end analysis, for maximum system performance efficiency.
A note on interpreting WS-PSNR results and their parallel with the more familiar PSNR metric may be appropriate here. The tile-level spherical distortion coefficients $w_{nm}$ are less than one and render the computed expected viewport distortion to be notably smaller. In our empirical evaluation, this has led to a consistent difference of around 11dB between the corresponding average PSNR and average WS-PSNR values. Moreover, a relative difference of a few dB in WS-PSNR may correspond to a difference of several dB in PSNR. Thus, the performance trade-offs of one method and its performance advances over another are highlighted more when interpreted through PSNR. Given the limited number of figures one can have, we opted to include the present collection of results under the WS-PNSR metric as most representative of the key performance aspects and advances introduced by the envisioned VR system concept.
\begin{figure}[htb]
\centering
\includegraphics[width=0.9\linewidth]{Figures/psnr_temporal_icc.eps}
\caption{Sample temporal viewport quality in our system. (The user load is 6.) Y-PSNR: User viewport luminance (Y) video signal PSNR.}
\label{fig:psnrtemporal}
\vspace{-15 pt}
\end{figure}
\underline{\em (ii) 6DOF VR:} Finally, we highlight the simultaneous immersion quality and application reliability benefits of the envisioned VR system, when streaming challenging 6DOF VR content to mobile users in the arena. The performance results presented here are obtained by running the multi-user system optimization as described in the respective section. We focus on the use of VLC as a representative xGen technology in this performance analysis example. Concretely, we examine the delivered viewport quality over time for a user, in comparison to that for state-of-the-art reference methods. {\em LiFi} is a cellular VLC system that dynamically assigns a moving user to the cell and its stationary transmitter that maximize her SNR \cite{zeng2019angle}. We also examined a variant of our system that uses only (single-link) high-frequency wireless transmission (VLC) to stream the content. We refer to this variant later as PropNoWiFi, to indicate that no traditional (sub-6 GHz) wireless connectivity link is used in parallel in that case, in the proposed system. There are six simultaneous users in the arena. To contrast the earlier performance analysis examples, here we measured the traditional PSNR of the delivered viewport for a user, to asses its quality.
We can observe from Figure~\ref{fig:psnrtemporal} that the viewport quality varies over time for all three compared methods, due to one of the following two events, both induced from time to time by rapid head and body navigation movements. Either the VLC link is transiently dropped or a brief mismatch takes place between the viewport knowledge used to construct the enhancement content layer at the server and the actual user viewport at the receiving client (see Section~\ref{sec:Scal360TilingViewportAdaptation}: “Scalable \tsd Video Tiling and Viewport-Driven Adaptation”). Still, the observed viewport quality variation is much lower for our system, which considerably increases its quality of experience and reliability, relative to {\em PropNoWiFi} and {\em LiFi} that experience an application downtime during such instances. Concretely, viewport quality gains of 5dB and 7dB and three times smaller standard deviation of viewport quality are enabled over the latter two methods, as seen from the two smaller inset graphs to the left. We can also observe that the user experienced 22$\%$ application downtime for {\em PropNoWiFi} and {\em LiFi}, compared to 0$\%$ downtime for our method, as shown in the smaller inset graph to the right. These benefits merit the technical advances of the envisioned VR system.
\begin{table}[htb]
\caption{Average user performance (six users in the system).}
\vspace{-15 pt}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
\textbf{Method} & \textbf{Data rate (T)} & $\boldsymbol{\sigma}_T$ & \textbf{Y-PSNR} & $\boldsymbol{\sigma}_{\textbf{PSNR}}$ \\
\hline
Prop. & 1020 Mbps & 164.83 Mbps & 50.37 dB & 5.79 dB \\
\hline
{\em LiFi} & 550 Mbps & 210.89 Mbps & 43.83 dB & 14.96 dB \\
\hline
\end{tabular}
\label{tab:tab3}
\end{center}
\vspace{-15 pt}
\end{table}
In Table~\ref{tab:tab3}, we summarize the average user performance in regard to viewport quality and delivered data rate. We can see that our system consistently outperforms {\em LiFi} across all performance metrics considered, enabled by its synergistic technical advances. We highlight that MPEG-DASH could only deliver inadequate viewport quality of 39 dB in this setting, as expected. Finally, we note that the proposed system enables another performance benefit over LiFi, which is higher robustness to the user load, where a more graceful degradation in performance is provided, as the number of users in the arena is increased \cite{ChakareskiK:21}. These results cannot be included here due to the limited space.
\section{Conclusions and the road ahead}
The two pedagogical aspects of the article, to educate and inspire the reader, have been necessarily intertwined in its presentation, given the nature of its topic and its objectives. We started by providing an overview of virtual reality and its key present challenges towards further advancement, highlighting its prospectively most exciting use cases for our society in the future, comprising high-fidelity remote scene immersion and untethered lifelike navigation. Then, a broad survey of related studies using traditional approaches and synergistic advances in other technologies, and a system-level primer on virtual reality and \tsd video technologies were integrated, to put these challenges in perspective, highlight the shortcomings of present implementations, and inspire new approaches.
Next, the article provided a contextual high-level review of several emerging technologies and unconventional techniques, highlighting that only by their synergistic integration we can aim to overcome the bottlenecks of hyper-intensive computation, ultra-high data rate, and ultra-low latency toward the envisioned future societal applications.
A novel 6DOF VR system concept that embodies this integration in an indoor setting and a rigorous analysis that captures the fundamental interplay among communication, computation, and scalable signal representation that arises herein, and that optimizes the system's end-to-end performance were presented subsequently. Finally, several representative performance analysis examples were provided to highlight these trade-offs and the benefits of the envisioned system. These outcomes motivate the system as a broad research platform for further investigations spanning actual implementation and deployment, new analyses and optimizations, and integration of further technologies and techniques. Moreover, many of the advances introduced by this feature article and such follow-up work could benefit related technologies, such as augmented reality and holograms.
\section{Authors}
{\bf Jacob Chakareski} (\href{mailto:<EMAIL>}{<EMAIL>}) completed his Ph.D. degree in electrical and computer engineering at Rice University and Stanford University. He is an Associate Professor in the College of Computing at the New Jersey Institute for Technology (NJIT), Newark, New Jersey, 07103, USA, where he holds the Panasonic Chair of Sustainability and directs the AI-Enabled Laboratory for Virtual and Augmented Reality Immersive Communications and Network Systems. Dr. Chakareski organized the first National Science Foundation (NSF) visioning workshop on future virtual and augmented reality communications and network systems in 2018 \cite{Chakareski:19b}. He has held research appointments with Microsoft, HP Labs, and Ecole Polytechnique F\'{e}d\'{e}rale de Lausanne (EPFL), and served on the advisory board of Frame, Inc (acquired in 2019 by Nutanix, Inc.). His research interests span next generation virtual and augmented reality systems, UAV-IoT sensing and networking, fast reinforcement learning, 5G wireless edge computing and caching, ubiquitous immersive communication, and societal applications. He received the Adobe Data Science Faculty Research Award in 2017 and 2018, the Swiss NSF Career Award Ambizione (2009), the AFOSR Faculty Fellowship in 2016 and 2017, and Best Paper Awards at ICC 2017 and MMSys 2021. His research has been supported by the NSF, NIH, AFOSR, Adobe, Tencent Research, NVIDIA, and Microsoft. For further information, please visit \url{www.jakov.org}.
{\bf Mahmudur Khan} received the B.Sc. degree in Electrical and Electronic Engineering from Bangladesh University of Engineering and Technology, in 2011, the M.S. degree in CSE from the University of Nevada Reno, in 2015, and the Ph.D. degree in CE from the University of Central Florida, in 2018, respectively. He was a postdoctoral fellow at NJIT and is presently an Assistant Professor at York College. His research interests include free-space-optical communications, wireless ad hoc networks, and UAV communications.
{\bf Murat Yuksel} received the BS degree in CE from Ege University, Izmir, Turkey, in 1996, and the MS and PhD degrees in CS from RPI, in 1999 and 2002, respectively. He is an Associate Professor with the ECE Department, University of Central Florida (UCF), Orlando. Prior to UCF, he was with the CSE Department, University of Nevada, Reno as a faculty member until 2016. He worked as a software engineer with Pepperdata, Sunnyvale, California, and a visiting researcher with AT\&T Labs and the Los Alamos National Lab. His research interests include networked, wireless, and computer systems with a recent focus on big-data networking, UAV networks, optical wireless, and network management.
\section*{Acknowledgements}
To Ilija Chakareski and Stojan Angjelov for their love. The work of Jacob Chakareski and Mahmudur Khan has been supported in part by the National Science Foundation (NSF) under awards CCF-2031881, ECCS-2032387, CNS-2040088, CNS-2032033, and CNS-2106150; by the NIH under award R01EY030470; and by the Panasonic Chair of Sustainability at NJIT. The work of Murat Yuksel has been supported in part by the NSF under awards CNS-2115215, CNS-2120421, and CNS-1836741. The authors are grateful to the editors Dr. Matthew McKay, Dr. Robert Heath, and Dr. Laure Blanc-Féraud, as well as to the numerous anonymous reviewers for their constructive guidance and comments that have considerably helped improve the quality of this feature article.
\bibliographystyle{F://ThinkPad/Jakov/Publications/_StyleFiles/IEEEtran}
\def\baselinestretch{0.8}
|
\section{Introduction}
The Cisco Annual Internet Report (2018-2023) forecasts a four-fold increase in the number of wireless fidelity (WiFi) hotspots between 2018 and 2023, to 628 million by the end of 2023 [\ref{Ref: Cisco White Paper}]. The dense deployment of WiFi hotspots would lead to intense competitions for available channels, due to the limited radio-frequency (RF) spectrum. This drives research on light fidelity (LiFi) [\ref{Ref: What is LiFi}], which operates in a way similar to WiFi but explores the extremely wide visible light spectrum ($\sim$300 THz). In contrast to WiFi, LiFi offers several prominent advantages including licence-free, availability in RF-restricted areas, and high security. Upon the existing light infrastructure, LiFi can also realize illumination and communication simultaneously, boosting the energy efficiency. Recent experimental results showed that LiFi is capable of rendering a link data rate up to 24 Gbps [\ref{Ref: VLC Speed 24Gbps}].
However, LiFi offers a relatively small coverage area with a single access point (AP), of usually 2-3 \rm{m} diameter. Also, LiFi is susceptible to channel blockage caused by opaque objects such as human bodies and furniture. Combining the high transmission speed of LiFi and the ubiquitous coverage of WiFi, the hybrid LiFi and WiFi network (HLWNet) is gaining research momentum recently. Such a network can greatly improve network capacity over stand-alone LiFi or WiFi networks [\ref{Ref: 7145863}]. Meanwhile, load balancing (LB) is identified as one of the key challenges in HLWNets [\ref{Ref: X. Wu Hybrid Network Survey}], due to the fact that WiFi APs have a larger coverage area but a lower system capacity than LiFi APs. When applying the signal strength strategy (SSS)\footnote{The SSS connects the user to the AP that provides the highest received signal power.} as if in a heterogeneous network, the users would be more likely to be served by WiFi than LiFi, making the WiFi system prone to overload. As a result, LB becomes essential and paramount to HLWNets, even though the users and their traffic demands are uniformly distributed in geography.
\subsection{Related Works}
\begin{table*}
\renewcommand{\arraystretch}{1.2}
\centering
\setlength\tabcolsep{1.5 mm}
\footnotesize
\caption{Comparison of Relevant Studies (MF: Max-Min Fairness; PF: Proportional Fairness; EPF: Enhanced PF).}
\begin{tabular}{|m{0.5cm}|m{1.2cm}|m{2.9cm}|c|c|c|c|m{4.2cm}|}
\hline
\textbf{Ref.} & \multicolumn{2}{l|}{\textbf{Method}} & \textbf{Optimality} & \textbf{Runtime} & \textbf{Fairness} & \textbf{LiFi AP density}\footnote{} & \textbf{Remark} \\
\hline
\multirow{2}{*}{[\ref{Ref: X. Li Cooperative LB TCOM}]} & \multicolumn{2}{l|}{Centralized optimization} & Optimal & Extremely High & \multirow{2}{*}{PF} & \multirow{2}{*}{16} & \multirow{2}{*}{Need to solve an NP-hard problem} \\
\cline{2-5} & \multirow{5}{*}{\vtop{\hbox{\strut Iterative}\hbox{\strut method}}} & Distributed optimization & \multirow{5}{*}{Near-optim.} & \multirow{5}{*}{High} & & & \\
\cline{1-1} \cline{3-3} \cline{6-8}
[\ref{Ref: Y. Wang LB Game TWC}] & & Game theory & & & MF, PF, EPF & 4, 16 & The number of iterations exponentially increases with the network size \\
\cline{1-1} \cline{3-3} \cline{6-8}
[\ref{Ref: College Admission Model TWC}] & & College admission model & & & $-$ & 1 & Solve a joint AP assignment and PA problem under constraints\\
\hline
[\ref{Ref: X. Wu AP Selection TCOM}] & \multirow{3}{*}{\vtop{\hbox{\strut Rule-based}\hbox{\strut decision} \hbox{\strut making}} } & Fuzzy logic (FL) & Sub-optim. & \multirow{3}{*}{Medium} & PF & 16 & Lack adaptiveness due to fixed rules \\
\cline{1-1} \cline{3-4} \cline{6-8}
[\ref{Ref: Globecom Conference}]& & Mixed FL and optim. & Near-optim. & & PF & $-$ & Optimize resource allocation in the decision-making procedure of FL \\
\hline
[\ref{Ref: RL for LB SJ}] & \multirow{3}{*}{\vtop{\hbox{\strut Learning}\hbox{\strut -aided} \hbox{\strut method}}} & Reinforcement learning & \multirow{3}{*}{Near-optim.} & \multirow{3}{*}{Low} & PF & 4 & Need retraining when the user number changes \\
\cline{1-1} \cline{3-3} \cline{6-8}
This work & & A-TCNN (DNN-aided) & & & PF & 4, 9 & Adapt to varying user numbers without the need for retraining \\
\hline
\end{tabular}\label{Table: Contributions}
\end{table*}
The conventional LB methods for HLWNets can be classified into three categories: centralized optimization [\ref{Ref: X. Li Cooperative LB TCOM}-\ref{Ref: Mobility-aware LB JOCN}], iterative methods [\ref{Ref: Y. Wang LB Game TWC}], [\ref{Ref: College Admission Model TWC}] and rule-based decision-making [\ref{Ref: Y. Wang Fuzzy Logic Handover ICC}-\ref{Ref: LB for Hybrid Networks}]. In [\ref{Ref: X. Li Cooperative LB TCOM}], a centralized optimization algorithm was proposed to achieve the proportional fairness (PF) between users. In [\ref{Ref: W. Ma Global Optimization LB CL}], the LB issue is formulated as mixed integer nonlinear programming (MINLP) problem considering users' data rate requirements and quality of service (QoS). In [\ref{Ref: Mobility-aware LB JOCN}], mobility-aware LB for both single transmission and multiple transmission modes was considered and then solved by the joint optimization method. Different from the above centralized optimization methods with their excessive computational complexity, autonomous optimization based on iterations can achieve near-optimal results, and to some extent reduce the complexity. In [\ref{Ref: Y. Wang LB Game TWC}], an iterative method based on evolutionary game theory (GT) was proposed for LB with multiple fairness functions considered. This iterative optimization method takes blockages, random orientation of LiFi receivers and users' data rate requirements into account. Taking power allocation (PA) into account, the authors in [\ref{Ref: College Admission Model TWC}] proposed college admission model based iterative method for joint AP assignment and power allocation problems in the three-tier hybrid VLC/RF network. However, the iterative method still requires a substantial amount of processing power. Rule-based decision-making method was proposed to further reduce the processing power [\ref{Ref: Survey on Fuzzy Logic}]. In [\ref{Ref: Y. Wang Fuzzy Logic Handover ICC}], the authors introduced a fuzzy logic (FL) based dynamic handover scheme for hybrid LiFi/RF network which considers user speed, data rate requirements, and time-varying signal-to-noise ratio (SNR) information, but without the optimization of throughput or user fairness. In [\ref{Ref: X. Wu AP Selection TCOM}], an FL based method was proposed to split the LB problem into two stages: i) determine the user which should be connected to WiFi; ii) allocate time resources for the remaining users in the stand-alone WiFi or LiFi network. Considering user mobility and light path blockage, the authors in [\ref{Ref: LB for Hybrid Networks}] proposed an FL based algorithm to reduce the complexity of the joint LB optimization problem in HLWNets. Although the rule-based decision-making method can handle the LB issue with a significant reduction of computational complexity, optimality is not achieved. Therefore, optimization based LB methods can provide optimal or near-optimal solutions, but centralized or iterative optimization algorithms require excessive computational complexity. In contrast, decision-making based methods, which demonstrate much lower complexity, compromise part of optimality.
\footnotetext[2]{The number of LiFi APs per coverage area of one WiFi AP.}
Unfortunately, the above conventional LB methods fail to achieve near-optimal performance and low computational
complexity at the same time, while the latter is crucial for meeting the zero-perceived latency requirement (0.1 ms) [\ref{Ref: 6G Vision}] in the sixth generation (6G) of wireless communication technology. For this reason, machine learning (ML) has recently enjoyed enormous attention to tackle the LB issue in heterogeneous networks [\ref{Ref: ML for Heterogeneous Survey}], and most of these approaches are based on reinforcement learning (RL). A few attempts have also been made to develop RL-aided LB methods for HLWNets [\ref{Ref: Q-learning for LB}-\ref{Ref: RL for LB SJ}]. In [\ref{Ref: Q-learning for LB}], a model-free and value-based Q-learning algorithm was proposed to maximize the network throughput. However, this method is infeasible to handle continuous state spaces, making it difficult to tackle densely deployed APs and users. In [\ref{Ref: DQN for LB}], Q-learning was combined with deep neural network (DNN) to jointly optimize bandwidth, power, and users association. Such a method is able to deal with discrete and continuous state spaces. In [\ref{Ref: RL for LB SJ}], a policy-based reinforcement learning (RL) algorithm was designed to determine an optimal AP assignment strategy, with the aim of maximizing the network throughput as well as the users' satisfaction and fairness. Compared with the value-based Q-learning methods, this policy-based method achieves a faster convergence. However, the above RL-aided methods all consider fixed numbers of users. When a new user joins the network or an existing user disconnects, these methods would need the retraining process to update the Q table or neural network, significantly limiting their practicability. Motivated by this, we aim to design an adaptive LB scheme for HLWNets which enables a state-of-the-art trade-off between optimality and complexity.
\subsection{Contributions}
In this paper, a novel DNN-aided LB method is proposed to tackle the LB issue in HLWNets, without the need for retraining when the number of users changes. The traditional DNN structure has fixed numbers of inputs and outputs, and thus is difficult to handle different numbers of users. To address this challenge, we develop a novel DNN structure named adaptive target-condition neural network (A-TCNN). The main contributions are explicitly contrasted to the existing literature in Table \ref{Table: Contributions}, and summarized as follows:
\begin{itemize}
\item A novel A-TCNN based LB approach is developed to tackle different numbers of users without the need for retraining. Specifically, the proposed A-TCNN introduces an adaptive mechanism to map a smaller user number to a larger one, followed by outputting the AP selection result for a target user based on the condition of other users.
\item Testing was carried out to validate the benefits of the proposed A-TCNN structure, which shows that compared with the method taken in test set, our A-TCNN method can achieve near equivalent performance in throughput and fairness. In addition, advantages of the proposed A-TCNN could be greater still when considering the effects of user clustering distributions.
\item A comprehensive comparison was conducted between the proposed A-TCNN and benchmarks, in terms of achievable throughput, fairness and computational complexity. Extensive results show that the proposed A-TCNN approach can achieve near-optimal network performance with a significantly reduced runtime up to 78 and 1400 times compared with two benchmarks.
\end{itemize}
The remainder of this paper is organized as follows. The system model is presented in Section II. The proposed method is elaborated in Section III. In Section IV dataset collection methods are summarized, and simulation results are shown in Section V. Finally, conclusion and future work are given in Section VI.
\textit{Notations:} Throughout the paper, we use lowercase italic letters to denote scalar variables, while bold lowercase and uppercase letters stand for vectors and matrices, respectively. The notations $\left|\cdot\right|$, $\left\| \cdot \right\|$ and $[\cdot]^{T}$ represent the absolute value, the Frobenius norm, and the transpose of a vector or matrix, respectively. $\mathbb{E}[\cdot]$ and $\mathrm{Var}[\cdot]$ are the expectation and variance values of a random variable. The operator $\lfloor \cdot \rfloor$ denotes the floor function, while ${\rm{mod}}(a, b)$ returns the remainder after a real number $a$ is divided by real number $b$. $\mathbb{R}^n$ represents the real coordinate space with a dimension of $n$. $\mathcal{N}(0, \sigma)$ denotes a Gaussian distribution with zero mean and standard variance $\sigma$.
\section{System Model}
The system model related to HLWNets is introduced in this section, including the network architecture, channel models and link capacity, and performance metrics.
\subsection{Network Architecture}
Fig. \ref{Fig:system model} shows an indoor HLWNet that consists of one WiFi AP and a number of LiFi APs. The WiFi AP is placed in the centre of the room on the ground, providing coverage to the entire room. The LiFi APs are arranged in a grid, with each AP embedded into a ceiling lamp, covering a confined area. Frequency reuse with a reuse factor of 4 is adopted to avoid inter-cell interference (ICI) among adjacent LiFi APs [\ref{Ref: Parallel LiFi}]. The ICI between further APs is trivial and can be neglected. The users are randomly located on the ground with a uniform distribution. The data rate requirements of the users are considered to be independent and identically distributed random variables. Without loss of generality, it is assumed that these variables follow a Gamma distribution with shape parameter $k$ and scale parameter $\theta$ [\ref{Ref: Gamma Distribution}]. Let $R_{j}$ denote the data rate required by the $j$-th user. The expected value of the average data rate requirement is $\bar{R} = k\theta$. Each user is connected to one AP, either WiFi or LiFi, while each AP can serve multiple users via time-division multiple access (TDMA).
\subsection{Channel Models and Link Capacity}
Quasi-static channels are considered here. For WiFi, the log-distance path loss model in [\ref{Ref: WiFi Model}] is adopted. As for LiFi, the channel consists of the line-of-sight (LoS) and first-order non-line-of-sight (NLoS) paths, as illustrated in Fig. \ref{Fig:system model}. The corresponding expressions can be found in [\ref{Ref: LoS and NLoS} eq.(10) and eq.(12)]. The capacity of WiFi is bounded by Shannon capacity. With respect to LiFi, a tighter bound can be found in [\ref{Ref: Tight Bound JLT}] due to the non-negative real signals in LiFi. Let $\gamma_{i,j}$ denote the signal-to-noise ratio (SNR) of the link between AP $i$ and user $j$. The link capacity $C_{i,j}$ can be expressed as:
\begin{equation}
C_{i,j} = \left\{
\begin{aligned}
&{{B_i}\over{2}}{\log_2}\big(1 + {{e}\over{2\pi}} \gamma_{i,j}\big), \:\:\: \rm{for \: LiFi} \\
&{B_i}{\log_2}(1 + \gamma_{i,j}) , \quad\quad \rm{for \: WiFi} \\
\end{aligned}
\right.
\end{equation}
where $e$ is the Euler’s number and $B_{i}$ denotes the bandwidth of AP $i$.
\begin{figure}[t]
\centering
\includegraphics[height=6cm,width=9cm]{ChannelModel.eps}
\caption{Schematic diagram of an indoor HLWNets.}
\label{Fig:system model}
\end{figure}
\subsection{Performance Metrics}
Let $\Gamma$ denotes the achievable throughput, which is given by:
\begin{equation}
\Gamma = \sum\limits_{i \in \mathbb{S}}\sum\limits_{j \in \mathbb{U}}{\rho_{i,j}}{\chi_{i,j}}{C_{i,j}},
\end{equation}
where $\mathbb{S} = { \{i|i\in [1, 2, ..., N_{\rm{t}}] \} }$ and $\mathbb{U} = { \{j|j\in[1, 2, ..., N_{\rm{u}}] \} }$ denote the sets of APs and users, while $N_{\rm{t}}$ and $N_{\rm{u}}$ are the total numbers of APs and users, respectively; $\chi_{i,j} = 1$ indicates there is a connection between AP $i$ and user $j$, and otherwise $\chi_{i,j}=0$; $\rho_{i,j} \in [0, 1]$ denotes the portion of time resource that is allocated by AP $i$ to user $j$.
Let $S_j$ denote the satisfaction degree of user $j$, which can be expressed as:
\begin{equation}
{S_{j}} = \min \bigg\{ \frac{ {\sum_{i \in \mathbb{S}}{{\rho_{i,j}}{\chi_{i,j}}{C_{i,j}}}} }{{R_j}}, 1\bigg\}.
\end{equation}
The fairness among users, denoted by $\xi$, is commonly measured by Jain's fairness index [\ref{Ref: Jain's Fairness}]:
\begin{equation}
\xi =\frac{\left(\sum_{j \in \mathbb{U}}S_j\right)^2}{N_{\rm{u}}\sum_{j \in \mathbb{U}}S_j^2}.
\end{equation}
\section{Proposed DNN-Aided Framework for LB}
\begin{figure*}[htb]
\centering
\includegraphics[width=18cm]{Adaptive_TCNN.eps}
\caption{Block diagram of the proposed A-TCNN.}
\label{Fig:Adaptive TCNN}
\end{figure*}
In this section, we present a novel DNN-aided LB method for HLWNets, which consists of two key components: the target-condition neural network (TCNN) and adaptive mechanism with respect to input user number. As illustrated in Fig. \ref{Fig:Adaptive TCNN}, the block diagram of the proposed A-TCNN framework includes two solid rounded rectangles. The right block represents the TCNN model, and the left one refers to the adaptive mechanism with respect to number of working users in the HLWNets. A detailed explanation on the proposed A-TCNN model for LB is given in subsections III-A and III-B. Finally, we introduce our training and evaluation procedure for the A-TCNN model.
\subsection{Target-Condition Neural Network}
For the conventional DNN model, a basic solution is to input the collected data including SNRs and data rate requirements for all users, and then to simultaneously estimate their optimal AP assignments, which is referred as DNN in this work. However, it does not support the adaptation to the number of users. When applying the adaptive mechanism in DNN model, the mirroring users split from the same user may result in different output results, which is conflicted in nature. Without an adaptive mechanism, the trained DNN network only working on a specific user case lacks of scalability. To solve this limitation, we focus our attention on a single user which needs to be estimated in the output and set all other users as condition, shown in the inner dashed rounded rectangle of Fig. \ref{Fig:Adaptive TCNN}. In our implementation, the TCNN is separated into three parts, namely \textit{target}, \textit{condition} and \textit{combiner}, respectively.
\subsubsection{Target}
For the target block, the focus is on user $j$. User $j$ contains two kinds of data: SNRs between candidate AP $i$ and user $j$ pairs, and the partial data rate requirement of user $j$ after the mapping. Therefore, the fed vector of target is expressed as $\boldsymbol{\tilde{x}}_j = [\gamma_{1,j}, \gamma_{2,j}, ..., \gamma_{N_{\rm{a}},j}, \tilde{R}_j]^T \in \mathbb{R}^{N_{\rm{a}}+1}$. Then, a fully-connection (FC) layer with 6 neurons is employed to process $\boldsymbol{\tilde{x}}_j$, which is illustrated in the rounded rectangle with yellow dash in Fig. \ref{Fig:Adaptive TCNN}. Let us denote the weights and bias of FC layer by $\mathbf{W}_1 \in \mathbb{R}^{(N_{\rm{a}}+1)\times 6}$ and $\boldsymbol{b}_1 \in \mathbb{R}^6$. Therefore, the output of target part can be expressed as:
\begin{equation}
\boldsymbol{u}_j = {\mathbf{W}_1}{\boldsymbol{\tilde{x}}_j} + \boldsymbol{b}_1.
\end{equation}
\subsubsection{Condition}
In the condition block, we deem all users as condition to input into the TCNN, which is an arrangement of entire SNR and data rate requirement parameters (including the focused user $j$ mentioned in the above target block). Let us denote the input of condition as $\boldsymbol{\bar{x}}_{\rm{C}}$, given as $\boldsymbol{\bar{x}}_{\rm{C}} = [{\boldsymbol{\tilde{x}}^T_1}, {\boldsymbol{\tilde{x}}^T_2}, ..., {\boldsymbol{\tilde{x}}^T_M}]^T \in \mathbb{R}^{M(N_{\rm{a}}+1)}$, where $M$ is a default maximum user number of input. For example, the maximum number of working users that can be estimated by the proposed TCNN is 50 if we fix $M$ as 50. To reduce the dimension of the condition while extracting the principal components, two FC layers with 64 and 6 neurons are utilized here, and the corresponding network parameters are denoted as $\mathbf{W}_2 \in \mathbb{R}^{64 \times M({N_{\rm{a}}}+1)}$, $\boldsymbol{b}_2 \in \mathbb{R}^{64}$ and $\mathbf{W}_3 \in \mathbb{R}^{6 \times 64}$, $\boldsymbol{b}_3 \in \mathbb{R}^6$, respectively. After that, the rectified linear unit (ReLU) is used as the activation function, defined as ${f_{\rm{ReLU}}}(\boldsymbol{x}) = {\rm{max}}({\rm 0}, \boldsymbol{x})$. Note that the eigenvalues of the FC layer's output would gradually approach the saturation interval of the activation function, causing gradients to tend to vanish during the training process. Therefore, the output from FC must to be normalized using a batch normalization (BN) layer to enable the eigenvalues distribution to stay within the standard normal distribution [\ref{Ref: Batchnorm}], which is more sensitive to the activation function. With this step, vanishing gradients can be avoided and training convergence rate improved. Here, the BN operation can be formulated as follows:
\begin{equation}
f_{\mathrm{BN}}(\boldsymbol{x}) = \frac{ \boldsymbol{x} - \mathbb{E}\left [ \boldsymbol{x} \right ]} {\sqrt{\mathrm{Var}\left [ \boldsymbol{x} \right ] + \epsilon }} \mu + \nu
\end{equation}
where the $\epsilon$ is a small constant close to zero to avoid division by zero, $\mu$ and $\nu$ are the trainable factors for scaling and shifting the distribution, which are set as 1 and 0 initially. Finally, the output of the condition block is as follows:
\begin{equation}
\small
\boldsymbol{u}_{\rm{C}} = {f_{\rm{ReLU}}}\left\{ {{f_{\rm{BN}}}\left[{{\mathbf{W_3}}\left\{ {{f_{\rm{ReLU}}}\left[ {{f_{{\rm{BN}}}}\left( {{\mathbf{W_2}}{{\bar{\boldsymbol{x}}}_{\rm{C}}} + {\boldsymbol{b_2}}} \right)} \right]} \right\} + \boldsymbol{b_3}} \right]} \right\}
\end{equation}
\subsubsection{Combiner}
In the combiner part, the subtracted features from the target and condition are concatenated (Cat) together in the aggregator and fed into the next layers. Here, the operation of Cat is expressed as ${\mathrm{Cat}}({\boldsymbol{u}_j}, {\boldsymbol{u}_{\rm{C}}}) = [{\boldsymbol{u}^T_j}, {\boldsymbol{u}^T_{\rm{C}}}]^T \in \mathbb{R}^{12}$. Similarly, an FC layer with weights matrix $\mathbf{W}_4 \in \mathbb{R}^{N_{\rm{a}} \times 12}$ and bias vector $\boldsymbol{b}_4 \in \mathbb{R}^{N_{\rm{a}}}$ is introduced here, followed by the probability-based activation function softmax. The softmax function is defined as ${f_{\rm{softmax}}}(\boldsymbol{x}) = e^{\boldsymbol{x}} / {\sum}e^{\boldsymbol{x}}$. Lastly, the focused user $j$ connects the corresponding AP $i^{\ast}$ with the largest possibility using a soft decision maker. The estimated optimal AP $i^{\ast}$ in the combiner can be calculated as:
\begin{equation}
{i^*} = \arg \mathop {\max}\limits_{i \in \mathbb{S}} \left\{ {{f_{{\rm{softmax}}}}\left[ {{{\mathbf{W}}_4}\left( {\rm{Cat}}({\boldsymbol{u}_j}, {\boldsymbol{u}_{\rm{C}}} \right) + {\boldsymbol{b}_4}} \right]} \right\},
\end{equation}
and the final estimated output is ${\boldsymbol{\hat{y}}_j} = [y_1, y_2, ..., {y_{N_{\rm{a}}}}]$, which satisfies that ${\boldsymbol{\hat{y}}}_j(i) = 1$ when $i = i^*$, otherwise ${\boldsymbol{\hat{y}}_j}(i) = 0$ for $\forall i \in \mathbb{S}$ and $\forall j \in \mathbb{U}$. Note, the estimated binary output ${\boldsymbol{\hat{y}}_j}(i)$ indicates the AP selection choice for each user $j$ with respect to AP $i$ in the HLWNets, where ${\boldsymbol{\hat{y}}_j}(i) = 1$ means there should be a connection, otherwise no connection.
\subsection{Adaptive Mechanism}
In the proposed TCNN, the estimation target is focused from the entire users to a single user, which is estimated in the combiner. However, the dimension of input fed into the condition block is unchangeable, i.e., the proposed TCNN can only work on $M$ users number case. To provide an adaptive functionality with respect to the estimated user number, we further present an adaptive mechanism in this subsection, which is illustrated in the left solid rounded rectangle in Fig. \ref{Fig:Adaptive TCNN}.
In general, the proposed adaptive mechanism can transform $N_{\rm{u}}$ users into $M$ mapped users. First, let us denote the user-centric data (including considerable SNRs $\gamma_{i,j}$ and data rate requirement $R_j$ for user $j$) as $\boldsymbol{x}_j = [\gamma_{1,j}, \gamma_{2,j}, ..., \gamma_{N_{\rm{a}},j}, R_j]^T \in \mathbb{R}^{N_{\rm{a}}+1}$. Then, a set of $\boldsymbol{x}_j$ are sent into the mapping operator, in which the mirroring users that equally share the data rate requirement $R_j$ while remaining the same SNR values are created. Here, we use $\boldsymbol{\tilde{x}}_j \in \mathbb{R}^{N_{\rm{a}}+1}$ indicates the mirroring user $j$, where $j$ is an integer ranging from 1 to $M$. For example, the user number ${{N_{\rm{u}}}}$ is 20 here and $M$ is set as 30. After mapping, the mirroring data matrix can be expressed as $\left [\boldsymbol{\tilde{x}}_1, \boldsymbol{\tilde{x}}_2, ..., \boldsymbol{\tilde{x}}_{20}, \boldsymbol{\tilde{x}}_{21}, ..., \boldsymbol{\tilde{x}}_{30} \right ]$, here $\boldsymbol{\tilde{x}}_j = [\gamma_{1,j}, \gamma_{2,j}, ..., \gamma_{N_{\rm{a}},j}, R_j/2]^T$ for $\forall j \in [1, 2, ..., 10]$ and $[21, 22, ..., 30]$, and $\boldsymbol{\tilde{x}}_j = [\gamma_{1,j}, \gamma_{2,j}, ..., \gamma_{N_{\rm{a}},j}, R_j]^T$ for $\forall j \in [11, 12, ..., 20]$. A general algorithm of the adaptive mechanism is shown in Alg. \ref{Alg:Adaptive_Mechanism}. Finally, an operation is conducted to separate the target and condition, and the corresponding streams are denoted as $\boldsymbol{\tilde{x}}_j$ and $\boldsymbol{\bar{x}}_{\rm{C}} = [{\boldsymbol{\tilde{x}}^T_1}, {\boldsymbol{\tilde{x}}^T_2}, ..., {\boldsymbol{\tilde{x}}^T_M}]^{T}$, respectively.
\begin{algorithm}[t]\label{Alg:Adaptive_Mechanism}
\SetAlgoLined
\KwIn{${\boldsymbol{x}_1}, {\boldsymbol{x}_2}, ..., {\boldsymbol{x}_{N_{\rm{u}}}}, {N_{\rm{u}}}, M$}
\KwOut{$\boldsymbol{\tilde{x}}_1, \boldsymbol{\tilde{x}}_2, ..., \boldsymbol{\tilde{x}}_M$}
Calculate $a =\lfloor{M/{{N_{\rm{u}}}}} \rfloor $, and $b = {\rm{mod}} (M, {{N_{\rm{u}}}}) $ \;
\For {$j = 1$ {\rm{to}} $M$}{
\eIf{$ j \leqslant b$}{
$\boldsymbol{\tilde{x}}_j \Leftarrow [\gamma_{1,j}, \gamma_{2,j}, ..., \gamma_{N_a+1,j}, R_j/(a+1)]$\;
}
{\eIf{$ j \leqslant {N_{\rm{u}}}$}
{$\boldsymbol{\tilde{x}}_j \Leftarrow [\gamma_{1,j}, \gamma_{2,j}, ..., \gamma_{N_a+1,j}, R_j/a]$\;}
{$j^\ast = {\rm{mod}}(j-1, N_{\rm{u}}) + 1 $ \;
$\boldsymbol{\tilde{x}}_j \Leftarrow [\gamma_{1,j^\ast}, \gamma_{2,j^\ast}, ..., \gamma_{N_a+1,j^\ast}, R_j/a]$\;}
}}
\caption{Adaptive Mechanism.}
\end{algorithm}
\subsection{Training and Testing}
In our training phase, one key feature is that adaptive user number can be chosen to train the network. Assuming that the user number $N_{\rm{u}}$ ranges from 5 to $M$ and dataset collection can be resolved in the next section, our training process can be introduced by two aspects.
\subsubsection{Fixed Numbers of Users} For the case of TCNN with specific user number of $N_{\rm{u}}$, we first collect $N$ batches for training and the batch size is $T$. In one batch, each column denotes a sample showing as $\boldsymbol{\bar{x}}_{\rm{C}} = [{\boldsymbol{\tilde{x}}^T_1}, {\boldsymbol{\tilde{x}}^T_2}, ..., {\boldsymbol{\tilde{x}}^T_M}]^{T}$. Note that the fed data $\boldsymbol{\bar{x}}_{\rm{C}}$ containing $\gamma_{i,j}$ and $R_j$ needed to be normalized separately, since $\gamma_{i,j}$ and $R_j$ are drawn from different distributions. Here, linear normalization is adopted for $\gamma_{i,j}$ while logarithmic normalization is utilized for $R_j$.\footnote{For SNRs in dB scale, most values follow a normal distribution, while $R_j$ with a gamma distribution is mainly centred around the mean value of $k\theta$. Providing that the minimum and maximum values for $R_j$ are 1 and 1000 Mbps respectively, and the mean value is 100 Mbps, linear normalization for $R_j$ would compress most $R_j$ to a small range, which would slow down learning. As a result, we adopted the logarithmic normalization, which can better scale most $R_j$ in range of [0, 1].} In addition, all batches are normalized with the same maximum and minimum values, which are obtained from the training dataset. After that, normalized $N$ batches are grouped together and fed into the TCNN. The ground truth labels can be denoted as $\boldsymbol{y}_j$, which is the correct AP assignment results for the input user $j$ and has been saved in the output of dataset. During training process, we use stochastic learning rather than batch learning as it can speed up learning, particularly on large redundant datasets [\ref{Ref: Efficient BackProp}]. We selected the mean square root (MSE) loss function to maximize the agreement between the estimation from the proposed TCNN and the label values, formulated as:
\begin{equation}\label{Equation:MSE Loss}
L_{\mathrm{MSE}}\left ( \theta \right )=\frac{1}{T}\sum_{j=1}^{T}\left\| {\boldsymbol{y}_j}-{\hat{\boldsymbol{y}}_j}\right\|^{2},
\end{equation}
where the $\theta$ is the trainable parameter. During training process, network parameter $\theta$ is updated and optimized using adaptive moment estimation (Adam) scheme [\ref{Ref: Adam}] for each batch over $T$ batch samples. As a result, the updating of $\theta$ can be formulated as follows:
\begin{equation}
\theta \Leftarrow \theta -\eta \nabla L_{\mathrm{MSE}}\left ( \theta \right ),
\end{equation}
where $\eta$ is the learning rate during training process, and $\nabla L_{\mathrm{MSE}}(\theta)$ means the gradient of the loss function with respect to $\theta$.
\subsubsection{Adaptive Numbers of Users}
To adapt to a variable number of users, we aim to convert the dataset consisting of $N_{\rm{u}}$ users into the mapped dataset with $M$ users so that the TCNN model with $M$ input users is still working for a smaller user count. The detailed illustration of adaptive mechanism can be referred to Alg. \ref{Alg:Adaptive_Mechanism}.
The testing phase includes loss testing and performance evaluation. As for the loss testing, 256 samples are collected for each user example. The computation process of testing is the same as the training, while the loss will not be back-propagated to update parameters of the neural network. The model can be verified if it overfits the training set through the MSE loss testing formulated in (\ref{Equation:MSE Loss}). In the evaluation stage, two network metrics are adopted in this work: \textit{accuracy} and \textit{performance gap}.
\begin{itemize}
\item \textit{Accuracy:} We define accuracy as the ratio between the number of users with correct AP connection estimated by A-TCNN model and the total number of tested users.
Note correct AP connection is evaluated based on the ground truth of dataset. Besides, users at different positions are sequentially fed into the TCNN model.
\item \textit{Performance gap:} In HLWNets, AP assignment results have a significant influence on the achievable throughput of hybrid network. Therefore, we define performance gap as the achievable network throughput gap between the estimated AP selection and the real AP assignment results, where the same optimal PA is achieved for comparison.
\end{itemize}
\section{Dataset Collection}
In this section, we introduce and discuss a number of potential dataset collection methods for the proposed A-TCNN. These dataset collection methods are four-fold: \textit{global optimization}, \textit{direct decision-making}, \textit{iteration based}, and \textit{mixed decision-making and optimization} methods.
\subsection{Global Optimization Method}
The aim of global optimization is to maximize the user satisfaction with proportional fairness [\ref{Ref: PF}].
Global optimization is an optimal solving algorithm which provides the upper bound for the problem formulation, thus it is also termed as exhaustive search or brute force search. The problem formulation of global optimization for LB in HLWNets can be expressed as:
\begin{equation}\label{eq:Problem}
\begin{array}{l}
\!\!\!\!{\max\limits_{\chi_{i,j},\rho _{i,j}}\quad\! \sum\limits_{j \in \mathbb{U}}{\log S_{j}}}\\
{{\rm{s}}{\rm{.t}}{\rm{.}}\quad\quad\! \sum\limits_{i \in \mathbb{S}} {\chi_{i,j}} = 1} ,\forall j \in \mathbb{U};\\
\quad\quad\quad \; \sum\limits_{j\in\mathbb{U}} {\chi_{i,j}{\rho_{i,j}} \le 1} ,\forall i \in \mathbb{S};\\
\quad\quad \quad\; 0 \le {\rho _{i,j}} \le 1,\forall\; i,j;\\
\quad\quad \quad\; {\chi_{i,j}} \in \{ 0,1\} ,\forall\; i,j,
\end{array}
\end{equation}
Here (\ref{eq:Problem}) is similar to problem formulation given in [\ref{Ref: X. Li Cooperative LB TCOM}], but the objective function is revised with consideration of users' satisfaction. As shown in (\ref{eq:Problem}), assigning AP selections is an NP-hard problem [\ref{Ref: NP-hard Problem}] and PA in a stand-alone network is a nonlinear programming problem. Although it has been shown that MINLP can be solved by the OPTI toolbox, nonlinear branch-and-bound algorithm based global optimization solvers (e.g. BONMIN, SCIP) take a long time to return global results [\ref{Ref: MINLP}]. Particularly in the case of dense deployment of APs and users, the computational complexity is exponentially increasing with regard to the network size, which is far from a wise deployment for a real-time and user-centric network.
\subsection{Direct Decision-Making Method}
For direct decision-making method, the object is to make an artificial but reasonable rule set according to the existing social or natural system. In [\ref{Ref: X. Wu AP Selection TCOM}], the authors presented an FL based LB method for HLWNets, where the ICI is solved by introducing SNR variance and activity of adjacent AP to define how likely the user is affected by the ICI. However, in this paper, we consider frequency reuse to avoid ICI [\ref{Ref: Parallel LiFi}]. As a result, we revise the original FL rules for a fair comparison with benchmarks and the new fuzzy rule is listed in Table \ref{Table: Fuzzy Rule}. Note in the FL method, we adopt SSS as the AP assignment method in the second stage of LB, thus only the nearest LiFi AP providing biggest SNR will be considered for each user in the fuzzy rule. Though this kind of heuristic decision-making method can significantly reduce the processing time, it is unknown how far the solution is from achieving the optimal optimization results. In addition, the designed rules cannot adapt to changes in the environment such as changing user number and/or network size, as is seen in HLWNets.
\begin{table}
\renewcommand{\arraystretch}{1.2}
\centering
\setlength\tabcolsep{1.6 mm}
\caption{Fuzzy Rules for the FL Method (Req.: Required Data Rate; Ava.: Availability of WiFi or LiFi AP Resource; Sel.: Selection)}
\scriptsize
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\textbf{Rule} & \textbf{Req.} & \textbf{WiFi SNR} & \textbf{LiFi SNR} & \textbf{WiFi Ava.} & \textbf{LiFi Ava.} & \textbf{WiFi Sel.} \\
\hline
1 & Low & $-$ & $-$ & \textit{not} Low & $-$ & High \\
\hline
2 & \textit{not} Low & $-$ & $-$ & \textit{not} Low & $-$ & Medium \\
\hline
3 & $-$ & $-$ & $-$ & Low & \textit{not} Low & Low \\
\hline
4 & $-$ & Low & \textit{not} Low & Low & Low & Low \\
\hline
5 & $-$ & Low & Low & Low & Low & Medium \\
\hline
6 & $-$ & \textit{not} Low & Low & Low & Low & High \\
\hline
7 & $-$ & \textit{not} Low & \textit{not} Low & Low & Low & Medium \\
\hline
8 & High & $-$ & High & $-$ & $-$ & Low \\
\hline
9 & \textit{not} High & High & $-$ & \textit{not} Low & $-$ & High \\
\hline
\end{tabular}\label{Table: Fuzzy Rule}
\end{table}
\subsection{Iterative Method}
Another typical strategy which is capable of solving LB problem utilizes a number of iterations to reach a steady state, which means that a near-optimal result is achieved. Here, we select GT based LB method [\ref{Ref: Y. Wang LB Game TWC}] as iteration based baseline to compare. During the GT process, each player occurs randomly and follows the rule that the player with lower value of payoff would be more likely to change its strategy, which is termed as `mutation and selection mechanism' rule [\ref{Ref: Y. Wang LB Game TWC}]. The Nash equilibrium is obtained when all players can not continue mutation and selection for one loop. However, it is not always optimal as no single player can improve its payoff does not guarantee that the entire players can further globally maximize their payoffs. Besides, this GT approach is highly sensitive to initial states.
\subsection{Mixed Decision-Making and Optimization Method}
In this paper, we extend our previous work [\ref{Ref: Globecom Conference}] to LB for HLWNets and introduce a revised FL based method which combines decision-making and optimization to provide a better trade-off between optimality and complexity. This method is termed as FL-OPT in this paper. The proposed mixed method for dataset collection considers user satisfaction maximization in each decision step. In general, the mixed LB method follows two principles: i) for users candidates, each user $j$ first chooses the AP providing the biggest $S_{j}$ as candidates; ii) for chosen AP candidates, connect a pair of AP and user with the highest output score yielded in fuzzy logic system. Therefore, the proposed dataset collection method has inherent low-complexity advantage of decision-making and performance enhancement characteristic of optimization method. Specifically, the satisfaction optimization problem providing a specific set of $\chi_{i,j}$ can be formulated as:
\begin{equation}\label{eq:Problem2}
\begin{array}{l}
\max\limits_{\rho_{i,j}} \;\quad\sum\limits_{j \in \mathbb{U}} {\log S_{j}}\\
\;\;{\rm{s}}{\rm{.t}}{\rm{. }}\quad\;\sum\limits_{j\in\mathbb{U}} {\chi_{i,j}{\rho_{i,j}} \le 1} ,\forall i \in \mathbb{S};\\
\quad\quad \quad\;\; 0 \le {\rho _{i,j}} \le 1,\forall\; i,j,\\
\end{array}
\end{equation}
By creating the Lagrangian function and using the Karush–Kuhn–Tucker (KKT) conditions [\ref{Ref: Karush}] and [\ref{Ref: Kuhn-Tucker}], the optimal PA results of $\rho_{i,j}$ for a given case of $\chi_{i,j}$ can be calculated as:
\begin{equation}\label{Equation:PA results}
{{\rho_{i,j}} \simeq { 1 \over {\sum\limits_{j\in\mathbb{U}}{\chi_{i,j}}}}}
,\quad \forall \; i.
\end{equation}
where $\simeq$ denotes the optimal PA results can be approximately equal to the right term in (\ref{Equation:PA results}). The proof of (\ref{Equation:PA results}) is provided in Appendix A. In the decision making process, the approximate PA result shown in (\ref{Equation:PA results}) is adopted. The fuzzified boundaries are set as (10, 70) dB and (0, 10000) Mbps, also fuzzified breakpoints are (25, 40, 44) dB and (0, $\bar{R}$, $2\bar{R}$) Mbps for SNR and data rate requirements respectively. For the sake of simplicity, the detailed technical descriptions and other parameter values are referred to our previous work [\ref{Ref: Globecom Conference}].
In this paper, choosing the global optimization method as a benchmark is infeasible due to the excessive computations. For the direct decision-making method, its optimality is compromised. The complexity and runtime of mixed decision-making and optimization method is lower than the iteration based method, thus we adopt the mixed decision-making and optimization method to generate the dataset for training and testing the A-TCNN, which is analysed in the next section.
\section{Simulation Results}
\subsection{Simulation Setup}
In this section, the Monte Carlo simulations are conducted to evaluate the proposed A-TCNN structure versus its benchmarks: SSS, GT [\ref{Ref: Y. Wang LB Game TWC}], and FL [\ref{Ref: X. Wu AP Selection TCOM}] methods, in which the network throughput and users' Jain's fairness are chosen as wireless network metrics for comparison. Here we consider two general square room sizes with side length of 5 m and 9 m, where LiFi AP number is 4 and 9, respectively. For each LiFi, the transmitted optical power is 3 W, and breakpoint distance for WiFi is 3 m. Simulation setup is coded on MATLAB R2021a. Other parameters can be referred to [\ref{Ref: Globecom Conference}].
The architecture of the A-TCNN has been shown in Fig. \ref{Fig:Adaptive TCNN}, where the main features of network (e.g. layer number, neutrons number and activation function) has been presented. For simplicity, here we consider 4 LiFi AP case for training discussion and presentation in Fig. \ref{Fig:CDNN_versus_Global} and Fig. \ref{Fig:Adaptive_versus_CDNN}. Also, $M$ is fixed as 50 for simplicity in this paper. During the offline training process, we set batch number $N$ as 1000 and batch size $T$ as 256. Note we pick user number as 5, 10, 15, ..., 50 (10 training cases equally share 1000 batch) for training. Value of momentum in Adam optimizer is 0.95 and learning rate is chosen as 0.0001. During the offline testing, batch size for validation is set as 256 to evaluate the testing loss and performance metrics. The training, testing and evaluation are coded in Python3, which is running in a PC with Linux operation system, Intel Core i7-10875H processor and NVIDIA 2060 GPU. Our training code is open-sourced at [\ref{Ref: Github}].
\begin{figure}[t]
\centering
\includegraphics[width=9cm]{TCNN_versus_Global.eps}
\caption{Training loss and estimation accuracy for the proposed TCNN versus benchmark of DNN, where $N_{\rm{u}}$ = 50.}
\label{Fig:CDNN_versus_Global}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[height=7.5cm,width=9cm]{Adaptive_versus_TCNN.eps}
\caption{Comparisons of accuracy for the proposed A-TCNN versus TCNN under various user numbers, where Case I picks two chosen user numbers (20 and 30) which has been used in the training process, while Case II adopts two unchosen user numbers (8 and 44).}
\label{Fig:Adaptive_versus_CDNN}
\end{figure}
\subsection{Learning Network Metrics}
In this subsection, we first to evaluate the training process between TCNN and DNN. Note DNN is referred to the structure of the conventional DNN. As can be shown in Fig. \ref{Fig:CDNN_versus_Global}, the training loss of TCNN and DNN decrease with the increase of epoch, and finally converge. Here, each epoch takes about 8.3 s runtime for training TCNN without evaluating performance metrics. Note, training time of each epoch for A-TCNN is the same as TCNN as they have the same training structure in nature.
Totally, runtime of convergence is about 83 s for the proposed (A-)TCNN model, which is similar to DNN benchmark. Besides, it shows that both networks can constantly learn the relationship between input and output well because the same dataset can be seen with epoch index increases. For online testing, the testing loss approaches training loss curve well, which illustrates no over-fitting occurs during training. In other words, the trained networks both have a good generalization ability over unseen data. As for the training accuracy, the TCNN outperform DNN with about 12\% improvement when training is finished. The advantage of accuracy is explained that the increased dimension of input and output vector space in DNN challenges estimation mission.
Next, another experiment is conducted for the A-TCNN and pure TCNN with different user number structure in Fig. \ref{Fig:Adaptive_versus_CDNN}. Specifically, two cases are picked here, where Case I represents the chosen user number (20 and 30) belongs to the trained user number (e.g. 5, 10, ..., 50), and Case II represents the chosen user number (8 and 44) are not trained. In this figure, three significant clues can be found:
\begin{figure}[t]
\centering
\includegraphics[height=7.5cm, width=9cm]{Adaptive_Gap.eps}
\caption{Achievable throughput for the proposed A-TCNN model compared with training dataset.}
\label{Fig:Adaptive_Gap}
\end{figure}
\begin{itemize}
\item Denote the different user number in TCNN as $M_{\rm{C}}$. We first conclude that with the increase of $M_{\rm{C}}$ from 8 to 44, the accuracy of TCNN rises from 50\% to 70\%. The higher accuracy is mainly because the condition part with the increased dimension in TCNN can be extracted more features for learning, which inspires us that the structure of TCNN (i.e. maximum user number of $M_{\rm{C}}$) plays a dominant role in accuracy wise. In addition, previous Fig. \ref{Fig:CDNN_versus_Global} shows that TCNN with 50 user number structure can reach about 72\% accuracy. Compared with 70\% when $M_{\rm{C}}$ = 44, it is also indicated that the accuracy improvement is not significant when user number is bigger than 50. This drive us to choose $M$ as 50 for the proposed A-TCNN.
\item Second, accuracy of A-TCNN remains stable at level of 70\% for different user numbers. This means that the different user numbers of A-TCNN can reach a near accuracy level as TCNN with 50 UE number since the maximum default UE number $M$ in A-TCNN is fixed as 50.
\item Last, another observation can be seen from comparison between Case I and Case II. For A-TCNN, both Case I (20 and 30 UE cases) and Case II (8 and 44 cases) have the same accuracy performance, which illustrates that the trained A-TCNN can not only estimate the seen user cases (e.g. 20 and 30), but also generates a good generalization ability over the unseen user cases. This indicates us no need to collect entire users' cases for training.
\end{itemize}
From the above observations, we show that the A-TCNN has a better accuracy and scalability. From communication network wise, we next evaluate its performance gap compared with the provided standard dataset. As presented in Fig. \ref{Fig:Adaptive_Gap}, we plot achievable throughput for the A-TCNN method using estimated AP assignments scheme versus dataset for different user number, where we evaluate performance gap in different HLWNets size. In general, the estimated curve of A-TCNN fits the standard points given in dataset well without big performance gap. It is worthy noting that smaller network size of 4 LiFi performs better than bigger size of 9 LiFi, which is because that input dimension of target and output dimension of condition in A-TCNN increase when more LiFi APs involve, and thus decrease our estimation accuracy slightly. Overall, less than 3\% performance loss is negligible for 9 LiFi case compared with dataset results. Based on this, we replace FL-based dataset method with A-TCNN and then compare the wireless network metrics versus various benchmarks in the next subsection.
\subsection{Wireless Network Metrics}
In this subsection, we evaluate two wireless network metrics: achievable throughput and Jain's fairness from the following three aspects: user number, user distribution and required data rate.
\begin{figure}[t]
\centering
\includegraphics[height=7.5cm,width=9cm]{Throughput_versus_UEnum.eps}
\caption{Achievable throughput versus user number for the proposed A-TCNN compared with benchmarks.}
\label{Fig:Throughput_versus_UEnum}
\end{figure}
\subsubsection{Effect of User Number} As shown in Fig. \ref{Fig:Throughput_versus_UEnum}, we consider two general networks having size of 4 LiFi and 9 LiFi with respect to adaptive user number, where $\bar{R}$ is fixed as 100 and 200 Mbps for 4 LiFi and 9 LiFi. Note for the A-TCNN method, the throughput is measured based on the estimated output in the trained network which reaches convergence at epoch times of 20. All results are averaged using Monte-Carlo simulation. When more users are associated with APs, they compete with each other and explore more capacity resources in HLWNets, bringing the increase of achievable throughput shown in Fig. \ref{Fig:Throughput_versus_UEnum}. However, competition between users stops when it approaches the network upper bound, which is validated by figure in which the achievable throughput curves become saturating when user number increases to 50 for 4 and 9 LiFi AP cases. Specifically, the A-TCNN performs best compared with benchmarks when user number is bigger than 25 with throughput improvement about 100 Mbps, and it also remains better than FL method when user number is smaller than 25. For GT-based benchmark, it performs best when user number is less than 25, while SSS benchmark is worst as it selects AP with consideration of separate SNR and without consideration of competition and data rate requirements among users. In total, network with 9 LiFi size can provide bigger throughput (about 1000 Mbps improvement) than 4 LiFi since it involves more APs as data transmitters.
Fig. \ref{Fig:Fairness_UEnum} depicts the Jain's fairness index versus adaptive user number, where network size is chosen as 9 LiFi AP for the sake of simplicity and $\bar{R}$ is also 200 Mbps. In general, Jain's fairness drops steadily when user number becomes larger as the competition among users degrades the satisfaction values of users due to the upper bound of capacity is fixed in HLWNets. Given a small number of user, e.g. 5, the A-TCNN achieves the same fairness 0.95 as GT and FL benchmarks. For larger user number, A-TCNN performs better than benchmarks and shows 0.55 fairness compared with 0.53 fairness of GT and FL benchmarks. However, when user number is less than 25, GT method has a bigger fairness which is similar with the throughput conclusion observed in Fig. \ref{Fig:Throughput_versus_UEnum}.
\begin{figure}[t]
\centering
\includegraphics[height=7.5cm,width=9cm]{Fairness_versus_UE.eps}
\caption{Comparisons of Jain’s fairness index versus user number.}
\label{Fig:Fairness_UEnum}
\end{figure}
\subsubsection{Effect of User Distribution}
In practical, users are not usually distributed randomly and sometimes there are a number of hot spots which serve multiple users at the same time, which is named as cluster. In Fig. \ref{Fig:Throughput_versus_UEnum}, FL method can approach the performance of A-TCNN when user number is 50. This is mainly because under the random distribution, the network upper bound has been reached with the increase of user number. As long as user number is big enough, each LiFi or WiFi AP would be overloaded, which does not show that this method owns its algorithm-wise advantage in nature.
To verify this point and take users distribution into account, Fig. \ref{Fig:Cluster} presents the comparison of achievable throughput versus cluster number, where we assume that connected users in each cluster are subject to bivariate normal distribution, shown as $\mathbf{X} \sim \mathcal{N}_2(0, \mathbf{\Sigma})$. Here $\mathbf{\Sigma}$ denotes the covariance matrix given as [0.25 0; 0 0.25], and the number of connected users for each cluster is set as 10. As can be seen in Fig. \ref{Fig:Cluster}, the effect of clustering distribution significantly degrades the performance of SSS and FL benchmarks, while GT and proposed A-TCNN have a good tolerance ability in this scenario. When cluster number is 5, SSS and FL methods show about 38.7 and 36.5\% network throughput loss, while this percentages are only about 10.7 and 10.6 for GT and A-TCNN methods.
\begin{figure}[t]
\centering
\includegraphics[height=7.5cm,width=9cm]{Cluster.eps}
\caption{Comparison of achievable throughput for A-TCNN versus benchmarks under effect of clustering distribution, where we consider 9 LiFi APs and 50 users for presentation.}
\label{Fig:Cluster}
\end{figure}
\subsubsection{Effect of Data Rate Requirement}
In Fig. \ref{Fig:Throughput_versus_Rb}, we fix user number as 20 and 40 for 4 and 9 LiFi APs respectively, and compare the achievable throughput for A-TCNN versus benchmarks with respect to the effect of data rate requirement from 10 to 400 Mbps. For 4 LiFi APs scenario, A-TCNN method improves throughput from about 200 Mbps given $\bar{R}$ of 10 Mbps to 1300 Mbps given $\bar{R}$ of 400 Mbps, where it meets saturation when $\bar{R}$ is set as about 150 Mbps. Before this threshold, A-TCNN performs slightly worse than GT based benchmark, which is negligible, while the achieved throughput gain is about 100 Mbps when $\bar{R}$ is equal to 400 Mbps. When network size is configured as 9 LiFi APs, the same tendency between benchmarks can be seen from Fig. \ref{Fig:Throughput_versus_Rb}, which illustrates that the proposed A-TCNN also has a good adaption ability to different hybrid network size. Among these benchmarks, SSS performs worst which is consistent with conclusions in Fig. \ref{Fig:Throughput_versus_UEnum} and Fig. \ref{Fig:Throughput_versus_UEnum}. For simplicity, we only present the achievable throughput metric here considering the effect of data rate requirement and no need for fairness metric presentation.
In summary, we prove that A-TCNN can perform well in terms of network throughput and fairness. It is worth noting that although GT based benchmark shows a similar performance to A-TCNN, it is a distributed algorithm that needs excessive iterations to reach the Nash equilibrium.
\begin{figure}[t]
\centering
\includegraphics[width=9cm]{Throughput_versus_Rb.eps}
\caption{Achievable throughput versus data rate requirement for the proposed A-TCNN load balancing method compared with benchmarks.}
\label{Fig:Throughput_versus_Rb}
\end{figure}
\begin{table}[t]
\renewcommand{\arraystretch}{1.1}
\setlength\tabcolsep{1.5 mm}
\centering
\caption{Analysis of Big-O Complexity.}
\begin{tabular}{|c|c|c|}
\hline
\textbf{Methods} & \multicolumn{2}{c|}{\textbf{Big-O Complexity}} \\
\hline
SSS & \multicolumn{2}{c|}{$\mathcal{O}({N_{\rm{a}}}{N_{\rm{u}}})$} \\
\hline
GT [\ref{Ref: Y. Wang LB Game TWC}] & \multicolumn{2}{c|}{$\mathcal{O}({N_{\rm{a}}}N_{\rm{u}}{I})$} \\
\hline
FL [\ref{Ref: X. Wu AP Selection TCOM}] & \multicolumn{2}{c|}{$\mathcal{O}(N_{\rm{u}}{K_1}{K_2})$} \\
\hline
FL-OPT [\ref{Ref: Globecom Conference}] & \multicolumn{2}{c|}{$\mathcal{O}({N_{\rm{u}}}{K_1}{K_2}) + \mathcal{O}({N_{\rm{a}}}{N_{\rm{u}}})$} \\
\hline
\multicolumn{1}{|c|}{\multirow{3}{*}{A-TCNN}} & Collection & $\mathcal{O}({N_{\rm{a}}}{N_{\rm{u}}}NT)$ \\
\cline{2-3}
\multicolumn{1}{|c|}{} & Training & $\mathcal{O}(N_{\rm{epoch}}K_3)$ \\
\cline{2-3}
\multicolumn{1}{|c|}{} & Implementation & \multicolumn{1}{c|}{$\mathcal{O}({N_{\rm{a}}}{N_{\rm{u}}}K_4 + K_5)$} \\
\hline
\end{tabular}\label{Table: Computational Complexity}
\end{table}
\subsection{Computational Complexity}
Table \ref{Table: Computational Complexity} summarizes the Big-O complexity of the proposed method and benchmarks. With the SSS method, each user selects the AP that provides the highest SNR, leading to a computational complexity of $\mathcal{O}({N_{\rm{a}}}{N_{\rm{u}}})$. For the GT method, the Big-O complexity is estimated to be $\mathcal{O}({N_{\rm{a}}}{{{N_{\rm{u}}}}}{I})$ [\ref{Ref: Y. Wang LB Game TWC}], where $I$ denotes the number of iterations required, which turns bigger with the increase of ${N_{\rm{u}}}$. The computational complexity of the FL method depends on the number of inputs $K_{1}$ and the number of rules $K_{2}$ in the fuzzy logic. The Big-O complexity of this method can be approximated to be $\mathcal{O}({N_{\rm{u}}}{K_1}{K_2})$ [\ref{Ref: X. Wu AP Selection TCOM}]. With respect to the FL-OPT method, the part of processing fuzzy rules demands a Big-O complexity similar to that of the FL method, and the part of decision making requires $\mathcal{O}({N_{\rm{a}}}{N_{\rm{u}}})$. The A-TCNN method consists of three components: dataset collection, training, and real-time implementation. Firstly, the complexity of dataset collection is linear with the collected batch number $N$ and batch size $T$, which can be expressed as $\mathcal{O}({N_{\rm{a}}}{N_{\rm{u}}}NT)$. Secondly, training process takes several epochs to reach convergence. In each epoch, the complexity of forward and backward propagation can be denoted as $K_3$ as it is difficult to formulate. The complexity of training process can be expressed as $\mathcal{O}(N_{\rm{epoch}}K_3)$. Thirdly, in the real-time implementation of A-TCNN, complexity can be roughly given as $\mathcal{O}({N_{\rm{a}}}{N_{\rm{u}}}K_4 + K_5)$, in which $\mathcal{O}(K_4)$ represents the complexity of the matrix addition and multiplication in an FC layer, $\mathcal{O}(K_5)$ means the complexity of other operations in the TCNN, including BN and activation functions. Note the BN operation plays a dominant role in the complexity of $\mathcal{O}(K_5)$. Therefore, the changes of $N_{\rm{a}}$ and $N_{\rm{u}}$ have a minor effect on the computational complexity of implementation.
Finally, we present a comparison of runtime for the proposed A-TCNN versus various benchmarks considering different network size and user number. The results are summarized in Table \ref{Table: runtime}.\footnote{Note Bog-O notation gives upper bound for the growth rate of complexity with respect to variables in the worst case. For example, for the runtime of SSS given in Table II, it is not linear with the increase of AP number. In a basic choosing maximum element in an array, the best complexity is $\mathcal{O}(1)$ while worst case is $\mathcal{O}(N)$ as it is related to whether the array is sorted. In SSS, WiFi normally has a bigger SNR than WiFi, thus the maximum value can always be chosen in the first access.} Note for a fair comparison, we convert our learning model from Python into MATLAB platform and then measure the averaged runtime for all methods on the same computer. As can be seen in Table \ref{Table: runtime}, SSS method with simplest decision rule regarding SNR shows the lowest runtime around 0.001 $\sim$ 0.005 ms among other benchmarks, while GT, FL and FL-OPT methods all cost a long runtime and all longer than 1 ms. Between them, the proposed A-TCNN only takes a tiny runtime at a level of 0.1 ms. Consider user number and LiFi AP number, advantages of A-TCNN are more outstanding. When user number turns bigger, the FL and FL-OPT methods show a linear increment of runtime and GT method costs a exponential increment of runtime, while the A-TCNN scheme presents a stable runtime around 0.1 ms as it is adaptive to different user number in principle. In a bigger number of LiFi AP, the proposed A-TCNN requires a slightly higher runtime because the input dimensions of target and condition become higher which costs an additional processing power. When more APs are involved, benchmarks consume more runtime except from SSS and FL. For example, runtime of A-TCNN is 0.0992 and 0.123 ms for 4 LiFi, 50 users and 9 LiFi, 50 users cases, which is 91, 74 times shorter than the FL benchmark, and 480, 1503 times shorter than the GT benchmark.
\begin{table}[t]
\centering
\renewcommand{\arraystretch}{1.2}
\setlength\tabcolsep{1.6 mm}
\scriptsize
\caption{Comparison of Runtime (in Milliseconds).}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\multicolumn{2}{|c|}{\diagbox[]{Methods}{User No.}} & 10 & 20 & 30 & 40 & 50 \\
\hline
\multirow{5}{*}{4 LiFi} & SSS & 0.00132 & 0.00222 & 0.00325 & 0.00426 & 0.00501 \\
\cline{2-7}
& GT [\ref{Ref: Y. Wang LB Game TWC}] & 1.74 & 6.29 & 15.5 & 29.7 & 47.7 \\
\cline{2-7}
& FL [\ref{Ref: X. Wu AP Selection TCOM}] & 2.10 & 4.00 & 5.76 & 7.40 & 9.16 \\
\cline{2-7}
& FL-OPT [\ref{Ref: Globecom Conference}] & 4.70 & 9.86 & 16.7 & 25.2 & 33.1 \\
\cline{2-7}
& \textbf{A-TCNN} & \textbf{0.0819} & \textbf{0.0872} & \textbf{0.0962} & \textbf{0.0989} & \textbf{0.0992} \\
\cline{2-7}
\hline
\multirow{5}{*}{9 LiFi} & SSS & 0.00152 & 0.00236 & 0.00347 & 0.00454 & 0.00501 \\
\cline{2-7}
& GT [\ref{Ref: Y. Wang LB Game TWC}] & 6.34 & 27.5 & 63.1 & 104 & 185 \\
\cline{2-7}
& FL [\ref{Ref: X. Wu AP Selection TCOM}] & 2.15 & 3.94 & 5.67 & 7.56 & 9.25 \\
\cline{2-7}
& FL-OPT [\ref{Ref: Globecom Conference}] & 5.62 & 11.3 & 19.0 & 28.6 & 39.9 \\
\cline{2-7}
& \textbf{A-TCNN} & \textbf{0.0867} & \textbf{0.101} & \textbf{0.121} & \textbf{0.122} & \textbf{0.123} \\
\cline{2-7}
\hline
\end{tabular}\label{Table: runtime}
\end{table}
\section{Conclusion and Future Works}
In this paper, we presented a novel adaptive target-condition neural network (A-TCNN) aided structure to tackle the LB in HLWNets. This new structure enables HLWNets being adaptive to different users and AP deployment by utilizing the generalization ability of DNN. We also introduced the dataset collection method by investigating and comparing a number of benchmarks: SSS, GT, FL, and the FL-OPT which mixes the decision-making and optimization to better balance the optimality and complexity. Training and testing results show that A-TCNN has a good convergence and scalability, together with a higher accuracy than DNN model. In addition, we verified that the trained A-TCNN can solve LB issue with near-zero performance gap and unique feature of adapting to different user number. Furthermore, simulation results illustrate that our learned A-TCNN structure can provide better network metrics in terms of achievable throughout and fairness among users. Last but not least, we analysed the computational complexity and compared the runtime for A-TCNN versus various benchmarks, in which shows that A-TCNN method can save runtime up to 458 and 1388 times compared with FL and GT benchmarks, respectively. To the best of the authors' knowledge, the proposed A-TCNN is the first work that tackles LB problem with near-optimal performance and sub-millisecond runtime in HLWNets, making the system promising for meeting the latency requirement in 6G.
Looking forward, we point out a number of interesting and promising directions for future research: 1) we consider quasi-static users in this paper for sake of simplicity while it is more realistic when consider mobility, blockage and handover in HLWNets; 2) although DNN owns generalization ability to adapt to different user number, it is unknown whether DNN-aided model can handle LB in HLWNets when consider mobile and time sequential scenario. Learning from recent progresses in ML, how to introduce a more advanced and robust model into HLWNets is an open issue, such as deep RL and generative adversarial network; 3) most works in HLWNets only focus on the physical layer while the research on network layer is also necessary and challenging. Future works will be carried out considering software-defined networking wise, which involves metrics of network packets, routing process, packet loss ratio and latency.
\appendices
\section{Proof of Optimal PA Results}
Consider AP assignments has been tackled, i.e. $\chi_{i,j}$ has been chosen as 1 in a fixed stand-alone network. Assuming the satisfaction index is relaxed as $S_j = {\sum_{i \in \mathbb{S}}{{\rho_{i,j}}{C_{i,j}}}}/{{R_j}}$, (\ref{eq:Problem2}) is equivalent to
\begin{equation}\label{eq:Problem3}
\begin{array}{l}
\max\limits_{\rho_{i,j}} \;\quad\sum\limits_{j \in \mathbb{U}} {\log S_j}\\
\;\;{\rm{s}}{\rm{.t}}{\rm{. }}\quad\;\sum\limits_{j\in\mathbb{U}} {{\rho_{i,j}} \le 1}, \forall i \in \mathbb{S},\\
\quad\quad \quad\;\; S_j \le 1, \forall\; j,\\
\quad\quad \quad\;\; 0 \le {\rho _{i,j}} \le 1, \forall\; i,j.\\
\end{array}
\end{equation}
Then, the Lagrangian function can be given as:
\begin{equation}
\mathcal{L}(\rho_{i,j}, \lambda_1, \lambda_2) = \sum\limits_{j \in \mathbb{U}} \log S_j + \lambda_1(1 - \sum\limits_{j \in \mathbb{U}}\rho_{i,j}) + \lambda_2(1 - S_j).
\end{equation}
The KKT conditions include stationarity condition, dual feasibility, complementary slackness, and primal feasibility. These four conditions are given as:
\begin{subequations}
\begin{align}
&\nabla_{\rho_{i,j}} \mathcal{L}(\rho_{i,j}, \lambda_1, \lambda_2) = 0 , \\
&\lambda_1 \geq 0,\; \lambda_2 \geq 0, \\
&\lambda_1(1 - \sum_{j \in \mathbb{U}}\rho_{i,j}) = 0, \; \lambda_2(1 - S_j) = 0, \\
&(1 - \sum_{j \in \mathbb{U}}\rho_{i,j}) \leq 0,\; 1 - S_j \leq 0,
\end{align}
\end{subequations}
\begin{itemize}
\item When $\lambda_1 = 0$ and $\lambda_2 > 0$, it can be derived that $S_j = 1$ for $\forall j$. Only number of users and $R_j$ are small enough, constraint of $\sum_{j\in \mathbb{U}} \rho_{i,j} \leq 1$ can be always satisfied. Thus, KKT conditions can be satisfied provided $\rho_{i,j} = R_j/C_{i,j}$ for $\forall j$.
\item When $\lambda_1 > 0$ and $\lambda_2 = 0$, similarly we have $\rho_{i,j} = 1/{\sum_{j\in \mathbb{U}} \chi_{i,j}}$ and $\rho_{i,j} \leq R_j/C_{i,j}$ to meet KKT conditions. This case is reasonable when number of users and $R_j$ are big enough.
\item When $\lambda_1 = 0$ and $\lambda_2 = 0$, the stationarity condition of (17a) can not be met, thus this case is meaningless.
\item When $\lambda_1 > 0$ and $\lambda_2 > 0$, we have $\rho_{i,j} = 1/{\sum_{j\in \mathbb{U}} \chi_{i,j}}$ and $\rho_{i,j} = R_j/C_{i,j}$ for $\forall j$, which is controversial for different users.
\end{itemize}
In total, (\ref{eq:Problem3}) has no closed-form formulation which can satisfy all KKT conditions for different network deployment. For case of user number and $R_j$ being small, $\rho_{i,j} = 1/{\sum_{j\in \mathbb{U}} \chi_{i,j}}$ means more resources are allocated to users, bringing maximum satisfaction index as $S_j = 1$. While the theoretical result of $\rho_{i,j} = R_j/C_{i,j}$ also brings satisfaction value being 1 as fewer users with small data rate requirements are easy to serve. For case of user number and $R_j$ being large, solution of $\rho_{i,j} = 1/{\sum_{j\in \mathbb{U}} \chi_{i,j}}$ can meet all KKT conditions. When user number and $R_j$ are in medium level, only part of users can be served with $S_j = 1$ so that no theoretical solution for (\ref{eq:Problem3}) can be derived. Finally, we can approximately choose $\rho_{i,j} = 1/{\sum_{j\in \mathbb{U}} \chi_{i,j}}$ as solution of PA problem.
|
\section{Introduction}
\label{sec:section_introduction}
With the rapid prevalence of smart devices \cite{rapid_prevalence_smart_devices_shafique2020internet} such as mobile phone and Internet-of-Things (IoT) devices \cite{8264678,iot_app,9698094}, a vast amount of data has been generated \cite{wu_theta_join} and the demand of computation resources has been boosted \cite{huge_data_generated_marjani2017big,li2021self_big_data}. Due to the limited computation, storage and energy capacity of these smart devices \cite{iot_limited_computation_shakarami2020survey}, the powerful \emph{cloud computing} has been leveraged to provide elastic on-demand services to cope with limitations of smart devices \cite{9691460,SHEN2022103140}. With the support of resource-rich cloud servers, processing and storage-intensive applications such as Augmented Reality (AR) \cite{ar_ren2019edge} and Virtual Reality (VR) \cite{vr_zhang2017towards} become feasible.
However, the fast growing of computation demands pose severe burden on cloud centres \cite{burden_on_cloud_centre_mao2017mobile,burden_on_cloud_centre_yu2017survey,burden_on_cloud_centre_shi2016edge}, and tremendous amount of data generated \cite{data_generated,wu_theta_join} congests the network with limited bandwidth \cite{burden_on_bandwidth_shi2016promise,burden_on_cloud_centre_yu2017survey,burden_on_cloud_centre_shi2016edge}, hence causing bottlenecks for the cloud-based computing paradigm. To relieve the pressure on cloud centres, the \emph{edge computing} concept has emerged \cite{edge_computing_survey_khan2019edge,burden_on_cloud_centre_yu2017survey}, which allows computation to be performed at the edge network. The \emph{Edge network} \cite{burden_on_cloud_centre_shi2016edge} refers to the computing and network resources sit along the path between data sources and cloud centres. The rationale of \emph{computation offloading} \cite{heto_du2020algorithmics,computation_offloading_wang2019edge} is to let the computation happen at closer proximity to the data sources, so that not only the load pressure of cloud centres can be lessened, but also the Quality of Service (QoS) can be improved as the edge computing can provide more efficient responses \cite{qos_mach2017mobile}.
\rw{To fully excavate the potential of edge-cloud computation offloading, several past research efforts investigated performance-influencing factors and proposed optimisation models to maximise the performance gain while not causing significant costs. }Wang et al., \cite{homogeneous_communication_limited_factor_wang2017computational} presented a Two-Phase Optimisation algorithm and an Iterative Improvement algorithm to jointly optimise the computation costs and latency under the mobile-edge setting. Li et al., \cite{limited_factor_li2001computation} constructed a cost graph to optimise the energy consumption of handheld computing devices during computation offloading to achieve considerable energy saving. Works in \cite{homogeneous_communication_limited_factor_wang2017computational,larac_juttner2001lagrange,homogeneous_communication_wu2016optimal,homogeneous_communication_dong2018computation} all considered optimising the communication cost under the edge-cloud task offloading setting. Specifically, they adopted a homogeneous communication model with two assumptions: cloud-cloud and edge-edge communication cost were ignored, and the edge-cloud and cloud-edge communications were assumed to have symmetric costs, irrespective of the communication direction and distance between nodes. To address the over-simplicity of homogeneous communication models in these methods, Du et al., \cite{heto_du2020algorithmics} attempted the heterogeneous communication model and proposed HETO algorithm that can jointly minimise the computation, communication and migration costs during computation offloading. Their work was the first to propose a heterogeneous communication model so that the deficiencies in existing researches can be overcame.
\rw{Despite various attempts of past researches to optimise the edge-cloud computation offloading problem}, these models still suffered from the following drawbacks:
\begin{itemize}
\item \rw{These methods were not fine-grained enough. Although some methods considered heterogeneous communication cost, these methods failed to leverage more fine-grained factors such as the distance between node pairs. These methods also only leveraged a homogeneous cost model for computation tasks, which did not reflect the task heterogeneity in real-world settings. }
\item \rw{During computation offloading, some methods did not pay attention to the load balancing, which can cause overloading on certain nodes. }
\item These methods were cost-oriented, which failed to jointly optimise the profit and cost during computation offloading.
\end{itemize}
To address these issues, we propose a novel edge-cloud computation offloading model which not only utilises the more realistic heterogeneous communication and computation cost model, but also considers the cost and profit heterogeneities of tasks. Hence, the proposed method jointly optimises the profit and cost yielded during computation offloading. The model is named as \emph{PECCO}, which stands for ``\ul{P}rofit and Cost-oriented \ul{E}dge-\ul{C}loud \ul{C}omputation \ul{O}ffloading''.
Considering this optimisation problem is hard in nature and the objective of the \emph{PECCO} model is not differentiable, we consider using the Moth-flame Optimisation (MFO) algorithm \cite{moth_flame_optimisation_mirjalili2015moth} to tackle the computation offloading problem. \rw{As a swarm-based algorithm, it is gradient-free and it balances exploration and exploitation. Besides, empirically it outperforms other swarm-based counterparts in terms of convergence speed \cite{moth_flame_optimisation_mirjalili2015moth}, etc., which make it suitable to be leveraged in this case. }We therefore propose an improved Moth-flame optimiser (\emph{PECCO-MFI}) that addresses several drawbacks of the original MFO and significantly boosts its optimisation effectiveness. Specifically, a density-aware Moth-flame initialiser is designed to fit under the edge-cloud computation offloading setting. A dynamic hierarchical flaming mechanism is applied to avoid the single flame matching which is more likely to cause local optima stagnation. Moreover, the lifetime of moths is introduced to promote exploration when the corresponding paired flame is eliminated.
In summary, this paper makes the following contributions:
\begin{itemize}
\item We construct a profit and cost-oriented edge-cloud computation offloading optimisation model \emph{PECCO} that jointly optimises both the heterogeneous profit of computation tasks and the heterogeneous cost produced during computation offloading.
\item \rw{We not only utilise the heterogeneous communication cost, but also consider the load balancing among nodes during optimisation. }
\item We realise the suitability to leverage the Moth-flame Optimiser, and propose an improved algorithm which tackles several deficiencies of the original MFO and hence boosts the effectiveness when solving the proposed computation offloading model.
\end{itemize}
The rest of the paper is organised as follows, Section \ref{sec:section_related_work} introduces some related works on both edge-cloud computation offloading models and optimisation algorithms. The research opportunities are then discussed. \rw{The background, suitability and room for improvements to the Moth-flame Optimisation algorithm are given in Section \ref{sec:section_background}. }Section \ref{sec:section_model_and_method} presents the details of the proposed model, as well as how the Moth-flame Optimisation algorithm is improved and integrated. Section \ref{sec:section_experiment} presents the experimental settings and results to verify the effectiveness of the proposed algorithm when tackling the proposed model. Section \ref{sec:section_conclusion} concludes the paper.
\section{Related Work}
\label{sec:section_related_work}
\rw{In this section, past research works on edge-cloud computation offloading models will be presented. Then, some well-known optimisers will be presented and compared. Finally, the research opportunities of our work are discussed. }
\subsection{Offloading Model}
\label{sec:section_edge_cloud_computation_offloading_model_related_works}
As a promising technique that can relieve the burden posed on cloud centres, edge-cloud computation offloading has drawn huge attention from both industry and academic community \cite{iot_limited_computation_shakarami2020survey,9261459}. Wu et al., \cite{homogeneous_communication_wu2016optimal} formulated the edge-cloud computation offloading problem into a graph min-cost partitioning problem, in which computation tasks will be partitioned to be run on either the cloud side or the edge side. The proposed Min-Cost Offloading Partitioning (MCOP) algorithm took both the execution time and energy consumption into account when deciding an optimal task partitioning strategy. Li et al., \cite{limited_factor_li2001computation} put forward a partition scheme to offload computation tasks on handheld devices. A cost graph was constructed and the partition scheme was applied to split computation programs into server tasks and client tasks with the aim to reduce the energy consumption. Juttner et al., \cite{larac_juttner2001lagrange} presented the Lagrange Relaxation based Aggregated Cost (LARAC) algorithm, which formulated a task graph and traversed the shortest path between nodes when considering the communication costs. The proposed algorithm was effective on delay-sensitive applications, justified by the simulation experiment they performed.
In works completed by Wang et al., \cite{homogeneous_communication_limited_factor_wang2017computational} and Dong et al., \cite{homogeneous_communication_dong2018computation}, they paid attention to the communication cost faced in the edge-cloud computation offloading problem. \rw{When modelling the communication cost, communications between nodes on the same side (cloud-cloud, or edge-edge) were assumed to be cost-free. Moreover, to simplify the model, communication costs were assumed to be symmetric, i.e., cloud-edge and edge-cloud communications have the same cost, irrespective of direction and the distance between nodes on different sides. The homogeneous communication model they leveraged is considered to be over-simplified and highly infeasible in real-world settings, as the cost can be asymmetric and distance-dependent. Therefore, Du et al., \cite{heto_du2020algorithmics} proposed a more fine-grained heterogeneous cost model, in which the symmetric assumption was relaxed, and the communication costs between nodes in a single side were no longer ignored. }They then formulated the problem as a graph partitioning problem and designed the HETO algorithm to find a sub-optimal offloading strategy. Experiments on PageRank datasets testified to the excellent performance of the HETO algorithm when minimising the communication, computation and migration costs.
Despite that various research efforts have been drawn to optimise the edge-cloud computation offloading, they still suffered from some drawbacks which need to be addressed:
\begin{itemize}
\item Although the heterogeneous communication cost has been considered in some works, they failed to leverage more fine-grained factors such as distance between node pairs.
\item When considering the cost during computation offloading, these methods utilised a homogeneous cost model for computation tasks, i.e., task heterogeneity was ignored.
\item During computation offloading, some methods did not take load balancing into consideration, i.e., some node may be overloaded.
\item These methods were cost-oriented, which failed to jointly optimise the profit and cost during computation offloading.
\end{itemize}
\subsection{Model Optimiser}
\label{sec:section_model_optimiser_related_work}
A suitable optimiser is indispensable to tackle the edge-cloud computation offloading problem and find out an excellent offloading strategy. Some well-known individual-based optimisation algorithms were proposed \cite{SHI2018316,lai2020analysis}. They only optimised a single search candidate, and hence enjoyed a lighter computation cost and required less function evaluations. For instance, \rw{Lawrence \cite{hill_climbing_davis1991bit} presented the Hill Climbing (HC) algorithm which iteratively improved a single search candidate by changing its variables. }The Iterated Local Search (ILS) algorithm proposed by Lourenco et al., \cite{ils_lourencco2003iterated} was an improvement towards the HC algorithm. The best solution obtained in each iteration was perturbed and utilised as the starting point of the next iteration. Despite the efficiency enjoyed by these algorithms, they suffered a lot from the local optima stagnation. These algorithms may encounter the premature convergence, which prevents them from converging towards the global optima. Some more advanced algorithms such as gradient descent \cite{gradient_descent_ruder2016overview} have also been widely applied, especially for the optimisation in the field of deep learning \cite{goodfellow2016deep,li2020simultaneous}. \rw{However, these methods required gradient information of the objective function, which made them not applicable when the objective function is not differentiable. }
In order to provide better local optima avoidance, some population-based optimisation algorithms have been proposed and became popular in the past few years. By utilising multiple search candidates and meanwhile balancing between exploration and exploitation, they provided higher possibilities to approach the global optima. \rw{As a sub-class of population-based algorithms, swarm-based algorithms \cite{swarm_keerthi2015survey} utilised multiple search candidates for the purpose of exploration. }These search candidates then iteratively evolve, and eventually the healthier individuals will survive, making the exploitation become possible. \rw{Kennedy et al., \cite{kennedy1995particle} presented the Particle Swarm Optimisation (PSO) algorithm that mimicked the behaviour of birds in a flock which keep track of their individual and global best positions. The PSO involved only primitive math operations and was computationally inexpensive. }Yang \cite{yang2013firefly} proposed the Firefly Optimisation algorithm (FFA), which was inspired by fireflies. During flying, fireflies are attracted by other fireflies with higher brightness. The effectiveness of the algorithm was verified on several test functions. \rw{A Whale Optimisation algorithm (WOA) was proposed by Mirjalili et al., \cite{mirjalili2016whale} which was inspired by the bubble-net hunting strategy of humpback whales. The WOA algorithm mathematically modelled this behaviour to guide optimisation. A Grey Wolf Optimiser (GWO) was also proposed by Mirjalili et al., \cite{2014Grey} which modelled the social hierarchy of grey wolves during hunting to guide the optimisation process. }Besides, Mirjalili \cite{moth_flame_optimisation_mirjalili2015moth} put forward the Moth-flame Optimisation algorithm (MFO), which was one of the most famous swarm-based optimisers. The MFO utilised a population of moths to act as search candidates so that the probability of approximating the global optima was increased. Inspired by the transverse orientation characteristic of moths, the location of moths will be updated based on their transverse oriented path, with extra parameters controlling the exploration and exploitation. Each search candidate was iteratively assessed by a fitness function and hence the MFO algorithm was gradient-free. After several generation of evolvements, the fittest moth will be regarded as the optimised result. \rw{Experiments on several benchmarks \cite{moth_flame_optimisation_mirjalili2015moth} demonstrated that compared with several counterparts, the MFO algorithm can achieve better optimisation results with statistical significance, while also converge in a fast manner. }
\subsection{Research Opportunity}
\label{sec:section_research_opportunity}
Considering that past edge-cloud computation offloading models suffered from these aforementioned drawbacks, we find it promising to propose an optimisation model that is both profit and cost-oriented. In terms of costs, the heterogeneous communication cost should be considered in a fine-grained manner. Moreover, a heterogeneous cost model should also be applied for tasks to make the model more practical. Besides, during computation offloading, load balancing should be taken care of to avoid computation node overloading.
A comprehensive optimisation model and an excellent optimiser are both indispensable to produce a better task offloading strategy. Given the suitability of the MFO algorithm such as a higher chance to converge towards the global optima and its gradient-free merit, \rw{we propose an improved Moth-flame optimiser that addresses some design flaws of the original MFO, which can boost the effectiveness when working on the proposed computation offloading model. }
\section{Background}
\label{sec:section_background}
In this section, we firstly introduce the background of the Moth-flame Optimisation (MFO) algorithm, including how it works, its advantages and its suitability to be utilised to solve the \emph{PECCO} model. Then, some deficiencies of the MFO algorithm are pointed out which \rw{provide room for improvements for its improved version. }
\subsection{The Moth-flame Optimisation Algorithm}
\label{sec:section_the_moth_flame_optimisation_algorithm}
\textbf{Motivation and Rationale} As a nature-inspired optimiser, the Moth-flame Optimisation algorithm is population-based as it involves a population of moths. The moth has a special navigation mechanism called \textit{transverse orientation}, which they use as a flight path maintaining method. As shown in upper portion of the conceptual Figure \ref{fig:figure_transverse_orientation}, the moth attempts to maintain a fixed angle (marked in pink) between its flying direction and the moon, so that they can fly in a relatively straight path since the moon is far away from the moth. However, as illustrated in the lower part of Figure \ref{fig:figure_transverse_orientation}, the moth can sometimes confuse the artificial light with the moon. Then, it will try to maintain the transverse orientation mechanism with the light which is much closer than the moon, leading to the entrapment towards the light and eventually hit it.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.26\textwidth]{figure/transverse_orientation.png}\\
\caption{Flight mechanism of the moth. The upper portion illustrates the \textit{transverse orientation} mechanism. The lower portion illustrates the artificial light entrapment. The moth is represented using the blue arrow. }
\label{fig:figure_transverse_orientation}
\end{center}
\end{figure}
\vspace{0cm}
Inspired by this phenomenon, the Moth-flame Optimisation regards moths as search candidates, and treats the lights (flames) as potential optimal solutions. The Moth-flame Optimisation algorithm mimics the transverse orientation mechanism and hopes that the moth can reach the most optimal flame, which is regarded as the approximation to the global optima. By utilising a population of moths instead of a single one, the Moth-flame Optimiser possesses higher chance to avoid local optima entrapment and hence better approximates the global optima.
\textbf{General Framework} \rw{The Moth-flame Optimisation algorithm works under the general framework of swarm-based algorithm \cite{genetic_algorithms_framework_mirjalili2019genetic}. The species population will firstly be initialised, then they will keep evolving, eliminating individuals with bad fitness and updating until the termination criteria are reached. Eventually, the fittest individual will survive and will be treated as the optimal solution. }
\textbf{Formulation} The Moth-flame Optimisation algorithm involves $n$ moths, each is a search candidate wandering in the search space. Each moth $M_n \in \mathbb{R}^d$ is a $d$ dimensional vector, where $d$ is the number of features to be optimised. Hence, it leads to the moth matrix $M$ with dimension $n \times d$, represented as follows:
\begin{equation}\label{equ:moth_matrix}
M = \begin{bmatrix}
M_1\\
M_2\\
\vdots\\
M_n
\end{bmatrix}
= \begin{bmatrix}
m_{1, 1} & m_{1, 2} & \ldots & m_{1, d}\\
m_{2, 1} & m_{2, 2} & \ldots & m_{2, d}\\
\vdots & \vdots & \ddots & \vdots\\
m_{n, 1} & m_{n, 2} & \ldots & m_{n, d}\\
\end{bmatrix}
\end{equation}
\rw{A fitness function $f$ is required to evaluate the fitness of each moth $M_n$ by taking it as input, and returns its fitness, i.e., the objective value. }The objective function has the following formulation:
\begin{equation}\label{equ:objective_function_formulation}
f: \mathbb{R}^d \rightarrow \mathbb{R}, f(M_n) = OM_n
\end{equation}
and hence, the corresponding fitness vector $OM$ is defined as follows:
\begin{equation}\label{equ:moth_fitness_vector}
OM = \begin{bmatrix}
f(M_1)\\
f(M_2)\\
\vdots\\
f(M_n)
\end{bmatrix}
= \begin{bmatrix}
OM_1\\
OM_2\\
\vdots\\
OM_n
\end{bmatrix}
\end{equation}
\rw{In the Moth-flame Optimisation algorithm, the flames are not the real flames in the real world. Instead, they are set to be moths with top $k$ highest fitness values that have the right to survive (as in line $9$ in Algorithm \ref{algo:the_prodra_mf_optimisation_algorithm}), hence the flame matrix $F$ has dimension $k \times d$. }Without prior knowledge about which moth location is better, initially, the Moth-flame Optimisation algorithm randomly initialises the flame matrix $F$ with $k = n$. During iterations, the $k$ will be gradually decreased as the population evolves. The flame matrix $F$ is represented as follows:
\begin{equation}\label{equ:the_flame_matrix}
F = \begin{bmatrix}
F_1\\
F_2\\
\vdots\\
F_k
\end{bmatrix}
= \begin{bmatrix}
f_{1, 1} & f_{1, 2} & \ldots & f_{1, d}\\
f_{2, 1} & f_{2, 2} & \ldots & f_{2, d}\\
\vdots & \vdots & \ddots & \vdots\\
f_{k, 1} & f_{k, 2} & \ldots & f_{k, d}\\
\end{bmatrix}
\end{equation}
and its corresponding fitness vector $OF$ is defined as follows:
\begin{equation}\label{equ:flame_fitness_vector}
OF = \begin{bmatrix}
f(F_1)\\
f(F_2)\\
\vdots\\
f(F_k)
\end{bmatrix}
= \begin{bmatrix}
OF_1\\
OF_2\\
\vdots\\
OF_k
\end{bmatrix}
\end{equation}
The details on how the Moth-flame Optimisation algorithm is integrated in the \emph{PECCO} model, i.e., what moth matrix $M$ stands for, etc., will be explained in Section \ref{sec:section_integration_of_the_moth_flame_optimiser}.
\rw{As is previously mentioned, if there is no knowledge about which initial position is better, then a random initialisation will be applied} to generate both the moth matrix and the flame matrix using the following random generator:
\begin{equation}\label{equ:random_initialiser}
m_{i,j} = (ub(i) - lb(i)) * random() + lb(i)
\end{equation}
where $ub$ and $lb$ are the upper and lower bound of the range constraint, respectively. The $random()$ function is a random number generator with range in $[0, 1]$.
\ul{Deficiency 1}: The Moth-flame Optimisation algorithm applies a random initialisation as it assumes there is no prior knowledge about which initial location is better. However, if prior knowledge presents, the random initialisation will degrade the performance. Besides, the random initialisation is not density-aware, i.e., the random initialisation may produce random vectors that are highly similar and hinder the diversity of the random population. An initial population with poor diversity will impair the benefit of population-based optimisers.
\textbf{Balancing Exploration and Exploitation} The Moth-flame Optimisation algorithm puts effort to balance between exploration and exploitation. Initially, there are $n$ moths and $n$ flames, each moth pursues its corresponding flame as illustrated by the solid arrows in Figure \ref{fig:figure_moth_flame_pairing}, which encourages exploration to avoid local optima stagnation as much as possible. During iterations, the moths will be sorted based on their fitness value in descending order, and the moths with top $k$ highest fitness value will survive while other moths will be eliminated as shown in Algorithm \ref{algo:the_prodra_mf_optimisation_algorithm}. The value of $k$ keeps decreasing based on the following formula during iterations so that the exploitation will be gradually emphasised:
\begin{equation}\label{equ:equation_step_round}
k = round(n - CI * \frac{n - 1}{MI})
\end{equation}
where $n$ is the initial number of moth/flame, $MI$ denotes the total number of iterations (Max iteration), $CI$ stands for the current iteration. Eventually, $k$ will decrease to $1$, the last survived flame is regarded as the optimal solution produced by the Moth-flame Optimisation algorithm. The decreasing trend of $k$ has been illustrated in Figure \ref{fig:figure_decreasing_k}.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.3\textwidth]{figure/figure_auxiliary_step_20220218140240.pdf}\\
\caption{Illustration of the decreasing trend of the number of flames $k$, i.e., Equation \ref{equ:equation_step_round}. }
\label{fig:figure_decreasing_k}
\end{center}
\end{figure}
\vspace{0cm}
The number of flames $k$ keeps decreasing while the number of moths $n$ remains unchanged, the Moth-flame Optimisation algorithm therefore designs a moth-flame pairing mechanism as presented in Figure \ref{fig:figure_moth_flame_pairing} so that the moths can decide which target flame is designated for them to pursue. At the beginning, the number of moths and flames are equal, i.e., $n$, hence each moth will pursue its corresponding flame, i.e., $M_i \rightarrow F_i$, as represented by the solid arrows in Figure \ref{fig:figure_moth_flame_pairing}. During iterations, the value of $k$ will keep decreasing, hence the number of flames will be less than the number of moths. Under the MFO moth-pairing mechanism, the moth $M_i$ will still pursue its corresponding flame $F_i$ if flame $F_i$ still survives, or otherwise $M_i$ will chase the last flame $F_k$ as represented by the dashed arrows in Figure \ref{fig:figure_moth_flame_pairing}.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.26\textwidth]{figure/moth_flame_pairing.png}\\
\caption{The original moth-flame pairing mechanism. The moths are represented using blue boxes while the flames are represented using orange boxes. }
\label{fig:figure_moth_flame_pairing}
\end{center}
\end{figure}
\vspace{0cm}
\ul{Deficiency 2}: \rw{During moth evolvement, at any given time, moth $M_i$ will always pursue a single designated flame, which increases the chance of being trapped in the local optima. }
\ul{Deficiency 3}: \rw{During moth evolvement, those moths that have their corresponding flame being eliminated will always pursue the last surviving flame, which is the flame with the worst fitness. Neither pursuing the worst flame nor letting lots of moths pursuing a single flame is a reasonable design. }
\rw{In terms of the moth updating mechanism, the Moth-flame Optimisation algorithm mimics the transverse orientation based on the following equation: }
\begin{equation}\label{equ:equation_spiral_update_equation}
U(M_i, F_j) = D_{i,j} \times e^{bt} \times cos(2 \pi t) + F_j
\end{equation}
where $F_j$ is the paired flame designated for $M_i$ to pursue, $t$ is a random number in range $[r, 1]$, $r$ is a random number that will linearly decrease from $-1$ to $-2$ during iterations, $b$ is the shape parameter, and $D_{i, j}$ denotes the $L1$-distance between $M_i$ and $F_j$ which is defined as follows:
\begin{equation}\label{equ:equation_distance_calculation}
D_{i,j} = |F_j - M_i|
\end{equation}
The shape of an example updating path, i.e., the spiral shape, has been illustrated in Figure \ref{fig:figure_transverse_orientation} and \ref{fig:figure_exploration_exploitation}.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.3\textwidth]{figure/exploration_exploitation.png}\\
\caption{Illustration of the exploration vs exploitation of the Moth-flame Optimisation algorithm. }
\label{fig:figure_exploration_exploitation}
\end{center}
\end{figure}
\vspace{0cm}
Specifically, the $t$ parameter decides how close to the flame will the moth's terminal position be. As illustrated in Figure \ref{fig:figure_exploration_exploitation}, a $t$ value that is closer to $1$ will let the moth ends up with a position that is farther from the flame (the blue shaded area), which emphasises exploration. On the other hand, a negative $t$ value will draw the moth closer towards its target flame as indicated by the green shaded area in Figure \ref{fig:figure_exploration_exploitation}, which encourages exploitation. Since t is within the range of $[r, 1]$, initially, $r$ has value $-1$ which promotes exploration by avoiding the moth to be too close to the flame. As the process evolves, $r$ linearly decreases from $-1$ to $-2$, which gradually encourages exploitation over exploration.
\textbf{Termination} \rw{The termination criterion is when there is only one flame remaining. It will be treated as the optimal solution. }
\textbf{Advantage and Applicability} In summary, the Moth-flame Optimisation algorithm has the following advantages which make it applicable in our case:
\begin{itemize}
\item Since the \emph{PECCO} model is hard in nature, applying this population-based algorithm with multiple search candidates while enabling the balance between exploration and exploitation will possess higher chance to approximate the global optima.
\item Since the objective function in the \emph{PECCO} model is not differentiable, the Moth-flame Optimiser becomes applicable as it evaluates each search candidate using the fitness function and therefore is gradient-free.
\item \rw{Compared with its counterparts, the Moth-flame Optimiser achieves superior optimisation results and converges in an efficient manner. }
\end{itemize}
\textbf{Room for Improvements} As is aforementioned, the Moth-flame Optimisation algorithm suffers from three deficiencies, which leave us with room for improvements. We propose three new mechanisms to tackle these deficiencies as follows:
\begin{itemize}
\item The profit, cost and density-aware initialiser $\rightarrow$ Deficiency 1
\item The dynamic hierarchical flaming mechanism $\rightarrow$ Deficiency 2
\item The lifetime-enabled moth-flame pairing strategy $\rightarrow$ Deficiency 3
\end{itemize}
Together, these mechanisms form the improved \emph{PECCO-MFI} algorithm. The details will be presented in Section \ref{sec:section_the_improvements}.
\section{Model and Method}
\label{sec:section_model_and_method}
\rw{In this section, the problem formulation will be provided, followed by the presentation of the proposed \emph{PECCO} optimisation model, in which the profit and cost component of the \emph{PECCO} model will be explained. }Then, we will explain how the Moth-flame Optimisation algorithm is improved and integrated to form our \emph{PECCO-MFI} algorithm.
\subsection{Problem Formulation}
\label{sec:section_problem_formulation}
\rw{In the edge-cloud environment, there are cloud nodes and edge devices (nodes), }with a connection topology to form a connected graph. Hence, we formulate the problem as a graph $G = (V, E)$ where $V$ stands for a set of cloud/edge nodes and $E$ represents a set of communication links. There are in total $N$ computing nodes, in which it contains $I$ cloud nodes and $J$ edge nodes, hence we have
\begin{equation}\label{equ:equation_v_definition}
\begin{split}
& V^C = \{V_1^C, V_2^C, \cdots, V_I^C\}, V^E = \{V_{I+1}^E, V_{I+2}^E, \cdots, V_{I+J}^E\}\\
& V = V^C \cup V^E, V^C \cap V^E = \emptyset, N = I + J
\end{split}
\end{equation}
\rw{Note that we can simplify the notation $V^X_n$, $X \in \{C, E\}$ to be $V_n$ as the range of subscript $n$ can tell whether the node belongs to the cloud or the edge. }
For each computing node $V_n$, it has the following properties. Firstly, each computing node is capable of handling certain capacity of computation tasks. Hence $Cap_{V_n}\_max$ denotes the maximum number of units of computation workload that node $V_n$ is capable of handling, while $Cap_{V_n}\_min$ stands for the minimum workload of node $V_n$ when it is idle. We assume that no node can be overloaded by computation tasks. By considering the capacity of each node, it can also indirectly model other performance factors such as power consumptions.
As for edges $E$ in the graph $G$, there are in total $Q$ edges, denoted as $E_q$, or interchangeably $E_{<V_s,V_t>}$, which stands for edge $E_q$ is an edge that starts from node $V_s$ and points to node $V_t$. To make the heterogeneous model more generalisable, the length of each edge is also considered instead of being ignored as in \cite{heto_du2020algorithmics,larac_juttner2001lagrange}, and is denotes as $L_{E_q}$ (or $L_{E_{<V_s,V_t>}}$ using the interchangeable notation).
In terms of tasks to be executed, there are in total $K$ of them, each task $T_k$ has a property $WL_{T_k}$ that represents how many units of computation workload does task $T_k$ have. Each task can only be allocated on either a cloud node or an edge node, and a task is allowed to stay if its initial allocation is good enough. Hence, we define $\mathcal{A}_{T_k}^I$ and $\mathcal{A}_{T_k}^O$ to be the initial and offloaded allocation of task $T_k$, which satisfies
\begin{equation}\label{equ:equation_ai_ao}
\mathcal{A}_{T_k}^I, \mathcal{A}_{T_k}^O \in V^X, X \in \{C, E\}
\end{equation}
Moreover, vector $\mathcal{A}^O$ is defined to represent the offloaded allocations of all $K$ tasks as follows:
\begin{equation}\label{equ:equation_A_O}
\mathcal{A}^O = \begin{bmatrix}
\mathcal{A}^O_{T_1}\\
\mathcal{A}^O_{T_2}\\
\vdots\\
\mathcal{A}^O_{T_K}
\end{bmatrix}
\end{equation}
Besides, the \emph{PECCO} model applies a heterogeneous cost model for tasks. Instead of applying a homogeneous task cost as in \cite{heto_du2020algorithmics}, for each task, it has different costs $C_{T_k}^C$ if it is executed on the cloud, or $C_{T_k}^E$ if being allocated to the edge. \rw{By utilising a heterogeneous cost model for each task, it can reflect that different tasks can have different costs when being allocated to different sides, which makes the model more realistic. }To make the model profit-oriented, each task is also associated with two profits, i.e., $P_{T_k}^C$ and $P_{T_k}^E$, \rw{which stand for the profit gained of completing task $T_k$ on the cloud and edge, respectively. }By jointly considering profit and cost, it makes the proposed \emph{PECCO} model become profit and cost-oriented.
\subsection{\emph{PECCO} Cost Model}
\label{sec:section_the_pecco_edge_cloud_computation_offloading_cost_model}
The \emph{PECCO} model is a multi-factored model that jointly considers both the generalised heterogeneous communication cost and the heterogeneous computation cost.
\subsubsection{Generalised Heterogeneous Communication Cost Model}
\label{sec:section_generalised_heterogeneous_communication_cost_model}
Considering that previously proposed communication cost models in past researches suffered from several drawbacks (e.g., applied the unrealistic symmetric and cost-free assumption, failed to consider communication distance, \rw{applied homogeneous communication costs between node pairs, }etc. ) which made them become hardly generalisable in practice, it naturally leads to the rationale of our generalised heterogeneous communication cost model.
There are in general four types of communication costs, i.e., $w^{CC}$, $w^{EE}$, $w^{CE}$ and $w^{EC}$. The $CE$ here for instance represents the communication from a cloud node to an edge node. Inside each type of communication cost, it can also have different costs between different nodes, which is denotes as $w^{XX}_{<V_s,V_t>}$. For example, $w^{EC}_{<V^E_j,V^C_i>}$ denotes the communication cost from edge node $V^E_j$ to the cloud node $V^C_i$. \rw{As different nodes may work under different conditions like being operated by different service providers, }the communication cost between node pairs can be different even if they are situated on the same side. \rw{Therefore, the model is more realistic in practice. }This kind of generalisation also offers convenience to represent communication failures for instance, by setting the edge-wise communication cost to be a large value. \rw{The communication cost function $C_{E_{<V_s,V_t>}}$ is formulated as follows: }
\begin{equation}\label{equ:equation_communication_indicator_function}
\begin{split}
& C_{E_{<V_s,V_t>}} = sum(L_{E_{<V_s,V_t>}}
\times [1_{V_s, V_t \in V^C}, 1_{V_s, V_t \in V^E}, 1_{V_s \in V^C, V_t \in V^E}, 1_{V_s \in V^E, V_t \in V^C}]
\odot [w^{CC}_{<V_s,V_t>}, w^{EE}_{<V_s,V_t>}, w^{CE}_{<V_s,V_t>}, w^{EC}_{<V_s,V_t>}])
\end{split}
\end{equation}
inside it, for instance, $1_{V_s \in V^C, V_t \in V^E}$ will return 1 if $V_s$ is a cloud node and $V_t$ is an edge node, and will return 0 otherwise, other indicator functions carry the similar meaning. The $\odot$ represents the element-wise multiplication operator and $\times$ is the scalar multiplication operator. Hence, the communication cost function $C_{E_{<V_s,V_t>}}$ will return the length of the inputted edge times the corresponding cost of that type of communication so that the heterogeneous communication costs between node pairs can be considered.
After defining the generalised heterogeneous communication cost model, the optimal cost path between any pairs of computing nodes can be pre-computed using the shortest path algorithm. The optimal cost path between node $V_i$ and $V_j$ is denoted as $OCP_{<V_i,V_j>}$, and therefore we can define the optimal communication cost from node $V_i$ to node $V_j$ as follows:
\begin{equation}\label{equ:generalised_node_wise_communication_cost}
COMM_{<V_i,V_j>} = \sum_{E_{<V_s,V_t>} \in OCP_{<V_i,V_j>}} C_{E_{<V_s,V_t>}}
\end{equation}
and therefore, the total communication cost is defined as follows:
\rw{\begin{equation}\label{equ:generalised_total_communication_cost}
COMM(G, T, \mathcal{A}^I, \mathcal{A}^O) = \argmin_{\mathcal{A}^O}\{\sum_{k=1}^K COMM_{<\mathcal{A}^I_{T_k}, \mathcal{A}^O_{T_k}>}\}
\end{equation}}
The optimisation algorithm should find an optimal offloading strategy $\mathcal{A}^O$ to offload task $T_k$ so that it can achieve a communication cost $COMM(G, T, \mathcal{A}^I)$ as low as possible. \rw{In summary, the proposed generalised heterogeneous communication cost model overcomes the drawbacks of previously proposed communication models and has the following benefits: }
\begin{itemize}
\item It no longer ignores the communication cost between nodes on the same side (i.e., cloud-cloud, edge-edge).
\item The asymmetry between communication costs is considered, cost from cloud to edge and from edge to cloud communication can be heterogeneous.
\item It considers distances when modelling communication cost between two nodes.
\item It allows different node pairs to have heterogeneous communication costs.
\end{itemize}
\subsubsection{Heterogeneous Computation Cost Model}
\label{sec:section_generalised_computation_cost_model}
\rw{Next, the heterogeneous computation cost model is defined which also takes the heterogeneities between computing tasks into account. }Generally, for each task $T_k$, it possesses cost $C_{T_k}^C$ and $C_{T_k}^E$, which is the cost of executing task $T_k$ on the cloud and edge, respectively. Due to the diversity of tasks, $C_{T_k}^C$ is not necessarily lower than $C_{T_k}^E$. The previously proposed homogeneous computation cost model for tasks is infeasible, it is unreasonable to assume that all tasks share exactly the same computation cost when being executed on a single side. Hence, the computation cost model for tasks we considered in the \emph{PECCO} is more generalisable. The overall computation cost is formulated as follows:
\rw{\begin{equation}\label{equ:computation_cost}
\begin{split}
& COMP(G, T, \mathcal{A}^O) =
\argmin_{\mathcal{A}^O} \{\sum^K_{k=1} (1_{\mathcal{A}^O_{T_k} \in V^C} \times C^C_{T_k} + 1_{\mathcal{A}^O_{T_k} \in V^E} \times C^E_{T_k})\}
\end{split}
\end{equation}}
where the indicator function $1_{\mathcal{A}^O_{T_k} \in V^C}$ will return 1 if the allocation for task $T_k$ $\mathcal{A}^O_{T_k}$ is a cloud node, and will return 0 otherwise, similar for $1_{\mathcal{A}^O_{T_k} \in V^E}$.
\subsection{\emph{PECCO} Profit Model}
\label{sec:section_pecco_profit_model}
Different from previously proposed works, the proposed \emph{PECCO} model is not only cost-oriented, but also profit-oriented. \rw{For each task $T_k$, it has profit $P_{T_k}^C$ and $P_{T_k}^E$ when being executed on the cloud and edge, respectively. The overall profit is formulated as follows: }
\rw{\begin{equation}\label{equ:profit}
\begin{split}
& PROFIT(G, T, \mathcal{A}^O) =
\argmin_{\mathcal{A}^O} \{\sum^K_{k=1} (1_{\mathcal{A}^O_{T_k} \in V^C} \times P^C_{T_k} + 1_{\mathcal{A}^O_{T_k} \in V^E} \times P^E_{T_k})\}
\end{split}
\end{equation}}
\subsection{Overall Optimisation Objective}
\label{sec:section_overall_objective_function}
\rw{Finally, the \emph{PECCO} optimisation model will integrate the aforementioned cost and profit model to become profit and cost-oriented. The objective function is defined as follows: }
\rw{\begin{equation}\label{equ:overall_objective}
\begin{split}
& Obj(G, T, \mathcal{A}^I) =
\argmin_{\mathcal{A}^O} \{(COMM(G, T, \mathcal{A}^I) + COMP(G, T, \mathcal{A}^O))
+ \lambda \times PROFIT(G, T, \mathcal{A}^O)\}
\end{split}
\end{equation}}
\rw{where $\lambda$ is a ratio parameter being set to a negative value to integrate the profit into the objective to be minimised. }By having $\lambda$, the objective function can minimise the cost and simultaneously maximise the profit.
By jointly optimising the profit and cost-oriented optimisation model \emph{PECCO}, we can find a solution that can jointly optimise costs and the profit as much as possible.
\subsection{The \emph{PECCO-MFI} Optimiser}
\label{sec:section_the_moth_flame_optimiser}
In this section, we will introduce the proposed improved Moth-flame Optimiser with detailed explanations to the improvements we made. Then, how the improved Moth-flame Optimisation algorithm is integrated to optimise the \emph{PECCO} model is explained, i.e., what moths stand for in the \emph{PECCO-MFI} algorithm, how are tasks offloaded based on the optimised allocation strategy $\mathcal{A}^O$, etc.
\subsubsection{Algorithm Improvement}
\label{sec:section_the_improvements}
\rw{To tackle the deficiencies mentioned in Section \ref{sec:section_the_moth_flame_optimisation_algorithm} and therefore boost the performance, }we propose an improved Moth-flame Optimiser called \emph{PECCO-MFI} with three improvements to tackle three deficiencies, respectively.
\begin{algorithm}[!ht]
\begin{algorithmic}[1]
\Require
\StateX Number of search candidates (moths) $nsa$,
\StateX Edge-cloud graph $G$,
\StateX Tasks $T_k \in T$,
\StateX Allocation upper bound $ub$,
\StateX Objective function $Obj()$ as defined in Equation \ref{equ:overall_objective},
\StateX Initial allocation $\mathcal{A}^I$
\Ensure Profit, cost and density-aware moth initialisation with dimension $nsa \times N$
\For{$T_k$ \textup{in} $T$}
\State Calculate costs based on Equation \ref{equ:generalised_total_communication_cost} and \ref{equ:computation_cost}
\State Calculate the profit based on Equation \ref{equ:profit}
\State Calculate the profit and cost-oriented objective based on Equation \ref{equ:overall_objective}
\EndFor
\For{$i$ \textup{in} $range(nsa \times 1.5)$}
\For{$k$ \textup{in} $range(K)$}
\If{allocate task $T_k$ to the cloud side yield a lower objective value}
\State Store random number in range $[0, \frac{ub}{2})$ into $\mathcal{A}^I_i$
\Else
\State Store random number in range $[\frac{ub}{2}, ub]$ into $\mathcal{A}^I_i$
\EndIf
\EndFor
\EndFor
\While{$len(\mathcal{A}^I) \neq nsa$}
\State Find pair $(\mathcal{A}^I_i, \mathcal{A}^I_j)$ with minimum intra-pair L$2$ distance
\State Add $\frac{\mathcal{A}^I_i + \mathcal{A}^I_j}{2}$ into $\mathcal{A}^I$
\State Remove both $\mathcal{A}^I_i$ and $\mathcal{A}^I_j$ from $\mathcal{A}^I$
\EndWhile
\State \Return $\mathcal{A^I}$
\end{algorithmic}
\caption{The profit, cost and density-aware moth initialiser $initialiser(nsa, G, T, ub, obj, \mathcal{A}^I)$ of the \emph{PECCO-MFI} Algorithm}
\label{algo:the_profit_aware_initialisation}
\end{algorithm}
\textbf{Improvement 1 (Profit, Cost and Density-aware Moth Initialiser)}: To tackle the \ul{Deficiency 1: trivial random initialisation}, we design a new moth initialiser that is profit, cost and density-aware, as shown in Algorithm \ref{algo:the_profit_aware_initialisation}.
\ul{Profit and Cost-awareness}: The newly designed moth initialiser will allocate tasks (elements in each moth) to cloud or edge side based on their profit and cost. It is natural to allocate services to the side in which they possess a lower profit-cost objective than the other side. \rw{Conversely, if the allocation is done in the reversed way, then this task is likely to be migrated during the optimisation process, which therefore incurs unnecessary costs. }The profit and cost-aware moth initialisation mechanism has been shown in line $7$ - $13$ in Algorithm \ref{algo:the_profit_aware_initialisation}. By utilising this prior knowledge, the improved Moth-flame Optimisation algorithm is reasonable to outperform its knowledgeless random counterpart.
\ul{Density-awareness}: The rationale of applying the population-based paradigm is to maximise the chance of approximating the global optima as close as possible. However, if some moths are initialised to have a close proximity, the benefit of the population-based paradigm will be greatly hindered. \rw{To ensure the initialised moths have rich diverity, }the newly designed initialiser will generate more moth vectors than required, then it will iteratively remove the closest pair of moths and keep the average of them. The procedure has been given in line $15$ - $19$ in Algorithm \ref{algo:the_profit_aware_initialisation}. The removal will be continued until the number of moth vectors is satisfied as required. By leveraging this mechanism, moth vectors that are initialised to be too close will be merged into a new one to prevent the performance degradation from happening. Therefore, the proposed moth initialiser is density-aware.
\begin{algorithm}[!ht]
\begin{algorithmic}[1]
\Require
\StateX Current iteration $CI$,
\StateX Max iteration $MI$
\State Define $F_\mathbb{1}$, $F_\mathbb{2}$, $F_\mathbb{3}$ as the flames with top $3$ fitness values, respectively
\State $\omega \leftarrow \frac{CI}{MI}$
\If{moth $M_i$'s corresponding flame $F_i$ still survives}
\If{there are $\geq 3$ flames survive}
\State $M_i$ will pursue $\frac{F_i + \omega \times F_\mathbb{1} + \omega \times F_\mathbb{2} + \omega \times F_\mathbb{3}}{1 + 3 \times \omega}$
\ElsIf{there are $2$ flames survive}
\State $M_i$ will pursue $\frac{F_i + \omega \times F_\mathbb{1} + \omega \times F_\mathbb{2}}{1 + 2 \times \omega}$
\EndIf
\Else
\State $\tau \leftarrow random(0, 1)$, where $\tau$ is the lifetime parameter
\If{$\tau$ > $0.8$}
\State $F_{\mathcal{I}} \leftarrow$ a randomly initialised flame
\Else
\State $\mathcal{I} \leftarrow random(0, k)$
\EndIf
\If{there are $\geq 3$ flames survive}
\State $M_i$ will pursue $\frac{F_{\mathcal{I}} + \omega \times F_\mathbb{1} + \omega \times F_\mathbb{2} + \omega \times F_\mathbb{3}}{1 + 3 \times \omega}$
\ElsIf{there are $2$ flames survive}
\State $M_i$ will pursue $\frac{F_{\mathcal{I}} + \omega \times F_\mathbb{1} + \omega \times F_\mathbb{2}}{1 + 2 \times \omega}$
\EndIf
\EndIf
\end{algorithmic}
\caption{\rw{The dynamic hierarchical flaming mechanism and the lifetime-enabled moth-flame pairer $enhanced\_pairer(CI, MI)$ of the \emph{PECCO-MFI} Algorithm}}
\label{algo:the_enhanced_moth_flame_pairing_mechanism}
\end{algorithm}
\textbf{Improvement 2 (Dynamic Hierarchical Flaming Mechanism)}: To deal with the \ul{Deficiency 2: Single moth-flame pairing}, a dynamic hierarchical flaming mechanism is applied. \rw{As pursuing a single flame will lead to higher risk of being trapped in local optima, inspired by the social hierarchy possessed in the moth species, the moths with top $3$ highest fitness values will be regarded as leaders, }which will provide guiding reference for other moths to pursue. Hence, instead of pursuing a single flame, in the newly designed algorithm, each moth will chase the linear combination of its designated flame and the leader flames, as shown in line $3$ - $8$ in Algorithm \ref{algo:the_enhanced_moth_flame_pairing_mechanism}.
\ul{Exploration and Exploitation Balance}: At the beginning of the training process, putting too much dependency on the top $3$ flames will incur risk of local optima stagnation as the performance of flames at the beginning is not promising enough. Hence, an adjusting factor $\omega$ is introduced which linearly increases from $0$ to $1$ during the training process as indicated in line $2$ in Algorithm \ref{algo:the_enhanced_moth_flame_pairing_mechanism}. \rw{After applying the adjusting factor $\omega$ as in line $5$ and $7$ in Algorithm \ref{algo:the_enhanced_moth_flame_pairing_mechanism}, exploration will be encouraged at the beginning by putting less emphasise on the top $3$ flames since initially the value of $\omega$ is small. }As the training progresses, $\omega$ will gradually increase, which will emphasise more exploitation, since the guiding reference of top $3$ flames will be gradually reinforced as the $\omega$ keeps growing. As such, the utilisation of adjusting factor $\omega$ in the newly designed dynamic hierarchical flaming mechanism will balance between exploration and exploitation.
\textbf{Improvement 3 (Lifetime-enabled Moth-Flame Pairing Strategy)}: Finally, to solve the \ul{Deficiency 3: naive moth-flame pairing}, we design a fairer pairing strategy as indicated in line $9$ - $21$ in Algorithm \ref{algo:the_enhanced_moth_flame_pairing_mechanism}. Instead of letting all moths whose corresponding flames are eliminated to chase the last surviving flame, we introduce a lifetime parameter $\tau$ with lifetime threshold set as $0.8$. Due to the elimination of their unpromising flames, these moths are not promising themselves and hence the lifetime parameter $\tau$ is used to decide whether certain moth will be re-initialised, i.e., starting a new lifetime. Hence, as indicated in line $11$ - $12$ in algorithm \ref{algo:the_enhanced_moth_flame_pairing_mechanism}, if the randomly generated lifetime parameter $\tau$ is higher than the lifetime threshold, the moth will start a new lifetime by pairing with a newly initialised flame. Otherwise, the moth will continue its lifetime, and the algorithm will let it to pair with a randomly selected survived flame to promote a fairer exploration. By randomly pairing with a survived flame, these moths will fairly explore all possible survived flames instead of all exploiting the worst-fitted flame. During the re-pairing process, the aforementioned dynamic hierarchical flaming mechanism will be utilised again to provide better exploration and exploitation balancing while enabling the guiding reference of top $3$ flames as in line $16$ - $20$ in Algorithm \ref{algo:the_enhanced_moth_flame_pairing_mechanism}. By utilising this enhanced moth-flame pairing strategy, the exploration of the algorithm will be further encouraged and hence leading to a higher chance to approach the global optima.
\subsubsection{Integration of the Moth-flame Optimiser}
\label{sec:section_integration_of_the_moth_flame_optimiser}
\begin{algorithm}[!ht]
\begin{algorithmic}[1]
\Require
\StateX Shape parameter $b$,
\StateX Number of search candidates (moths) $nsa$,
\StateX Allocation upper bound $ub$,
\StateX Objective function $Obj()$ as defined in Equation \ref{equ:overall_objective}
\Ensure Resource allocation strategy $\mathcal{A}^O$ of computation tasks, which is the best flame $F$
\State Initialise moth matrix $M \leftarrow initialiser(nsa, G, T, ub, obj, \mathcal{A}^I)$ (in Algorithm \ref{algo:the_profit_aware_initialisation})
\State $OM \leftarrow Obj(M)$
\While{$len(F) \neq 1$}
\State Update $k$
\State $OM \leftarrow obj(M)$
\If{$CI$ = 1}
\State $F \leftarrow M.sortBy(OM)$
\Else
\State $F \leftarrow M.sortBy(OM)[0:k]$
\EndIf
\State Update moth-flame pairing using $enhanced\_pairer(CI, MI)$ (Algorithm \ref{algo:the_enhanced_moth_flame_pairing_mechanism})
\For{$i$ \textup{in} $range(nsa)$}
\For{$j$ \textup{in} $range(K)$}
\State Update $r$ and $t$
\State Calculate $D$ with respect to the paired moth and flame using Equation (\ref{equ:equation_distance_calculation})
\State Update moth position using Equation (\ref{equ:equation_spiral_update_equation})
\EndFor
\EndFor
\EndWhile
\State $OF \leftarrow obj(F)$
\State \Return $F$, $OF$
\end{algorithmic}
\caption{Workflow of the \emph{PECCO-MFI} Algorithm}
\label{algo:the_prodra_mf_optimisation_algorithm}
\end{algorithm}
\rw{We apply the improved Moth-flame Optimiser \emph{PECCO-MFI} to optimise the \emph{PECCO} model. The pseudocode of the algorithm has been presented in Algorithm \ref{algo:the_prodra_mf_optimisation_algorithm}. }In the \emph{PECCO-MFI} algorithm, each moth vector $M_i$ is a $1 \times K$ vector, where $K$ is the number of tasks waiting to be allocated. The values in the moth vector $M_i$ are within the range of $[0, ub]$, where $ub$ is a constant. If the value is in range $[0, \frac{ub}{2})$, it indicates that this task will be executed to the cloud side, otherwise, this task will be offloaded to the edge side. Then, the algorithm will find the computing node with the cheapest communication cost in the designated side and allocate the task to that computing node. If the computing node will be overloaded by taking this task, the algorithm will find the node at the designated side with the second cheapest communication cost and so on. \rw{Eventually, the task will either be allocated to a computing node without causing overloading, or it will not be satisfied due to workload unavailability. }
\section{Experiment}
\label{sec:section_experiment}
In this section, we will introduce our experimental setup and the dataset we utilised during experiments. Then, experimental results will be presented and explained to testify to the superiority of the proposed method. Specifically, objective values yielded by different methods are compared, followed by the comparison of profit, cost and profit-cost ratio to demonstrate the effectiveness of the \emph{PECCO-MFI} algorithm in terms of joint profit and cost optimisation. Finally, to demonstrate the \emph{PECCO-MFI} algorithm offloads computation tasks wisely, the task allocation and the resource utilisation are compared and analysed.
\subsection{Dataset, Parameter and Experimental Setup}
\label{sec:section_experimental_setup_and_dataset}
\textbf{Dataset}. The dataset we use to simulate the edge-cloud environment \cite{huang2019dp_greedy} is the Sydney train station parking dataset obtained from the Open Data Portal provided by the New South Wales Government Department of Transportation \cite{nsw_parking}. The dataset contains the parking lot availability information at each train station in Sydney, Australia. Train stations in City of Sydney are treated as cloud nodes, and suburban train stations act as edge nodes. Each station has a parking with limited available parking lots, which represents the capacity of the node. The length of connected edges are the length of the shortest road between train stations. The dataset contains communication heterogeneity as paths between different nodes can have different charges due to factors such as toll roads, etc. Parking requests are simulated as tasks which will be allocated to nodes. We pick $20$ stations from the City of Sydney to act as cloud nodes, and $30$ stations from Sydney suburban areas as edge nodes, numbered from $1$ to $20$, and from $21$ to $50$, respectively. We utilise $200$ tasks, each with certain amount of parking requests that will be treated as workloads. Parking in the city or in the suburban area can yield different parking fares, which serves as the cost of the task when being executed on the cloud side and edge side, respectively. Finally, successfully allocating each parking task will earn certain profit.
\textbf{Parameter Setting}. We now introduce parameter settings in two parts: \emph{PECCO} optimisation model parameter settings, as well as improved Moth-flame Optimiser parameter settings.
\ul{\textit{PECCO Optimisation Model Parameter Setting}}. To reflect the communication heterogeneity, we set the communication cost $w^{CC}$, $w^{CE}$, $w^{EC}$ and $w^{EE}$ to have an average of $1$, $2$, $4$ and $6$, respectively. Thanks to the powerful network infrastructure equipped in cloud centres, the intra-cloud communication should be the cheapest among communication directions. The download cost should be cheaper than the upload cost and hence the cost is set to reflect this pattern. Finally, due to the limited network capacity between edge devices, it is natural to possess the highest communication cost.
The ratio parameter $\lambda$ that balances between cost and profit is set to be $-8$ so that the algorithm can jointly minimise the cost and profit times this negative ratio.
\ul{\textit{PECCO-MFI Parameter Setting}}. We set the default allocation upper bound $ub$ to be $1$, and the Moth-flame shape parameter $b$ to be $1$ to comply with $ub$. If the shape parameter $b$ is set to be too small, then the shape of the spiral will be very tight and it will never allocate tasks to some nodes. On the other hand, if $b$ is too large, then the spiral will be too wide and it may generate offloading strategy which does not make sense. The default number of search candidates, i.e., moths, is set to be $30$, and the number of iterations is set to be $100$ to make the algorithm efficient. The default value of the threshold of the lifetime parameter $\tau$ is set to be $0.8$.
\textbf{Experimental Setup and Hardware Configuration}. We compare the \emph{PECCO-MFI} algorithm with two edge-cloud computation offloading algorithms, including the LARAC algorithm, which traverses the shortest path during communication and optimises the computation cost, as well as GREEDY, which allocates each task to the side that will yield lower objective value, then greedily select the node which has the shortest distance from the initial location of the task. \rw{The effectiveness of the \emph{PECCO-MFI} is also compared with $9$ swarm-based optimisers, }including Bat Algorithm (BAT) \cite{yang2012bat}, Sine Cosine Algorithm (SCA) \cite{mirjalili2016sca}, Whale Optimisation Algorithm (WOA) \cite{mirjalili2016whale}, Cuckoo Search Algorithm (CS) \cite{yang2010engineering}, Firefly Algorithm (FFA) \cite{yang2013firefly}, Particle Swarm Optimisation (PSO) \cite{kennedy1995particle}, Grey Wolf Optimisation (GWO) \cite{2014Grey}, Differential Evolution (DE) \cite{price2013differential} and the original Moth-flame Optimisation Algorithm (MFO) \cite{moth_flame_optimisation_mirjalili2015moth}. To make the experimental results concrete, all experiments are repeated $10$ times and the average results are reported.
We implement the method using python $3.8$, and conduct all experiments on a server equipped with Intel Core i9 9900K CPU and 32GB of memory.
\begin{table*}[!ht]
\centering
\begin{tabular}{c|cccccc}
\hline
\backslashbox{Value}{Method} & LARAC & Greedy & BAT & SCA & WOA & CS \\ \hline
Overall Objective & -19253.93 & -22615.6 & -41317.38 & -41496.97 & -42489.74 & -44088.8 \\
Profit & 2650.29 & 3054.72 & 5388.28 & 5407.45 & 5531.59 & 5737.87 \\
Cost & 1948.38 & 1822.16 & 1788.83 & \textbf{1762.64} & 1762.98 & 1814.15 \\
Profit/Cost Ratio & 1.36 & 1.67 & 3.02 & 3.07 & 3.14 & 3.17 \\
\hline
\backslashbox{Value}{Method} & FFA & PSO & GWO & DE & MFO & MFI \\ \hline
Overall Objective & -44741.63 & -45546.13 & -46494.95 & -46598.59 & -45953.91 & \textbf{-48069.48} \\
Profit & 5814.99 & 5915.65 & 6039.85 & 6050.26 & 5969.91 & \textbf{6229.31} \\
Cost & 1778.32 & 1779.07 & 1823.88 & 1803.52 & 1805.37 & 1765.0 \\
Profit/Cost Ratio & 3.28 & 3.34 & 3.32 & 3.36 & 3.32 & \textbf{3.53} \\
\hline
\end{tabular}
\caption{Objective value, profit, cost and the profit-cost ratio of different algorithms. Note the MFO stands for the original MFO algorithm, while the MFI stands for the improved \emph{PECCO-MFI} algorithm. }
\label{tab:objective_profit_cost_pcratio}
\end{table*}
\subsection{Comparison of Objective Optimisation between Algorithm}
\label{sec:comparison_of_objective_optimisation_between_algorithm}
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=0.26\textwidth]{figure/figure_objective_reduction20220413170341.pdf}\\
\caption{The percentage of objective value reduction achieved by the \emph{PECCO-MFI} algorithm compared with other methods. Since the objective value reduction yielded by the \emph{PECCO-MFI} is $149.7\%$ and $112.6\%$ compared with LARAC and GREEDY, respectively. Therefore, for better visualisation, these two methods are omitted from the plot. }
\label{fig:figure_objective_reduction}
\end{center}
\end{figure}
To verify the effectiveness of the \emph{PECCO-MFI} algorithm in terms of objective optimisation, the objective results of the \emph{PECCO-MFI} and $11$ compared methods are listed in Table \ref{tab:objective_profit_cost_pcratio}. As we can notice, the proposed \emph{PECCO-MFI} algorithm achieves the lowest objective value among all compared methods. As we can observe from Figure \ref{fig:figure_objective_reduction}, the \emph{PECCO-MFI} algorithm produces a $4.6\%$ and $3.16\%$ objective value reduction compared with the original Moth-flame Optimisation algorithm and the best-performed Differential Evolution algorithm when tackling this optimisation problem, demonstrating the effectiveness of the \emph{PECCO-MFI} algorithm. The significant performance improvement achieved by the \emph{PECCO-MFI} over the original Moth-flame Optimiser also verifies the effectiveness of improvements we made on the MFO.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=0.5\textwidth]{figure/figure_profit_cost_pcratio_reduction20220413170341.pdf}\\
\caption{The percentage of increase on profit and profit-cost ratio achieved by the \emph{PECCO-MFI} algorithm compared with other methods are presented in sub-figure (a) and (b), respectively. Since the profit achieved by the \emph{PECCO-MFI} is $135\%$ and $103.9\%$ higher than LARAC and GREEDY, respectively, and the profit-cost ratio is $159.6\%$ and $111.4\%$ higher than LARAC and GREEDY, respectively. Therefore, for better visualisation, these two methods are omitted from plots. }
\label{fig:figure_profit_pcratio_increase}
\end{center}
\end{figure}
\vspace{0cm}
\subsection{Comparison of Profit and Cost between Algorithms}
\label{sec:comparison_of_profit_and_cost_between_algorithms}
After investigating the total objective, we now look at the profit and cost component. As indicated in Table \ref{tab:objective_profit_cost_pcratio} and Figure \ref{fig:figure_profit_pcratio_increase}(a), the \emph{PECCO-MFI} algorithm achieves the highest profit among all comparing methods. Specifically, the profit achieved is $4.35\%$ and $2.96\%$ higher than the original MFO algorithm, as well as the best-performed Differential Evolution, respectively. Besides, the \emph{PECCO-MFI} achieves the second lowest cost during computation offloading, which is only $0.1\%$ higher than the SCA algorithm, who has the lowest cost. However, when it comes to the profit-cost ratio, the \emph{PECCO-MFI} yields the best performance. As indicated in Figure \ref{fig:figure_profit_pcratio_increase}(b), the \emph{PECCO-MFI} algorithm achieves significant profit-cost ratio boost compared with all other methods. The higher the profit-cost ratio is, the more profit will be yielded by spending one unit of cost, i.e., the computation offloading is wiser as it can achieve higher profit by spending unit amount of cost. According to Figure \ref{fig:figure_profit_pcratio_increase}(b), the \emph{PECCO-MFI} has a profit-cost ratio that is $6.33\%$ and $5.06\%$ superior than the original MFO and Differential Evolution counterparts, respectively, which demonstrates the effectiveness of the \emph{PECCO-MFI} in terms of profit and cost-oriented offloading optimisation. It also indicates that the \emph{PECCO-MFI} can draw a computation offloading strategy that can utilise cost wisely to produce excellent profit.
\begin{table*}[!ht]
\centering
\begin{tabular}{c|cccccc}
\hline
\backslashbox{Value}{Method} & LARAC & Greedy & BAT & SCA & WOA & CS \\ \hline
\#Allocation & 64.9 & 92.9 & 175.3 & 170.7 & 177.2 & 176.7 \\
Profit/Allocation Ratio & \textbf{40.96} & 31.69 & 30.74 & 31.68 & 31.22 & 32.48 \\
Cost/Allocation Ratio & 30.02 & 19.61 & 10.20 & 10.33 & 9.95 & 10.27 \\
\hline
\backslashbox{Value}{Method} & FFA & PSO & GWO & DE & MFO & MFI \\ \hline
\#Allocation & 178.6 & 179.2 & 179.3 & 177.6 & 179.1 & \textbf{179.3} \\
Profit/Allocation Ratio & 32.57 & 33.69 & 33.66 & 34.06 & 33.33 & 34.75 \\
Cost/Allocation Ratio & 9.96 & 9.93 & 10.17 & 10.15 & 10.08 & \textbf{9.84} \\
\hline
\end{tabular}
\caption{Number of computation tasks being allocated, the profit-allocation ratio and the cost-allocation ratio of different algorithms. }
\label{tab:profit_allocation_paratio}
\end{table*}
\vspace{0cm}
\subsection{Comparison of Task Allocation between Algorithms}
\label{sec:comparison_of_profit_and_task_completion_between_algorithms}
We now focus on task allocation done by different algorithms. If offloading in an unwise manner, some computing nodes will be overloaded, causing some tasks failed to be allocated. As we can see from Table \ref{tab:profit_allocation_paratio}, the proposed \emph{PECCO-MFI} algorithm achieves the highest task allocation number. A high number of tasks being allocated after computation offloading indicates that the \emph{PECCO-MFI} algorithm can offload tasks wisely without causing severe overloading. Hence, most number of tasks can be successfully allocated and completed instead of being stuck on overloaded computing nodes.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=0.5\textwidth]{figure/figure_profit_allocation_paratio_reduction20220413170341.pdf}\\
\caption{The percentage of profit-allocation ratio increase and cost-allocation ratio decrease achieved by the \emph{PECCO-MFI} algorithm compared with other methods are shown in (a) and (b), respectively. The extreme case LARAC is omitted for better visualisation. }
\label{fig:figure_paratio_increase}
\end{center}
\end{figure}
\vspace{0cm}
In terms of the profit-allocation ratio, both Table \ref{tab:profit_allocation_paratio} and Figure \ref{fig:figure_paratio_increase} show that except for the extreme case LARAC due to poor task allocation, the \emph{PECCO-MFI} algorithm produces the highest profit-allocation ratio. \rw{The higher the profit-allocation ratio is, the more profit will be yielded by completing each task. }As visualised in Figure \ref{fig:figure_paratio_increase}(a), the \emph{PECCO-MFI} algorithm achieves $4.26\%$ and $2.03\%$ higher profit-allocation ratio compared with the original MFO and the best-performed Differential Evolution, respectively. Besides, according to Table \ref{tab:profit_allocation_paratio}, the \emph{PECCO-MFI} yields the lowest cost-allocation ratio, which indicates the \emph{PECCO-MFI} causes the lowest cost to satisfy each task, demonstrating its cost-effectiveness. Hence, by achieving the highest number of task allocation, a high profit-allocation ratio and a low cost-allocation ratio, the effectiveness of the offloading strategy yielded by the \emph{PECCO-MFI} algorithm is verified.
\begin{table*}[!ht]
\centering
\begin{tabular}{c|cccccc}
\hline
\backslashbox{Value}{Method} & LARAC & Greedy & BAT & SCA & WOA & CS \\ \hline
Utilisation & \textcolor{grey}{121\%} & \textcolor{grey}{112\%} & 95\% & 92\% & 95\% & 95\% \\
Profit/Utilisation Ratio & 22.12 & 32.51 & 56.97 & 59.02 & 57.95 & 60.28 \\
Cost/Utilisation Ratio & \textbf{16.10} & 16.27 & 18.83 & 19.16 & 18.56 & 19.10 \\
\hline
\backslashbox{Value}{Method} & FFA & PSO & GWO & DE & MFO & MFI \\ \hline
Utilisation & 95\% & 96\% & 95\% & 94\% & 96\% & 95\% \\
Profit/Utilisation Ratio & 60.82 & 61.89 & 63.23 & 64.02 & 62.45 & \textbf{65.2} \\
Cost/Utilisation Ratio & 18.72 & 18.53 & 19.20 & 19.19 & 18.81 & 18.58 \\
\hline
\end{tabular}
\caption{Average computing node workload utilisation, the profit-utilisation ratio and the cost-utilisation ratio of different algorithms. }
\label{tab:profit_utilisation_puratio}
\end{table*}
\vspace{0cm}
\subsection{Comparison of Resource Utilisation between Algorithms}
\label{sec:comparison_of_profit_and_computing_node_utilisation_between_algorithms}
Finally, the computing node workload resource utilisation, the profit-utilisation and cost-utilisation ratio are indicated in Table \ref{tab:profit_utilisation_puratio}. As we can observe, except the LARAC and GREEDY algorithm which overloads some computing nodes, all other methods produce offloading strategy that is free from overloading.
As we can see from Figure \ref{fig:figure_puratio_increase}, the profit-utilisation ratio produced by the \emph{PECCO-MFI} algorithm is significantly higher than all other compared methods. Specifically, the \emph{PECCO-MFI} algorithm achieves a $4.4\%$ and $1.8\%$ increase in terms of profit-utilisation ratio compared with the original MFO and Differential Evolution, respectively. The higher the profit-utilisation ratio is, the more profit will be yielded by utilising one unit of computing node resource. \rw{On the other hand, the \emph{PECCO-MFI} yields a relatively low cost-utilisation ratio, which means the algorithm will not incur a high cost by utilising one unit of computation resource. }Hence, the excellent profit-utilisation and cost-utilisation ratio indicate the effectiveness of the \emph{PECCO-MFI} algorithm, i.e., producing a computation offloading strategy that can utilise computation resource wisely to achieve a high profit and a low cost, without overloading any computing nodes.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=0.5\textwidth]{figure/figure_profit_utilisation_puratio_reduction20220413170341.pdf}\\
\caption{The percentage of profit-utilisation ratio increase and cost-utilisation ratio decrease achieved by the \emph{PECCO-MFI} algorithm compared with other methods are shown in (a) and (b), respectively. The extreme cases LARAC and GREEDY are omitted from the plot. }
\label{fig:figure_puratio_increase}
\end{center}
\end{figure}
\vspace{-1cm}
\section{Conclusion}
\label{sec:section_conclusion}
In this paper, we propose a profit and cost-oriented edge-cloud computation offloading model \emph{PECCO} which jointly considers the heterogeneous communication and computation cost, as well as the profit yielded after computation offloading. \rw{An improved Moth-flame Optimisation algorithm with three improvements is proposed which addresses several deficiencies of the original MFO and is then integrated to produce an optimised edge-cloud computation offloading strategy, }forming the \emph{PECCO-MFI} algorithm. Comprehensive experiments are conducted and the \emph{PECCO-MFI} algorithm is compared with several other baseline methods to testify to the effectiveness of the \emph{PECCO-MFI} algorithm when optimising the edge-cloud computation offloading model, \rw{as well as the effectiveness of the improvements made over the original MFO. }
\section*{Acknowledgement}
\label{sec:section_acknowledgement}
This work is supported in part by Key-Area Research and Development Program of Guangdong Province (2020B010164002) and Zhejiang Provincial Natural Science Foundation of China (LZ22F020002).
\section*{Data Availability Statement}
\label{sec:section_data_availability_statement}
The data that support the findings of this study are available from the corresponding author upon reasonable request.
\begin{spacing}{0.85}
|
\section{Method}\label{sec:method} In this section, we describe the metrics
used, how these metrics are measured, and our experiment setup.
\subsection{Metrics}
\paragraph{Availability.} We are interested in determining which DoH
resolvers are still active and responding to queries. We define a resolver as
unresponsive from a given vantage point if we fail to receive \emph{any}
response to the queries issued from a particular server.
\paragraph{Latency.} We
performed network latency measurements for each recursive resolver. Each time
we issued a set of DoH queries to a resolver, we also issued four ICMP ping
messages and computed the average round-trip time. This enabled us to explore
whether there was a consistent relationship between high query response times
and network latency.
\paragraph{DNS query response time.} We define DNS query response time as the
end-to-end time it takes for a client to initiate a query and receive a
response. To measure query response times with various DoH resolvers, we
extended the open-source DNS measurement tool developed by Hounsel et
al.~\cite{hounsel2020comparing} The tool enables researchers to issue
traditional DNS, DoT, and DoH queries. It utilizes the \texttt{getdns}
library for traditional DNS and DoT, and \texttt{libcurl} for DoH.
For each DoH resolver, the tool establishes a TCP connection and an HTTPS session,
encodes a DoH query, sends the query to the DoH resolver, and reports the
response time to the client. Importantly, the tool includes the time it took
to establish a TCP connection and an HTTPS session.
We note that \texttt{libcurl} attempts to utilize TLS 1.3 when a recursive
resolver supports it, and otherwise falls back to an older version
(e.g., TLS 1.2).
It also attempts to utilize HTTP/2, falling back to an older version when a
recursive resolver does not support it.
We modified the tool to support continuous DoH response time measurements
across multiple days. We also modified the tool to enable clients to provide
a list of DoH resolvers they wish to perform measurements with, preventing
clients from needing to re-run the executable for each resolver (link
anonymized for review). After a set of measurements complete with a list of
DoH resolvers and domain names, the tool writes the results to a JSON file.
\subsection{Experiment Setup} To provide a comparative assessment of DNS
performance across DoH resolvers, we perform measurements across 75 DoH
resolvers, grouped by their geographical locations—17 in North America, 22 in
Asia, and 36 in Europe~\cite{dnscrypt}. We performed our measurements between
October 15--25, 2021. We also took the four highest
performing resolvers (\texttt{Google}, \texttt{Cloudflare}, \texttt{Quad9},
\texttt{Hurricane Electric}) located in North America and measured their
performance in Europe and Asia to better understand how they compare in
farther vantage points. We employed MaxMind's GeoLite2 databases to geolocate
each DoH resolver~\cite{maxmind}.
\paragraph{Vantage Points.} We performed our measurements from three global
vantage points through Amazon EC2~\cite{amazon_ec2}. We deployed one server
in each of the Ohio, Frankfurt, and Seoul EC2 regions. We chose to perform
measurements from multiple global vantage points to understand how DoH
performance varies not only by which resolver is used, but also which
geographic region the client is located in. Each server utilized 16 GB of RAM
and 4 virtual CPU cores (the \texttt{t2.xlarge} instance type), and they each
used Debian 10~\cite{amazon_ec2_instance_types}.
\paragraph{Resolvers and Domain Names.} \Fref{sec:resolvers} lists each of
the DoH resolvers we measured. These resolvers were
scraped from a list of public DoH resolvers provided by the \texttt{DNSCrypt}
protocol developers~\cite{dnscrypt-public-resolvers}. Previous work has
largely studied major DNS providers in use by web browsers; in contrast, we
measure the performance of a larger set of encrypted DNS resolvers~\cite{hounsel2020comparing,hounsel2021can,hoang2020k,lu2019end-to-end}.
We issued queries for two domains to each resolver: \texttt{google.com} and
\texttt{netflix.com}. We chose these domains based on their popularity, but
other domain names would have likely sufficed. We do not expect our choice of
domain names to unfairly skew our performance comparisons between resolvers.
\paragraph{Measurement Procedure.} We performed the following steps to measure
the performance of each of the encrypted DNS resolvers, from each of our three vantage points:
\begin{enumerate}
\item For each
resolver that we aim to measure, establish an HTTPS session and
send a DoH query, measuring the query response time for two
domain names.
\item For each resolver, issue four ICMP ping
probes and compute the average round-trip latency.
\end{enumerate}
\paragraph{Limitations.} Our work has several limitations.
First, we do not measure how encrypted DNS affects application
performance, such as web page load time. Ultimately, an assessment of the
effects of encrypted DNS performance on application performance, including web
page load time, across the full set of encrypted DNS resolvers, could provide
a more comprehensive understanding of the effects of encrypted DNS on
application performance. Another potential limitation of our work is that we
perform measurements exclusively from Amazon EC2 instances, which are located
in data centers. Future work could explore similar measurements from a wider
variety of access networks, including cellular networks and broadband access
networks. Furthermore, although we do not expect it to affect conclusions, it may
be informative to perform measurements from a larger set domain names; our
measurements perform DNS lookups to just two domain names.
Finally, since DoH can re-use TCP connections and TLS sessions, future work
should report connection setup times separately.
\section{Background and Related Work}\label{sec:background}
In this section, we provide background on encrypted DNS protocols, including
the current deployment status of encrypted DNS, as well as various related
work on measuring encrypted DNS.
\subsection{Background: Encrypted DNS}
The Domain Name System (DNS) translates human-readable domain names into
Internet Protocol (IP) addresses, which are used to route
traffic~\cite{rfc1035}. These queries have typically been unencrypted, which
enables on-path eavesdroppers to intercept queries and manipulate responses.
\paragraph{Encrypted DNS.}
Protocols for encrypting DNS traffic have been proposed, standardized, and
deployed in recent years, including DNS-over-HTTPS (DoH) and DNS-over-TLS
(DoT). Zhu et al. proposed DoT in 2016 to address the eavesdropping and
tampering of DNS queries~\cite{zhu2015connection}. It uses a dedicated port (853)
to communicate with resolvers over a TLS connection. In contrast,
DoH---proposed by Hoffman et al. in 2018---establishes HTTPS sessions with
resolvers over port 443~\cite{rfc8484}. This design decision enables
DoH traffic to use HTTPS as a transport, facilitating deployment as well as
making it difficult for network operators and eavesdroppers to intercept DNS
queries and responses~\cite{boettger2019empirical}. DoH can function in many
environments where DoT is easily blocked.
\paragraph{Moving the privacy threat.} Encrypting DNS queries and responses hides queries
from eavesdroppers but the recipient of the queries---the DNS resolver---can
see the queries~\cite{IEEEfight}. By design, recursive resolvers receive
queries from clients and typically need to perform additional queries to a
series of authoritative name servers to resolve domain names. For these
resolvers to determine the additional queries they need to perform (or
determine if the query can be answered from cache), they must be able to see
the queries that they receive from clients. Thus, although DoT and DoH make
it difficult for eavesdroppers along an intermediate network path to see DNS
traffic, recursive resolvers can still observe (and potentially, log) the
queries that they receive from clients. The fact that many mainstream DoH
providers (e.g., Google) already collect significant information about users
potentially raises additional privacy concerns and makes it appealing for
users to have a large number of encrypted DNS resolvers that are reliable and
perform well. For this reason, users may wish to have more control over the
recursive resolver that they use to resolve encrypted DNS queries. Having
a reasonable set of choices that perform well in the first place is thus
important, and determining whether such a set exists is the focus of this
paper.
\paragraph{Status of browser DoH deployments.}
Most major browsers currently support DoH, including Brave, Chrome, Edge,
Firefox, Opera, and Vivaldi. Operating systems have also announced
plans to implement DoH, including iOS, MacOS, and
Windows~\cite{ffSettings,operaEdgeSettings,vivaldiSettings,iosSettings,jensen2020windows}.
In this paper, we focus on DoH because it is more widely deployed
than DoT. Each of these browsers and operating systems either
currently support or have announced support for DoH (but, to our knowledge, not
DoT).
\subsection{Related Work}
\paragraph{Previous measurement studies of encrypted DNS.}
Previous studies have typically measured DoT and DoH response times the
protocols from the perspective of a few commonly used
resolvers~\cite{lu2019end-to-end}; in contrast, in this paper, we study a much
larger set of encrypted DNS resolvers, many of which are not available as
default options in major browsers. Zhu et al. proposed DoT to encrypt DNS
traffic between clients and recursive resolvers~\cite{zhu2015connection}.
They modeled its performance and found that DoT's overhead can be largely
eliminated with connection re-use. Böttger et al. measured the effect of DoT
and DoH on query response times and page load times from a university
network~\cite{boettger2019empirical}. They find that DNS generally
outperforms DoT in response times, and DoT outperforms DoH. They also find
that much of the performance cost for DoT and DoH can be amortized by re-using
TCP connections and TLS sessions. Hounsel et al. also measure response times
and page load times for DNS, DoT, and DoH using Amazon EC2
instances~\cite{hounsel2020comparing}. They compare the recursive resolvers
for Cloudflare, Google, and Quad9 to the local recursive resolvers provided by
Amazon EC2 from five global vantage points in Ohio, California, Seoul, Sydney,
and Frankfurt. They find that despite higher response times, page load times
for DoT and DoH can be \emph{faster} than DNS on lossy networks. Lu et al.
utilized residential TCP SOCKS networks to measure response times from 166
countries and found that, in the median case with connection re-use, DoT and
DoH were slower than conventional DNS over TCP by 9 ms and 6 ms,
respectively~\cite{lu2019end-to-end}.
In contrast to previous work, our focus in this paper is not to measure the
DoH protocol itself or its relative performance to unencrypted DNS; instead,
our goal is to compare the performance of encrypted DNS resolvers {\em to
each other}, to understand the extent to which this larger set of DNS
resolvers could be used by clients and applications in different regions.
\paragraph{Studies and remedies for the centralization of encrypted DNS.} Other work has studied the centralization
of the DNS and proposed various techniques to address it. Foremski et al.
find that the top 10\% of DNS recursive resolvers serve approximately 50\% of
DNS traffic~\cite{foremski2019dns-observatory}. Moura et
al.~\cite{moura2020clouding} measured DNS requests to two country code
top-level domains (ccTLD) and found that five large cloud providers being
responsible for over 30\% of all queries for the ccTLDs of the Netherlands and
New Zealand. Hoang et al.~\cite{hoang2020k} developed K-resolver,
which distributes queries over multiple DoH recursive resolvers, so that no
single resolver can build a complete profile of the user and each recursive
resolver only learns a subset of domains the user resolved. Hounsel et al.
also evaluate the performance of various query distribution strategies and
study how these strategies affect the amount of queries seen by individual
resolvers~\cite{hounsel2021encryption}.
This line of research complements our work---these previous studies in many
ways motivate an enable the use of multiple encrypted DNS resolvers, but
designing a system to take advantage of multiple recursive resolvers must be
informed about how the choice of resolver affects performance.
\paragraph{Other DNS performance studies.} Researchers have also studied how
DNS performance affects application performance. Sundaresan et al. used an
early FCC Measuring Broadband America (MBA) deployment of 4,200 home gateways
to identify performance bottlenecks for residential broadband
networks~\cite{sundaresan2013measuring}. This study found that page load
times for users in home networks are significantly influenced by slow DNS
response times. Allman studied conventional DNS performance from 100
residences in a neighborhood and found that only 3.6\% of connections were
blocked on DNS with lookup times greater than either 20 ms or 1\% of the
application's total transaction time~\cite{allman2020putting}. Otto et al.
found that the ability of a content delivery network to deliver fast page load
times to a client could be significantly hindered when clients choosing
recursive resolvers that are far away from CDN caches~\cite{otto2012content};
a subsequent proposal, namehelp, proxied DNS queries for CDN-hosted content
and sent them directly to authoritative servers. Wang et al. developed WProf,
a profiling system that analyzes various factors that contribute to page load
times~\cite{wang2013demystifying}; this study demonstrated that queries for
uncached domain names at recursive resolvers can account for up to 13\% of the
critical path delay for page loads. We do not measure the performance of a
large number of encrypted DNS resolvers from residential broadband access
networks in this paper nor do we study the effect of the choice of a broad
range of encrypted DNS resolvers on page load time, but doing so would be a
natural direction for future work.
\section{Discussion}\label{sec:discussion}
\section{Conclusion}\label{sec:conclusion}
The increased deployment of encrypted DNS, including DNS-over-HTTPS~(DoH) has
been accompanied both with ``mainstream'' DoH deployments in major browser
vendors, as well as a much broader deployment of encrypted DNS servers around
the world that are not among the common choices for resolvers in major
browsers. Understanding the viability of this larger set of encrypted DNS
resolvers is important, particularly given that a lack of diversity of viable
resolvers potentially could create new privacy concerns, if only a small
number of organizations provided good performance. We find that many
non-mainstream resolvers have higher median response times than mainstream
ones, particularly if the resolvers are not local to the region; in contrast,
most mainstream resolvers appear to be replicated and provide better response
times across different geographic regions. Some non-mainstream resolvers
exhibit particularly poor performance, exceeding four times the network
round-trip latency between the client and resolver. In some cases, however, a
local non-mainstream resolver can exhibit equivalent performance as compared
mainstream resolvers ({\it e.g.}, {\tt ordns.he.net} in North America, {\tt
dns.alldns.com} in Europe, and {\tt doh.libredns.gr} in Europe). These results
suggest both good news and room for improvement in the future: On the one
hand, viable alternatives to mainstream encrypted DNS resolvers do exist. On
the other hand, users need easy ways of finding and selecting these
alternatives, whose availability and performance may be more variable over
time than mainstream resolvers. There is also an opportunity
to invest in encrypted
DNS infrastructure operated by a greater diversity of organizations.
\section{Results}\label{sec:results}
In this section, we describe the results of our experiments. We explore which
non-mainstream resolvers are available; how
the performance of mainstream resolvers compares to that of non-mainstream
resolvers; and whether (and to what extent) encrypted DNS response time
correlates to high network latency. As described
in Section~\ref{sec:method}, our results represent measurements
from October 2021; future work could perform ongoing
measurements based on the measurement framework we have developed to validate
and extend these results.
\subsection{Are Non-Mainstream Resolvers Available?}
\if 0
\begin{table}[t!]
\centering
\begin{tabular}{lllll}
\hline
\textbf{Location} & \textbf{Resolver} & \textbf{Responsive?} & & \\
& & \textbf{Ohio} & \textbf{Seoul} & \textbf{Frankfurt} \\
\midrule
North America & \begin{tabular}[c]{@{}l@{}}dnscrypt.ca-1-doh\\ dnscrypt.ca-2-doh\\ doh-cleanbrowsing\\ doh.post-factum.tk\end{tabular} & \begin{tabular}[c]{@{}l@{}}\\ \\ \\ \end{tabular} & \begin{tabular}[c]{@{}l@{}}\\ \\ \\ \end{tabular} & \begin{tabular}[c]{@{}l@{}}\\ \\ \\ \end{tabular} \\ \midrule
Asia & \begin{tabular}[c]{@{}l@{}}doh.tiarap.org\\ jp.tiarap.org\\ doh.linuxsec\end{tabular} & \begin{tabular}[c]{@{}l@{}}\\ \\ \end{tabular} & \begin{tabular}[c]{@{}l@{}}\\ \\ \end{tabular} & \begin{tabular}[c]{@{}l@{}}\checkmark\ \checkmark\ \end{tabular} \\ \midrule
Europe & \begin{tabular}[c]{@{}l@{}}doh.bortzmeyer\\ doh.appliedprivacy\\ doh.chewbacca.meganerd.nl\\ doh.powerdns\end{tabular} & \begin{tabular}[c]{@{}l@{}}\\ \\ \\ \end{tabular} & \begin{tabular}[c]{@{}l@{}}\\ \\ \\ \end{tabular} & \begin{tabular}[c]{@{}l@{}}\\ \\ \\ \end{tabular} \\ \bottomrule
\end{tabular}
\caption{Resolvers that failed to respond from each vantage point.}
\label{tab:unresponsive}
\end{table}
\fi
\begin{table}[t!]
\centering
\begin{small}
\begin{tabular}{l|rr}
\toprule
\textbf{Error} & \textbf{Count} & \textbf{\% of All Responses} \\
\midrule
Couldn't Connect to Server & 47,377 & 7\% \\
HTTP Error Status & 38,475 & 5.7\% \\
Couldn't Decode Response & 26,686 & 4\% \\
SSL Connect Error & 17,720 & 2.6\% \\
Couldn't Resolve the Resolver's Domain Name & 8,864 & 1.3\% \\
SSL Certificate Error & 4,465 & 0.7\% \\
Other Error & 234 & $<$ 1\% \\
SSL Timeout & 27 & $<$ 1\% \\
Error in the HTTP/2 Framing Layer & 2 & $<$ 1\% \\
\midrule
\textbf{Successful Responses} & \textbf{531,528} & \textbf{78.7\%} \\
\textbf{All Errors} & \textbf{143,848} & \textbf{21.3\%} \\
\bottomrule
\end{tabular}
\end{small}
\caption{Errors when trying to perform DoH queries from all vantage points.}
\label{tab:errors}
\end{table}
We first aimed to study the availability of encrypted DNS
resolvers.
We received responses from most resolvers that we queried. Table~\ref{tab:errors}
shows the most common errors we received from attempting to communicate with
the resolvers.\footnote{Due to a bug, we did not record one rare error type.}
The most common errors were related
to a failure to establish a connection. It is likely
that in many of these cases, the resolver itself was simply no longer
operational; additional longitudinal measurements could confirm this
hypothesis. We did not receive a response from eleven resolvers:
\texttt{dns1.dnscrypt.ca, dns2.dnscrypt.ca, doh.cleanbrowsing.org,
doh.post-factum.tk, doh.linuxsec.org, doh.tiar.app, jp.tiar.app,
doh.appliedprivacy.net, doh.bortzmeyer.fr, doh.chewbacca.meganerd.nl,
doh.powerdns.org}. However, we only performed measurements at a single point
in time, and thus we lack enough measurements to draw general conclusions
about the availability of these resolvers. Future work could extend this work
to perform longitudinal measurements of availability of these
resolvers.\footnote{We will release our measurement framework and data upon
publication of this work.}
\begin{figure}[t!]
\hspace*{-1in}
\begin{minipage}{1.35\textwidth}
\subfigure[North America (Local).]{%
\label{fig:Ohio_NA}%
\includegraphics[width=0.32\linewidth]{Ohio_North_America.pdf}}%
\hfill%
\subfigure[Asia.]{%
\label{fig:Ohio_Asia}%
\includegraphics[width=0.32\linewidth]{Ohio_Asia.pdf}}%
\hfill%
\subfigure[Europe.]{%
\label{fig:Ohio_Europe}%
\includegraphics[width=0.32\linewidth]{Ohio_Europe.pdf}}%
\caption{The DNS response time and ICMP ping time distributions for
encrypted DNS resolvers measured from a vantage point in the United States
(Ohio). Mainstream resolvers are shown in boldface across all three
sub-figures.}
\label{fig:dns-us}
\end{minipage}
\end{figure}
\subsection{\mbox{How Do Non-Mainstream Resolvers Perform?}}
Given the large number of non-mainstream resolvers that have not been
previously studied, we aimed to study how
the performance of these encrypted DNS resolvers compared to mainstream ones.
As previously mentioned, one of our motivations in doing so is to better
understand the global extent of encrypted DNS resolver deployment, as existing
lists of public encrypted DNS resolvers~\cite{dnscrypt-public-resolvers} do
not provide any overview of either reliability or performance. Additionally,
given that most organizations who have deployed mainstream resolvers are based
in the United States, we also wanted to explore how the performance of a
broader set of encrypted DNS resolvers varied by geography.
Figures~\ref{fig:dns-us}--\ref{fig:dns-europe} show the distributions of DNS
response times and ICMP ping times across encrypted DNS resolvers, as measured
from vantage points in the United States, Asia, and Europe, respectively. The
plots show distributions for both DNS response times and ICMP round-trip latency.
Although some distributions extend beyond 500~ms, we have truncated the plots
for ease of exposition, since responses beyond this range will not result in
good application performance.
Certain resolvers did not respond to our ICMP ping probes; for those
resolvers, no latency data is shown.
As expected, most mainstream resolvers outperformed non-mainstream resolvers
from most vantage points.
Non-mainstream resolvers also exhibited higher variability of
median query response times.
From North America, we observe that aside from the five resolvers with the
lowest encrypted DNS response time,
median query response times ranged from
60~ms to 323~ms, all considerably above typical DNS lookup response times. From Asia, median response times outside
of the top five were extremely variable, ranging from 114~ms to 31.6~seconds.
Although the slowest resolver in Asia as measured from Asia was an extreme
outlier, we saw that several resolvers in Asia had slow median response times,
even when queried from Asia. For example,
the second-slowest resolver also exhibited slow performance, with a median
response time of about one second from our vantage point in Asia.
We observe more
consistent performance for resolvers in Europe, but we still
see median query response times outside the top five range from 19 ms to
156 ms.
In some cases, however, a particular non-mainstream resolver would outperform
the mainstream DoH resolvers. As expected, \texttt{dns.quad9.net},
\texttt{dns.google}, and \texttt{dns.cloudflare.com} were among the top five
highest performing DoH resolvers in North America, Europe, and Asia.
Interestingly, however, \texttt{ordns.he.net}---a DoH resolver hosted by
Hurricane Electric, a global Internet service provider (ISP)---managed to
outperform \texttt{dns.google} and \texttt{dns.cloudflare.com} from all three
vantage points. From Frankfurt, \texttt{dns-family.adguard.com} and
\texttt{ordns.he.net} are the top two highest performing resolvers; from
Seoul, \texttt{doh-jp.blahdns.com} and \texttt{ordns.he.net} outperforms \texttt{dns.google}.
\begin{figure}[t!]
\hspace*{-1in}
\begin{minipage}{1.35\textwidth}
\centering
\subfigure[North America.]{%
\label{fig:Seoul_NA}%
\includegraphics[width=0.32\linewidth]{Seoul_North_America.pdf}}%
\hfill%
\subfigure[Asia (Local).]{%
\label{fig:Seoul_Asia}%
\includegraphics[width=0.32\linewidth]{Seoul_Asia.pdf}}%
\hfill%
\subfigure[Europe.]{%
\label{fig:Seoul_Europe}%
\includegraphics[width=0.32\linewidth]{Seoul_Europe.pdf}}%
\caption{The DNS response time and ICMP ping time distributions for
encrypted DNS resolvers measured from a vantage point in Asia (Seoul, South
Korea). Mainstream resolvers are shown in boldface across all three
sub-figures.}
\label{fig:dns-asia}
\end{minipage}
\end{figure}
\if 0
\begin{table}[t!]
\centering
\begin{tabular}{ll}
\hline
\textbf{Location} & \textbf{Resolver} \\ \hline
North America & ordns.he.net \\ \hline
Asia & \begin{tabular}[c]{@{}l@{}}ordns.he.net\\ doh-jp.blahdns.com\end{tabular} \\ \hline
Europe & \begin{tabular}[c]{@{}l@{}}ordns.he.net\\ dns-family.adguard.com\\ doh.libredns.gr\end{tabular} \\ \hline
\end{tabular}
\caption{Non-mainstream resolvers that outperformed mainstream resolvers.}
\label{tab:outperformed_resolvers}
\end{table}
\fi
\begin{figure}[t!]
\hspace*{-1in}
\begin{minipage}{1.35\textwidth}
\centering
\subfigure[North America.]{%
\label{fig:Frankfurt_NA}%
\includegraphics[width=0.32\linewidth]{Frankfurt_North_America.pdf}}%
\hfill%
\subfigure[Asia. ]{%
\label{fig:Frankfurt_Asia}%
\includegraphics[width=0.32\linewidth]{Frankfurt_Asia.pdf}}%
\hfill%
\subfigure[Europe (Local).]{%
\label{fig:Frankfurt_Europe}%
\includegraphics[width=0.32\linewidth]{Frankfurt_Europe.pdf}}%
\caption{The DNS response time and ICMP ping time distributions for
encrypted DNS resolvers measured from a vantage point in Europe (Germany). Mainstream resolvers are shown in boldface across all three
sub-figures.}
\label{fig:dns-europe}
\end{minipage}
\end{figure}
\begin{table}[t!]
\hspace*{-0.75in}
\begin{minipage}{1.25\linewidth}
\parbox{0.45\linewidth}{
\begin{scriptsize}
\centering
\begin{tabular}{l|rr}
\toprule
\textbf{Resolver} & \multicolumn{2}{c}{\textbf{Vantage Point}} \\
& \textrm{Seoul (ms)} & \textrm{Frankfurt (ms)} \\
\midrule
dns.twnic.tw & 31606.61 & 32319.90 \\
doh-jp.blahdns.com & 109.62 & 773.47 \\
sg.yepdns.com & 238.01 & 548.67 \\
public.dns.iij.jp & 200.82 & 496.74 \\
doh-sg.blahdns & 214.44 & 534.51 \\
\bottomrule
\end{tabular}
\end{scriptsize}
\caption{Median DNS response times for non-mainstream resolvers (Asia).}
\label{tab:UnconvAsia}
}\hfill
\centering
\parbox{0.45\linewidth}{
\begin{scriptsize}
\begin{tabular}{l|rr}
\toprule
\textbf{Resolver} & \multicolumn{2}{c}{\textbf{Vantage Point}} \\
& \textrm{Frankfurt (ms)} & \textrm{Seoul (ms)} \\
\midrule
doh.ffmuc.net & 108.92 & 1309.88 \\
ibksturm.synology.me & 132.66 & 1254.47 \\
dns-doh-no-safe-search.dnsforfamily & 135.88 & 1159.70 \\
dnsforge.de & 31.86 & 1043.03 \\
dns-doh.dnsforfamily & 155.81 & 1162.27 \\
\bottomrule
\end{tabular}
\end{scriptsize}
\caption{Median DNS response times for non-mainstream resolvers (Europe).}
\label{tab:UnconvEur}
}
\end{minipage}
\end{table}
To better understand the extent to which certain encrypted DNS resolvers
perform well for clients in some regions but not others, we identified
resolvers that exhibited low DNS response times in for clients in one region
but not another.
Tables~\ref{tab:UnconvAsia} and~\ref{tab:UnconvEur} show
the five encrypted DNS resolvers for Europe and Asia that exhibit the
largest differences in median DNS response times when
queried from a remote vantage point (queries of resolvers in Asia from Europe,
and of resolvers of Europe from Asia, respectively). In both cases,
Table~\ref{tab:UnconvAsia} shows that non-mainstream resolvers located
in Asia perform better from the vantage point in Seoul than the one in
Frankfurt. Similarly, as expected, Table~\ref{tab:UnconvEur} shows that the
median response times of non-mainstream resolvers in Europe are much lower when
measured from Frankfurt than the response times of those same resolvers
measured from Seoul.
\subsection{Does Latency to Resolver Correlate with Performance?}
Figure~\ref{fig:ping_qrt_scatters} compares median ICMP ping latencies to median
query response times for each resolver.
Higher network latency naturally translates to higher DNS response
times (as encrypted DNS queries do require several network round trips):
absent connection re-use, a DoH lookup will require at least three network
round trips: one for the TCP three-way handshake, one for the TLS handshake
(piggybacked on the third part of the three-way handshake), and one for the
DNS query and response. {\tt libcurl} uses TLS 1.3 when available on the
server, and otherwise falls back to an older version (e.g., TLS 1.2), which requires an additional
network round trip.
It also attempts to use HTTP/2, falling back to an older version when necessary.
Thus, we can expect that ``normal'' DoH lookup times would
be about 3--4 times the network latency.
\begin{figure*}[t]
\hspace*{-0.75in}
\begin{minipage}{1.35\linewidth}
\subfigure[Ohio (Local).]{%
\label{fig:Ohio_ping_qrt_scatter}
\includegraphics[width=0.3\linewidth]{Ohio_ping_qrt_scatter.pdf}
}
\hfill
\subfigure[Seoul (Local).]{%
\label{fig:Seoul_ping_qrt_scatter}
\includegraphics[width=0.3\linewidth]{Seoul_ping_qrt_scatter.pdf}
}
\hfill
\subfigure[Frankfurt (Local).]{%
\label{fig:Frankfurt_ping_qrt_scatter}
\includegraphics[width=0.3\linewidth]{Frankfurt_ping_qrt_scatter.pdf}
}
\caption{Median DoH query response times vs. median round-trip network
latency for each local DoH resolver, from each vantage point.}
\label{fig:ping_qrt_scatters}
\end{minipage}
\end{figure*}
In some cases, certain resolvers performed worse than might
be expected as a result of the latency from the vantage point to the
resolvers.
For example, from North America, we
observed a median query response time of 872 ms to \texttt{doh.xfinity.com}
despite a network round-trip latency of merely 166~ms.
Many non-mainstream resolvers in Europe and Asia exhibit encrypted DNS
response times that are more than 4x the round-trip network latency.
These discrepancies between network latency and DNS response times may be due
to a number of factors, including both caching and server load.
From Frankfurt, we observed two resolvers with median DoH query response
times that are less than 3--4 times the median network latency to these resolvers:
\texttt{dns.switch.ch} and \texttt{odvr.nic.cz}.
We hypothesize that this could be attributed to anomalous ICMP ping data and
network optimizations that enable DoH queries to be completed in fewer
round-trips.
Two resolvers---\texttt{doh.this.web.id} and
\texttt{dns.twnic.tw}---are not shown in Figure~\ref{fig:Seoul_ping_qrt_scatter}
because their median query response times are greater than 1,000 ms.
For reference, the next slowest resolver from Seoul has a median query response
time of 456~ms.
\if 0
\begin{figure}[t!]
\hspace*{-0.75in}
\begin{minipage}{1.25\textwidth}
\centering
\subfigure[North American resolvers measured from Ohio, USA]{%
\label{fig:Local_Ohio_NA}%
\includegraphics[width=0.32\linewidth]{Ohio_North_America.pdf}}%
\hfill%
\subfigure[Asian resolvers measured from Seoul, South Korea]{%
\label{fig:Local_Seoul_Asia}%
\includegraphics[width=0.32\linewidth]{Seoul_Asia.pdf}}%
\hfill%
\subfigure[European resolvers measured from Frankfurt, Germany]{%
\label{fig:Local_Frankfurt_Europe}%
\includegraphics[width=0.32\linewidth]{Frankfurt_Europe.pdf}}%
\caption{Local-to-local resolvers.}
\end{minipage}
\end{figure}
\fi
\section{Introduction}\label{sec:intro}
The Domain Name System (DNS) is a critical component of the Internet's
infrastructure that translates human-readable domain names ({\it e.g.},
\texttt{google.com}) into Internet Protocol (IP) addresses~\cite{rfc1035}.
Most Internet communications begin with a client device sending DNS queries to
a recursive resolver, which in turn queries one or more name servers, which
ultimately refer the client to a server who can map the domain to an IP
address.
The \emph{response times} of these queries---the time
to contact a recursive resolver, query various name servers, and return
the results---is important because the DNS underlies virtually all
communication on the Internet. For example, loading a
web page, a browser must first resolve the domain names for each object on the
page before the objects themselves can be retrieved and rendered. Thus, the
performance of DNS lookup is of utmost importance to application performance
such as web performance, as slow DNS lookup times will lead to slow web page
loads.
DNS did not originally take privacy and security into account: DNS queries
have historically been unencrypted, leaving users susceptible to
eavesdropping~\cite{schmitt2019:odns:pets}; queries can also be intercepted and
manipulated~\cite{jones2016detecting}. To
address these types of privacy and security vulnerabilities, encrypted DNS
protocols have been developed and deployed, including DNS-over-HTTPS
(DoH)~\cite{rfc8484}, which is now deployed---and even enabled by default---in
many browsers. DoH enables clients to communicate with recursive resolvers
over HTTPS, providing privacy and security guarantees that DNS previously
lacked.
For better or worse, most contemporary deployments of DoH have occurred in
browsers that provide limited options for
resolvers~\cite{ffChoices,chromeResolvers}. Although DoH
protects against on-path eavesdropping, it does not prevent resolvers
\emph{themselves} from seeing the contents of DNS queries. Thus, some have
argued that browser-based DoH deployments shift privacy concerns from
eavesdroppers to potential misuse by major DNS providers~\cite{vixie}.
\begin{table}[t]
\centering
\begin{tabular}{l|cccccc}
\hline
Browser & Cloudflare & Google & Quad9 & NextDNS & CleanBrowsing & OpenDNS
\\
\midrule
Chrome & \checkmark & \checkmark & & \checkmark & \checkmark & \checkmark \\
Firefox & \checkmark & & & \checkmark & & \\
Edge & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark \\
Opera & \checkmark & \checkmark & & & & \\
Brave & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark \\
\bottomrule
\end{tabular}
\caption{Modern browsers provide only a few choices for encrypted DNS
resolver, which we define as {\bf mainstream} resolvers.}
\label{tab:SupportedResolvers}
\end{table}
\Fref{tab:SupportedResolvers} shows the DoH resolvers that have been deployed
to users of major browsers as of October 20,
2021~\cite{bravebrowser,edgebrowser,ffbrowser,chromebrowser,operabrowser}. We
define the resolvers listed in~\Fref{tab:SupportedResolvers} as {\em
mainstream}.
Yet, many other DoH resolvers have been deployed that are currently
not in use by major browser deployments~\cite{dnscrypt}---in other words,
there are many non-mainstream DoH resolver deployments.
Previous studies have measured encrypted DNS performance, but they have mostly focused on mainstream DNS resolvers~\cite{hounsel2020comparing,hounsel2021can,hoang2020k,lu2019end-to-end}.
In this paper, we expand on these previous studies, exploring the performance
of all encrypted DNS resolvers---from a variety of global
vantage points, as opposed to simply characterizing the mainstream DoH
providers from well-connected vantage points.
Towards this goal, we make the following contributions:
\begin{enumerate}
\itemsep=-1pt
\item We measure DoH response times a large list of resolvers, including
both mainstream DoH resolvers that are included in major browser
vendors {\em and} a large collection of non-mainstream resolvers.
\item We study how the performance of various DoH resolvers differ based
on vantage point.
\end{enumerate}
\noindent
To our knowledge, this paper presents the first study of DoH performance
measurements for non-mainstream resolvers, as well as the first comparison of
DoH performance across a variety of vantage points, for a large number of
resolvers.
To perform these experiments, we developed and released an open-source
tool for measuring encrypted DNS performance to replicate and extend these
results, and to support further research on DoH performance.
The rest of the paper is organized as follows. Section~\ref{sec:background}
provides background on DNS, including the origin of encrypted DNS and related
standards, and discusses related work. Section~\ref{sec:method} details our research questions, the
experiments we conducted to study these questions, and the limitations of the
study. Section~\ref{sec:results} presents the results of these experiments.
Section~\ref{sec:conclusion} concludes with a discussion of the implications
of these results and possible directions for future work.
\section{Appendix}
\subsection{Additional Figures}
\begin{figure}[h!]
\hspace*{-1in}
\begin{minipage}{1.35\textwidth}
\centering
\subfigure[North America (Local).]{%
\label{fig:Ohio_NA_Appendix}%
\includegraphics[width=0.32\linewidth]{Ohio_North_America_Full.pdf}}%
\hfill%
\subfigure[Asia (Local).]{%
\label{fig:Seoul_Asia_Appendix}%
\includegraphics[width=0.32\linewidth]{Seoul_Asia_Full.pdf}}%
\hfill%
\subfigure[Europe (Local).]{%
\label{fig:Frankfurt_Europe_Appendix}%
\includegraphics[width=0.32\linewidth]{Frankfurt_Europe_Full.pdf}}%
\caption{Resolvers measured from local vantage points (including
mainstream resolvers). These plots do not
truncate response times at 500~ms.}
\end{minipage}
\end{figure}
\subsection{Resolvers}\label{sec:resolvers}
\begin{itemize}
\item https://dns.google/dns-query
\item https://dns.aa.net.uk/dns-query
\item https://adfree.usableprivacy.net/dns-query
\item https://dns.adguard.com/dns-query
\item https://dns-family.adguard.com/dns-query
\item https://doh.in.ahadns.net/dns-query
\item https://doh.la.ahadns.net/dns-query
\item https://doh.nl.ahadns.net/dns-query
\item https://dns.alidns.com/dns-query
\item https://dnsnl-noads.alekberg.net/dns-query
\item https://dnsnl.alekberg.net/dns-query
\item https://dns.arapurayil.com/dns-query
\item https://dohtrial.att.net/dns-query
\item https://dnses.alekberg.net/dns-query
\item https://doh.bortzmeyer.fr/dns-query
\item https://dns.circl.lu/dns-query
\item https://doh.opendns.com/dns-query
\item https://dns.cloudflare.com/dns-query
\item https://family.cloudflare-dns.com/dns-query
\item https://security.cloudflare-dns.com/dns-query
\item https://odvr.nic.cz/dns-query
\item https://dns.digitale-gesellschaft.ch/dns-query
\item https://dns.digitale-gesellschaft.ch/dns-query
\item https://dns1.ryan-palmer.com/dns-query
\item https://doh.sb/dns-query
\item https://dns.therifleman.name/dns-query
\item https://dns1.dnscrypt.ca/dns-query
\item https://dns2.dnscrypt.ca/dns-query
\item https://dns-doh.dnsforfamily.com/dns-query
\item https://dns-doh-no-safe-search.dnsforfamily.com/dns-query
\item https://dnsforge.de/dns-query
\item https://dns.dnshome.de/dns-query
\item https://doh.pub/dns-query
\item https://doh-ch.blahdns.com/dns-query
\item https://doh.cleanbrowsing.org/dns-query
\item https://doh.cleanbrowsing.org/dns-query
\item https://doh.cleanbrowsing.org/dns-query
\item https://doh.crypto.sx/dns-query
\item https://doh-de.blahdns.com/dns-query
\item https://doh-fi.blahdns.com/dns-query
\item https://ibksturm.synology.me/dns-query
\item https://doh-jp.blahdns.com/dns-query
\item https://doh-sg.blahdns.com/dns-query
\item https://doh.appliedprivacy.net/dns-query
\item https://doh.ffmuc.net/dns-query
\item https://doh.tiarap.org/dns-query
\item https://ordns.he.net/dns-query
\item https://doh.tiar.app/dns-query
\item https://public.dns.iij.jp/dns-query
\item https://doh.this.web.id/dns-query
\item https://jp.tiar.app/dns-query
\item https://jp.tiarap.org/dns-query
\item https://doh.libredns.gr/dns-query
\item https://doh.libredns.gr/dns-query
\item https://doh.linuxsec.org/dns-query
\item https://doh.linuxsec.org/dns-query
\item https://adblock.doh.mullvad.net/dns-query
\item https://doh.chewbacca.meganerd.nl/dns-query
\item https://doh.mullvad.net/dns-query
\item https://dns.nextdns.io/dns-query
\item https://dns.njal.la/dns-query
\item https://doh.post-factum.tk/dns-query
\item https://draco.plan9-ns2.com/dns-query
\item https://doh.powerdns.org/dns-query
\item https://doh.seby.io/dns-query
\item https://doh-2.seby.io/dns-query
\item https://dns.twnic.tw/dns-query
\item https://dns9.quad9.net/dns-query
\item https://dns9.quad9.net/dns-query
\item https://dnsse.alekberg.net/dns-query
\item https://dns.switch.ch/dns-query
\item https://dns.t53.de/dns-query
\item https://unicast.uncensoreddns.org/dns-query
\item https://anycast.uncensoreddns.org/dns-query
\item https://sg.yepdns.com/dns-query
\item https://doh.xfinity.com/dns-query
\end{itemize}
|
\section{Introduction}
\label{s:intro}
Parental origin effects exist when the genetic effects of alleles inherited from parents depend on their parental origin. Parental origin effects play an important role in certain neonatal diseases, such as Beckwith-Wiedemann, Prader-Willi, and Angelman syndrome \citep{falls1999genomic}. There is evidence that parental origin effects of fetal genes may also be involved in the development of maternal pregnancy-related diseases \citep{kanayama2002deficiency, petry2007does, saftlas2005immunogenetic, wangler2005factors}.
Most studies for testing parental origin effects used traditional family designs because detecting these effects requires familial information, including the transmission disequilibrium test \citep{weinberg1999methods,purcell2007plink,orton2011vitamin,wang2012ror2}. Due to high cost of recruiting father-mother-child trios \citep{yang2013robust}, it is an intermediate solution to use a mother-child pair design, which is widely used in the studies of neonatal diseases \citep{fu2013testing,mendoncca2019epigenetic,van2020maternal}. There have been several studies that used mother-child data to assess parental origin effects \citep{ainsworth2011investigation,weinberg2009genetics,howey2012premim} by applying a log-linear model to single nucleotide polymorphism (SNP) data. The major difficulty in studying parental origin effects is that the parental origin of the children alleles cannot be unambiguously determined when the genotypes of both mother and child are heterozygous. The power for testing parental origin effects could be substantially improved by exploiting linkage disequilibrium information between the target SNP and its adjacent SNPs since such information can assist to determine parental origin of child alleles \citep{browning2009unified,kong2009parental,howey2015increased,zhang2020incorporating}. Using mother-child paired genotype data, \cite{lin2013multi} developed a likelihood based efficient statistical inference method by fully utilizing the information from adjacent SNPs. Furthermore, some studies considered using data from case-parent trios \citep{weinberg1998log,cordell2004case} or other types of family structures \citep{han2013joint,zhang2016optimum,zhang2019imprinting}.
Some studies use discrete covariates for stratification \citep{gjessing2006case,gjerdevik2018parent}, but none of the above methods can adjust for continuous covariates in a general way. In this paper, we propose to model genetic effects through haplotypes related to SNPs tightly linked to the target SNP, adjusting for environmental factors. Considerable efficiency gain can be achieved since the information from adjacent SNPs is utilized. We derive a modified profile likelihood function for the observed data by treating the joint distribution of maternal genotype and covariates to be fully nonparametric, so that our method is robust to some extent. By treating the haplotypes associated with multiple SNPs as missing data, we extend the conventional expectation-maximization algorithm to maximize the modified profile likelihood function, and show that the algorithm converges under some mild conditions. The resultant estimator is shown to be asymptotically efficient in the sense of \cite{bickel1993efficient}, and it could be more powerful than the existing methods that do not incorporate covariates in some situations. Furthermore, our expectation-maximization algorithm can be naturally extended to handle missing child genotypes.
\section{Method}
\label{s:method}
\subsection{Model and notations}
Let $Y$ denote binary case-control status ($Y = 1$: case; $Y = 0$: control), ($g^m$, $g^c$) denote the target SNP genotypes for a mother-child pair, and $X$ denote a $p$-dimensional vector of environmental factors (covariates) collected from the mother. Let $A$ and $a$ denote the two alleles of the target SNP, where $a$ is the minor allele (frequency $\le0.5$). For the sake of simplicity, we assume that all SNPs follow an additive mode of inheritance by coding the genotypes $AA$, $Aa$, and $aa$ as $0$, $1$, and $2$, respectively. Let $g^c_m$ and $g^c_p$ denote children alleles at the target SNP inherited from the mother and father, respectively, coded as $0$ and $1$ for $A$ and $a$, respectively. The parental origin information can be unambiguously inferred if and only if at least one of $g^m$ and $g^c$ is not 1. In the situation where the parental origin information cannot be unambiguously inferred using the target SNP genotypes, one can use the genotypes of those SNPs tightly linked with the target SNP to improve inference accuracy. Such SNPs are called adjacent SNPs hereafter. Let $G^m$ and $G^c$ denote the genotype vectors at the target SNP and adjacent SNPs for the mother and child, respectively.
Let $n$ denote the total number of families. For $u=1,\ldots,n$, let the observed data for the $u$th mother-child pair be denoted by $(Y_u, X_u, G_u^m, G_u^c)$. We assume the following penetrance function:
\begin{equation} \label{equ:penetrance}
\mbox{pr}(Y=1 \mid g^{m}, g_{m}^{c}, g_{p}^{c}, X)= \mbox{expit}\{H(g^{m}, g_{m}^{c}, g_{p}^{c}, X ; \bm{\beta})\},
\end{equation}
where $\mbox{expit}(\cdot) = e^\cdot/(1 + e^\cdot)$ is the expit function, $H(g^{m}, g_{m}^{c}, g_{p}^{c}$, $X$ ; $\bm{\beta})$ can be a linear combination function of $g^m$, $g^c$, $g^c_m-g^c_p$, $X$, and their products, and $\bm{\beta}$ is a $p$-vector of associated parameters. For example, a penetrance function involving main genetic effects, parental origin effects, and gene-environment interaction effects is
\begin{align}
H\left(g^{m}, g_{m}^{c}, g_{p}^{c}, X ; \bm{\beta}\right)= &\beta_0 + \beta_{g^{m}} g^{m} + \beta_{g^{c}} g^{c} + \beta_{i m}\left(g_{m}^{c}-g_{p}^{c}\right) + \beta_{X}^{{ \mathrm{\scriptscriptstyle T} }} X \notag\\
&+ \beta_{g^m X}^{{ \mathrm{\scriptscriptstyle T} }} g^{m} X + \beta_{g^c X}^{{ \mathrm{\scriptscriptstyle T} }} g^{c} X,\label{equ:penetrance1}
\end{align}
where $a^{ \mathrm{\scriptscriptstyle T} }$ stands for the transpose of vector $a$.
We are interested in detecting the parental origin effect $\beta_{im}$.
\subsection{A retrospective likelihood function for observed data}
Let $(Y_u,X_u,G^m_u,G^c_u,g^m_u,g^c_u,g^c_{mu},g^c_{pu})$, $u=1,\ldots,n$, be independent copies of the random vector $(Y,X,G^m,G^c,g^m,g^c,g^c_{m},g^c_{p})$.
Throughout this paper, we assume that the population prevalence $f$ is known \textit{a priori}. Then, under the case-control design, the retrospective likelihood function can be written as
\[\prod_{u=1}^{n} \mbox{pr}\left(G_{u}^{m}, G_{u}^{c}, X_{u}\mid Y_{u}\right)\]
under the constraint $\mbox{pr}(Y=1) = f$.
Recall that $g^m_u$, the maternal genotype at the target SNP, is a component of the maternal genotype vector $G^m_u$ so that $\mbox{pr}(G^m_u,g^m_u)=\mbox{pr}(G^m_u)$. Then the retrospective likelihood function is equivalent to
\begin{equation} \label{equ:retro}
\prod_{u=1}^{n} \mbox{pr}\left(Y_{u},G_{u}^{m}, G_{u}^{c}, X_{u} \right)
=\prod_{u=1}^{n}\mbox{pr}(Y_u\mid G_u^m,G_u^c,X_u)\mbox{pr}(G_u^m,G_u^c\mid g_u^m,X_u)\mbox{pr}(g_u^m,X_u)
\end{equation}
under the constraint $\mbox{pr}(Y=1) = f$.
In our method, we leave the joint distribution of $(g^m_u,X_u)$ to be totally unspecified. That is, the empirical likelihood approach \citep{owen2001empirical} is adopted by introducing probability parameters $\pi_u=\mbox{pr}(g^m_u,X_u)$ subject to the constraint $\sum_{u=1}^{n}\pi_u = 1$. Denote $\bm{\pi}=(\pi_1,\dots,\pi_n)^{ \mathrm{\scriptscriptstyle T} }$. We assume that $(G^m,G^c)$ and $X$ are conditionally independent given $g^m$, which holds true if the considered SNPs are associated with $X$ only through the target SNP \citep{chen2012semiparametric, zhang2020efficient}. Under this conditional independence assumption, we have
\begin{equation} \label{equ:geno}
\mbox{pr}(G^m,G^c\mid g^m,X)=\frac{\mbox{pr}(G^m,G^c)}{\mbox{pr}(g^m)}.
\end{equation}
Note that $\mbox{pr}(Y\mid G^m,G^c,X)$ depends on $(G^m,G^c,X)$ only through the penetrance function $\mbox{pr}(Y=1\mid g^m,g^c_m,g^c_p,X)$. However, $(g^c_m,g^c_p)$ cannot be inferred only using the genotypes of mother-child pair at the target SNP when both of the genotypes are heterozygous. Instead, we can apply the Law of Total Probability to calculate $\mbox{pr}(Y\mid G^m,G^c,X)$ by going through all possible haplotype combinations compatible with the observed genotypes. Assume the joint genotypes are availabe for the target SNP and its $K-1$ adjacent SNPs, and any two of the $K$ SNPs are close enough so that the recombination probability between them is nearly 0. Let $\{h_1,\dots,h_S\}$ denote all possible haplotypes, and let $\bm{\mu} = (\mu_1,\dots,\mu_S)^{ \mathrm{\scriptscriptstyle T} }$ denote the corresponding haplotype frequencies with $\sum_{s=1}^{S} \mu_s = 1$. Let $h_i^m h_j^m$ denote an unordered diplotype (haplotype pair) for the mother, and $(h_w^c,h_l^c)$ denote an ordered diplotype for the child,
where $h_w^c$ is the haplotype inherited from the mother so that $w=i$ or $w=j$, and $h_i^m$, $h_j^m$, $h_w^c$, and $h_l^c$ belong to $\{h_1,\dots,h_S\}$. The retrospective likelihood function \eqref{equ:retro} can be rewritten as
\begin{equation} \label{equ:haplo}
\prod_{u=1}^{n}\sum_
{\substack{h_{iu}^{m}h_{ju}^{m}\in{H}_u^m\\(h_{wu}^c,h_{lu}^c)\in{H}_u^c}}
\mbox{pr}(Y_u\mid h_{iu}^{m}h_{ju}^{m},h_{wu}^{c},h_{lu}^{c},X_u)\mbox{pr}(h_{iu}^{m}h_{ju}^{m},h_{wu}^{c},h_{lu}^{c}\mid g_u^m,X_u)\mbox{pr}(g_u^m,X_u).
\end{equation}
Here, $H_u^{m}=\{h_{i u}^{m} h_{j u}^{m}\mid i,j=1,\ldots,S\}$ is the set of unordered diplotypes compatible with the joint maternal genotype $G_{u}^{m}$ for the $u$th family, and $H_u^{c}=\{(h_{w u}^{c}, h_{l u}^{c})\mid {w,l}=1,\ldots,S\}$
is the set of ordered diplotypes compatible with the joint maternal and child genotypes $(G_{u}^{m},G_{u}^{c})$ for the $u$th family.
Let $\bm{\Theta} = (\bm{\beta}^{ \mathrm{\scriptscriptstyle T} },\bm{\mu}^{ \mathrm{\scriptscriptstyle T} })^{ \mathrm{\scriptscriptstyle T} }$ denote the vector of all unknown model parameters. Similar to \eqref{equ:geno}, we have
\[\mbox{pr}(h_i^m h_j^m,h_w^c,h_l^c\mid g^m,X)=\frac{\mbox{pr}(h_i^m h_j^m,h_w^c,h_l^c)}{\mbox{pr}(g^m)}, \]
so that \eqref{equ:haplo} can be written as
\begin{equation} \label{equ:semi}
L(\bm{\Theta},\bm{\pi}) = \prod_{u=1}^{n}\sum_
{\substack{h_{iu}^{m}h_{ju}^{m}\in{H}_u^m\\(h_{wu}^c,h_{lu}^c)\in{H}_u^c}}
\mbox{pr}(Y_u\mid h_{iu}^{m}h_{ju}^{m},h_{wu}^{c},h_{lu}^{c},X_u)\frac{\mbox{pr}(h_{iu}^m h_{ju}^m,h_{wu}^c,h_{lu}^c)}{\mbox{pr}(g_u^m)}\pi_u.
\end{equation}
We assume that the outcome depends on the haplotypes and diplotypes only through the genotypes at the target SNP, i.e., $\mbox{pr}(Y_u\mid h_{iu}^{m}h_{ju}^{m},h_{wu}^{c},h_{lu}^{c},X_u) = \mbox{pr}(Y_u\mid g_u^m,g_{mu}^c,g_{pu}^c,X_u)$, which is a function of $\bm{\beta}$. The probability of any possible haplotype combination can be obtained under the assumptions of Mendelian law and Hardy-Weinberg equilibrium for haplotypes, i.e., $\mbox{pr}(h_{iu}^m h_{ju}^m,h_{wu}^c,h_{lu}^c) = \mu_i \mu_j \mu_l$, which is a function of $\bm{\mu}$. The target SNP minor allele frequency $\theta$ is a function of $\bm{\mu}$, so $\mbox{pr}(g_u^m)$ is also a function of $\bm{\mu}$. In the next subsection, we will derive a modified profile likelihood to avoid estimating the high-dimensional nuisance parameters $\bm{\pi}$.
\subsection{An efficient estimator based on a modified profile log-likelihood}
\label{sec:modified}
Unknown parameters can be estimated by maximizing the likelihood function \eqref{equ:semi} under constraints $\mbox{pr}(Y=1)=f$ and $\sum_{u=1}^{n}\pi_u = 1$, or equivalently, maximizing the log-likelihood function
\[\log L(\bm{\Theta},\bm{\pi}) = l_1(\bm{\Theta}) + \sum_{u=1}^{n}\log \pi_u \]
under the same constraints, where $l_1(\bm{\Theta})$ is defined as
\begin{equation}\label{l1}
l_1(\bm{\Theta}) = \sum_{u=1}^{n}\log\sum_{\substack{h_{iu}^{m}h_{ju}^{m}\in{H}_u^m\\(h_{wu}^c,h_{lu}^c)\in{H}_u^c}} \left[\mbox{pr}(Y_u\mid h_{iu}^{m}h_{ju}^{m},h_{wu}^{c},h_{lu}^{c},X_u)\frac{\mbox{pr}(h_{iu}^{m}h_{ju}^{m},h_{wu}^{c},h_{lu}^{c})}{\mbox{pr}(g_u^m)}\right].
\end{equation}
The first constraint can be rewritten as
\begin{equation}
\sum_{u=1}^{n} L_u(\bm{\Theta}) \pi_{u}=f,
\end{equation}
where
\begin{equation}\label{Hu}
L_u(\bm{\Theta}) := \mbox{pr}(Y=1\mid g_u^m,X_u) = \sum_{(g_{mu}^c,g_{pu}^c)} \mbox{pr}(Y=1\mid g^m_u,g_{mu}^c,g_{pu}^c,X_u)\mbox{pr}(g_{pu}^c)\mbox{pr}(g_{mu}^c\mid g^m_u).
\end{equation}
By virtue of the Lagrange multiplier method, the above constrained optimization problem can be shown to be equivalent to maximizing the profile log-likelihood
\begin{equation} \label{equ:profile}
l_p(\bm{\Theta}) = l_0(\bm{\Theta},\lambda_{\bm{\Theta}} ) = l_1(\bm{\Theta}) - \sum_{u=1}^{n} \log \left[n\left\{1+\lambda_{\bm{\Theta}}\left(L_u(\bm{\Theta})-f\right)\right\}\right],
\end{equation}
where $\lambda_{\bm{\Theta}}$ is the solution to the equation
\begin{equation}
\sum_{u=1}^{n} \frac{L_u(\bm{\Theta})-f}{1+\lambda\left(L_u(\bm{\Theta})-f\right)}=0 \end{equation}
with respect to $\lambda$. The derivation of \eqref{equ:profile} is postponed to Appendix A (Supplementary Material).
We can solve the score equations
\begin{equation}
\frac{\partial l_0(\bm{\Theta}, \lambda)}{\partial \lambda}=0 \ \mbox{and}\ \frac{\partial l_0(\bm{\Theta}, \lambda)}{\partial \bm{\Theta}}=0
\end{equation}
to obtain the maximizer of $l_p(\bm{\Theta})$ defined in \eqref{equ:profile}. However, $l_0(\bm{\Theta}, \lambda)$ is not a log-likelihood function, and the solution to the above score equations is a saddle point, as can be shown similarly to \cite{zhang2018efficient}. As in \cite{zhang2018efficient,zhang2020efficient}, to resolve this numerical problem, we propose to modify the profile log-likelihood function \eqref{equ:profile} by replacing the multiplier $\lambda_{\bm{\Theta}}$ with its limiting value
\begin{equation}
\lambda_0 =\frac{n_1}{nf} - \frac{n_0}{n(1-f)}.
\end{equation}
Accordingly, the maximization problem is simplified to maximizing the following modified profile log-likelihood function:
\begin{equation}\label{equ:modified}
l_{\operatorname{mp}}(\bm{\Theta}) =l_1(\bm{\Theta}) - l_2(\bm{\Theta},\lambda_0),
\end{equation}
where
\begin{equation}\label{equ:l2}
l_2(\bm{\Theta},\lambda_0)=\sum_{u=1}^{n} \log \left[n\left\{1+\lambda_0\left(L_u(\bm{\Theta})-f\right)\right\}\right].
\end{equation}
An estimator $\hat\bm{\Theta}_{\operatorname{mp}}$ can be obtained by solving the corresponding score equation
\begin{equation}\label{equ:scoreequation}
\frac{\partial l_{\operatorname{mp}}(\bm{\Theta})}{\partial \bm{\Theta}} = 0.
\end{equation}
In what follows, we establish some large sample properties of $\hat\bm{\Theta}$. Let $\bm{\Theta}_{0}$ denote the true value of $\bm{\Theta}$. Denote $Z = (G^m,G^c,X)$, $W = (g^m,X)$, $p_{Z}(Z ; \bm{\Theta})=\mbox{pr}_{\bm{\Theta}}(Y=1\mid Z)$. We need the following regularity conditions:
\begin{description}
\item[] (C1) The parameter space is $\{(\bm{\beta}, \bm{\mu})\mid \bm{\beta} \in \mathbb{R}^p; \bm{\mu}=(\mu_1, \ldots, \mu_S)^{ \mathrm{\scriptscriptstyle T} } \text{ with }\mu_1, \ldots, \mu_S > 0$ and $\sum_{s=1}^{S} \mu_s = 1\}$, and the true parameter vector $\bm{\Theta}_0$ is an interior point of the parameter space.
\item[] (C2) The expectations
$$E\bigg[\bigg|\frac\partial{\partial \bm{\Theta}} p_{Z}(Z ; \bm{\Theta})\bigg|_{\bm{\Theta}=\bm{\Theta}_{0}}\bigg|\bigg]$$
and
$$E\bigg[\bigg|\frac{\partial}{\partial \bm{\Theta}} \log\mbox{pr}_{\bm{\Theta}}(G^m,G^c\mid W)\bigg|_{\bm{\Theta}=\bm{\Theta}_{0}}\bigg|\bigg]$$
exist.
\item[] (C3) The sample size ratio $n_1/n_0$ is fixed as long as the total sample size $n=n_0+n_1$ varies. That is, there exists a constant $a_0\in(0,1)$ independent of $n$, such that $n_0 = a_0 n$ and $n_1 = a_1 n$, where $a_1=1-a_0$.
\item[] (C4) The expectation $E[|l_{\operatorname{mp}}(\bm{\Theta})|]$ exists for any $\bm{\Theta}$ in some neighborhood of $\bm{\Theta}_0$.
\item[] (C5) Minus the expected Hessian matrix
$$-E\bigg[\bigg|\frac{\partial^{2}l_{\operatorname{mp}}(\bm{\Theta})}{\partial \bm{\Theta} \partial \bm{\Theta}^{ \mathrm{\scriptscriptstyle T} }} \bigg|\bigg]\bigg|_{\bm{\Theta}=\bm{\Theta}_{0}} $$
is positive definite.
\item[] (C6) For any $\bm{\Theta}$ in some neighborhood of $\bm{\Theta}_0$ and the $i$th element $\Theta_i$ and $j$th element $\Theta_j$ of $\bm{\Theta}$, there exists a non-negative random variable $C$ with finite expectation, such that
$$\bigg|\frac{\partial^{2}l_{\operatorname{mp}}(\bm{\Theta})}{\partial \Theta_i \partial \Theta_j} - \frac{\partial^{2} l_{\operatorname{mp}}(\bm{\Theta})}{\partial \Theta_i \partial \Theta_j}\bigg|_{\bm{\Theta}=\bm{\Theta}_{0}}\bigg| \le C\|\bm{\Theta} - \bm{\Theta}_0\|.$$
\item[] (C7) The so called ``Fisher information matrix''
$$\operatorname{cov}\bigg(\frac{\partial l_{\operatorname{mp}}(\bm{\Theta})}{\partial \bm{\Theta}}\bigg)\bigg|_{\bm{\Theta}=\bm{\Theta}_{0}}$$
exists.
\item[] (C8) The following technical condition holds:
$$E\bigg[\frac{\partial^{2} l_0(\bm{\Theta}, \lambda)}{\partial \bm{\Theta} \partial \lambda^{{ \mathrm{\scriptscriptstyle T} }}}\bigg]\bigg|_{\bm{\Theta}=\bm{\Theta}_{0},\lambda=\lambda_{0}}=0.$$
\end{description}
In what follows, we first show in Lemma \ref{lemma1} that the function $l_{\operatorname{mp}}(\bm{\Theta})$ shares a property with the conventional likelihood score, then present our main asymptotic results in Theorem \ref{theorem1}. Refer to Appendices B and C (Supplementary Material) for proofs of Lemma \ref{lemma1} and Theorem \ref{theorem1}.
\begin{lemma}
\label{lemma1}
Under the regularity condition (C2), we have
$$
\left.E\left[\frac{\partial l_{\operatorname{mp}}(\bm{\Theta})}{\partial \bm{\Theta}}\right]\right|_{\bm{\Theta}=\bm{\Theta}_{0}}=0.
$$
\end{lemma}
\begin{theorem}\label{theorem1}
As the total sample size $n$ goes to infinity, we have the following asymptotic properties:
\begin{description}
\item [](i) Under condition (C1), with probability tending to 1, $l_{\operatorname{mp}}(\bm{\Theta})$ takes the value of negative infinity on the boundary of the parameter space, and
there exists a solution to \eqref{equ:scoreequation}, denoted by $\hat{\bm{\Theta}}_{\operatorname{mp}}$, in the interior of the parameter space.
\item [](ii) Under conditions (C1)-(C5), $\hat{\bm{\Theta}}_{\operatorname{mp}}$ converges in probability to $\bm{\Theta}_0$.
\item [](iii) Under conditions (C1)-(C7),
$$n^{1/2}(\hat{\bm{\Theta}}_{\operatorname{mp}}-\bm{\Theta}_0)\rightarrow N\left(\bf{0},A_{\operatorname{mp}}^{-1}(\bm{\Theta}_{0}) \Sigma_{\operatorname{mp}}\left(\bm{\Theta}_{0}\right) A_{\operatorname{mp}}^{-1}\left(\bm{\Theta}_{0}\right)\right)\text{ in distribution},$$
where
$$A_{\operatorname{mp}}(\bm{\Theta}_{0})=\frac1n E\bigg[\frac{\partial^{2} l_{\operatorname{mp}}(\bm{\Theta})}{\partial \bm{\Theta} \partial \bm{\Theta}^{ \mathrm{\scriptscriptstyle T} }}\bigg]\bigg|_{\bm{\Theta}=\bm{\Theta}_{0}}$$
and
$$\Sigma_{\operatorname{mp}}(\bm{\Theta}_{0})=\frac1n \operatorname{cov}\bigg( \frac{{\partial l_{\operatorname{mp}}(\bm{\Theta})}}{\partial \bm{\Theta}}\bigg)\bigg|_{\bm{\Theta}=\bm{\Theta}_{0}}.$$
\item [] (iv) Under conditions (C1)-(C8), $\hat{\bm{\Theta}}_{\operatorname{mp}}$ is asymptotic equivalent to the conventional maximum likelihood estimator, which means that $\hat{\bm{\Theta}}_{\operatorname{mp}}$ is asymptotically efficient in the sense of \cite{bickel1993efficient}.
\end{description}
\end{theorem}
Note that $A_{\operatorname{mp}}(\bm{\Theta}_{0})$ can be consistently estimated by
$$\hat{A}_{\operatorname{mp}}(\hat{\bm{\Theta}}_{\operatorname{mp}}) := \frac1n\frac{\partial^{2} l_{\mathrm{\operatorname{mp}}}(\bm{\Theta})}{\partial \bm{\Theta} \partial \bm{\Theta}^{{ \mathrm{\scriptscriptstyle T} }}}\Big|_{\bm{\Theta}=\hat{\bm{\Theta}}_{\operatorname{mp}}},$$
and $\Sigma_{\operatorname{mp}}(\bm{\Theta}_{0})$ can be consistently estimated by a weighted average of two sample variance-covariance matrices corresponding to the subgroups of cases and controls. Consequently, $n^{1/2}(\hat{\bm{\Theta}}_{\operatorname{mp}}-\bm{\Theta}_0)$ is approximately multivariate normal with zero expectation and variance-covariance matrix $\hat{A}_{\operatorname{mp}}^{-1}(\hat{\bm{\Theta}}_{\operatorname{mp}}) \hat{\Sigma}_{\operatorname{mp}}(\hat{\bm{\Theta}}_{\operatorname{mp}})\hat{A}_{\operatorname{mp}}^{-1}(\hat{\bm{\Theta}}_{\operatorname{mp}})$.
Confidence intervals and significance tests of $\bm{\Theta}$ can be constructed based on this normality approximation.
\begin{remark}
Condition (C2) is a sufficient condition for the existence of $E[\partial l_{\operatorname{mp}}(\bm{\Theta}_0) / \partial \bm{\Theta}]$, which is similar to that for the conventional likelihood function. Condition (C3) is commonly used to establish large sample properties for case-control studies \citep{bickel1993efficient}. Conditions (C4)-(C7) ensure the consistency and asymptotic normality of $\hat{\bm{\Theta}}_{\operatorname{mp}}$, which are also similar to those for the conventional likelihood function.
\end{remark}
\begin{remark}
Condition (C8) is used in the proof of the asymptotic equivalence of $\hat{\bm{\Theta}}_{\operatorname{mp}}$ and the asymptotically efficient estimator (i.e., the maximizer of the profile likelihood function \eqref{equ:profile}), so that $\hat{\bm{\Theta}}_{\operatorname{mp}}$ is also asymptotically efficient. We demonstrated that this condition is nearly satisfied in some simulation situations, refer to Appendix D and Table S1 (Supplementary Material) for details.
\end{remark}
\subsection{An expectation-maximization algorithm for the modified profile log-likelihood}\label{EMalgorithm}
Note that the profile log-likelihood \eqref{equ:modified} has a complex form, whose maximization is computationally intensive. By treating haplotypes as missing data, the maximization of \eqref{equ:modified} can be effectively simplified by applying an expectation-maximization algorithm. Specifically, the expectation step is to calculate the conditional expectation of the modified profile log-likelihood function given the observed data (genotype data, disease status, and covariates) under initial parameters or those estimated in the previous step, and the maximization step is to maximize the resultant conditional expectation with respect to unknown parameters. This expectation-maximization algorithm is detailed as follows.
Let $\bm{\Theta}^{(r)}$ denote the estimator of $\bm{\Theta}$ obtained at the $r$th iteration of the expectation-maximization algorithm, and let $\mathcal{D}$ denote all observed data, $\mathcal{H}$ denote the missing haplotype data. Then, the modified profile log-likelihood for the complete data ($\mathcal{D}$, $\mathcal{H}$) is
\begin{equation}\label{equ:lmpc}
l_{\operatorname{mp}}(\bm{\Theta};\mathcal{D},\mathcal{H})=l_1(\bm{\Theta};\mathcal{D},\mathcal{H})-l_2(\bm{\Theta},\lambda_0),
\end{equation}
where
\begin{equation} \label{equ:EMcom}
l_{1}(\bm{\Theta};\mathcal{D},\mathcal{H}=h) = \sum_{u=1}^{n} \log \mbox{pr}(Y_{u}, G_{u}^{c}, G_{u}^{m},h_u\mid g_u^m, X_{u}) = \sum_{u=1}^{n}\log\bigg[\mbox{pr}(Y_u\mid X_u,h_u)\frac{\mbox{pr}(h_u)}{\mbox{pr}(g_u^m)}\bigg],
\end{equation}
and $h_u$ is the haplotype vector for the $u$th family, which is compatible with the observed joint mother-child genotypes under some regularity conditions such as tight linkage, random mating, and Mendelian inheritance.
Noting that $L_u(\bm{\Theta})$ does not depend on the missing haplotype data, we have $E_{\bm{\Theta}^{(r)}}[l_2(\bm{\Theta},\lambda_0)\mid \mathcal{D}]=l_2(\bm{\Theta},\lambda_0)$. Consequently, in the $(r+1)$th iteration of the expectation-maximization algorithm, the expectation step is to compute
\begin{equation} \label{equ:estep}
E_{\bm{\Theta}^{(r)}}[l_{\operatorname{mp}}(\bm{\Theta};\mathcal{D},\mathcal{H})\mid \mathcal{D}] = E_{\bm{\Theta}^{(r)}}[l_{1}(\bm{\Theta};\mathcal{D},\mathcal{H})\mid\mathcal{D}] - l_2(\bm{\Theta},\lambda_0),
\end{equation}
where
$$E_{\bm{\Theta}^{(r)}}[l_{1}(\bm{\Theta};\mathcal{D},\mathcal{H})\mid\mathcal{D}]=\sum_{u=1}^n \sum_{h_u} pr_{\bm{\Theta}^{(r)}}(h_u\mid Y_u,X_u,G_u^m,G_u^c)\log\bigg[\mbox{pr}(Y_u\mid X_u,h_u)\frac{\mbox{pr}(h_u)}{\mbox{pr}(g_u^m)}\bigg].$$
The maximization step in the $(r+1)$th iteration of the expectation-maximization algorithm is to maximize \eqref{equ:estep} with respect to $\bm{\Theta}$ using any non-linear optimization algorithm.
The above expectation-maximization algorithm is different from the conventional expectation-maximization algorithm since our objective function is a modified profile log-likelihood function instead of a log-likelihood function. Nevertheless, we still have the following convergence result for our expectation-maximization algorithm (refer to Appendix E in Supplementary Material for a proof):
\begin{theorem} \label{theorem2}
If $l_{mp}(\bm{\Theta}_1) \neq l_{mp}(\bm{\Theta}_2)$ for any two stationary points $\bm{\Theta}_1$ and $\bm{\Theta}_2$ of the modified profile log-likelihood, then $\bm{\Theta}^{(r)}$ defined below \eqref{equ:l2} converges to a stationary point of the modified profile log-likelihood as $r\to\infty$.
\end{theorem}
A starting value $\bm{\Theta}^{(0)}$ is required for the proposed expectation-maximization algorithm. In practice, we propose to use a prospective logistic regression to be introduced in Section~\ref{s:simu} to obtain a reasonably good starting value $\bm{\Theta}^{(0)}$, which can speed up the convergence of the expectation-maximization algorithm. The maximization step involves unknown parameters $\bm{\Theta}=(\bm{\beta},\bm{\mu})$. To simplify the maximization problem, we propose to update only a part of parameters ($\bm{\beta}$ or $\bm{\mu}$) at a time \citep{meng1993maximum}.
Our proposed method can be naturally extended to handle missing child genotype data for a subset of families by slightly modifying \eqref{equ:modified}. That is, we allow the haplotype combinatorial sets $H_u^m$ and $H_u^c$ in \eqref{l1} to be compatible with the observed maternal genotype data only.
\subsection{Implementation of the proposed method}
We outline our inference procedure as follows.
\begin{enumerate}
\item [(1)] Specify the prevalence $f$. The prevalence can be either estimated from extra cohort studies or extracted from public databases. As demonstrated in our simulation study in Section~\ref{s:simu}, misspecifying $f$ only affects the estimation of intercept parameter $\beta_0$ and does not affect the other parameters of our interest.
\item [(2)] Specify possible haplotypes associated with the target SNP and available adjacent SNPs, which can be obtained by any existing method such as the one implemented in the function \texttt{haplo.em} in the \texttt{R} package \texttt{haplo.stats}.
\item [(3)] Specify an initial value of haplotype frequency vector $\bm{\mu}$. The outputs of the \texttt{R} function \texttt{haplo.em} include an estimate of $\bm{\mu}$, which can serve as an initial value.
\item [(4)] Specify an initial value of regression parameter vector $\bm{\beta}$. A prospective logistic
regression method to be introduced in Section~\ref{s:simu} can be applied to produce an estimate of $\beta$, which can serve as an initial value of $\bm{\beta}$.
\item [(5)] Apply the expectation-maximization algorithm described in Section \ref{EMalgorithm} to estimate both $\bm{\mu}$ and $\bm{\beta}$, with the initial values being specified in steps (3) and (4), and estimate the variance matrix of $\hat{\bm{\beta}}$ as described in Section~\ref{sec:modified}.
\item [(6)] Obtain confidence intervals (z-intervals) and p-values of significance test (z-tests) for $\bm{\beta}$.
\end{enumerate}
For convenience, we refer to our robust method based on the modified profile log-likelihood function as ROB-HAP. Here, ROB is the shorthand for robust, emphasizing that our method is robust in the sense that it does not impose any constraint on the relationship between the maternal genotype and the covariates; HAP is the shorthand for haplotype, emphasizing that our method uses haplotypes to determine the parental origins.
Our proposed method ROB-HAP has been implemented in an \texttt{R} function \texttt{assessPOE}, which is available at \url{https://github.com/yatian20/APOE}. The inputs of \texttt{assessPOE} include the case-control status $Y$, maternal covariates $X$, genotypes of mother-child pairs $(G^m,G^c)$, prevalence $f$, and target SNP ID. The outputs of \texttt{assessPOE} include point estimates, confidence intervals, and significance test p-values for $\bm{\beta}$. In our preliminary numerical study, we found that it was quite computationally intensive to run our expectation-maximization algorithm when many SNPs were involved. To speed up the computation, users can adopt the \texttt{R} package \texttt{LDheatmap} \citep{shin2006ldheatmap} to select a limited number of SNPs with strong linkage disequilibrium among the candidate SNPs.
\section{Simulation Studies}\label{s:simu}
\subsection{Data generation} \label{GENdata}
Simulation studies were conducted to evaluate the performance of our proposed method ROB-HAP. The general data generation process is described in this subsection, then specific parameter settings are given in the subsequent subsections. We generated genotype data for five SNPs using seven identified haplotypes, which were extracted from published haplotype data in the genomic region GPX1. The configurations and frequencies of the haplotypes can be found in Table 1 of \cite{chen2004haplotype}. Diplotypes were then generated for $N = 10^7$ mother-child pairs under Hardy-Weinberg equilibrium. Specifically, three haplotypes $h_i^m$, $h_j^m$, $h_l^c$ were randomly sampled for each mother-child pair according to the distribution given in Table S2, then $h_w^c$ was generated according to the Mendelian law. Maternal genotype $g^m$ at the target SNP was extracted from the maternal diplotype. Environmental risk factor $X$ was generated through the following linear model:
\begin{equation} \label{equ:linear model}
X = \eta\{g^m - E(g^m)\} + e,
\end{equation}
where $E(g^m)$ was the expectation of $g^m$, the random error $e$ was independent of $g^m$ and followed the standard normal distribution, and $\eta$ was a regression parameter. The intercept $\beta_0$ was determined by the above distributions, the prevalence of phenotype and log-odds ratio (log-OR) parameters. Unless otherwise specified, we fixed $\eta = \log(3.0)$ in model \eqref{equ:linear model} and $\beta_{g^m} = \log(1.8)$, $\beta_{g^c} = \log(1.5)$, $\beta_{im} = \log(1.5)$, $\beta_{X} = \log(1.8)$, and $\beta_{g^mX}=\beta_{g^cX}=0$ in model \eqref{equ:penetrance1}. Hardy-Weinberg equilibrium was adopted in generating diplotypes unless otherwise specified. The phenotypes $Y$'s were generated according to the penetrance model \eqref{equ:penetrance} together with \eqref{equ:penetrance1}. Data were randomly sampled for $n_0 = 200$ control families and $n_1 = 200$ case families with $Y=0$ and $Y=1$, respectively.
\subsection{Considered methods}
For the comparison purpose, we considered several alternative methods to ROB-HAP. The first alternative method is similar to ROB-HAP but it directly uses the parental origin information that could not be completely inferred in practice, which is referred to as ROB-COM (COM is the shorthand for complete). The second alternative method is a special version of ROB-HAP that only uses the genotypes from the target SNP instead of all available SNPs to infer the parental origins, which is referred to as ROB-SNP. The third alternative method is a simplified version of ROB-HAP by further incorporating independence assumption between maternal genotype and covariates, which is referred to as IND-HAP (IND is the shorthand for independent). In terms of inference efficiency, we can intuitively rank the four methods mentioned so far: ROB-COM $>$ IND-HAP $>$ ROB-HAP $>$ ROB-SNP. However, one should be aware that ROB-COM is not applicable in practice if the parental origins cannot be unambiguously inferred for some families and IND-HAP could produce biased results if the independence assumption is violated. We also considered the prospective logistic regression, which is referred to as LOGIT-HAP (the fourth alternative method). In LOGIT-HAP, first, the joint diplotypes of mother-child pairs are inferred using the function \texttt{haplo.em} in the \texttt{R} package \texttt{haplo.stats} with the adjacent SNP genotypes, then, the parental origins are determined by the joint diplotypes and only those families with uniquely inferred parental origins are kept. Since LOGIT-HAP is based on a prospective likelihood, it is not efficient by ignoring the information on Hardy-Weinberg equilibrium, Mendelian law between maternal genotype and child genotype, and the conditional independence between covariates and child genotype given maternal genotype. Finally, we considered two existing methods in the literature, i.e., the fifth alternative method P-HAP \citep{lin2013multi} and the sixth alternative method EMIM \citep{ainsworth2011investigation}. Both P-HAP and EMIM can evaluate parental origin effects using case-control mother-child pair data but neither of them can incorporate covariate information. The \texttt{R} package HAPLIN can also be used to detect parental origin effects, which is based on a log-linear model \citep{gjessing2006case}. It has been shown that EMIM and Haplin have comparable powers in detecting genetic effects through extensive simulation studies \citep{gjerdevik2019haplin}, and EMIM is more often adopted due to its computational advantage. We therefore do not show the simulation results for HAPLIN in the next sections.
\subsection{Comparison of prospective and retrospective likelihoods}\label{simu1}
We conducted a simulation study to demonstrate the advantage of the retrospective likelihood (ROB-HAP) over the prospective likelihood (LOGIT-HAP). The third of the five SNPs was set as the target SNP unless specified otherwise. Estimation results based on 2,500 simulated datasets are summarized in the upper panel of Table \ref{tab1}.
\begin{table}
\caption{
Estimation results with $g^m$ and $X$ being correlated ($\eta = \log 3$; $f = 0.01$)
}\label{tab1}
\centering
\begin{threeparttable}
\begin{tabular*}{\hsize}{@{}@{\extracolsep{\fill}}ccccccccccc@{}}
\hline
& & \multicolumn{4}{c}{ROB-HAP$^a$} & & \multicolumn{4}{c}{LOGIT-HAP$^b$} \\ \cline{3-6} \cline{8-11}
Log-OR & True$^c$ & Bias$^d$ & SE$^e$ & SEE$^f$ & CP$^g$ & & Bias$^d$ & SE$^e$ & SEE$^f$ & CP$^g$ \\ \hline
$\beta_{g^m}$ & 0.588 & 0.017 & 0.259 & 0.261 & 0.955 & & 0.029 & 0.304 & 0.304 & 0.951 \\
$\beta_{g^c}$ & 0.405 & -0.012 & 0.177 & 0.174 & 0.948 & & 0.018 & 0.245 & 0.237 & 0.945 \\
$\beta_{im}$ & 0.405 & -0.003 & 0.188 & 0.187 & 0.949 & & -0.013 & 0.270 & 0.265 & 0.952 \\
$\beta_X$ & 0.588 & 0.011 & 0.124 & 0.124 & 0.953 & & 0.017 & 0.126 & 0.125 & 0.951 \\
\hline
& & \multicolumn{4}{c}{ROB-SNP$^i$} & & \multicolumn{4}{c}{ROB-COM$^j$} \\
\cline{3-6} \cline{8-11}
Log-OR & True$^c$ & Bias$^d$ & SE$^e$ & SEE$^f$ & CP$^g$ & & Bias$^d$ & SE$^e$ & SEE$^f$ & CP$^g$ \\ \hline
$\beta_{g^m}$ & 0.588 & -0.009 & 0.266 & 0.270 & 0.958 & & 0.006 & 0.256 & 0.257 & 0.953 \\
$\beta_{g^c}$ & 0.405 & 0.002 & 0.179 & 0.178 & 0.953 & & 0.002 & 0.176 & 0.172 & 0.946 \\
$\beta_{im}$ & 0.405 & 0.011 & 0.216 & 0.218 & 0.954 & & 0.006 & 0.174 & 0.172 & 0.951 \\
$\beta_X$ & 0.588 & 0.007 & 0.124 & 0.125 & 0.957 & & 0.011 & 0.126 & 0.125 & 0.947 \\
\hline
\end{tabular*}
\begin{tablenotes}
\item $^a$Our proposed method using haplotypes; $^b$the conventional logistic regression method using haplotypes; $^c$the true value of the log-OR parameter; $^d$difference between the mean estimate and true parameter value; $^e$empirical standard error; $^f$mean estimated standard error; $^g$empirical coverage probability of the 95$\%$ confidence intervals; $^i$a simplified verion of ROB-HAP only using the genotypes at the target SNP; $^j$an ideal method utilizing the parental origin information that could not be completely inferred in practice.
\end{tablenotes}
\end{threeparttable}
\end{table}
Evidently, both ROB-HAP and LOGIT-HAP were virtually unbiased, though the bias of LOGIT-HAP was slightly larger. The empirical standard errors (SEs) appeared to be close to the averages of estimated SEs (SEEs). ROB-HAP was uniformly more efficient than LOGIT-HAP in terms of SEs. For example, the SE of $\hat{\beta}_{im}$ was $30\%$ lower than that of LOGIT-HAP. Panel A of Fig. \ref{fig1} showcases the powers and type-I error rates for testing parental origin effects, where each of the five SNPs was treated in turn as the target SNP while the genotypes from other SNPs were used to infer the parental origins. Both LOGIT-HAP and ROB-HAP had perfect controls of type-I error rates. ROB-HAP were uniformly more powerful than LOGIT-HAP, with power gains ranging from 0.21 to 0.28.
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{figure11.eps}
\caption{Panel A: The type-I error rates and powers for testing parental origin effects with confounding covariate $X$ incorporated ($\eta=\log3$). Various target SNPs in gene GPX1 were considered. Data were generated under $\beta_{im} = 0$ (for assessing type-I error rates) or $\beta_{im} = \log1.5$ (for assessing powers). The other underlying parameters were $f = 0.01$, $\beta_{g^m} = \beta_{X} = \log1.8$, and $\beta_{g^c} = \log1.5$. Panel B: The type-I error rate and power trends against the length of haplotypes for testing parental origin effects. The target SNP was SNP1 in gene GPX1. Here Inf corresponds to ROB-COM with the true parent origins being used. Data were generated under $\beta_{im} = 0$ (for assessing the type-I error rates) or $\beta_{im} = \log1.5$ (for assessing the powers). The other underlying parameters were $f = 0.01$, $\eta = \log3$, $\beta_{g^m} = \beta_{X} = \log1.8$, and $\beta_{g^c} = \log1.5$.}
\label{fig1}
\end{figure}
\subsection{The impact of parental origins on the power of parental origin effect detection}\label{simu2}
Determining the parental origins is a key step in testing parental origin effects. We evaluated the power improvement of the proposed method ROB-HAP over ROB-SNP, where ROB-SNP only uses the target SNP to infer the parental origins. We also included ROB-COM for comparison, which uses the true parental origin information that could not be completely inferred in practice. ROB-SNP can be regarded as a special case of ROB-HAP with the number of adjacent SNPs being zero.
Results based on 2,500 simulated datasets are summarized in the lower panel of Table \ref{tab1}.
Both ROB-SNP and ROB-COM appeared to be unbiased. As expected, ROB-HAP was more efficient than ROB-SNP, especially for the parental origin effects. In fact, the SE of $\hat{\beta}_{im}$ by ROB-HAP was $13\%$ lower compared with ROB-SNP. Of course, ROB-COM was most efficient, with a further SE reduction of $6\%$. Efficiency gain of ROB-HAP over ROB-SNP could depend on the number of adjacent SNPs, minor allele frequencies, and so on. As shown in Panel B of Fig. \ref{fig1}, the power for testing parental origin effects steadily increased with the length of haplotypes. LOGIT-HAP had the same power trend, but it was uniformly inferior to ROB-HAP.
\subsection{Comparison with existing methods that do not incorporate covariates}\label{simu3}
There are quite a few methods in the literature for testing parental origin effects, but none of them can incorporate covariates in a general way. Among these methods, P-HAP \citep{lin2013multi} can use haplotype information to improve estimation/test efficiency for parental origin effects; EMIM \citep{ainsworth2011investigation} has been widely used in testing maternal, imprinting, and interaction effects. We conducted a simulation study to compare their performances with our proposed method ROB-HAP. Because P-HAP and EMIM cannot incorporate covariates, they could be invalid if the covariates are associated with both maternal genotypes and the phenotype of interest. We also considered a slightly modified version of ROB-HAP, which imposes the independence constraint between the covariates and maternal genotypes and is referred to as IND-HAP.
Simulation data were generated as in Section \ref{simu1}, except that the covariate was independent of the maternal genotypes (i.e., $\eta=0$). The estimation results for IND-HAP and P-HAP are summarized in Table \ref{tab3}. Because EMIM was derived under a different model and the estimated parameters were different from those in model \eqref{equ:penetrance}, we only show the powers and type-I errors of EMIM with respect to the parental origin effects, but not its estimation results. Evidently, both IND-HAP and P-HAP were virtually unbiased, the estimated standard errors were close to the empirical ones, and the 95\% confidence intervals had coverage probabilities close to the nominal level. IND-HAP and P-HAP appeared to be comparable in terms of SEs.
The type-I error rates and powers for testing parental origin effects are presented in Panel A of Fig. \ref{fig2} for IND-HAP, ROB-HAP, P-HAP, and EMIM.
The type-I error rates were well controlled by all the four methods. IND-HAP and P-HAP appeared to have comparable powers, EMIM was least powerful, and ROB-HAP was intermediate.
\begin{table}
\centering
\caption{Estimation results with $g^m$ and $X$ being independent ($\eta = 0$, $f = 0.01$)}\label{tab3}
\begin{threeparttable}
\begin{tabular*}{\hsize}{@{}@{\extracolsep{\fill}}ccccccccccc@{}}
\hline
& & \multicolumn{4}{c}{IND-HAP$^a$} & & \multicolumn{4}{c}{P-HAP$^b$} \\ \cline{3-6} \cline{8-11}
Log-OR & True$^c$ & Bias$^d$ & SE$^e$ & SEE$^f$ & CP$^g$ & & Bias$^d$ & SE$^e$ & SEE$^f$ & CP$^g$ \\ \hline
$\beta_{g^m}$ & 0.588 & -0.015 & 0.196 & 0.200 & 0.954 & & 0.018 & 0.199 & 0.198 & 0.948 \\
$\beta_{g^c}$ & 0.405 & 0.013 & 0.153 & 0.153 & 0.947 & & 0.009 & 0.155 & 0.153 & 0.946 \\
$\beta_{im}$ & 0.405 & -0.006 & 0.162 & 0.166 & 0.954 & & -0.022 & 0.168 & 0.166 & 0.942 \\
$\beta_X$ & 0.588 & 0.014 & 0.108 & 0.111 & 0.947 & & - $^h$ & - & - & - \\ \hline
\end{tabular*}
\begin{tablenotes}
\item $^a$A simplified version of ROB-HAP with independence assumption between $X$ and $g^m$; $^b$a parental origin effect testing method without incorporating covariates \citep{lin2013multi}; $^c$true value of the log-OR parameter; $^d$difference between the mean estimate and true parameter value; $^e$empirical standard error; $^f$mean estimated standard error; $^g$empirical coverage probability of the 95$\%$ confidence interval; $^h$covariates not incorporated in P-HAP.
\end{tablenotes}
\end{threeparttable}
\end{table}
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{figure21.eps}
\caption{Panel A: The type-I error rates (OR = 1) and powers (OR = 1.5) for testing parental origin effects with non-confounding covariate $X$ ($\eta=0$). The other underlying parameters were $f = 0.01$, $\beta_{g^m} = \beta_{X} = \log1.8$, and $\beta_{g^c} = \log1.5$. Panel B: Type-I error rates for testing parental origin effects in the presence of a confounding covariate $X$. The target SNP was SNP1 in gene GPX1. Data were generated under a non-linear model relating the covariate $X$ and maternal genotype $g^m$: $X = \eta \{(g^m)^2 - E(g^m)^2\} + e$, where $e$ was standard normal. The other underlying parameters were $f = 0.01$, $\beta_{im} = 0$, $\beta_{g^m} = \beta_{X} = \log1.8$, and $\beta_{g^c} = \log1.5$.}
\label{fig2}
\end{figure}
When $g^m$ and $X$ was correlated ($\eta\not=0$), the three methods not incorporating $X$ or assuming $g^m$-$X$ independence could be either liberal or conservative, depending on $\eta<0$ or $\eta>0$ (Panel B of Fig. \ref{fig2}).
\subsection{Sensitivity analyses with model misspecification}
Some model assumptions are required for ROB-HAP, such as Hardy-Weinberg equilibrium for haplotypes, penetrance model, and disease prevalence. We conducted sensitivity analyses to examine the robustness of ROB-HAP with respect to each of these model assumptions.
First, we assessed the robustness of ROB-HAP with respect to Hardy-Weinberg equilibrium. We generated data under model
\[
\mbox{pr}(h_{i}h_{j})= \left\{\begin{array}{cl}
F \mu_i+(1-F) \mu_i^{2}, & \text { for } i=j ,\\
2(1-F) \mu_i \mu_j, & \text { for } i \neq j,
\end{array}\right.
\]
where $F$ is a fixation index and $h_{i}h_{j}$ is an unordered diplotype. As shown in Table S3 (Supplementary Material), ROB-HAP had well controlled type-I error rates even when Hardy-Weinberg equilibrium was seriously violated ($F>0.5$). This was not surprising since Hardy-Weinberg equilibrium is required only in the ratio of $\mbox{pr}(h_{iu}^{m}h_{ju}^{m},h_{wu}^{c},h_{lu}^{c})$ to $\mbox{pr}(g^m_u)$ in the modified profile likelihood \eqref{equ:modified}, which is partially free of the fixation index $F$. In fact, when $g_u^m = 1$, we have
\[
\frac{\mbox{pr}(h_{iu}^{m}h_{ju}^{m},h_{wu}^{c},h_{lu}^{c})}{ \mbox{pr}(g_u^m)} = \frac{(1-F)\mu_i \mu_j \mu_l }{2(1-F)\theta (1-\theta)} = \frac{\mu_i \mu_j \mu_l }{2\theta (1-\theta)},
\]
which is free of the fixation index $F$; when $g_u^m \neq 1$, the parental origins can be directly derived from the genotypes of each mother-child pair. Therefore, the modified profile likelihood function \eqref{equ:modified} is only weakly dependent of $F$.
Second, we studied the impact of misspecifying the penetrance model on the performance of ROB-HAP. Data were generated under the penetrance model \eqref{equ:penetrance} with $\beta_{g^m X} = \beta_{g^c X} = \log1.2$ or $\beta_{g^m X} = \beta_{g^c X} = \log1.5$. The resultant underlying model had an interaction effect, but it was ignored in ROB-HAP. By doing so, we could assess whether or not ROB-HAP is sensitive to the presence of interaction effects. As shown in Table \ref{tab4}, the estimation of $\beta_{g^m}$, $\beta_{g^c}$, and $\beta_{X}$ were biased, but this was not the case for $\beta_{im}$. This indicates that ROB-HAP is robust to the penetrance model misspecification to some extent when assessing the parental origin effects.
\begin{table}
\centering
\caption{Estimation results of ROB-HAP with the penetrance model being misclassified $(f = 0.01, \eta = \log3)$}\label{tab4}
\begin{threeparttable}
\begin{tabular*}{\hsize}{@{}@{\extracolsep{\fill}}ccccccccccc@{}}
\hline
& & \multicolumn{4}{c}{$\beta_{g^mX}=\beta_{g^cX}=\log1.2$$^a$} & & \multicolumn{4}{c}{$\beta_{g^mX}=\beta_{g^cX}=\log1.5$$^a$} \\ \cline{3-6} \cline{8-11}
Log-OR & True$^b$ & Bias$^c$ & SE $^d$ & SEE $^e$ & CP$^f$ & & Bias$^c$ & SE $^d$ & SEE$^e$ & CP$^f$ \\ \hline
$\beta_{g^m}$ & 0.588 & 0.143 & 0.298 & 0.298 & 0.931 & & 0.418 & 0.386 & 0.387 & 0.854 \\
$\beta_{g^c}$ & 0.405 & 0.245 & 0.187 & 0.185 & 0.744 & & 0.699 & 0.225 & 0.223 & 0.100 \\
$\beta_{im}$ & 0.405 & -0.012 & 0.191 & 0.197 & 0.957 & & 0.001 & 0.231 & 0.239 & 0.960 \\
$\beta_X$ & 0.588 & 0.272 & 0.150 & 0.150 & 0.576 & & 0.710 & 0.201 & 0.203 & 0.024 \\ \hline
\end{tabular*}
\begin{tablenotes}
\item $^a$With a nonzero interaction effect but ignored in ROB-HAP fitted model; $^b$true parameter value; $^c$difference between the mean estimate and true parameter value; $^d$empirical standard error; $^e$mean estimated standard error; $^f$empirical coverage probability of the 95$\%$ confidence interval.
\end{tablenotes}
\end{threeparttable}
\end{table}
Finally, we assessed the robustness of ROB-HAP with respect to the disease prevalence $f$. We generated data with $f=0.01$ but misspecified it to be 0.002 or 0.05. The estimation results shown in Table \ref{tab5} demonstrate that the misspecification had little impact on the odds-ratio parameters of interest. Such robustness was also observed previously \citep{zhang2020efficient}.
\begin{table}
\centering
\caption{Estimation results of ROB-HAP with the prevalence being misclassified $(f = 0.01, \eta = \log3)$}\label{tab5}
\begin{threeparttable}
\begin{tabular*}{\hsize}{@{}@{\extracolsep{\fill}}ccccccccccc@{}}
\hline
& & \multicolumn{4}{c}{Misspecified $f = 0.002^a$} & & \multicolumn{4}{c}{Misspecified $f = 0.05$$^a$} \\ \cline{3-6} \cline{8-11}
Log-OR & True$^b$ & Bias$^c$ & SE$^d$ & SEE$^e$ & CP$^f$ & & Bias$^c$ & SE$^d$ & SEE$^e$ & CP$^f$ \\ \hline
$\beta_{g^m}$ & 0.588 & 0.023 & 0.260 & 0.260 & 0.946 & & -0.009 & 0.270 & 0.267 & 0.945 \\
$\beta_{g^c}$ & 0.405 & -0.005 & 0.180 & 0.170 & 0.934 & & 0.023 & 0.187 & 0.187 & 0.952 \\
$\beta_{im}$ & 0.405 & -0.011 & 0.177 & 0.183 & 0.954 & & 0.016 & 0.205 & 0.203 & 0.950 \\
$\beta_X$ & 0.588 & 0.011 & 0.125 & 0.125 & 0.952 & & 0.001 & 0.123 & 0.125 & 0.957 \\ \hline
\end{tabular*}
\begin{tablenotes}
\item $^a$Misspecified prevalence in ROB-HAP; $^b$true parameter value; $^c$difference between the mean estimate and true parameter value; $^d$empirical standard error; $^e$mean estimated standard error; $^f$empirical coverage probability of the 95$\%$ confidence intervals.
\end{tablenotes}
\end{threeparttable}
\end{table}
\subsection{Handling missing child genotypes}\label{sim:missing}
In real data analyses, some individuals could be subject to missing genotypes. In our method ROB-HAP, the children missing genotypes can be naturally handled through an expectation-maximization algorithm. A na\"{i}ve approach to handling missing genotypes is to simply ignore those mother-child pairs with missing genotype at any SNP. The efficiency loss of this na\"{i}ve approach could be considerable if the number of SNPs is large. A small-scale simulation study was conducted to demonstrate the efficiency improvement of incorporating mother-child pairs with missing child genotypes. Simulation data were generated in the same way as in Section \ref{GENdata}, and the genotype of each SNP was then randomly dropped with a probability of 0.01. Three versions of ROB-HAP were applied to the simulated data. The first version (ROB-HAP-com) used complete genotypes, the second version (ROB-HAP-EM) incorporated missing genotypes through the expectation-maximization algorithm, and the third version (ROB-HAP-del) simply deleted those mother-child pairs with missing genotypes. Estimation results for parental origin effects by the three versions are graphically displayed in Fig. S1 (Supplementary Material). The three versions were shown to be virtually unbiased. ROB-HAP-EM and ROB-HAP-com were nearly comparable, and they had smaller standard errors than ROB-HAP-del. These results evidently demonstrate that incorporating mother-child pairs with missing data can significantly improve the estimation efficiency.
\section{Application to the Jerusalem Perinatal Study}
\label{s:apply}
We applied the considered methods for assessing the parental origin effects in the gene \emph{PPARGC1A} on low birth weight in the Jerusalem Perinatal Study, a historical birth cohort study including 17,003 births in Jerusalem residents from 1974 to 1976. Among 8,238 eligible children with mothers' prepregnancy body mass index (pp-BMI) smaller than 25, there were 297 cases of low birth weight (below 2.5 kg). Genotype data were obtained for 691 eligible mother-child pairs (125 case pairs and 566 control pairs). We incorporated pp-BMI in applicable methods as a covariate because it was shown to be strongly associated with infant nutrition during pregnancy. The low birth weight prevalence $f$ was set to be the proportion of low birth weight children in the original data, i.e., 0.036 ($=297/8,238$).
The genotypes on 24 SNPs in gene \emph{PPARGC1A} were available, five of which (i.e., rs3774921, rs3755863, rs8192678, rs2970849, and rs1472095) showed strong linkage disequilibrium using the \texttt{R} package \texttt{LDheatmap} based on allelic correlation $r^2$ and the physical distance between the SNPs. Exploratory analyses of the five candidate SNPs can be found in Table S9 of \cite{zhang2021covariate}. It was shown previously that SNP rs8192678 may have a parental origin effect on birth weight \citep{lin2013multi} when pp-BMI was not adjusted for. We therefore applied eight methods (i.e., ROB-HAP, ROB-SNP, LOGIT-HAP, LOGIT-SNP, IND-HAP, IND-SNP, P-HAP, and P-SNP) to re-assess parental origin effect of SNP rs8192678 on low birth weight adjusting for pp-BMI. Methods imposing independence assumption between rs8192678 and pp-BMI (i.e., IND-HAP and IND-SNP) are applicable since this SNP was not significantly associated with pp-BMI \citep{zhang2020efficient}. Among the 691 mother-child pairs, 15 had incomplete genotype data and 58 had incompatible genotypes at any of the five SNPs. Those child genotypes incompatible with their mothers' genotypes were treated as missing. Among the $691-15-58=618$ families with incomplete or incompatible genotypes, 145 had heterozygous genotypes at the SNP rs8192678 for both mothers and children. Among the 145 families, the parental origins of 73 families were unambiguously determined using the joint genotypes at the five SNPs. It happened that all of the children in these 73 families inherited the minor allele from their mothers.
Both ROB-HAP and IND-HAP could accommodate missing child genotypes via our proposed expectation-maximization algorithm. The other methods only used those families with complete genotypes.
\begin{table}
\centering
\caption{Parental origin effect estimates for SNP rs8192678 in the Jerusalem Perinatal Study}\label{tab6}
\begin{threeparttable}
\begin{tabular*}{\hsize}{@{}@{\extracolsep{\fill}}ccccc@{}}
\hline
Method$^a$ & EST$^b$ & SE$^c$ & 95\% CI$^d$ & P-value$^e$ \\ \hline
IND-SNP & 0.612 & 0.256 & {[}0.111, 1.114{]} & 0.017 \\
ROB-SNP & 0.597 & 0.258 & {[}0.092, 1.102{]} & 0.020 \\
LOGIT-SNP & 0.827 & 0.373 & {[}0.096, 1.558{]} & 0.027 \\
P-SNP & 0.611 & 0.256 & {[}0.110, 1.112{]} & 0.017 \\
IND-HAP & 0.368 & 0.236 & {[}-0.096, 0.831{]} & 0.120 \\
IND-HAP$^f$ & 0.320 & 0.226 & {[}-0.122, 0.763{]} & 0.156 \\
ROB-HAP & 0.340 & 0.238 & {[}-0.127, 0.807{]} & 0.154 \\
ROB-HAP$^f$ & 0.305 & 0.230 & {[}-0.145, 0.755{]} & 0.184 \\
LOGIT-HAP & 0.415 & 0.247 & {[}-0.069, 0.899{]} & 0.093 \\
P-HAP & 0.363 & 0.236 & {[}-0.100, 0.826{]} & 0.124 \\ \hline
\end{tabular*}
\begin{tablenotes}
\item $^a$Various methods mentioned in Section \ref{s:simu}; $^b$estimated parental origin effects; $^c$estimated standard errors for parental origin effects; $^d$95$\%$ Confidence intervals for parental origin effects; $^e$p-values for testing parental origin effects; $^f$families with missing data used.
\end{tablenotes}
\end{threeparttable}
\end{table}
Parental origin effect estimates are summarized in Table \ref{tab6}. All methods produced positive parental origin effect estimates, indicating the minor allele (T) of the child with low birth weight was preferentially transmitted from the mother, though the estimated effect sizes varied to some extent. The estimation differences between the HAP and SNP methods could be due to the fact that all of the children in the 73 families with hetorozygous genotypes at the SNP rs8192678 inherited the minor allele from their mothers. The prospective likelihood methods LOGIT-SNP and LOGIT-HAP were uniformly less efficient than the retrospective likelihood based methods. For example, the SE by LOGIT-SNP was 0.373, compared with 0.256, 0.258, and 0.256 by IND-SNP, ROB-SNP, and P-SNP. This verified the simulation results in Section \ref{simu1}.
Furthermore, utilizing haplotypes could effectively improve the estimation efficiency for the parental origin effects. In fact, the SEs were 0.256, 0.258, 0.383, and 0.256 by IND-SNP, ROB-SNP, LOGIT-SNP, and P-SNP, compared with 0.236, 0.238, 0.247, and 0.236 by IND-HAP, ROB-HAP, LOGIT-HAP, and P-HAP. This result was consistent with the simulation results in Section \ref{simu2}. Furthermore, utilizing families with missing or incompatible genotypes slightly improved estimation efficiency, that is, the SE was reduced from 0.238 to 0.230 with ROB-HAP and from 0.236 to 0.226 with IND-HAP. Finally, P-HAP and ROB-HAP had comparable SEs, which was consistent with the simulation results in Section \ref{simu3}.
\section{Discussion}
\label{s:discuss}
Testing parental origin effects is an important genetic problem, and it is a major challenge to develop an inference procedure of high efficiency utilizing case-control mother-child pair data. Under this study design, we develop a modified retrospective-likelihood-based statistically efficient inference method, ROB-HAP, for testing parental origin effects while controlling for the confounding effects of covariates. Genotypes from genetic markers tightly linked with the target genetic marker are used to accurately infer the parental origin. ROB-HAP can fully exploit genetic information in the available data, including Hardy-Weinberg equilibrium, Mendelian law between maternal genotype and child genotype, and the conditional independence between covariates and child genotype given maternal genotype. ROB-HAP is statistically robust to some extent in the sense that no parametric assumption is imposed on the relationship between the maternal genotypes and covariates. A modified profile likelihood function is adopted to ensure computational robustness, without sacrificing statistical efficiency. Simulation results demonstrated that ROB-HAP was very robust to the violation of Hardy-Weinberg equilibrium and misspecification of disease prevalence. It was also shown to be robust to the presence of unaccounted gene-environment interaction effects in testing parental origin effects.
The proposed method ROB-HAP can be improved from several aspects.
First, ROB-HAP assumes an additive mode of inheritance for the genetic effects, but this assumption can be straightforwardly relaxed by introducing dominance effects. Second, the robustness of ROB-HAP and the effciency of IND-HAP can be balanced by adopting a shrinkage estimator similar to the one proposed in \cite{chen2009shrinkage}. Specifically,
let $\hat{\bm{\beta}}_\textrm{ROB}$ and $\hat{\bm{\beta}}_\textrm{IND}$ denote the estimators of regression parameter vector $\bm{\beta}$ using ROB-HAP and IND-HAP, respectively, and let $\beta_\textrm{ROB}$ and $\beta_\textrm{IND}$ denote their limiting values. The parameter $\psi = \beta_\textrm{IND} - \beta_\textrm{ROB}$ can be used to describe the degree of deviation between the true model and the independence model. A shrinkage estimator $\hat{\bm{\beta}}_\textrm{EB}$ can be derived under the empirical Bayesian framework: $\hat{\bm{\beta}}_\textrm{EB} = \hat{\bm{\beta}}_\textrm{ROB} + K(\hat{\bm{\beta}}_\textrm{IND} - \hat{\bm{\beta}}_\textrm{ROB})$, where $K = V(V + \hat{\psi}\hat{\psi}^\mathrm{\scriptscriptstyle T})^{-1}$, $\hat\psi=\hat{\bm{\beta}}_\textrm{IND} - \hat{\bm{\beta}}_\textrm{ROB}$, and $V$ is an appropriate consistent estimator of $\textrm{cov}(\hat{\psi})$. Third, the proposed method can be extended to family data of other types. In some studies, genotype data can be obtained from some mother-father-child triads. The paternal genotype is helpful in determining the parental origin, thus improving statistical inference efficiency. ROB-HAP can be immediately extended to handle such families by adding paternal genotypes in the retrospective likelihood \eqref{equ:retro}. Statistical efficiency can be further improved when genotypes are available from multiple children in some families. Various types of family structures may be present in real studies, and it deserves further investigation to combine these data.
In practice, we recommend using four adjacent SNPs in ROB-HAP and ignoring haplotypes with small estimated frequencies (e.g. less than 0.01) since involving more adjacent SNPs has very limited power gain in testing parental origin effects but significantly increase computational burden. In some studies, the collected families may have multiple ethnicities. If the ancestry information is known for each family, a stratified analysis can be easily conducted by utilizing Fisher's P-value combination method \citep{fisher1992statistical,ray2016usat}; otherwise, principle components extracted from genomewide data can be used to adjust for population structure.
\section*{Acknowledgement}
The work of HZ was partially supported by the Natural Science Foundation of China (No. 12171451), and the work of JC was partially suported by the grants from the National Institutes of Health (No. R21-ES020811 and R01-ES016626).
\section*{Supplementary material}
\label{SM}
Supplementary material
includes derivations or proofs of the theoretical results and additional simulation results. (https://github.com/yatian20/APOE).
\bibliographystyle{apalike}
|
\section{Introduction}
Precision agriculture is a farming practice that utilizes sensor networks to help improve the use of agronomic inputs (e.g., water, fertilizers, pesticides)~\cite{ZHANG2002113}. Robotics research in precision agriculture has largely focused on remote sensing via ground or aerial robots (e.g.,~\cite{maes2019perspectives,radoglou2020compilation,kim2019unmanned}). Besides remote sensing, an increasing number of works has begun addressing interactions with the crop. Such works consider primarily robotic harvesting in both row (e.g., corn and soybean) and tree crops (e.g., citrus and avocado). For example, robots have been deployed to pick peppers, apples, citrus, and tomatoes by wrapping the fruit and twisting it off the stem with either a soft gripper~\cite{lehnert2017autonomous,hohimer2019design,chowdhary2019soft}, rigid gripper~\cite{mehta2014vision,mehta2016robust,de2011design,davidson2017dual,davidson2016proof,nguyen2013task,uppalapati2020berry}, or vacuum~\cite{schupp2011preliminary,baeten2008autonomous,zhang2020system}. Some robots can pick strawberries, cucumbers, citrus, and peppers by cutting the stem~\cite{hayashi2010evaluation,van2009optimal,van2003field,aloisio2012next,arad2020development,r2018research}.
This paper focuses on interaction with tree crops and addresses a conceptually-related yet less explored topic compared to robotic harvesting: \emph{robotic leaf sampling}. Leaf sampling is important in agriculture since remote sensing typically provides field-level information without sufficient resolution to accurately diagnose problems. Agronomists utilize specialized instruments that can be difficult to transport to the field and thus rely upon sample retrieval for later lab analysis. While this has been mostly a manual process to date, some work has been performed using aerial and ground robots. Mueller-Sim et al. demonstrated a robotic platform for rapid phenotyping and capable of manipulating leaves for in-situ measurements~\cite{mueller2017robotanist,abel2018fieldrobotic}. Orol et al. developed a tele-operated aerial robot for cutting and collecting leaves from trees~\cite{VijayICUAS2017}. Ahlin et al. presented an algorithm for selecting and grasping tree leaves using a robotic arm~\cite{ahlin2016autonomous}. The latter work demonstrates a high level of control using monoscopic depth analysis (MDA) and image-based visual servoing, but focuses on grasping and pulling the leaf instead of cleanly cutting the stem of the leaf, which is the focus of our work.
\begin{figure}[!t]
\vspace{6pt}
\centering
\includegraphics[trim={0cm 5cm 0cm 3cm },clip,width=0.95\linewidth]{IMG_1856.jpg}
\vspace{-6pt}
\caption{We develop a custom-built end-effector attached to an off-the-shelf 6-DOF robotic arm and a visual perception algorithm to detect, localize and cut leaves at their stem. (The supplementary video demonstrates the end-effector's operation and overall system testing.)
}
\label{fig:eefonarm}
\vspace{-18pt}
\end{figure}
Our work is motivated by the need to perform leaf water potential measurements, an important process performed by agronomists to estimate tree stress levels and hence optimize irrigation patterns~\cite{UCANRpressureIrrigation}. %
A leaf cut at its stem is placed inside a pressurized chamber instrument with its cut end exposed, and then the pressure at which water begins to escape from the cut stem is used to determine the leaf water potential~\cite{scholander1965sap,tyree1972measurement}. Agronomists use this measurement as a proxy for tree stress levels to optimize irrigation patterns.
Though effective, these instruments can be tedious and potentially dangerous to operate.\footnote{\url{www.pmsinstrument.com/maintenance/safety/}}
As a result, a single tree is often used to quantify the health of the entire orchard leading to infrequent measurements and undersampled regions. Enabling robotic leaf sampling (this paper's focus) for future use in robotic leaf water potential analysis can help improve measurement coverage and frequency while reducing human fatigue, and risk of bodily injury. Our work joins a growing body of works on robotic means for monitoring crop health and improving irrigation management practices~\cite{RAPIDBookChapter, CARPINCASE2108B}.
Compared to existing robotic leaf sampling methods~\cite{mueller2017robotanist, abel2018fieldrobotic, VijayICUAS2017, ahlin2016autonomous} and harvesting systems that cut the stem of a fruit/vegetable~\cite{hayashi2010evaluation, van2009optimal, van2003field, aloisio2012next, arad2020development, r2018research}, we are interested in performing clean cuts at leaves' stems and retaining leaves for stem water potential analysis.
As with related works, we also incorporate a visual perception component (to identify and localize a leaf) and an actuation component (to move the end-effector toward the leaf, and then cut it). Collecting a leaf sample from a tree presents unique challenges in perception and actuation, distinct from robotic fruit harvesting.
Similarly, finding a motion plan to retrieve a physical sample needs to account for the presence of other leaves and branches which can also interfere with the extraction process. Yet, identifying a leaf sample involves not only segmenting the canopy, but also selecting an unblemished leaf suitable for stem water potential analysis~\cite{UCANRpressureIrrigation}.
To this end, we propose a leaf-cutting end-effector combined with a visual perception system that detects the center of a leaf and estimates its 6D pose (Fig.~\ref{fig:eefonarm}). The end-effector can cut and capture leaves of several common tree crops, such as avocado, clementine, grapefruit, and lemon. Unlike the MDA approach~\cite{ahlin2016autonomous}, we use a depth camera and a 3D point cloud to identify the centroid of the leaf and then estimate its 6D pose. This paper outlines our perception and actuation process to detect, localize, and cut leaves at their stem while retaining them, to enable future automated leaf water potential analysis in tree crops.
\section{Related Works}
Development of harvesting end-effectors is an active area of research due to the wide variety of crops. While there are some commonalities across approaches, differences in size, weight, shape, texture, and firmness of specialty crops have led to unique solutions. Apples and citrus require a specific motion to grasp, twist, and pull from the tree without damage~\cite{bu2020experimental,aloisio2012next}. Bell peppers and cucumbers can be directly cut and harvested~\cite{lee2019vision,arad2020development,van2009optimal,van2003field}. More delicate crops like strawberries call for manipulators with force feedback and flexible pneumatic actuators~\cite{Simonton1991,hayashi2010evaluation,xiong2020autonomous,bao2015flexible}.
Despite their unique applications, harvesting end-effectors generally have three primary components: the gripping mechanism (mechanical, pneumatic or hybrid), the removal mechanism (mechanical or electrical), and the sensing modality (monocular camera, stereo camera, time-of-flight)~\cite{morar2020robotic,r2018research}.
At the same time, there has been development of perception techniques to monitor
crop growth~\cite{ZHU201628,sadeghi2017automated}, help prevent disease through early detection~\cite{akram2017towards,8706936},
assist with quality control~\cite{su2018potato,JAHNS200117}, and help automate harvesting~\cite{bu2020experimental}.
Success of these tasks depends on the visual perception subsystem's ability to provide precise and accurate information about the target crop and relevant environmental context~\cite{2021}, including
segmentation and localization of targets of interest.
Most approaches have focused on fruit/vegetable targets by harnessing distinct colors and/or shapes~\cite{ahlin2016autonomous,FU2020105687,chen2017counting,NGUYEN201633,7942724}.
In this paper we are targeting identification and pose estimation of individual tree-crop leaves. This presents similar yet unique challenges compared to fruit (and broader canopy) identification.
Instead of filtering out the leaves to focus on the fruit, our objective is to retain the leaves and segment the tree canopy further to obtain individual leaf poses.
Leaf segmentation has been considered in current research using both classical computer vision tools~\cite{Chen_2019_CVPR_Workshops,MIAO2021106310,ELNASHEF201951} as well as machine learning~\cite{Guo2021LeafMaskTG, 9025429, Scharr2015LeafSI}. However, classical methods are prone to changes in the environment, such as light, occlusions or overlapping surfaces, whereas learning-based methods require large training datasets and may still generalize poorly as environmental factors vary~\cite{app11010228}.
Furthermore, these techniques have rarely been employed online on onboard computers as part of a robotic manipulation system to identify, localize and physically cut the leaf.
Although a leaf's 3D position can be readily obtained, it is not sufficient to successfully accomplish the task as orientation plays an important role as to how a robotic arm approaches the leaf to cut it.
Thus, obtaining at least an estimate of the 6D pose (position and orientation) is critical.
Traditional 6D pose estimation approaches usually perform local keypoint detection and feature matching, and then a RANSAC-based PnP algorithm on the established 3D-to-2D correspondences to estimate the pose of an object ~\cite{Michel2017GlobalHG, Brachmann2016UncertaintyDriven6P}. Still, they typically fail to perform with heavily occluded and poorly textured objects. On the other hand, learning-based methods use a deep neural network (DNN) to obtain the correspondences between 3D objects points and their 2D image projections~\cite{Hu2019SegmentationDriven6O,Hu2020SingleStage6O, Park2019Pix2PosePC}. Use of synthetic data generators~\cite{Giuffrida2017ARIGANSA, Zhu2018DataAU} can relieve in part the challenge of acquiring large labeled datasets; however, it requires realistic models that take into account the variations of the detected object e.g., shape, size, orientation or curvature which can be hard to develop.
Our developed end-effector focuses on the actuation and perception techniques to cut and retain a leaf. This task has received much less attention by existing robotic harvesting/leaf sampling technology yet an important aspect toward enabling future robotic leaf water potential measurements.
\section{Technical Approach}
Picking a leaf requires two key components: actuation and perception. For actuation, we design a custom-built leaf-cutting end-effector (Section~\ref{sec:ta_leaf_cutting}) and retrofit it on a mobile manipulation base platform (Kinova Gen-2 six degree of freedom [6-DOF] robot arm mounted on a Clearpath Robotics Husky wheeled robot). For perception, we utilize point cloud data from a depth camera (Intel RealSense D435i) for the leaf detection and localization algorithm developed herein (Section~\ref{sec:ta_leaf_perception}). The point cloud data is processed using Open3D~\cite{Zhou2018} running on an Intel i7-10710U CPU, without any additional GPU acceleration. Figure~\ref{fig:systemdiagram} highlights how our contributions interact in a leaf-cutting system, which is further evaluated in Section~\ref{sec:exp_methods}.
\begin{figure}[!t]
\vspace{6pt}
\centering
\includegraphics[trim={0 0 0 0 },clip,width=0.975\linewidth]{systemdiagram2.png}
\vspace{-10pt}
\caption{Our approach jointly considers perception and actuation. The perception module processes point cloud data to segment leaves and deposit leaf candidates into a queue. Candidate leaves are then passed to the robot arm controller to actuate the end-effector. If a cut is successful, the routine ends. If unsuccessful, the arm controller requests the next leaf in the queue.
}
\label{fig:systemdiagram}
\vspace{-18pt}
\end{figure}
Identified and segmented leaves serve as target for the arm to move and align the end-effector along a viable leaf (to be defined in Section~\ref{sec:exp_methods}), at an offset position from the center of the leaf. The offset distance is equivalent to the length of the leaf. Once at the offset position, the arm moves linearly toward the leaf to capture it. When the leaf is enclosed, the end-effector cuts the leaf. Then, the arm returns home.
\subsection{Actuation}
\label{sec:ta_leaf_cutting}
The stem-cutting end-effector developed herein utilizes two 4-bar linkages to actuate a set of sliding gates, one of which contains a razor blade to remove the leaf from the tree (Fig.~\ref{fig:eef_cad}). The gates also help retain the leaf within the end-effector's chamber after removal from the tree. These 4-bar mechanisms are connected via a gear train to achieve synchronized motion. A low-cost, high-torque R/C servo (FEETECH FT5335M) drives the gear train while being amenable to position control. An Arduino Due microcontroller controls the servo motor and receives serial commands from a ROS control node. A breakout board connected to the Arduino contains a ``safe/armed" switch along with LED indicators to reduce the risk of accidental injury
Stem water potential analysis requires the test leaf's stem to be cleanly cut; a damaged specimen would negatively impact the analysis~\cite{scholander1965sap}. Organic matter such as leaf stems exhibit visco-elastic properties. Based on visco-elastic material principles, faster cuts will require less force and result in less deformation of the leaf stem. Our prototype end-effector is able to cut leaf stems with a design target force of $20$\;N at $1.1$\;m/s. The end-effector's chamber has an opening of $110$\;mm by $45$\;mm and a depth of $185$\;mm to accommodate typical avocado leaves. The end-effector is constructed with miniature aluminum extrusions, lightweight 3D printed parts, and laser-cut acrylic panels. The assembly weighs $1.091$\;kg, which is $42$\% of the robotic arm's $2.6$\;kg payload. The end-effector is powered separately from the arm to enable stand-alone testing with a $7.4$\;V 2S LiPo battery.
\begin{figure}[!t]
\vspace{6pt}
\centering
\includegraphics[trim={0 0 0 0 },clip,width=0.90\linewidth]{gripper_cad_render.jpg}
\vspace{-9pt}
\caption{The end-effector contains the components necessary to cut a leaf from a tree. The servo motor (red) actuates a double four-bar mechanism (yellow) that closes a set of gates (blue) with a razor blade to cut and capture a leaf. An Intel RealSense camera D435i is mounted on the top of the end-effector for perception. A microcontroller is mounted on the arm for controlling the motor. This end-effector can be mounted to a robotic arm using an adaptor plate (green). (Figure best viewed in color.)
}
\label{fig:eef_cad}
\vspace{-18pt}
\end{figure}
To determine the types of leaves that can be cut by the mechanism, we performed testing with a variety of trees in a local orchard. The end-effector was manually placed around leaves and activated. Four different crops were selected (avocado, clementine, grapefruit, and lemon) for evaluation. For each crop, ten cutting attempts were performed. Results are shown in Table~\ref{tab:leaf_cutting_agops}. The end-effector was able to cut 95\% of the leaves (38 out of 40). Lower success rates were observed for the lemon and grapefruit leaves. This is due to these particular leaves having shorter stems which made it harder to position the end-effector around the stem without interference from branches or other leaves. The end-effector worked consistently on clementine and avocado leaves.
\begin{table}[!h]
\centering
\caption{Leaf Cutting Tests}
\label{tab:leaf_cutting_agops}
\vspace{-6pt}
\begin{tabular}{cccc}
\toprule
Crop & Successful Cuts & Attempts & Rate \\
\midrule
\midrule
Avocado & 10 & 10 & 100\% \\
Clementine & 10 & 10 & 100\% \\
Grapefruit & 9 & 10 & 90\% \\
Lemon & 9 & 10 & 90\% \\
\midrule
Total & 38 & 40 & 95\% \\
\bottomrule
\end{tabular}
\vspace{-12pt}
\end{table}
\subsection{Perception}
\label{sec:ta_leaf_perception}
We propose a leaf detection and localization algorithm using 3D point cloud and processed through the Open3D library. Our approach is outlined in Fig.~\ref{fig:systemdiagram}.
The detection phase seeks to obtain the 3D bounding box of leaves candidates from point cloud captured from the depth camera. First, we remove outliers considered as noise resulting from sensor measurement inaccuracies and segment out the background at a specific distance threshold from the camera frame. Then, downsampling is applied to optimize the performance of the upcoming step. Next, we group the remaining point cloud segments into clusters using the Density Based Spatial Clustering of Applications with Noise (DBSCAN) approach~\cite{Ester1996ADA}. It relies on two parameters, the minimum distance between two points to be considered as neighbors (\textit{eps}) and the number of minimum points to form a cluster (\textit{MinPoints}).
Each resulting cluster is considered a potential leaf and described by a 3D bounding box defined by center $C = [c_x, c_y, c_z]^{T}$, dimensions $D = [h, w, d]$, and orientation $R(\theta, \Phi, \alpha)$. Then, filtering is applied on the clusters using geometric features of the bounding box: number of points, volume, leaf ratio. Finally, the pose of the center of each bounding box is returned as the 6D pose of a potential leaf.
\begin{figure}[!t]
\vspace{6pt}
\centering
\includegraphics[trim={0 0 0 0 },clip,width=0.95\linewidth]{point_cloud_outdoor.jpg}
\vspace{-12pt}
\caption{Key steps in our proposed leaf detection and localization process. The sample here corresponds to an outdoor point cloud: (a) corresponding RGB image of the tree, (b) raw point cloud, (c) distance filtered ROI, (d) downsampled point cloud, (e) segmented clusters, and (f) detected candidate leaves without 6D pose bounding boxes.
}\label{fig:pointcloudoutside}
\vspace{-18pt}
\end{figure}
To validate our approach, we conducted offline tests for detection and localization separately. For the detection step, ROSbags were collected both in indoor and outdoor settings.
Indoors (lab with constant light conditions), we used the Kinova arm with the camera placed at different distances ($0.2-0.3$\;m) from a potted tree.
Outdoors (local orchard with varying light conditions), we collected data manually. We considered a wide range ($0.5-1.6$\;m) of distances from trees; an example is shown in Fig.~\ref{fig:pointcloudoutside}.a.
A total of 25 point clouds were collected (10 indoor and 15 outdoor). and tested offline with different combinations for \textit{eps} and \textit{MinPoints} parameters, to determine optimal values for later use.
Table~\ref{tab:leaf_pcd_detection} shows the outcome of our experiments on the 10 indoor point clouds and 15 outdoor point clouds.
We attain an average of 80.0\% of detection with a maximum of 90\% for indoor dataset, and an average of 79.8\% with a maximum 85\% for outdoor.
Further, we observed that the distance between the camera and the tree impacts the optimal values for the point cloud processing. The greater the distance from the camera, the higher \textit{eps} while \textit{MinPoints} decreases.
\newcolumntype{L}{>{\centering\arraybackslash}m{1.25cm}}
\begin{table}[!ht]\vspace{-6pt}
\centering
\caption{Leaf Point Cloud Detection}
\label{tab:leaf_pcd_detection}
\vspace{-6pt}
\begin{tabular}{LLLLL}
\toprule
& Point Clouds & Total \# Leaves & Average Detection & Percentage\\
\midrule
\midrule
Indoor & 10 & 20 & 16 & 80.0\% \\
\midrule
Outdoor & 15 & 99 & 79 & 79.8\%\\
\bottomrule
\end{tabular}
\end{table}
To validate the localization phase, we compare several 6D poses obtained via our proposed approach against ground truth data obtained from a VICON motion capture camera system. Retroreflective markers were placed around the center of leaves, as shown in Fig.~\ref{fig:leafvicondots}, to estimate their pose.
\begin{figure}[!t]
\vspace{6pt}
\centering
\includegraphics[trim={0 0 0 0 },clip,width=0.80\linewidth]{vicon_leaf_arm_origin.jpg}
\vspace{-6pt}
\caption{We used motion capture to establish a ground truth for determining the leaf 6D pose. Markers were placed on a target leaf (left) with origin at the base of our 6-DOF robot (right). (A real avocado tree was used.)
}
\label{fig:leafvicondots}
\vspace{-18pt}
\end{figure}
Table~\ref{tab:leaf_pose_error} summarizes the results obtained for 12 random leaves positions. Our approach provides an estimation with mean error of 8.28 mm, 14.38 mm, and 15.54 mm along x-axis, y-axis, and z-axis, respectively,for avocado leaves of width ranging between $24-86$\;mm and length ranging between $54-150$\;mm. Based on the average leaf size ($48\times91$\;mm), estimation errors represent nearly 15\% of the width and 17\% of the length. We evaluated the orientation by calculating the Euclidean distance between the two provided values using the definition in~\cite{Huynh2009MetricsF3}. We obtained a mean error of 5.3$\deg$.
We observe that the obtained 6D pose may drift from the physical center of the leaf mainly on the y-axis and z-axis due to human-induced error and the non-rigid nature of the leaf which impacts marker placement.
\begin{table}[!ht]
\centering
\caption{Leaf 6D Pose Error}
\label{tab:leaf_pose_error}
\vspace{-6pt}
\begin{tabular}{ccccccc}
\toprule
Error & $\Delta x$ (mm)& $\Delta y$ (mm) & $\Delta z$ (mm)& Orientation ($\deg$) \\
\midrule
\midrule
Mean & 8.28 & 14.38 & 15.54 & 5.3\\
\midrule
Std dev & 7.46 & 5.46 & 6.69 & 15.5 \\
\bottomrule
\end{tabular}
\end{table}
The proposed approach provides an initial 6D pose along useful information of potential leaves using a processed 3D point cloud and obtained up to 80\% of detection and a mean error less than 16mm and 5.3 $\deg$. Both detection and localization steps were performed without the need of collection or storage of large data including 3D models, and training process. Furthermore, all tests were run using a CPU configuration, without any additional GPU acceleration.
\section{Experimental Validation of Leaf Cutting}
\label{sec:exp_methods}
To evaluate our overall leaf detection, localization and cutting pipeline, we tested with a real potted avocado tree indoors (lab). The mobile manipulator and end-effector system was positioned at random poses near the base of the tree so that the end-effector was at distances ranging between $0.2-0.3$\;m from the edge of the tree canopy. An experimental trial consisted of collecting a point cloud, storing the identified and localized potential leaves in a queue, and then sending the queued leaves to the arm for a retrieval attempt. Each trial concluded once the queue was depleted and the tree was repositioned for the next trial. Figure~\ref{fig:trialsequence} outlines this process.
\begin{figure*}[!t]
\vspace{6pt}
\centering
\includegraphics[trim={0cm 0cm 0cm 0cm },clip,width=0.95\linewidth]{leaf_cut_vid_frames.jpg}
\vspace{-6pt}
\caption{Overall leaf retrieval process. During the perception phase, (a) the point cloud is processed to determine a potential leaf. If a viable leaf is detected, (b) the arm will move to an offset position. (c) The arm will then perform a linear motion to capture the leaf. Once in position, (d) the arm will cut the leaf and (e) the leaf will fall into the enclosed chamber. (f) After completing the cut, the arm will return to the home position.
}
\label{fig:trialsequence}
\vspace{-18pt}
\end{figure*}
For each retrieval attempt, leaf candidates and viable leaves are determined. \emph{Leaf candidates} are leaves that have a pose within the arm's workspace. \emph{Viable leaves} are leaf candidates that have a retrieval path within the arm's workspace. For testing our point cloud detection, we are interested in monitoring both successful captures and successful cuts of the leaf. A \emph{successful capture} occurs when the end-effector is placed around a viable leaf while a \emph{successful cut} occurs when the enclosed leaf is removed from the tree. A \emph{clean cut} occurs when the leaf is severed cleanly at the stem such that it could be used for stem water potential analysis.
Out of 46 trials, 63 potential leaves were detected by the point cloud. (Note that each point cloud in the trial could produce a variable amount of leaves, hence a higher number of potential leaves than trials.) After filtering the potential leaves to remove the leaves outside of the work space, 39 viable leaves remained. Out of these leaves, 27 were captured successfully (69.2\%) while 21 of the 27 captured leaves were cut (77.8\%). Table~\ref{tab:succ_rates} summarizes retrieval results while Table~\ref{tab:proc_time} highlights the process times. The mean point cloud processing (perception) time was $5.6$\;sec and the mean cutting (actuation) time was $10.6$\;sec. The mean total retrieval time was $16.2$\;sec.
\begin{table}[!ht]
\centering
\caption{Leaf Retrieval Numbers \& Rates}
\label{tab:succ_rates}
\vspace{-6pt}
\begin{tabular}{lrr}
\toprule
Stage & \multicolumn{1}{l}{Number} & \multicolumn{1}{l}{Rate} \\
\midrule
\midrule
Potential Leaves & 63 & \multicolumn{1}{l}{N/A} \\
Candidate Leaves & 51 & 81.0\% \\
Viable Leaves & 39 & 76.5\% \\
Successful Captures & 27 & 69.2\% \\
Successful Cuts & 21 & 77.8\% \\
\midrule
Clean Cuts & 4 & 19.0\% \\
Near Misses & 7 & 30.0\% \\
\bottomrule
\end{tabular}
\end{table}
Our system was able to remove a total of 21 leaves from the tree. However, not all leaves were clean cuts on the stem; four were classified as clean cuts for use in stem water potential analysis. The majority of the leaves were severed at the top of the leaf and not at the stem (Fig.~\ref{fig:leafcutarray}). Our system produced seven near-misses where the leaf was cut within an average of $9.58$\;mm from the stem (std dev: $6.1$\;mm). The remaining 10 leaves were severed closer to the middle of the leaf, largely due to collisions with the branches. Similar branch interference also lead to four out of the six missed cuts from the captured leaf. These two problems could be solved in future work through a refined end-effector design, more robust path planning to account for branches, and implementing visual servoing for continuous stem alignment as the end-effector approaches a viable leaf.
\begin{figure}[!t]
\vspace{6pt}
\centering
\includegraphics[trim={0 0 0 0 },clip,width=0.8\linewidth]{leaf_cut_results2.jpg}
\vspace{-6pt}
\caption{Sample leaves cut from our avocado tree during automated indoor tests. (a) The four leaves represent clean cuts suitable for stem water potential analysis. (b) The system also cut seven leaves that were classified as near-misses, which removed the leaf without the stem. (c) The remaining leaves were cut closer to the center, due to interference between the end-effector and the branches. (d) In two cases, collateral damage occurred when a second leaf was removed along with the target leaf. These instances were classified as a single successful cut, but not a clean cut since the two leaves would need to be separated for stem water potential analysis.
}
\label{fig:leafcutarray}
\vspace{-18pt}
\end{figure}
\begin{table}[!ht]
\centering
\caption{Leaf Retrieval Performance Time (Seconds)}
\label{tab:proc_time}
\vspace{-6pt}
\begin{tabular}{lrrr}
\toprule
Metric & Perception Part & Actuation Part & Overall Retrieval \\
\midrule
\midrule
Min & 0.5 & 4.6 & 6.1 \\
Max & 11.0 & 61.7 & 62.5 \\
Mean & 5.6 & 10.6 & 16.2 \\
Median & 7.7 & 8.1 & 15.3 \\
Std dev & 3.9 & 10.4 & 10.2 \\
\bottomrule
\end{tabular}
\end{table}
\section{Conclusions}
Our work develops a co-designed actuation and perception method for leaf identification, 6D pose estimation and cutting. Our developed leaf-cutting end-effector can cut leaves of various types of trees (avocado, clementine, grapefruit and lemon) cleanly at their stem with a 95\% success rate on average. Our proposed 3D point cloud technique can be successful for detecting an average of 80.0\% of leaves indoors and 79.8\% outdoors, and localizing them with less than 17\% error along the leaf's length or width.
Experimental testing of the overall proposed framework for leaf cutting reveals that our system can capture 69.2\% of viable leaves and cut 77.8\% of those captured leaves.
These results offers a promising initial step toward automated stem water potential analysis, nonetheless several steps remain and are exciting avenues for future work.
The end-effector can effectively cut the leaves, but its size presents a challenge when cutting certain leaves like those from lemon and grapefruit trees which in turn calls for further design optimization. The current path planning approach works well for leaves that are on the periphery of the tree's canopy. Alternate path planning strategies can be explored to reach leaves within the canopy closer to the trunk, and integrated with visual servoing to better align the cutter with the stem of the leaf as it is about to cut it.
Furthermore, the system will need to be robust to disturbances such as wind before deployment in an outdoor orchard environment.
Finally, to enable automated stem water potential analysis, the captured leaf will need to be transferred from the end-effector into a pressure chamber.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Massive galaxy clusters that show (strong and weak) gravitational lensing effects are ideal targets to investigate the large scale distribution of dark matter and the relation between the dark matter and the baryonic matter (e.g., \citep{lot17,fox22} and references therein). The baryonic matter in galaxy clusters are composed mainly of galaxies and intracluster medium (ICM). Primary components in the ICM are X-ray emitting hot gas, intracluster light (ICL), intracluster globular clusters (GCs), and intracluster planetary nebulae. Among them, the intracluster GCs are, because they are compact and abundant, very useful to study the relation between dark matter and baryonic matter at large scales.
Wide field imaging has been a main tool for the survey of intracluster GCs in the nearby galaxy clusters \citep{lee10,dur14,pen11, iod17,har17, hil18, mad18, har20}.
High resolution power of the HST enables us to probe the intracluster GCs in the more distant galaxy clusters, and study the relation between the intracluster GC and dark matter distributions \citep{ala13,ala17,lee16a}.
Using the deep ACS/F814W and WFC3/F105W images of the HST Frontier Fields \citep{lot17}, \citet{lee16a} presented a survey of GCs including ultracompact dwarfs (UCDs) in the massive merging cluster Abell 2744 (known as the Pandora cluster) at
$z=0.308$. They found a rich population of GCs and UCDs in the cluster, a significant fraction of which are intracluster GCs that spread in the wide area of the cluster. They showed that the number density map of the intracluster GCs in Abell 2744 is similar to the dark matter mass map derived from CATS strong lensing models \citep{jau15}.
In this study, we search for GCs of SMACS J0723.3-7327 (called SMACS 0723 hereafter), a massive cluster at $z=0.39$, which is farther than Abell 2744, using the first JWST/NIRCam images released on July 12, 2022 as part of the early released JWST images \citep{pon22,rig22}.
Recently \citet{mow22} found five GCs that are considered to be associated with the Sparkler galaxy at $z=1.38$ which is gravitationally lensed by SMACS 0723 from the JWST NIRCam images and HST/ACS images.
Our primary goal is to find and study the GCs that belong to SMACS 0723.
Basic properties of SMACS 0723 are listed in {\bf Table \ref{tab:info}}.
HST observations of this cluster were included in the HST Snapshot survey \citep{rep18} and the RELICS program \citet{coe19},
providing the magnitudes of the brightest cluster galaxy (BCG), F814W = 19.14 ($M_{F814W} = -22.9$) mag
and F606W=20.35 ($M_{F606W}=-22.0$) mag.
We checked first the HST data of SMACS 0723 available in the archive, but found that the images are too shallow to detect GCs in this cluster. The JWST/NIRCam data go much deeper for these GCs which are composed of old stars, enabling us to detect them.
SMACS 0723 shows a large number of gravitationally lensed images of background sources, which
were used for tracing dark matter in the cluster \citep{pas22,gol22, cam22,mah22}.
The mass of SMACS 0723 based on strong lensing models is $M (R<400 {\rm kpc})\sim 3 \times 10^{14} M_\odot$ \citep{mah22}, and the mass derived from PLANCK observations is $M({\rm PLANCK})=8.39 \times 10^{14} M_\odot$ \citep{pas22}.
\citet{mah22} presented a mean redshift of 26 cluster members located in the central region of SMAC 0723, $z=0.3877$, which is slightly smaller than the redshift of the BCG, $z=0.3912$. In the following analysis, we adopt $z=0.39$ and the position of the BCG as the cluster center.
We adopt the cosmological parameters:
$H_0=$ 67.8 km s$^{-1}$ Mpc$^{-1}$, $\Omega_{\Lambda}=0.7$, and $\Omega_{\Lambda}=0.3$.
At the adopted distance of SMACS 0723 ($d_L=2170$ Mpc, $(m-M)_0=41.69$), one arcsec (arcmin) corresponds to 5.453 (327.16) kpc.
This paper is organized as follows. Data and data reduction are described in Section 2, and main results are given in Section 3. The main results are discussed in Section 4, and are summarized in the final section.
\section{Data}
In the data set of the first released JWST images of SMACS 0723, we use NIRCam images obtained with
three short wavelength (SW) filters (F090W, F150W, F200W) and three long wavelength (LW) filters (F277W, F356W, F444W) on June 7, 2022 (see the field locations in Figure 1 of \citet{pon22}). The total exposure time is 7537 sec for each filter.
The SW filter images have an image scale of $0\farcs03$ per pixel, and the LW images have
a twice larger image scale of $0\farcs06$ per pixel.
The released images include two fields: a cluster central field that covers the central region of the cluster, and a parallel field that covers the outskirt of the cluster, as shown in {\bf Figure \ref{fig:finder}(a)}.
The field of view for each of the NIRCam images is $129\arcsec \times 129\arcsec$, and the gap between the central field and the parallel field is $44\arcsec$. The parallel field is $\sim2\farcm9$ to the southwest from the central field. The radial coverage of the parallel field is 550 kpc $<R<$ 1.3 Mpc, where $R$ is the projected radial distance from the cluster center.
In {\bf Figure \ref{fig:finder}(a)} we display gray scale maps of F200W images of the two fields.
We mark the brightest galaxies (G2, G3, G4, and G5) including the BCG in the central field.
The edge regions of each field marked by burgundy color have low signal-to-noise ratios in the combined images so they are not used for the final analysis.
In {\bf Figure \ref{fig:finder}(b,c,d, and e)}
we also display the zoomed-in color image of the central field, showing the BCG region, and the zoomed-in images of the BCG region that show GCs clearly, which will be described in the following.
\section{Data Reduction}
We use F200W images, which have the highest sensitivity among the six NIRCam images, as a reference image for point source detection\footnote{JWST user documentation, https://jwst-docs.stsci.edu}.
We adopt the AB magnitudes in this study. The relations between the AB and Vega magnitudes in the
JWST CRDS\footnote{https://jwst-crds.stsci.edu} are:
mag(AB)--mag(Vega)=0.504, 1.243, 1.706, 2.315, 2.824, and 3.242 for F090W, F150W, F200W, F277W, F365W, and F444W.
Mean effective radii of typical globular clusters in the Milky Way are about 3 pc and the FWHM values of the point sources in the JWST/NIRCam/SW images are about 2.4 pixels
($0\farcs072$=0.39 kpc).
Thus typical GCs appear as point sources in the JWST/NIRCam images of SMACS 0723 (see also \citet{fai22}).
We select the point sources from which GCs are segregated as follows.
\subsection{Photometry of the Extended Sources}
First we obtain the photometry of the sources (including both extended and point sources) detected with 10$\sigma$ threshold in the images using Source Extractor \citep{ber96}. We carry out dual mode photometry with
the F277W image
as a reference image for source detection.
As the pixel scales for LW and SW detector are different, the SW images were reprojected and aligned with respect to the F277W image using the wregister task in IRAF \citep{tod86, sts12} and astroalign \citep{ber20}, respectively. For aligning the images, we use the positions of several stars.
We use LW images for the analysis of galaxies because the normal red sequence of galaxies in galaxy clusters is better distinguishable in the LW color-magnitude diagrams (CMDs), than in the SW CMDs, as shown in the following.
{\bf Figure \ref{fig:rhmag}} display effective radius $r_h$ vs.
F277W (AUTO) magnitudes of the sources detected with the Source Extractor.
We also mark the sources with small stellarity $<0.4$ (red circles), and the sources with large stellarity $ >0.8$ (blue circles).
The prominent narrow slanted sequence dominated by blue circles represent the point sources with small effective radii.
Some of the bright sources with large effective radii are found to have a large value of stellarity $>0.8$ so we consider that effective radius is a better indicator to distinguish extended sources than stellarity.
In this study we select extended sources using the effective radius criterion:
$r_h >1.8$ pixels in the F277W image.
The extended sources include galaxies and gravitationally-lensed sources.
We adopt AUTO magnitudes for the extended sources provided by the Source Extractor, and use them for the analysis of photometric properties of the galaxies in the following.
\subsection{Photometry of the Point Sources}
For better detection of the point sources in the images that include a number of bright galaxies, we subtract the diffuse background light from the original images.
First, we derive a smooth background image of the original F200W image, using a ring median filter with $r_{in}=18 $ pixel and $r_{out}=20$ pixel.
Then we subtract this smooth background image from the original image, and use the background-subtracted image for the point source detection.
\subsubsection{Point Spread Function Fitting Photometry}
We obtain the point spread function (PSF) fitting photometry of the sources in the F150W and F200W images of the central and parallel fields using DAOPHOT \citep{ste87}.
We derive the PSF using the isolated bright stars in the images.
Among the detected sources we select the point sources using the sharpness parameter and photometric errors provided by DAOPHOT.
We calibrate the instrumental PSF-fitting magnitudes
using the total magnitudes of bright stars given in the source catalog of SMACS 0723 included in the released data.
We use only the bright stars $21<F150W, F200W<24$ mag which have small errors.
However, it is noted that this zero-point calibration is preliminary. There could be a substantial (systematic) error associated with the pre-flight version of the JWST catalog \citep{boy22}.
The catalog of the point source photometry in this study will be publicly available at GitHub\footnote{https://github.com/hanlbomi/SMACS-J0723.3-7327-GCs}.
We calculate the extinction values for the JWST filters from the extinction value for $V$-band given by \citet{sch98,sch11} using the extinction law \citep{fit99,ind05} and the filter information in the SVO filter profile service website.\citep{rod12, rod20}.
The extinction values are listed in Table \ref{tab:info}.
We apply the foreground extinction correction to the apparent magnitudes,
use the subscript 0 for the foreground-extinction-corrected magnitudes.
\subsubsection{Completeness Tests}
We estimate the photometric completeness of the point sources using the artificial star experiments with DAOPHOT/ADDSTAR. We added ~2500 artificial stars with a range of magnitude and color to each of F150W and F200W images. We repeat this process 400 times for each set of images so that the total number of added artificial stars in each band is one million. We obtain the photometry of these sources using the same procedure as used for the original images.
Then we select the recovered/added stars with the GC color range, and estimate the completeness from the number ratio of the recovered stars to the added stars as a function of magnitude. Figure 3 displays the completeness vs. magnitude for F150W and F200W bands. We fit the data using the interpolation function given in \citet{har09}:
$ f(m) = 0.5 (\beta-\alpha (m-m_0)/\sqrt{1+\alpha^2(m-m_0)^2} ) $
where $\beta$ is the maximum completeness, $m_0$ is the magnitude for $\beta/2$,
and $\alpha$ represents the slope around the $m_0$, as shown by the solid lines in the figure.
The limiting magnitudes for 50\% photometric completeness are F105W=30.44 mag and F220W=30.11 mag.
\section{Results}
\subsection{CMDs of the Galaxies}
In {\bf Figure \ref{fig:galaxyCMD}}
we display the LW CMD (F277W vs (F277W--F356W)) and SW CMD (F200W vs (F150W--F200W)) (in AUTO magnitudes) of the extended sources in the central field of SMACS 0723.
We also mark the cluster member galaxies (N=26) confirmed with spectroscopic observations \citep{mah22}, the selected bright galaxies (including the BCG), and the known gravitational lens image sources in \citet{mah22}.
In the F277W vs (F277W--F356W) CMD, two prominent features are noted: an LW-blue vertical sequence with $(F277W-F356W)_0 \approx -0.6$,
and an LW-red slanted sequence with $(F277W-F356W)_0 \approx +0.1$.
The known cluster members
are mostly located in the bright part of the LW-blue sequence.
The known cluster members are mostly early-type galaxies, showing
that this sequence corresponds to the normal red sequence
in the optical CMDs of galaxy clusters.
So these member galaxies were plotted in red circles.
The selected bright galaxies including the BCG
are also located in the bright part of this sequence.
The presence of a large number of galaxies located in this LW-blue sequence shows that SMACS 0723 is indeed a rich cluster.
In contrast, the known gravitational lens image sources are mostly located along the LW-red sequence, but with a larger scatter than the LW-blue sequence. They look red in the released color images based on the combination of all SW and LW filter images. The gravitational lens image sources have much higher redshifts than that of SMACS 0723 so they are not the cluster members. They are background galaxies most of which show some star formation activity (i.e., late type galaxies having different colors compared with early-type galaxies).
However, the F200W vs (F150W--F200W) CMD
shows only one notable slanted sequence at (F150W--F200W)$_0\approx +0.2$. Both the cluster member galaxies and the lens image sources are located mostly in the same sequence, so they are hardly distinguished in this sequence.
Therefore, as the red sequence galaxies which are mostly early-type cluster members, we select those located along the LW-blue sequence as marked in the LW CMD:$-0.65 < (F277W - F356W)_0 < -0.45$.
We use them to check any correlation between the GCs and the galaxies in the following analysis.
\subsection{CMDs of the Point Sources: Detection of the GCs}
{\bf Figure \ref{fig:pointCMD}}
displays F200W vs (F150W--F200W) CMDs of the point sources in the central and parallel fields.
The CMD of the central field shows clearly a strong concentration of the point sources with GC colors ($-0.2<(F150W-F200W)_0 <0.6$)
in the magnitude range of $26.5<F200W_0 < 30$ mag.
In contrast, a much smaller number of the sources are located in the same region of the CMD of the parallel field.
This indicates that most of these sources in the central field belong to SMACS 0723.
We plot also the GCs and other compact source candidates in the Image 2 field of
Sparkler galaxy ($z=1.38$) located at $R=15\arcsec$ (82 kpc) east of the BCG in \citet{mow22}.
We cross-matched the compact sources in \citet{mow22} with our photometry and found that all the sources (N=12) were successfully recovered. However, only five of them passed our point source selection criteria. This indicates that a significant fraction of the compact sources in \citet{mow22} are at least marginally resolved or do not follow the clean JWST/NIRCam PSFs.
These sources are located in the similar GC color range, and the matched GCs occupy the brightest part of the GCs ($27<F200W_0 <28$ mag).
Note that there are dozens more sources in the similar CMD region
($26.5<F200W_0 <28$ mag)
of the central field, but few are in the CMD of the parallel field. They are probably UCDs, but some of them may be candidates for
background GCs that were gravitationally lensed like the Sparkler galaxy GCs.
The brightest of them are as bright as F200W$_0 \approx 26.5$ mag.
In {\bf Figure \ref{fig:pointCMD}} we also plot the color histograms of the point sources with
$27.5 < F200W_0 <29.5$
mag in the central (thin red line) and parallel (thick black line) fields. Assuming the parallel field as the background, we subtract the contribution of the background from the color histogram of the central field. The resulting net color histogram of the central field (thick red line) shows a significant excess component that is close to Gaussian with a peak at $(F150W-F200W)_0 \approx 0.2$.
This indicates that these point sources are mostly genuine GCs which belong to SMACS 0723.
We select the GC candidate sample using the cyan box marked in the figure ($-0.2<(F150W-F200W)_0 <+0.6$ and $27.5<F200W_0 <29.5$ mag) for the following analysis.
We conservatively selected the GC candidates with photometric completeness higher than ~70\% ($F200W_0 < 29.5$ mag).
The number of the selected GC candidates in the two fields is about 2500.
{\bf Figure \ref{fig:finder}(c,d, and e)}
display the zoomed-in images of the sample section of the BCG region that show these GCs better after subtracting the smooth background image from the original image.
The apparent magnitude range
($27.5<F200W_0 <29.5$ mag) of these sources at the cluster distance corresponds to the absolute magnitude range of
$-14.2<M_{F200W}<-12.2$ mag,
so these GCs belong to the bright end of the GCs, which include the UCDs.
Considering the redshift of SMACS 0723, the ages of these GCs are younger than 9.5 Gyr.
The ages of most Milky Way GCs range from 11 Gyr to 13 Gyr, with a mean value of $12.3\pm0.4$ Gyrs \citep{oli20,kan21}).
Thus the SMACS 0723 GCs may be about 8 Gyrs old, assuming that they formed as the Milky Way GCs.
We select two representative regions to check any difference between galaxy GCs and intracluster GCs: the BCG region (at $R<15\arcsec$ from the BCG center), and the intracluster GC region (with a radius of $15\arcsec$) that covers the western loop \citep{mah22,pas22,mon22} and no bright galaxies, as marked in {\bf Figure \ref{fig:finder}(a)}.
In {\bf Figure \ref{fig:GCCMD}(a,b)}
we display F200W vs (F150W--F200W) CMDs
of the point sources in the BCG region and the intracluster GC region.
We also plot the color histograms of both regions in {\bf Figure \ref{fig:GCCMD}(c)}.
We derived the color histograms for the BCG and intracluster GC regions after subtracting the background contribution based on the parallel field.
Note that the color distribution of the intracluster GC region shows a strong peak at $(F150W-F200W)_0 =0.05$ and a weaker component at $(F150W-F200W)_0 \approx 0.4$.
In contrast the color distribution of the BCG region shows a strong peak at $(F150W-F200W)_0 =0.15$, which is 0.1 redder than the peak in the intracluster GC region.
We calculate the blue GC fractions, obtaining f(blue GC)=65\% for the intracluster GC region and 54\% for the BCG region.
Thus the intracluster region has a higher fraction of blue GCs than the BCG region.
The color distribution for the intracluster GC region is broad across the peak color, showing that there are not only blue GCs but also red GCs, although the blue GC fraction is higher. The presence of both subpopulations in this region implies that the origin of the intracluster GCs is dual: the dominant progenitors of the intracluster GCs are low mass dwarf galaxies, and the minor progenitors are massive galaxies, as discussed in the case of Virgo intracluster GCs by \citet{lee10}.
Considering these, we divide the GCs in the entire field into two subpopulations according to their colors:
the blue (metal-poor) GCs with $-0.2<(F150W-F200W)_0 <+0.2$ and the red (metal-rich) GCs with $+0.2<(F150W-F200W)_0 <+0.6$.
We use these blue and red GCs for the following analysis to study any difference between the two subpopulations.
\subsection{Spatial Distributions of the GCs}
{\bf Figure \ref{fig:GCmap}(a)} shows the spatial distribution of the GC candidates selected in the central field.
In {\bf Figure \ref{fig:GCmap}(b)} we plot the number density contours and color gradient map of these sources on the gray scale map of the F200W image. The contours were smoothed with a Gaussian smoothing scale, $\sigma_G \simeq3\arcsec$.
The lowest level in the contours starts with $2\sigma_{bg}$ where $\sigma_{bg}$ is estimated from the background fluctuation of the selected GC candidates in the low density region of the parallel field: $\sigma_{bg} = 0.00052 \pm 0.000029$ per kpc$^2$.
The step between contours is $4\sigma_{bg}$.
We also mark the locations of the known cluster members, the selected brightest galaxies (including the BCG), the gravitational lens image sources \citep{mah22}, and the red sequence galaxies
selected in this study,
to check any spatial correlation with the GCs.
Several features are noted in this figure.
First, the overall distribution of the GC candidates shows clearly a central concentration around the BCG that is located close to the galaxy cluster center. This implies that most of the selected GC candidates are the members of SMACS 0723, indeed.
The strongest peak of the GC number density is slightly offset to the east direction from the BCG center, and this can be explained as follows.
There are two galaxies close to the BCG center in the west direction so it is difficult to subtract completely the stellar light of these galaxies in the central region. This decreases the detection rate of the GCs in that region, causing the slight offset from the BCG center. However the center of the contours at $R>10\arcsec$ is consistent with the BCG center.
Second, the overall distribution of the GCs is elongated along the major axis of the BCG.
Third, the spatial distribution of the GCs is much more extended than that of the stellar light of the BCG.
Fourth, strong peaks are found at the position of some brightest galaxies. The second strongest peak of the GC number density is seen at the position of G2, a bright elliptical member galaxy
($z=0.3841$)
at $R\approx 190$ kpc in the northwest of the BCG.
Another weaker peak is seen at the position of G3, a bright galaxy close to G2.
A recognizable clumping of GCs is seen also at the position of G4, another elliptical member galaxy ($z=0.3845$).
There is an isolated weak GC clump at $R\approx 300$ kpc south of the BCG, at the center of which one bright red sequence galaxy is located.
The rest of the known cluster member galaxies do not show any strong clumps of GCs.
Fifth, there is a strong concentration of GCs at $R\approx 200$ kpc in the east of the BCG (called as the GC Clump 1 (GCC 1)). Interestingly there are no recognizable bright galaxies which can be a host of these GCs.
Similarly there is a weak concentration of GCs at $R\approx 100$ kpc in the west of the BCG (called as the GC Clump 2 (GCC 2)). There are no recognizable bright galaxies which can be a host of these GCs.
Sixth, there is seen a concentration of GCs extended to the southwest from the second strongest G2 peak.
This can be a part of the large western loop-like structure, which will be described in detail later.
We selected the intracluster GC region to cover this substructure, as marked by the large circle in the figure.
Finally, the GCs located outside the few bright red sequence galaxies (including the BCG) are considered to be mostly intracluster GCs.
In {\bf Figure \ref{fig:GCmapBR}} we display the spatial distributions of the blue GC and red GC candidates selected in the central field.
This figure shows that the central concentration of the red GCs around the BCG is more prominent than that of the blue GCs, which will be also shown by their radial number density profiles in the following.
In addition, these maps show a loop-like structure with a radius of 100 kpc at $R\approx 200$ kpc west of the BCG. The structure is seen better in the red GC map.
This will be compared with the substructure in the ICL map in the following.
\subsection{Radial Number Density Profiles of the GCs}
Adopting the BCG center as the cluster center we derive the radial number density profile of the GCs, displaying it in {\bf Figure \ref{fig:GCRDP}(a)}.
The number density profiles of the GCs derived before background subtraction show a flattening at $R>90\arcsec$ in the parallel field.
We estimate the background level from the sources at $R>90\arcsec$ in the parallel field,
obtaining $\Sigma_{bg}= 0.000831\pm0.000075$ per kpc$^2$.
In the figure the number density of the GCs deceases as the cluster-centric distance increases, confirming again that these GCs mostly belong to the galaxy cluster. The radial number density of the GCs keeps decreasing until $R\approx 600$ kpc.
The central excess component represents the BCG GCs, while the outer part is dominated by the intracluster GCs.
In {\bf Figure \ref{fig:GCRDP}(b)} we also plot the radial number density profiles of the blue and red GCs.
The radial profile of the red GCs shows a stronger excess in the central region than the blue GCs,
and it shows an opposite trend in the outer region.
We fit the radial number density profiles of the GCs
with the two component S\'ersic law (the BCG component plus the intracluster GC component): $\Sigma_{GC}(R) = \Sigma_{\rm eff} \exp{ \{-b_{n}[(R/R_{\rm eff})^{1/n} - 1]\} }$
and $(b_{n} = 2n - 0.3271)$ \citep{ser63,gra05}.
We also fit the data with the power law for the outer region of the intracluster GC component ($158<R<631$):
$\Sigma_{GC}(R)\propto R^\alpha$.
Table \ref{tab:fitRDP} lists a summary of the fitting results.
The radial profile of all GCs is fit well with the S\'ersic parameters,
$R_{\rm eff,1}=28.5\pm3.2$ pc and $n_1=0.1\pm0.1$ for the BCG component, and $R_{\rm eff,2}=224.5\pm22.1$ pc, $n_2=1.3\pm0.3$ for the intracluster component.
The outer region
is also fit well with a power law index $\alpha=-2.0\pm0.2$.
The radial profiles of the blue GCs and red GCs are also fit with similar parameters:
in the case of the blue GC, $R_{\rm eff,1}=29.5\pm6.5$ pc, $n_1=0.1\pm0.2$ for the BCG component, and $R_{\rm eff,2}=213.7\pm22.1$ pc, $n_2=1.1\pm0.3$ for the intracluster GC component, and $\alpha=-2.2\pm0.2$ for the outer region;
in the case of the red GCs, $R_{\rm eff,1}=26.7\pm3.4$ pc, $n_1=0.1\pm0.1$ for the BCG component $R_{\rm eff,2}=241.7\pm48.6$ pc, $n_2=1.5\pm0.6$ for the intracluster GC component, and $\alpha=-1.9\pm0.3$ for the outer region.
Thus the effective radius of the intracluster GC system is about eight times larger than that of the BCG GC system.
The outermost region of the intracluster GC system
is also represented well by the power law with an index of $\alpha = -2.0\pm0.2$. This profile is slightly steeper than that of the intracluster GC system in the Virgo cluster, $\alpha= -1.5\pm0.1$\citep{lee10}.
\section{Discussion}
\subsection{Comparison with ICL Distribution}
From the JWST/NIRCam images of SMACS 0723, \citet{pas22} and \citet{mah22} derived a smooth map of the diffuse light using a median box car filter with $21 \times 21$ pixels and $100 \times 100$ pixels, respectively, showing the distribution of the ICL.
They noted that the diffuse ICL is elongated along the major axis of the BCG, but in a much more extended way. Also they found two interesting substructures of the ICL: a large scale loop in the west of the BCG and a large lobe-like feature in the east component.
To compare the distributions of the GCs and ICL, we derive similarly a smooth map of the diffuse light applying a median boxcar smoothing with $51 \times 51$ pixels to the F277W image,
and display it in
{\bf Figure \ref{fig:ICLmap}(a)}.
We overlay the GC number density contour maps (all GCs, blue GCs, and red GCs) to the smooth map of the ICL
in {\bf Figure \ref{fig:ICLmap}(b,c, and d)}.
We also mark the positions of
bright stars, known cluster member galaxies, and gravitational lens image sources \citep{mah22} in the figure.
{\bf Figure \ref{fig:ICLmap}} shows several distinguishable features.
First, the spatial distributions of the GCs and ICL are consistent in general.
However, the GCs are found farther than the ICL boundary.
Second, the western large scale loop structure seen in the GC number density map
is consistent with that of the ICL. A number of GCs are located along the loop of the diffuse ICL. %
Third, the GC number density map shows also an extended structure in the east of the BCG, which is consistent with the lobe-like diffuse component of the ICL. However, the GC Clump 1 (the strongest clump of GCs at $R\approx 200$ kpc east of the BCG) does not have a bright counterpart in the ICL map, although it is close to the boundary of the east loop.
The GC Clump 1 is better seen in the blue GC map, indicating that it is composed mainly of the blue GCs.
\subsection{Formation of SMACS 0723}
The existence of a number of substructures with varying scales (in addition to the BCG) in the GC map and ICL map indicates that a large number of galaxy groups with diverse masses are falling to the cluster center, and they are in the middle of merging process now.
The elongated structure seen in the maps of the GCs and ICL indicates that most galaxies are falling to the cluster center along the east-west direction.
These maps illustrate the distribution of the stars and GCs that are stripped off from their host galaxies or are the remnant of disrupted galaxies.
Considering the distributions of the GCs and ICL, G2 is probably a main host of a galaxy group that includes G3 (called the G2 group).
{\bf Figure \ref{fig:ICLmap}} shows that the G2 group and the BCG are connected not only in the GC map but also in the ICL map. The existence of the GCs and diffuse light between the G2 group and the BCG implies that both systems are physically interacting. This is similar to the case of the NGC 4839 group that is known to be falling to the Coma center, as shown in the observations of galaxies and various ICM including the GCs
\citep{bri92, whi93,col96,lys19,chu21,oh22}.
The loop structure might have formed when the G2 group is falling to the cluster center, which needs to be investigated further in the future.
\subsection{Comparison with X-ray Brightness Distribution}
Chandra X-ray brightness map of SMACS 0723 is given in \citet{mah22} (their Figs. 1 and 9).
Comparing the GC number density map in this study with the X-ray map of SMACS 0723 given in \citet{mah22},
we find both maps display generally similar structures in the sense that both show a large scale elongated distribution.
Interestingly the X-ray contour map of the central region of SMACS 0723 (Fig. 1 in \citet{mah22}) shows
a distinguishable component at the position of the western loop structure in addition to the strongest peak at the BCG center.
The location of this component is consistent with the center of the western loop in the GC number density map and ICL map.
\subsection{Comparison with Dark Matter Distribution}
Dark matter mass density maps (and radial profiles) of SMACS 0723 derived from the strong lensing analysis were presented by \citet{mah22,pas22,cam22, gol22} who used the JWST/NIRCam images
and HST/ACS/WFC3 images.
In {\bf Figure \ref{fig:compmassmap}} we overlay the GC number density contour maps on the dark matter surface mass density map (the convergence ($\kappa$=surface mass density normalized to the critical mass density) map) provided by \citep{mah22} (pseudo color maps).
The GC maps and dark matter mass maps show remarkably similar structures in the sense that both show a large scale structure that is elongated along the major axis of the BCG, but much more extended than the BCG. They show also similar substructures.
\citet{mah22} present a comparison of the dark matter radial density profiles of SMACS 0723 they derived from the JWST/NIRCam images with those derived using different lensing models in the literature:
RELICS-lenstool \citep{coe19}, RELICS-GLAFIC \citep{coe19}, LTM \citep{gol22} (their Fig. 6).
In this comparison, the first three models produce similar profiles extending out to $R\approx 1$ Mpc, while the LTM profile is much more steeply declining in the outer region ($R\approx 500$ kpc).
In {\bf Figure \ref{fig:compRDP}} we present a comparison of the radial number density profiles of the GCs with the dark matter mass density profiles based on the lensing analysis \citep{mah22}.
This figure shows that the radial density profile of the GCs is very similar to those of the dark matter in the inner range of $R=50-200$ kpc. However it becomes steeper in the outer region at $R=200-600$ kpc, deviating from the three lens models (RELICS-lenstool, RELICS-GLAFIC, and \citet{mah22}) and getting closer to the LTM model. However, the outermost GC profile is between the two model groups.
These results are consistent with those for the GCs found in the deep HST images of Abell 2744, another massive cluster at $z=0.308$ that is known as one of the most GC-abundant galaxy clusters \citep{lee16a}.
Similar trends are also seen in the radial profiles of the intracluster GC number density and the dark matter mass derived from the
simulated galaxy clusters selected from the cosmological simulations \citep{ram18, ram20}.
\section{Summary}
In the search of the GCs in the massive galaxy cluster SMACS 0723 using the early released JWST/NIRCam images, we find a large population of GCs which spread over the cluster. These GCs are mostly intracluster GCs as well as the BCG GCs.
The ages of these GCs, considering the cluster redshift, are younger than 9.5 Gyr. These GCs are about 8 Gyr old, if they formed as the Milky Way GCs.
Primary results are summarized as follows.
\begin{enumerate}
\item The F200W magnitudes of these GCs, $26.5<F200W_0 <29.5$ mag, correspond to $-15.2<M_{F200W}<-12.2$ mag, showing that they belong to the brightest GCs (including the UCDs)).
\item The spatial distributions of these GCs show a megaparsec-scale structure that is elongated along the major axis of the brightest cluster galaxy, but is much extended than the stellar light of the BCG.
In addition, they show a large number of substructures with various scales,
some of which are consistent with the substructures in the map of the diffuse ICL.
\item The existence of the GCs and ICL between the G2 group and the BCG indicates that the G2 group is falling to the cluster center, as in the case of the NGC 4839 group in the Coma cluster.
The large scale loop structure in the west of the BCG might have formed when the G2 group is falling to the cluster center.
\item The GC number density map is, in general, consistent with that of the dark matter mass density map based on the strong lensing analysis in the literature in the sense both show a large-scale elongated structure.
\item The radial number density profiles of the GCs
are fit reasonably by the two component S\'ersic law: the BCG component and the intracluster GC component. The effective radius of the intracluster GC system is about eight times larger than that of the BCG GC system.
\item The radial number density profile of the GCs
in the inner region at $R=50-200$ kpc is very similar to the dark matter mass profile obtained from strong lensing models.
However it becomes steeper in the outer region at $R=200-600$ kpc, deviating from the three lens models
and getting closer to the LTM model. However, the outermost GC profile is between the two model groups.
\end{enumerate}
All these results, based on the excellent data obtained with the JWST/NIRCam, show that the intracluster GCs are an excellent independent tool to probe the dark matter distribution in galaxy clusters as well as to reveal the cluster assembly history. SMACS 0723 is only the first case of distant galaxy clusters showing the power of the JWST, and many more will come in the JWST era.
\acknowledgments
The authors are grateful to the anonymous referee for useful comments. This work was supported by the National Research Foundation grant funded by the Korean Government (NRF-2019R1A2C2084019).
We thank Brian S. Cho for improving the English in the original manuscript.
This work is based on observations made with the
NASA/ESA/CSA James Webb Space Telescope.
The data
were obtained from the Mikulski Archive for Space Telescopes
at the Space Telescope Science Institute, which is
operated by the Association of Universities for Research
in Astronomy, Inc., under NASA contract NAS 5-03127
for JWST.
\vspace{15mm}
\facilities{JWST}
\software{ Numpy \citep{har20b}, Matplotlib \citep{hun07}, Scipy \citep{vir20}, Astropy \citep{ast13, ast18}, IRAF\citep{tod86}, PyRAF \citep{sts12},
Photutils \citep{bra21}, Astroalign \citep{ber20}, DAOPHOT \citep{ste87}, and SExtractor \citep{ber96} }
Some/all of the data presented in this paper were obtained from the Mikulski Archive for Space Telescopes (MAST) at the Space Telescope Science Institute. The specific observations analyzed can be accessed via \dataset[DOI]{https://doi.org/DOI}.
\clearpage
|
\section{Introduction}
Federated Learning (FL) is a distributed learning paradigm for training a global model from data scattered across different clients~\cite{konevcny2015federated}.
During the training process, all the clients need to operate data locally and transfer the model updates between servers and themselves back and forth.
Such a training process may raise many challenges, with communication cost often being the critical bottleneck~\cite{kairouz2021advances}.
Many studies have found that different clients might transfer similar (or redundant) model updates to the server, which is a waste of communication costs~\cite{balakrishnan2022diverse,pmlr-v139-fraboni21a,karimireddy2020scaffold}.
Client selection schemes are widely adopted to reduce the waste of communication costs, e.g., Federated Averaging (\texttt{FedAvg}) where only the randomly-selected subset of clients transfer their model update to the server instead of yielding all clients involved~\cite{mcmahan2017communication:Lixto}.
\texttt{FedAvg} can maintain the learning efficiency with reduced communication costs, since the random selection schemes can reduce the redundant update transmission.
When \texttt{FedAvg} meets client sets with low similarity, the server can not accurately pick out the representative clients.
The large variance of the update gradient aggregated from the unrepresentative subset directly slows the training convergence~\cite{pmlr-v139-fraboni21a,katharopoulos2018not}.
To accelerate the training convergence by variance reduction, many client selection criteria have been proposed in recent literature, e.g., importance sampling, where the probabilities for clients to be selected is proportional to their importance measured by the norm of update~\cite{chen2020optimal}, data variability~\cite{rizk2021optimal}, and test accuracy~\cite{mohammed2020budgeted}.
However, importance sampling could not effectively capture the similarities among the clients.
As shown in Figure~\ref{Difference}\gdhBlue{(a)}, applying importance sampling could cause learning inefficiency as the clients transfer excessive important yet similar updates to the server.
To make better use of the similarities among clients, some researchers propose raw gradient-based cluster sampling schemes that group the clients with similar gradient together~\cite{pmlr-v139-fraboni21a,muhammad2020fedfast}, shown in Figure~\ref{Difference}\gdhBlue{(b)}.
If the clustering effect is good enough, cluster sampling can easily pick out the representative clients.
Unfortunately, the effect of raw gradient-based clustering methods faces the following problems:
\textbf{(i)~Poor Effect}, the high dimension gradient of a client is too complicated to be an appropriate cluster feature, which can not bring a good clustering effect.
\textbf{(ii) Effect Fluctuation}, due to the limitations of the clustering algorithm, when operating the client set with different parameters settings, e.g., different numbers of clusters, the clustering effect tends to fluctuate greatly.
After clustering with poor and fluctuant effect, there always exists some clusters with low client similarity.
Applying client selection in the clusters with low client similarity may slow the overall training convergence.
Besides, both the importance and cluster sampling methods require all the clients to return the raw gradient information for better selection, which runs against the communication reduction objective.
\begin{figure}[H]
\centering
\includegraphics[width=12.8cm]{./Difference.pdf}
\caption{The differences among client selection schemes. The heights represent the norm of the gradients. The different colors denote different clusters. We highlight the main difference by red bold font.}
\label{Difference}
\end{figure}
\vspace{-1em}
We propose \textbf{\texttt{HCSFed}}, a novel \underline{\textbf{H}}ybrid \underline{\textbf{C}}lustering-based client \underline{\textbf{S}}election scheme for heterogeneous \underline{\textbf{Fed}}erated learning that further accelerates the FL convergence.
To improve the \textbf{poor clustering effect}, different from using the high dimension gradient, we adopt the compressed gradient as the cluster feature.
The different components of the gradient are grouped based on their numerical value, only the center of the group are retained to form the compressed gradient.
Adopting the compressed gradient as cluster feature not only allows better clustering effect but also is communication-efficient, since the compression operation filters out the redundant information within each gradient.
Besides, to control the \textbf{effect fluctuation}, we reset the number of clients to be selected in each cluster based on the client similarity and the size of the cluster, which we term as sample size re-allocation.
As shown in Figure~\ref{Difference}\gdhBlue{(c)}, sample size re-allocation scheme makes the cluster with low similarity have more number of clients to be selected.
The sampling effect becomes stable because the poor effect of random sampling in the clusters with low similarity is greatly mitigated by sample size re-allocation scheme.
Theoretically, we demonstrate the improvement of the proposed scheme in variance reduction.
Extensive experiments show the exceed efficiency of the proposed scheme compared to the cutting-edge FL client selection criteria in both convex and non-convex settings.
\section{System Setup and Prior Work}
\paragraph{The Federated Learning Optimization Settings.}
In federated learning, a total number of $N$ clients aim to jointly solve the following distributed optimization model:
\begin{eqnarray} \label{eq1}
\min_{\mathbf{w}\in \mathbb{R}^{d}}\left\{F(\mathbf{w}):=\sum^{N}_{k=1}\omega_k \left(F_k(\mathbf{w}):=\frac{1}{n_k}\sum^{n_k}_{j=1}f(\mathbf{w};x_{k,j})\right)\right\},
\end{eqnarray}
where $\omega_k$ is the weight of the $k^{th}$ client, $s.t.~\omega_k \ge 0$ and $\sum_{k=1}^N\omega_k=1$.
Suppose the $k^{th}$ client possesses $n_k$ training data: $\{x_{k,1},~x_{k,2},~...~,~x_{k,n_k}\}$.
The local objective function $F_k(\mathbf{w})$ is defined as the average of $f(\mathbf{w};x_{k,j})$, which is a user-specified loss function (possibly non-convex) made with model parameters $\mathbf{w}\in \mathbb{R}^{d}$ and training data $x_{k,j}$.
\paragraph{\texttt{FedAvg} Description.}
In the $t^{th}$ train round of the \texttt{FedAvg}, the server broadcasts the latest global model $\mathbf{w}_{t}$ to all the clients.
The $k^{th}$ client sets $\mathbf{w}_{t}^{k} \leftarrow \mathbf{w}_{t}$ and performs local training for $E$ epochs:
\begin{equation} \label{eq2}
\mathbf{w}_{t+1}^{k} \longleftarrow \mathbf{w}_{t}^{k}-\eta_{t} \nabla F_{k}\left(\mathbf{w}_{t}^{k}, \xi_{t}^{k}\right),
\end{equation}
where $\eta_{t}$ is the learning rate and $\xi_{t}^{k}$ is a data sample uniformly selected from the local data.
In the communication round, the server aggregates the local model update to produce the new global model $\mathbf{w}_{t+E}$.
In fact, to reduce the communication cost, \texttt{FedAvg} randomly generates a subset $\mathcal{S}_t$ consisting of $m$ clients from the entire client set and aggregates the model update, $\{\mathbf{w}_{t+E}^{1}, \mathbf{w}_{t+E}^{2}, \cdots, \mathbf{w}_{t+E}^{m}\}$,
\begin{equation}
\mathbf{w}_{t+E} \longleftarrow \underbrace{\sum_{k=1}^{N} \omega_{k} \mathbf{w}_{t+E}^{k}}_{\mathrm{Full~Participate}}\longleftarrow \underbrace{\sum_{k \in \mathcal {S}_{t}} \frac{N}{m}\omega_{k} \mathbf{w}_{t+E}^{k}}_{\mathrm{Partial~Participate}}.
\end{equation}
\texttt{LAG}~\cite{chen2018lag} is a improved aggregation scheme where it triggers the reuse of outdated gradients for the non-selected clients, but the variance of the aggregated gradient is still large, which directly slows the FL convergence.
To accelerate the convergence by variance reduction, many client selection criteria have been proposed in recent literature, including importance-based and clustering-based methods.
\vspace{-1em}
\paragraph{Importance Sampling.}Importance sampling is a non-uniform client selection method, where the probabilities for clients to be chosen are proportional to their importance~\cite{katharopoulos2018not,zhao2015stochastic}.
Different methods to measure the importance are proposed, including data variability~\cite{rizk2021optimal}, norm of update~\cite{chen2020optimal}, test accuracy~\cite{mohammed2020budgeted}, local rounds~\cite{singh2019sparq}, and local loss~\cite{cho2020client}.
Such methods can yield better selection by activating the clients with more importance.
However, they could not capture the similarities among the updates of clients and, therefore, trigger the communication redundancy.
\vspace{-1em}
\paragraph{Cluster Sampling.}To make better use of similarities among clients, some researchers propose raw gradient-based cluster sampling schemes that group the clients with similar gradients together~\cite{pmlr-v139-fraboni21a,muhammad2020fedfast}.
However, the effect of cluster sampling strongly relies on the performance of the clustering process.
Unfortunately, the high dimension gradient of a client is too complicated to be an appropriate cluster feature, which can not bring a good clustering effect.
Besides, meticulous one-by-one client selection~\cite{balakrishnan2022diverse,dieuleveut2021federated} is introduced to FL. It could achieve better selection outcomes but with extra expensive computation and communication costs for solving each noncommittal submodular maximization.
\vspace{-1em}
\paragraph{Difference from \texttt{SCAFFOLD}.}
While \texttt{SCAFFOLD}~\cite{karimireddy2020scaffold} appears to be similar to our method, there are fundamental differences.
\texttt{SCAFFOLD} augments the updates with extra “control-variate” item that is also transmitted along with the updates to reduce the variance, while we reduce the variance by selecting a more representative subset.
The former focuses on update control, while the latter focuses on better update selection, indicating that they are orthogonal and compatible with each other.
\section{\texttt{HCSFed}: Proposed Client Selection Scheme in Federated Learning}
\subsection{Analysis for the Optimization Objective of Client Selection Scheme}\label{sec31}
Inspired by the convergence analysis work~\cite{katharopoulos2018not}, we define the convergence speed as follows,
\begin{equation}\label{eq4}
\begin{split}
\mathrm{Speed}\triangleq&-\mathbb{E}\left[\left\|\mathbf{w}_{t+1}-\mathbf{w}^{*}\right\|_{2}^{2}-\left\|\mathbf{w}_{t}-\mathbf{w}^{*}\right\|_{2}^{2}\right]\\
=&-\mathbb{E}\left[\left(\mathbf{w}_{t}-\eta_{t} \nabla F_{k}\right)^{T}\left(\mathbf{w}_{t}-\eta_{t} \nabla F_{k}\right)+2 \eta_{t} \nabla F_{k}^{T} \mathbf{w}^{*}-\mathbf{w}_{t}^{T} \mathbf{w}_{t}\right] \\
=& \underbrace{\ 2\eta_{t}\left(\mathbf{w}_{t}-\mathbf{w}^{*}\right) \mathbb{E}\left[G_{t}\right]-\eta_{t}^{2} \mathbb{E}\left[G_{t}\right]^{T} \mathbb{E}\left[G_{t}\right]}_{Constant}-
\underbrace{\eta_{t}^{2} \operatorname{Tr}\left(\mathbb{V}\left[G_{t}\right]\right)}_{Changeable},
\end{split}
\end{equation}
where $G_{t}=\nabla F\left(\mathbf{w}_{t}, \mathcal{S}_{t}\right)$ and $\mathbb{E}\left[G_{t}\right]=\sum_{k=1}^{N} \nabla F_{k}\left(\mathbf{w}_{t}^{k}, x^{k}\right)$ is the unbiased estimate of the true update.
The first two terms are constant, since $\mathbb{E}\left[G_{t}\right]$ has a constant value.
Therefore, gaining a convergence speedup is equivalent to reducing the variance of gradient, i.e., $\mathbb{V}\left[{G}_{t}\right]$.
Consider a partial participation FL~\cite{li2020federated}, selecting $m$ clients can be regarded as selecting one client $m$ times by $m$ different discrete probability distributions $\left\{P_i\right\}_{i=1}^{m}$.
For each distribution, we have $P_{i} = \left\{ p_{k}^{i}\right\}_{k=1}^N$, where $p_{k}^{i}$ denotes the probability for the $k^{th}$ client to be selected in $i^{th}$ distribution.
For faster convergence, we seek to minimize the variance of model update $\mathbf{w}(\mathcal{S}_{t})$ aggregated from the subset that is equivalent to minimizing the variance of gradient corresponding to the following equation:
\begin{equation}
\min_{p_{k}^{i}}\left\{
\mathbb{V}\left[{G}_{t}\right]=\mathbb{V}\left[ \mathbf{w}(\mathcal{S}_t) \right] \right\}, s.t., \sum_{k=1}^{N} p_{k}^{i}=1 \text { with } p_{k}^{i} \geq 0~,~\left\Vert \mathcal{S}_t \right\Vert=q \cdot N=m,
\end{equation}
where we use $q$ and $N$ to denote the sampling ratio and the total number of all clients, respectively.
\subsection{Proposed Client Selection Scheme \texttt{HCSFed} for Variance Reduction}
How to explore and make use of the similarities among clients with acceptable resource costs is the key issue in constructing client selection criteria.
We propose \texttt{HCSFed}, a client selection scheme that exploits three components (cluster sampling, sample size re-allocation, importance sampling) guaranteeing convergence speedup by variance reduction.
\begin{wrapfigure}[18]{R}{0.50\textwidth}
\flushright
\vspace{-2.2em}
\begin{minipage}{1\linewidth}
\SetInd{0.3em}{0.4em}
\IncMargin{1em}
\begin{algorithm}[H]
\caption{ClientClustering}
\label{alg:algorithm1}
\KwIn{compressed gradient of all the clients in the $t^{th}$ round $\{X_{t}^{k}\}_{k=1}^{N}$}
\KwIn{The number of the clusters $\mathrm{H}$}
\textbf{Initialize} Randomly select H clients as cluster centers $\{\boldsymbol{\mu}_{1},\boldsymbol{\mu}_{2},\ldots, \boldsymbol{\mu}_{\mathrm{H}}\}$\;
\textbf{Initialize} $C_{i}=\varnothing\ (1 \leq i \leq \mathrm{H})$\;
\Repeat{$\forall\ i = \{1,2,\ldots, \mathrm{H}\}, \boldsymbol{\mu}_{i}^{\prime}=\boldsymbol{\mu}_{i}$}
{\For{each client k=1,2,\ldots,N}
{$\lambda_{k} = \arg \min _{i \in \{1,2,\ldots,\mathrm{H}\}} \left\|{X}_{t}^{k}-\boldsymbol{\mu}_{i}\right\|_{2}$\; $C_{\lambda_{k}}=C_{\lambda_{k}}\bigcup\left\{k^{th}~\mathrm{client~with}~X_{t}^{k}\right\}$\;
}
\For{each cluster i=1,2,\ldots,$\mathrm{H}$}
{New center $\boldsymbol{\mu}_{i}^{\prime}=\frac{1}{\left|C_{i}\right|} \sum_{{X_{t}^{k}} \in C_{i}}{X_{t}^{k}}$\;
{$\boldsymbol{\mu}_{i} \leftarrow \boldsymbol{\mu}_{i}^{\prime}$\;}
}
}
\KwOut{$ \mathcal{G}=\left\{C_{1}, C_{2}, \ldots, C_{\mathrm{H}}\right\}$\;}
\end{algorithm}
\end{minipage}
\end{wrapfigure}
\begin{figure}[H]
\vspace{-1em}
\centering
\includegraphics[width=13cm]{./PCA.pdf}
\caption{The visualizations of the raw gradient (left) and the clustered compressed gradient (right). Each point denotes the update gradient of one client. The $g_{i}$ is the $i^{th}$ component of the update gradient. The $c_{i}$ ci denotes the $i^{th}$ group of component. }
\label{fig1}
\end{figure}
\vspace{-1em}
\paragraph{The Compressed Gradient-Based Cluster Selection.}
To roughly explore and make use of the similarity among clients, we develop cluster selection that groups the similar clients together based on the client characteristic.
The following is a mathematical description of cluster selection.
Generating a subset $\mathcal{S}_t$ consisting of $m$ clients by clustering selection can be regarded as selecting $m$ clients using $\mathrm{H}$ different probability distributions $\left\{P_h \right\}_{h=1}^\mathrm{H}$~where~$P_h = \left\{p_{k}^{h} \right\}_{k=1}^N$.
For example, if we select a client from the $h^{th}$ cluster, the probability distribution $P_h$ will be used, and the $p_{k}^{h}$ comes as follows:
\begin{equation} \label{eq5}
p_{k}^{h} =
\begin{cases}
0, & \text{if }k^{th} \text{ client} \notin h^{th} \text{ cluster} \\
\frac{1}{N_{h}}, & \text{if }k^{th} \text{ client} \in h^{th} \text{ cluster}
\end{cases}
\end{equation}
where $N_h$ denotes the number of clients in the $h^{th}$ cluster $s.t. \sum_{h=1}^\mathrm{H} N_h = N$.
Prior cluster selections~\cite{pmlr-v139-fraboni21a} prefer to develop clustering based on the raw high dimension gradient (shown in Figure~\ref{fig1}), which is too complicated to distinguish clients and brings unacceptable communication costs running against the original objective.
Different from the previous raw gradient-based clustering methods, we develop clustering based on compressed gradient \cite{han2015deep} which is the information-preserving representation of the model update.
As shown in Figure~\ref{fig1}, different components of the gradient are grouped based
on their numerical value, only the center of the group are retained to form the compressed gradient.
The compression rate $R$ can be obtained by evaluating the quotient of $d$ and $d'$, i.e., $\frac{d'}{d}$, where $d'$ and $d$ are the total dimension of the compressed gradient and raw gradient, respectively.
Adopting the compressed gradient as cluster feature not only allows better clustering effect but also is
communication-efficient, since the compression operation filters out the redundant information within
each gradient.
The pseudo-code of Gradient Compress (GC) can be found in Appendix \ref{app:A.2}.
Algorithm \ref{alg:algorithm1} summarizes the main steps of client clustering.
Ideally, after clustering, the clients in the same cluster are similar to each other.
However, as shown in Figure~\ref{fig1}, some clusters always exist with low similarity (purple cluster), since the effect of the clustering process fluctuates, which may also lead to performance degradation.
\begin{wrapfigure}[25]{R}{0.51\textwidth}
\flushright
\vspace{-2em}
\begin{minipage}{1\linewidth}
\SetInd{0.3em}{0.4em}
\IncMargin{1em}
\begin{algorithm}[H]
\caption{\texttt{HCSFed}}
\label{alg:algorithm2}
\KwIn{updates in the $t^{th}$ round of all clients $\{G_{t}^{k}\}_{k=1}^{N}$}
\textbf{Initialize} $\mathbf{w}_{0}$\;
\textbf{Initialize} $\mathcal{S}_{t}=\varnothing$\;
\textbf{Server executes}:\\
\For{each round $t = 1,2, \cdots$}
{$m \leftarrow \max(qN,1)$\;
$\mathcal{G} \leftarrow \text {ClientClustering}\left(\mathbf{X}_{t},\mathrm{H}\right)$\;
\For{each cluster $h=1,2,\cdots,\mathrm{H}$}
{$m_h \leftarrow \frac{N_{h} S_{h}}{\sum_{h=1}^\mathrm{H} N_{h} S_{h}} \cdot m$\;
compute $\{p_{t}^{k}\}_{k \in N_{h}}$ using Equation \ref{eq7}\;
$\mathcal{S}_{t} = \mathcal{S}_{t} \bigcup m_{h}$ clients selected with $P_{h}$\;
}
\For{client $\in \mathcal{S}_{t}$ \textbf{in parallel}}
{$\mathbf{w}_{t+1}^k \leftarrow\;
\operatorname{ClientUpdate}(k,\mathbf{w}_{t}^{k})$\;
}
$\mathbf{w}_{t+1} \leftarrow \sum_{{k} \in \mathcal{S}_{t}} \frac{N}{m} \omega_{k} \mathbf{w}_{t+1}^{k}$\;
}
\textbf{ClientUpdate($k, \mathbf{w}_{t}^{k}$)}: \\
\For{each local epoch $i$ from $1\ \text{to}\ E$}
{$\mathbf{w}_{t+i+1}^{k}\longleftarrow\mathbf{w}_{t+i}^{k}-\eta \nabla F_{k}\left(\mathbf{w}_{t+i}^{k}, \xi_{t+i}^{k}\right)$\;}
\eIf{$k \in \mathcal{S}_{t}$}{\textbf{return} $\mathbf{w}^k_{t+1} \leftarrow \mathbf{w}^k_{t+E}$,
$X_{t}^{k} \leftarrow \mathrm{GC}(G_{t}^{k})$\;}
{\textbf{return} $X_{t}^{k} \leftarrow \mathrm{GC}(G_{t}^{k})$\;}
\end{algorithm}
\end{minipage}
\end{wrapfigure}
\paragraph{Sample Size Re-allocation Scheme.}Concentrating on the diverse heterogeneity of clusters, we make the first attempt to develop a sample size re-allocation scheme that pays more attention to the cluster with great heterogeneity.
\texttt{HCSFed} redetermines the sample size $m_{h}$, namely, the number of clients sampled from the $h^{th}$ cluster, by considering both the cluster size $N_h$ and variability.
\begin{equation}
\begin{split}
m_h=&\frac{N_{h} S_{h}}{\sum_{h=1}^\mathrm{H} N_{h} S_{h}} \cdot m, \\
\mathrm{where}~~S_h =&\frac{1}{N_h-1}\sum_{j=1, i \neq j}^{N_h} \left\| X_t^{i}-X_{t}^{j} \right\|_{2}^{2},
\end{split}
\label{eq6}
\end{equation}
$m_h$ denotes the number of clients in the subset $\mathcal{S}_t$ from the $h^{th}$ cluster $s.t. \sum_{h=1}^\mathrm{H}m_h=m.$
$S_h$ denotes the variability of the $h^{th}$ cluster where we use Cluster Cohesion based on the compressed model update to approximate.
The introduction of sample size re-allocation can further reduce the variance by assigning more sample chance to those clusters with greater heterogeneity.
Due to the limited number of clients to be selected, sometimes some clusters with great heterogeneity still can not have an adequate sample chance, which indeed need extra care.
\paragraph{Importance Selection.}
We introduce importance selection to optimize the probabilities for clients to be selected in the same cluster, which is based on the norm of the compressed gradient, i.e., $\mathrm{GC}(G_{t}^k)=X_{t}^{k}$.
We re-determine the probability for each client to be sampled in $t^{th}$ round as follows,
\begin{equation} \label{eq7}
p_{t}^{k} = \frac{\left\|\mathrm{GC}(G_{t}^k)\right\|}{\sum_{k=1}^{N_{h}}\left\|\mathrm{GC}(G_{t}^k)\right\|} = \frac{\left\|X_{t}^k\right\|}{\sum_{k=1}^{N_{h}}\left\|X_{t}^k\right\|},
\text{if}\ k^{th}\ \text{client} \in h^{th}\ \text{cluster in}~t^{th}~\text{round}.
\end{equation}
After completing the importance selection, the client with a higher norm of gradient (more importance) will have a higher probability to be sampled, which brings representative outcomes under an inadequate sampling ratio.
Importance selection provides a fine-grained optimization over cluster selection with sample size re-allocation, i.e., assigning more attention to the more representative clients.
These three ideas of selection discussed in this subsection together constitute the \texttt{HCSFed} scheme to optimize the client selection in FL.
\texttt{HCSFed} selects representative clients via allocating appropriate probability for each client to be selected, with the promise of reducing the variance between the model update aggregated from the sampled subset $\mathcal{S}_{t}$ and the entire set $\mathcal{K}$.
As we mentioned in Section \ref{sec31}, the reduced variance leads to faster convergence.
Therefore, \texttt{HCSFed} can achieve communication cost reduction, as it requires fewer rounds to approach the target accuracy.
Algorithm~\ref{alg:algorithm2} summarizes the main steps of \texttt{HCSFed}.
\section{Theoretical Analysis}\label{sec4}
\subsection{Variance Relationship among Different Selection Schemes}
\begin{theorem}[Variance Reduction]\label{th1}
If the population is large compared to the subset, $\frac{m}{N}$, $\frac{m_h}{N_h}$, $\frac{1}{m_h}$ and $\frac{1}{N}$ are negligible, then the selection (or cross-client) variance of different selection schemes satisfies
$$
\mathbb{V}\left({\mathbf{w}}_{hybrid}\right) \leq \mathbb{V}\left({\mathbf{w}}_{cludiv}\right) \leq \mathbb{V}\left({\mathbf{w}}_{cluster}\right) \leq \mathbb{V}\left({\mathbf{w}}_{rand}\right),
$$
where $\mathbf{w}_{hybrid},\mathbf{w}_{cludiv},\mathbf{w}_{cluster},\mathbf{w}_{rand}$ denote the model update aggregated from the subset $\mathcal{S}_t$ that generated by \texttt{HCSFed}, clustering selection scheme under re-allocation, clustering selection scheme under plain allocation and simple random selection scheme, respectively.
\end{theorem}
\paragraph{Proof Sketch.}
Below we provide a proof sketch to reveal the relationship among the cross-client variance of different client selection schemes, and then we naturally state when our \texttt{HCSFed} scheme could achieve variance reduction. We defer the details of proof to Appendix \ref{app:B}.
\textbf{Comparison of Random and Clustering Selection.}
We derive the Equation~\ref{eq9} to show the relationship between $\mathbb{V}\left({\mathbf{w}}_{cluster}\right)$ and $\mathbb{V}\left({\mathbf{w}}_{rand}\right)$:
\begin{equation}\label{eq9}
\mathbb{V}\left({\mathbf{w}}_{rand}\right) = \mathbb{V}\left({\mathbf{w}}_{cluster}\right) + \sum_{h=1}^{\mathrm{H}} \frac{N_h\left\|\mathbf{W}_{h} - \mathbf{W}(\mathcal{K})\right\|_{2}^2}{mN},
\end{equation}
where $\mathcal{K}$ denotes the set of all the clients. We have $\mathbb{V}\left(\mathbf{w}_{cluster}\right)<\mathbb{V}\left(\mathbf{w}_{rand}\right)$, unless $\forall h \in \{1, \cdots, \mathrm{H}\}, \mathbf{W}_{h} = \mathbf{W}(\mathcal{K}) $, i.e., each cluster has the same averaged model update with the entire population, which indicates that all the clusters are homogeneous in terms of the mean model update.
\textbf{Comparison of Plain Clustering and Clustering with Re-allocation.}
We derive the Equation~\ref{eq10} to show the relationship between $\mathbb{V}\left({\mathbf{w}}_{cludiv}\right)$ and $\mathbb{V}\left({\mathbf{w}}_{cluster}\right)$:
\begin{equation}\label{eq10}
\mathbb{V}\left({\mathbf{w}}_{cluster}\right) = \mathbb{V}\left({\mathbf{w}}_{cludiv}\right) + \sum_{h=1}^{\mathrm{H}} \frac{N_h({S_h} - {S})^2}{mN}.
\end{equation}
We have $\mathbb{V}\left(\mathbf{w}_{cludiv}\right) < \mathbb{V}\left(\mathbf{w}_{cluster}\right)$, unless $\forall h\in \{1, \cdots, \mathrm{H}\}, S_h={S}$, i.e., each cluster has equal variability, which indicates that all the clusters are homogeneous in terms of the variability.
\textbf{Comparison of Clustering with Re-allocation and \texttt{HCSFed}.}
We derive the Equation~\ref{eq11} to show the relationship between $\mathbb{V}\left({\mathbf{w}}_{cludiv}\right)$ and $\mathbb{V}\left({\mathbf{w}}_{hybrid}\right)$:
\begin{equation}\label{eq11}
\mathbb{V}\left({\mathbf{w}}_{cludiv}\right) = \mathbb{V}\left({\mathbf{w}}_{hybrid}\right) +\frac{\left(\sum_{i=1}^{N_{h}}\left\|G_{i}\right\|_{2}\right)^{2}}{N_{h}} \sum_{i=1}^{N_{h}}\left(I_{i}-\frac{1}{N_{h}}\right)^{2}.
\end{equation}
We have $\mathbb{V}\left(\mathbf{w}_{hybrid}\right) < \mathbb{V}\left(\mathbf{w}_{cludiv}\right)$, unless $\forall i\in \{1, \cdots, N_h\}, I_{i}=\frac{\left\|G_{i}\right\|_{2}}{\sum_{i=1}^{N_{h}} \left\|G_{i}\right\|_{2}}=\frac{1}{N_{h}}$, i.e., each client in $h^{th}$ cluster has equal norm of model update.
\begin{remark}
Theorem \ref{th1} verifies the capability of \texttt{HCSFed} in variance reduction.
%
The proof sketch indicates that the variance difference of different selections will vanish only if all the clusters and the norm of the gradients are identical.
%
The proposed approach can achieve variance reduction, since each cluster never has identical update mean, variability, and norm in heterogeneous FL.
\end{remark}
\subsection{Convergence Behavior of \texttt{HCSFed}}
We emphasize that the convergence analysis of \texttt{FedAvg} with random selection is \textbf{NOT} our contribution, we just follow the previous analysis work~\cite{li2019convergence} to verify that \texttt{HCSFed} could maintain the same convergence guarantees as a random selection in the worst case and in most cases, \texttt{HCSFed} will have tighter convergence compared with random selection.
We next state the assumptions used in our theorem and proof, which are common in the federated optimization literature, e.g.,~\cite{zhou2017convergence,li2019convergence,wang2021cooperative,stich2018local,haddadpour2019convergence,haddadpour2019local}.
Assume each function $F_{k}(k \in[N])$ is $\mu$-strongly convex and $L$-smooth.
Suppose that for all $k \in[N]$ and all $t$, the variance and expectation of stochastic gradients in each client on random samples $\xi$ are bounded by $\sigma_{k}^{2}$ and $G^2$, i.e., $\mathbb{E}\left[\left\|\nabla F_{k}\left(\mathbf{w}_{t}^{k}, \xi\right)-\nabla F_{k}\left(\mathbf{w}_{t}^{k}\right)\right\|_{2}^{2}\right] \leq \sigma_{k}^{2} $ and $\mathbb{E}\left[\left\|\nabla F_{k}^{2}\left(\mathbf{w}_{t}^{k}, \xi\right)\right\|_{2}^{2}\right] \leq G^{2}$, respectively. And $\Gamma$ is used to quantify the heterogeneity, where $\Gamma=F^{*}-\sum_{k=1}^{N} p_{k} F_{k}^{*}$.
\begin{theorem}[Convergence Bound]\label{th2}
Let Assumptions above hold and $L,\mu,\sigma_{k},G$ be defined therein. Consider \texttt{FedAvg} when sampling $m$ clients, then \texttt{HCSFed} satisfies:
\begin{align}
\mathbb{E}\left[F\left(\mathbf{w}_{T}\right)\right]-F^{*} \leq \underbrace{\mathcal{O}\left(\frac{\sum_{k=1}^{N} p_{k}^{2} \sigma_{k}^{2}+ E^{2} G^{2}+\gamma G^{2}}{\mu T}\right)
+ \mathcal{O}\left(\frac{L \Gamma}{\mu T}\right)}_{Bounding~\texttt{FedAvg}~with~Full~Participation} +\underbrace{\mathcal{O}\left(\frac{ E^{2} G^{2}}{m\mu T}\right)}_{Bounding~Variance}.
\end{align}
\end{theorem}
\begin{remark}
Theorem \ref{th2}~\cite{li2019convergence} gives the convergence upper bound of \texttt{FedAvg} with hybrid client selection.
%
The first two terms are used to bound the \texttt{FedAvg} with full client participant, while the last term is used to bound the variance of the model update aggregated from the subset. In Theorem~\ref{th1}, we demonstrate the effectiveness of \texttt{HCSFed} in variance reduction. \texttt{HCSFed} achieves a lower variance, which indicates that \texttt{HCSFed} could enjoy a tighter convergence bound.
\end{remark}
\begin{remark}
Theorem \ref{th2}~\cite{li2019convergence} uses Lemma \ref{lemma1}-\ref{lemma3} to bound the error of \texttt{FedAvg} with full participation corresponding to the first two terms, while using Lemma \ref{lemma4} to verify the unbiased property and Lemma \ref{lemma5} to bound the variance resulting from the client selection.
Theorem \ref{th1} shows the effectiveness of \texttt{HCSFed} to reduce selection variance, which also indicates that the variance of \texttt{HCSFed} is always lower than the mentioned selections unless the population is homogeneous.
Naturally, \texttt{HCSFed} satisfies the Lemma \ref{lemma5}.
The dependency between the Lemmas and Theorem \ref{th2} indicates that the proof the Lemma \ref{lemma4} is enough to maintain the convergence bound. We defer the proof to Appendix \ref{app:A.3}.
\end{remark}
We demonstrate the convergence guarantee of \texttt{HCSFed} under $\mu$-strongly convex assumption. As for the non-convex loss function, we refer to the previous proved proposition that \texttt{FedAvg} with any unbiased selections maintains the same FL convergence bound of random selection~(Theorem 2 in \cite{pmlr-v139-fraboni21a}).
\section{Experiment} \label{sec5}
\subsection{Experimental Setup}
We run logistic regression (convex) and a fully connected network with one hidden layer of 50 nodes (non-convex) on MNIST~\cite{lecun1998mnist}.
As for CIFAR-10~\cite{krizhevsky2009learning} and FMNIST~\cite{xiao2017fashion}, we use the same classifier of \texttt{FedAvg}~\cite{mcmahan2017communication:Lixto} composed of 3 convolutional and 2 fully connected layers.
We partition the dataset into 100 clients under both IID and non-IID data distribution.
IID data partition strategy employs the idea of random split to create a uniform federated dataset.
As for non-IID data partition, we use Dirichlet distribution to partition the entire client set.
Dirichlet distribution, i.e., $\operatorname{Dir}(\alpha)$, gives to each client the respective partitioning across labels by changing the value of $\alpha$.
Specifically, the lower the value of $\alpha$, the more heterogeneous the dataset is.
The schematic table of the Dirichlet distribution is presented in Appendix \ref{app:A.1}.
As for the hyperparameters, $N$ is the number of all clients, $q$ is the sampling ratio, $nSGD$ is the times of SGD running locally, $\eta$ is the learning rate, $B$ is the batch size, $T$ is the terminate round. For better comparison, we set all the hyperparameters following the popular FL optimization work \texttt{FedProx}~\cite{li2020federated}.
\begin{table}[bp]
\vspace{-1em}
\centering
\caption{Required rounds for different methods with convex model (logistic regression) to achieve 80$\%$ accuracy on non-IID MNIST and FMNIST. 200$+$ indicates 80 $\%$ accuracy was not reached after 200 rounds.}
\setlength{\tabcolsep}{0.3mm}
\scalebox{0.93}{
\begin{tabular}{llrcrcrc}
\toprule
\multirow{2}*{~} &\multirow{2}*{Methods} &\multicolumn{2}{c}{Sampling Ratio 10\%} &\multicolumn{2}{c}{Sampling Ratio 30\%}
&\multicolumn{2}{c}{Sampling Ratio 50\%}\\
\cmidrule(r){3-4} \cmidrule(r){5-6} \cmidrule(r){7-8}
~ &~ &\makecell[l]{Num. of Rounds} &\makecell[l]{Speedup}
&\makecell[l]{Num. of Rounds} &\makecell[l]{Speedup}
&\makecell[l]{Num. of Rounds} &\makecell[l]{Speedup}
\\ \midrule
\multirowcell{6}{MNIST} &Random & 96 \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./1.png}}
\end{minipage} & \color{gray}(1.0$\times$)
& 42 \begin{minipage}[b]{0.12\columnwidth}
\centering
\raisebox{-.25\height}{\includegraphics[width=\linewidth]{./6.png}}
\end{minipage}& \color{gray}(1.0$\times$)
& 28 \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./11.png}}
\end{minipage} & \color{gray}(1.0$\times$)
\\
&\texttt{SCAFFOLD} & 74 \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./2.png}}
\end{minipage} & \color{gray}(1.3$\times$)
& 36 \begin{minipage}[b]{0.12\columnwidth}
\centering
\raisebox{-.25\height}{\includegraphics[width=\linewidth]{./7.png}}
\end{minipage}& \color{gray}(1.2$\times$)
& 21 \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./12.png}}
\end{minipage} & \color{gray}(1.3$\times$)
\\
&Importance & 35 \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./3.png}}
\end{minipage} & \color{gray}(2.7$\times$)
& 24 \begin{minipage}[b]{0.12\columnwidth}
\centering
\raisebox{-.25\height}{\includegraphics[width=\linewidth]{./8.png}}
\end{minipage}& \color{gray}(1.8$\times$)
& 22 \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./13.png}}
\end{minipage} & \color{gray}(1.3$\times$)
\\
&Cluster & 31 \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./4.png}}
\end{minipage} & \color{gray}(3.1$\times$)
& 16 \begin{minipage}[b]{0.12\columnwidth}
\centering
\raisebox{-.25\height}{\includegraphics[width=\linewidth]{./9.png}}
\end{minipage}& \color{gray}(2.6$\times$)
& 8 \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./14.png}}
\end{minipage} & \color{gray}(3.5$\times$)
\\
&\textbf{\texttt{HCSFed}} & \color{red}\textbf{9} \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./5.png}}
\end{minipage} & \color{gray}(\textbf{10.7}$\times$)
& \color{red}\textbf{8} \begin{minipage}[b]{0.12\columnwidth}
\centering
\raisebox{-.25\height}{\includegraphics[width=\linewidth]{./10.png}}
\end{minipage}& \color{gray}(\textbf{5.2}$\times$)
& \color{red}\textbf{8} \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./15.png}}
\end{minipage} & \color{gray}(\textbf{3.5}$\times$)
\\ \cmidrule{2-8}
\multirowcell{6}{FMNIST} &Random & 200$+$ \begin{minipage}[b]{0.12\columnwidth}
\centering
\raisebox{-.25\height}{\includegraphics[width=\linewidth]{./f_1.png}}
\end{minipage}& \color{gray}(1.0$\times$) & 200$+$ \begin{minipage}[b]{0.12\columnwidth}
\centering
\raisebox{-.25\height}{\includegraphics[width=\linewidth]{./f_6.png}}
\end{minipage}& \color{gray}(1.0$\times$)
& 158 \begin{minipage}[b]{0.12\columnwidth}
\centering
\raisebox{-.25\height}{\includegraphics[width=\linewidth]{./f_11.png}}
\end{minipage}& \color{gray}(1.0$\times$) \\
&\texttt{SCAFFOLD} & 180 \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./f_2.png}}
\end{minipage} & \color{gray}(>1.1$\times$)
& 144 \begin{minipage}[b]{0.12\columnwidth}
\centering
\raisebox{-.25\height}{\includegraphics[width=\linewidth]{./f_7.png}}
\end{minipage}& \color{gray}(>1.4$\times$)
& 115 \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./f_12.png}}
\end{minipage} & \color{gray}(1.4$\times$)
\\
&Importance & 180 \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./f_3.png}}
\end{minipage} & \color{gray}(>1.1$\times$)
& 116 \begin{minipage}[b]{0.12\columnwidth}
\centering
\raisebox{-.25\height}{\includegraphics[width=\linewidth]{./f_8.png}}
\end{minipage}& \color{gray}(>1.7$\times$)
& 100 \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./f_13.png}}
\end{minipage} & \color{gray}(1.6$\times$)
\\
&Cluster & 145 \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./f_4.png}}
\end{minipage} & \color{gray}(>1.4$\times$)
& 104 \begin{minipage}[b]{0.12\columnwidth}
\centering
\raisebox{-.25\height}{\includegraphics[width=\linewidth]{./f_9.png}}
\end{minipage}& \color{gray}(>1.9$\times$)
& 75 \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./f_14.png}}
\end{minipage} & \color{gray}(2.1$\times$)
\\
&\textbf{\texttt{HCSFed}} & \color{red}\textbf{81} \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./f_5.png}}
\end{minipage} & \color{gray}(>\textbf{2.5}$\times$)
& \color{red}\textbf{80} \begin{minipage}[b]{0.12\columnwidth}
\centering
\raisebox{-.25\height}{\includegraphics[width=\linewidth]{./f_10.png}}
\end{minipage}& \color{gray}(>\textbf{2.5}$\times$)
& \color{red}\textbf{75} \begin{minipage}[b]{0.12\columnwidth}
\centering \raisebox{-.25\height}{\includegraphics[width=\linewidth]{./f_15.png}}
\end{minipage} & \color{gray}(\textbf{2.1}$\times$)
\\
\bottomrule
\end{tabular}
}
\label{table1}
\end{table}
\begin{figure}[tp]
\begin{minipage}[t]{0.33\linewidth}
\centering
\includegraphics[width=4.33cm]{./rounds_5.pdf}
\end{minipage}
\begin{minipage}[t]{0.33\linewidth}
\centering
\includegraphics[width=4.33cm]{./rounds_10.pdf}
\end{minipage}
\begin{minipage}[t]{0.33\linewidth}
\centering
\includegraphics[width=4.33cm]{./rounds_8.pdf}
\end{minipage}
\quad
\begin{minipage}[t]{0.33\linewidth}
\centering
\includegraphics[width=4.33cm]{./rounds_6.pdf}
\end{minipage}
\begin{minipage}[t]{0.33\linewidth}
\centering
\includegraphics[width=4.33cm]{./rounds_11.pdf}
\end{minipage}
\begin{minipage}[t]{0.33\linewidth}
\centering
\includegraphics[width=4.33cm]{./rounds_9.pdf}
\end{minipage}
\vspace{-1.0em}
\caption{Required rounds for random, cluster, importance sampling, \texttt{SCAFFOLD}, and \texttt{HCSFed} to achieve 60$\%$ accuracy with $q=0.1$, $N=100$, $nSGD=50$, $\eta=0.01$, $B=50$ on MNIST and FMNIST, with $\alpha \in \{0.01, 0.001\}$, $q=0.1$, $N=100$, $nSGD=80$, $\eta=0.05$, $B=50$ on CIFAR-10.}
\label{round1}
\vspace{-1.0em}
\end{figure}
\subsection{Main Results}
Our evaluation target is to compare the convergence speed of different methods in convex and non-convex FL context.
We choose image classification as our downstream task and compare \texttt{HCSFed} with simple random sampling~\cite{mcmahan2017communication:Lixto}, norm-based importance sampling~\cite{chen2020optimal}, cluster sampling~\cite{pmlr-v139-fraboni21a}, and \texttt{SCAFFOLD}~\cite{karimireddy2020scaffold}.
For better representation of the convergence speed, we follow the previous FL acceleration work setting~\cite{karimireddy2020scaffold} and report the required rounds for different methods with convex model to reach the target test accuracy in Table \ref{table1}.
We can observe that \texttt{HCSFed} consistently achieves the fastest convergence (has the fewest required rounds to reach the target accuracy) across all settings, surpassing not only the client selection schemes but also the SOTA variance-reduced method \texttt{SCAFFOLD}~\cite{karimireddy2020scaffold}.
We also run \texttt{HCSFed} with non-convex model on different datasets.
The result of non-convex experiment is presented in Figure~\ref{round1}.
\texttt{HCSFed} consistently requires the fewest rounds to reach the target accuracy.
All the results reveal that \texttt{HCSFed} can greatly accelerate the convergence, especially in the low sampling ratio and date heterogeneity case.
As we emphasized in the theoretical analysis, the improvement of \texttt{HCSFed} depends on the sampling ratio and the client heterogeneity.
The lower the sampling ratio and the more heterogeneous the dataset, the more improvement \texttt{HCSFed} can enjoy.
Naturally, such great improvement in convergence speed can be attributed to the two key components in \texttt{HCSFed}:
(i) Different from the previous raw gradient-based methods, we introduce gradient compression to improve the clustering effect by using the more expressive cluster feature representation.
(ii) To further guarantee the robustness of the sampling effect, we introduce a sample size re-allocation scheme and extra importance sampling to control the sampling effect fluctuation in the cluster with low client similarity.
For more details, please refer to Appendix \ref{app:C}.
\subsection{Parameter Sensitivity Study}
\paragraph{The Number of Clusters~(H).} This experiment demonstrates the sensitivity of \texttt{HCSFed} to the different numbers of clusters.
We run convex model-based \texttt{HCSFed} on non-IID MNIST with different numbers of clusters, i.e. with $\mathrm{H}\in\left\{5, 6, \cdots, 10\right\}$.
We report the best classification accuracy and the required rounds for \texttt{HCSFed} with different numbers of clusters to reach the best test accuracy in Figure \ref{Number of Clusters}.
\texttt{HCSFed} achieves stable classification accuracy and similar convergence speeds (measured by the required rounds to achieve the best accuracy), which highlights the robust effect and efficiency of \texttt{HCSFed}.
The results also reveal that the clustering effect of \texttt{HCSFed} is robust regardless of the number of clusters.
Such merit might attribute to the sample size re-allocation scheme or the importance sampling scheme that we use to control the clustering effect fluctuation.
\begin{figure}[htbp]
\centering
\subfigure[The impact of different numbers of clusters]{
\centering
\includegraphics[width=6.75cm]{./pic_sensitivity_2.pdf}
\label{Number of Clusters}
}
\subfigure[The impact of different compression rates]{
\centering
\includegraphics[width=6.75cm]{./pic_sensitivity_1.pdf}
\label{Compression rate}
}
\vspace{-1.0em}
\caption{Visualization of Sensitivity Study Results on non-IID MNIST.}
\label{Sensitivity}
\vspace{-1.0em}
\end{figure}
\paragraph{The Compression Rate~(R).} We also evaluate how different compression rates affect the global model performance.
We run convex model-based \texttt{HCSFed} on non-IID MNIST with different compression rates (R), i.e. with $\mathrm{R} \in \left\{5\%, 10\%, 20\%, 40\%, 100\% \right\}$.
We report the best accuracy and the required rounds of \texttt{HCSFed} with different compression rates to reach the best test accuracy in Figure~\ref{Compression rate}.
The results show that \texttt{HCSFed} achieves different convergence speed as the compression rate changes.
The noticeable information is that both too high and too low compression rates can not achieve good learning efficiency.
Intuitively, the low compression rate can not ensure the integrity of gradient information while the high compression rate is too complicated to learn efficiently.
When we set the compressed rate to $100\%$, i.e., without gradient compression, the results show that both the convergence speed and final accuracy of the model are not satisfactory, which confirms the effectiveness of gradient compression.
In practice, following the previous compression work~\cite{han2015deep}, we set the compression rate with the highest within-group sum of squares to get a good sampling effect.
\begin{figure}[htbp]
\centering
\subfigure[{The test accuracy of individual components}]{
\centering
\includegraphics[width=6.6cm]{./pic_ablation_revision_1.pdf}
\label{Ablation.1}
}
\hspace{-0.2cm}
\subfigure[{The impact of individual components}]{
\centering
\includegraphics[width=7cm]{./pic_ablation_revision_2.pdf}
\label{Ablation.2}
}
\vspace{-1.0em}
\caption{Visualization of Ablation Study Results on non-IID MNIST.}
\label{Ablation}
\vspace{-1.0em}
\end{figure}
\subsection{Ablation Study}
We run logistic regression (convex) on non-IID MNIST with $q = 0.1, nSGD = 50$ to analyse the individual contributions of different components of \texttt{HCSFed} in both final accuracy and convergence speed.
In Figure~\ref{Ablation.1}, \texttt{HCSFed} is represented as the green line starting from the accuracy of 10$\%$ and ending at 91$\%$, whereas \texttt{FedAvg} is the yellow line starting also from the same position as \texttt{HCSFed} but ending at 81$\%$ after 200 rounds, i.e., 10$\%$ less.
We report the required rounds for different methods to achieve 80$\%$ accuracy in Figure~\ref{Ablation.2}.
Compared with the plain \texttt{FedAvg}, \texttt{FedAvg} combined with cluster~(blue) or importance sampling~(purple) can accelerate the convergence speed (measured by the required rounds to achieve the target accuracy) by $2.6\times$ and $2.1\times$ times.
The experimental results confirm the important role of the two in improving the learning efficiency.
Impressively, compared with importance sampling, the improvement of the clustering process is more obvious.
This is because clustering can make better use of the correlation among clients, thereby selecting more representative clients to participate in model training.
We also compare \texttt{FedAvg} combined with cluster, \texttt{FedAvg} combined with cluster and re-allocation, \texttt{FedAvg} combined with cluster and importance sampling, and \texttt{HCSFed}.
All the latter methods can achieve a faster convergence in model training than the former, which validates the effectiveness of re-allocation and extra importance sampling.
These results indicate that each component of \texttt{HCSFed} is complementary to each other, and also show the effectiveness of the variance reduction theory in Section~\ref{sec4}.
\section{Concluding Remarks}\label{sec:section6}
In this paper, we propose \texttt{HCSFed}, a novel hybrid clustering-based selection scheme to accelerate the training convergence by variance reduction.
As we mentioned in Section \ref{sec31}, the reduced variance leads to faster convergence.
Therefore, \texttt{HCSFed} can achieve communication cost reduction, as it requires fewer rounds to approach the target training accuracy.
Theoretically, we demonstrate the improvement of the proposed scheme in variance reduction.
We present the convergence guarantee of our proposed approach under the convex assumption.
Experimental results demonstrate the superiority and the effectiveness of our method with both the convex and non-convex models.
In the future, the co-variance among different clients and unavailable client settings will be considered, since it may also lead to performance degradation in FL. In the theoretical aspect, following the latest convergence analysis~\cite{khaled2020tighter}, we plan to loose the widely used but strict assumption “bounded gradient” and convexity for more general analysis of the faster convergence achieved by \texttt{HCSFed}.
\clearpage
|
\section{Introduction}
Dipolar quantum gases, and especially dipolar Bose-Einstein condensates (BEC),
are gaining again significant attention \cite{arxive2201.02672}
since self-bound droplets consisting of $^{164}$Dy \cite{luPRL11dysprosium,Kadau2016,Schmitt2016,Ferrier2016, Ferrier2016_2,Ferrier2018}
and $^{166}$Er \cite{chomazPRX16} were realized in experiments.
In contrast to Bose mixtures, the competition between attractive
and repulsive parts of the interaction does not originate from the
interaction between the components of the mixture, but rather
from the dipole-dipole interaction itself, which
in general has repulsive and attractive regions.
As in Bose mixtures droplets, quantum fluctuations are the driving force behind
the stabilization of dipolar droplets, as confirmed by theory \cite{Wachtler_2016,bailliePRA16,bombinPRL17}.
More recently, even droplets consisting of dipolar mixtures have been realized
in experiments \cite{Trautmann2018,Durastante2020} and described with
beyond mean-field methods \cite{Bisset2021,Smith2021}.
In such droplets the components are not necessarily miscible, but can demix while staying self-bound.
In experiments the dipole moments of all atoms are aligned in parallel
by an external magnetic field of well-controlled strength
and direction. This provides a means to modify the anisotropy of the dipole-dipole interaction and
triggers the transition to a density-modulated, self-organized stripe phase, which shows supersolid properties \cite{Leonard2017,Tanzi2019supersolid,Zhang2019,Roccuzzo2019,Hertkorn2021,Tanzi2021science}.
Such a transition is also visible in the excitation spectrum of a dipolar BEC, where a so-called roton minimum emerges \cite{Santos2003,odellPRL03,chomazNatPhys18,Natale2019,Schmidt2021,Blakie2020}.
Just like a droplet, a density modulation is a state that is not stable in
in a mean-field approximation~\cite{Komineas2007}, but rather
stabilized by quantum fluctuations~\cite{Wenzel2018}.
In previous theoretical studies \cite{maciaPRL12} we observed density modulations
in the form of stripes in a two-dimensional dipolar
Bose gas with the polarization axis tilted with respect to the
perpendicular direction. In these studies,
where the tilt angle $\theta$ was small enough that
the dipole interaction stayed purely repulsive,
a very high density was required to reach the stripe phase, and no self-binding was involved.
In this work we investigate the formation of
stripes at much lower densities which can be realized in experiments with magnetic dipole moments.
We achieve this by increasing the tilting angle $\theta$ beyond a critical angle, where the projection of the dipole-dipole interaction
on the 2D plane becomes attractive, see fig.~\ref{FIG:dipole}, such that self-binding is possible. Density oscillations were observed experimentally \cite{Tanzi2019droplet,Bottcher2019stripes,Chomaz2019,Ilzhofer2021}.
Conventional mean-field theories are not capable of describing such situations,
and more powerful methods like extended mean-field with Lee-Huang-Yang corrections~\cite{Ferrier2016,chomazPRX16},
or Quantum Monte Carlo (QMC) techniques
\cite{Macia2014,Macia2016,bombinPRL17,Bottcher2019droplet,Kora2019} have to be applied. In this work we employ
the hypernetted-chain Euler-Lagrange (HNC-EL) method \cite{Kro86,KroTrieste,QMBT00Polls},
which incorporates correlations beyond
mean-field approaches especially for strongly-correlated and self-bound systems
\cite{hebenstreitPRA16,Staudinger2018} and requires a much lower computational effort than QMC.
\begin{figure}[h!]
\begin{center}
\includegraphics*[width=0.48\textwidth]{dipole.pdf}
\end{center}
\caption{
Sketch of the dipolar interaction in 2D,
with tilting angle $\theta$ and hard core parameter $C_\mathrm{h}$. The direction of the dipole moment
is shown by the black arrow and the hard core is implied by the black circle.
The repulsive and attractive regions of the compound interaction are
shown with red and blue colors, respectively.}
\label{FIG:dipole}
\end{figure}
\section{Methodology}
In the following we consider a dipolar Bose gas that is so tightly trapped in the
$z$-direction that we can assume particles are restricted to move in two dimensions,
taken to be the $xy$-plane. The Hamiltonian reads
\begin{align}
\hat{H} = - \frac{\hbar^2}{2m} \sum\limits_{i=1}^N \Delta_i
+ \sum\limits_{i<j} v(\Vec{r}_i-\Vec{r}_j)~,
\end{align}
with the interaction being the sum of the dipolar term and a respulsive core
\begin{align}
v(\Vec{r}) =\frac{C_\mathrm{{dd}}}{4 \pi}\left[\frac{1}{\left| \Vec{r} \right|^{3}} - \frac{3 \left(x \sin \theta \right)^2}{\left| \Vec{r} \right|^{5}} \right]
+ \frac{E_0 C_\mathrm{h}^{12}}{\left| \Vec{r} \right|^{12}}~.
\label{eq::dipole-dipole interaction}
\end{align}
In this expression $C_{dd}$ sets the strength of the dipolar interaction and is proportional
to the square of the (electric or magnetic) dipole moment. It is useful to define the characteristic length scale
$r_0=m C_\mathrm{dd}/\left( 4 \pi \hbar^2 \right)$
and the associated energy scale $E_0=\hbar^2/\left( mr_0^2 \right)$, which serve as units for our calculations.
All dipoles are polarized along a direction in the $xz$-plane that forms an angle $\theta$ with respect to
the $z$-axis, see Fig.~\ref{FIG:dipole}. With this geometry, the dipole-dipole interaction is
repulsive around the $y$-direction (red regions in fig. \ref{FIG:dipole}), but for large enough
$\theta>\theta_\mathrm{c}=\arcsin \left( 1/\sqrt{3} \right) \approx 0.61548$ an attractive region appears in the $x$-direction
(blue regions in fig. \ref{FIG:dipole}). In this work we
explore the highly tilted polarization regime $\theta>\theta_\mathrm{c}$
where the purely dipolar gas is unstable. In order to prevent collapse, we add a
short range repulsive interaction
$C_\mathrm{h}/r^{12}$-potential with the short-range repulsion parameter $C_\mathrm{h}$
as shown in Eq.~(\ref{eq::dipole-dipole interaction}).
As a check of universality of this model we compare the results with those obtained with a $1/r^6$-potential.
We describe the ground state using a variational Jastrow-Feenberg ansatz~\cite{Feenberg} of the form
\begin{align}
\Psi(\Vec{r}_1, \cdots, \Vec{r}_N) = \exp \left[\frac{1}{2} \sum\limits_{i < j} u(\Vec{r}_{i}-\Vec{r}_{j}) \right] \ ,
\label{eq:jastrowansatz}
\end{align}
which includes pair correlations $u(\Vec{r})$ and accounts for quantum fluctuations. To obtain the optimal
ground state we solve the Euler-Lagrange equation
\begin{equation}
{\delta e \over \delta\sqrt{g(\Vec{r})}} = 0
\label{eq:EL}
\end{equation}
where $e$ is the energy per particle
\begin{align}
e(\rho_0) \equiv \frac{E}{N} = \frac{\rho_0}{2} \int \mathrm{d}^2r~ g(\Vec{r}) \left[ v(\Vec{r}) - \frac{\hbar^2}{4m} \nabla^2 u(\Vec{r}) \right]~.
\end{align}
The pair distribution function $g(\Vec{r})$ is given in terms of the wave function
in Eq.~(\ref{eq:jastrowansatz}) as
\begin{align}
g(\Vec{r}_1-\Vec{r}_2) = \frac{N(N-1)}{\left<\Psi | \Psi \right> \rho_0^2} \int\! \mathrm{d}^2 r_3\dots \mathrm{d}^2 r_N \left| \Psi(\Vec{r}_1,\dots,\Vec{r}_N) \right|^2,
\end{align}
Closure is provided by the HNC relation between $g(\Vec{r})$ and $u(\Vec{r})$ \cite{Hansen}.
In the following we restrict ourselves to the HNC-EL/0 approximation,
where the so-called elementary diagrams are neglected
in the cluster expansion \cite{Hansen}. We have calculated the leading
contribution of the elementary diagrams to the total energy but found it to be
less than 3\% for densities $\rho_0 r_0^2 \leq 1$ (see appendix \ref{sec::Elementary Diagrams}).
In the HNC-EL/0 framework, Eq.~(\ref{eq:EL}) can be cast as
\begin{equation}
\Big[-{\hbar^2\over m}\Delta + v(\Vec{r})+w_\mathrm{I}(\Vec{r})\Big]\sqrt{g(\Vec{r})}=0
\label{eq:HNCEL0}
\end{equation}
which has the form of an effective 2-body zero-energy scattering equation
with the bare potential $v$ and an additional induced many-body potential $w_\mathrm{I}$,
which is defined via its Fourier transform
\begin{equation*}
w_\mathrm{I}(\Vec{k})=-{\hbar^2k^2\over 4m}\Big(1-{1\over S(\Vec{k})}\Big)^2(2S(\Vec{k})+1)
\end{equation*}
in terms of the static structure factor
\begin{equation}
S(\Vec{k}) = 1 + \mathrm{FT} \left[ g(\Vec{r}) -1 \right]
\label{Sk}
\end{equation}
where $\mathrm{FT}$ denotes the Fourier transformation multiplied with the density $\rho_0$.
We note that Eq.(\ref{eq:HNCEL0}) is not a simple linear differential
equation because the induced potential $w_\mathrm{I}$ depends on $g$ itself.
The details on how to solve Eq.~(\ref{eq:HNCEL0})
iteratively can be found elsewhere~\cite{KroTrieste,QMBT00Polls}.
From experience with other systems \cite{Clements94,Campbell96,hebenstreitPRA16,raderPRA17,Staudinger2018},
solving the HNC-EL/0 equations is straightforward
for systems with a stable or metastable ground state, but fails to converge if the system is unstable
against infinitesimal perturbations
(e.g. spinodal instability of a system with homogeneous density \cite{Campbell96,Staudinger2018}). Inspection of structural
quantities like $g(\Vec{r})$ and $S(\Vec{k})$ provides clues as to the nature of the instability (e.g.
long-ranged fluctuations in $g(\Vec{r})$ in the case of a spinodal instability).
More quantitative information on that is provided by a stability analysis of the solution of
the HNC-EL/0 equation~\cite{Castillejo1979}. For this purpose, we evaluate
the second functional derivative of the energy $e$ with respect to the pair distribution function,
$K(\Vec{r},\Vec{r}')=\delta^2 e/\delta\sqrt{g(\Vec{r})}\delta\sqrt{g(\Vec{r}')}$.
If this operator is
positive definite, the solution of the HNC-EL/0 equation (\ref{eq:EL}) is stable against infinitesimal
perturbations of $g(\Vec{r})$. This is guaranteed if all eigenvalues $\lambda_i$ in the equation
\begin{equation}
\int \mathrm{d}^2r'\,K(\Vec{r},\Vec{r}') f_i(\Vec{r}') = \lambda_i f_i(\Vec{r})
\label{eq:ev1}
\end{equation}
are positive. Conversely, if the lowest eigenvalue $\lambda_0$ is close to zero, the system approaches
an instability. More importantly, the eigenvector $f_0(\Vec{r})$ provides information about the nature
of the instability as shown by our results below.
The explicit form of $K(\Vec{r},\Vec{r}')$ is easily
calculated in the HNC-EL/0 approximation. Following the notation of Ref.~\cite{Castillejo1979}, the
eigenvalue problem can be written as
\begin{equation}
\Big[-{\hbar^2\over m}\Delta + v(\Vec{r})+w_\mathrm{I}(\Vec{r})+\hat W\Big] f_i(\Vec{r}) = \lambda_i f_i(\Vec{r})
\label{eq:ev2}
\end{equation}
where the $\hat W$ operator is defined as
\begin{equation}
\hat W f_i(\Vec{r}) =
\rho_0 \int \mathrm{d}^2r'\,\sqrt{g(\Vec{r})}W(\Vec{r}-\Vec{r}')\sqrt{g(\Vec{r}')} f_i(\Vec{r}')
\label{eq:Wdef}
\end{equation}
$W$ is given in in momentum space as
\begin{equation*}
W(\Vec{k})=-{\hbar^2k^2\over m}\left(1-{1\over S(\Vec{k})^3}\right) \ .
\end{equation*}
Since we only need the lowest eigenvalue to assess the stability, we solve eq.(\ref{eq:ev2}) by
imaginary time propagation, see appendix \ref{sec::Stability Analysis} for details.
\section{Energy and Stability}
\label{sec::Energy and Stability}
In the liquid phase, as opposed to the gas phase, a system is self-bound: the energy per particle
is negative and attains its minimum at an equilibrium density $\rho_\mathrm{eq}$.
Furthermore, below the spinodal density a homogeneous liquid
becomes unstable against long wavelength density fluctuations, and then
breaks into droplets. In this section we analyze
the ground state energy for various short range repulsion strengths $C_\mathrm{h}$,
dipole tilt angles $\theta$ and densities $\rho_0$ in order to check whether
the system is in a liquid or in a gas phase.
We also assess the stability against density fluctuations. Instead of reaching
a spinodal instability typical of isotropic liquids, we find a transition to a density
wave in the $y$-direction, i.e.\ a stripe phase.
\begin{figure}[h!]
\begin{center}
\includegraphics*[width=0.48\textwidth]{energy_cut_HC.pdf}
\end{center}
\caption{Energy per particle $e(\rho_0)$ as a function of the density $\rho_0$ for $C_\mathrm{h}=0.33 r_0$
and $\theta=1.02-1.16$. The last converging point at the lowest density in each curve
is marked with a blue open circle and the equilibrium energy $e_\mathrm{eq}$ with a red
filled circle. The insets show a magnification of the result for $\theta=1.10$ and $1.14$,
which illustrate the behavior of the equation of state near the stripe phase.}
\label{fig::energy cut constant C}
\end{figure}
We first fix $C_\mathrm{h}=0.33 r_0$ and vary the tilt angle between $\theta=1.02$ and $\theta=1.16$.
The results are depicted in fig.~\ref{fig::energy cut constant C}, where we show the energy
per particle $e(\rho_0)$ as a function of the density.
In each case we start the calculation at the largest density considered and
solve Eq.~(\ref{eq:EL}) iteratively. The resulting pair distribution function is then used
as an input to solve the same equation at the next lower density in the grid. This procedure
ensures rapid convergence.
Depending on the tilting angle $\theta$ three different cases can occur, corresponding
to three different phases: a gas, a homogeneous liquid, or a striped liquid.
For $\theta\leq 1.04$ the energy per particle is positive and approches zero as $\rho\to 0$ where
it attains its minimum value $e=0$. The system is then in a gas phase and the corresponding pressure is always positive.
Beyond $\theta=1.04$ the system enters a different phase where the energy becomes negative as the density is lowered.
The system is self-bound and thus in a liquid phase. However, as long as $\theta\leq 1.12$, the
calculation ceases to converge before the energy per particle attains a minimum; the density $\rho_\mathrm{c}$
where this happens is indicated with open blue circles in fig. \ref{fig::energy cut constant C}.
The inset for $\theta=1.1$ shows this more clearly, where one can see that the liquid phase stops being stable before
reaching a homogeneous equilibrium density $\rho_\mathrm{eq}$
The homogeneous HNC-EL/0 equations are know to cease to converge
at a continuous phase transition~\cite{maciaPRL12,Staudinger2018}.
In the following we show that the dipolar system undergoes a transition to a self-organized stripe phase
where the density exhibits a spatial modulation.
\begin{figure}[h!]
\begin{center}
\includegraphics*[width=0.48\textwidth]{S2D_stripes.pdf}
\end{center}
\caption{Static structure factor $S(0,k_\mathrm{y})$ for $\theta=1.08$ and $C_\mathrm{h}=0.33 r_0$
as a function of $k_\mathrm{y}$.
As the density (color coded) is lowered approaching a critical density
$\rho_\mathrm{c}=0.2023/r_0^2$, a large peak develops.
$S(k_\mathrm{x},0)$, shown in the inset, is almost independent on $\rho_0$ and does not develop a peak.}
\label{fig::S dipoles stripes}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.48\textwidth]{structure_factor_3D.pdf}
\caption{Static structure factor $S(k_\mathrm{x},k_\mathrm{y})$ for $\theta=1.08$, $C_\mathrm{h}=0.33 r_0$
at the critical density $\rho_\mathrm{c}=0.2023/r_0^2$.}
\label{fig::S3D dipoles stripes}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.48\textwidth]{pair_distribution_3D.pdf}
\caption{Pair distribution function $g(x,y)$ for $\theta=1.08$, $C_\mathrm{h}=0.33 r_0$ at the critical density $\rho_\mathrm{c}=0.2023/r_0^2$.}
\label{fig::g3D dipoles stripes}
\end{figure}
Fig. \ref{fig::S dipoles stripes} shows the static structure factor $S(k_x=0,k_y)$ along the $y$-direction,
for $\theta=1.08$ and $C_\mathrm{h}=0.33 r_0$, and densities spanning the range from $0.2023/r_0^2$ to $1/r_0^2$.
A pronounced peak appears when the density approaches a critical value $\rho_\mathrm{c}=0.2023/r_0^2$,
thus signaling increasing spatial order in that direction.
This suggests a self-organized long-range order below $\rho_\mathrm{c}$, where homogeneous
HNC-EL/0 does not converge anymore, as it happened for tilted dipoles in
Ref.~\cite{maciaPRL12}, except that here the density is three orders of magnitude lower and the dipoles are in a self-bound phase.
From the position $k_\mathrm{p}$ of the peak in $S(k_x=0,k_y)$ at $\rho_\mathrm{c}$
we can predict the wave number of the density oscillation. In contrast, the structure factor in $x$-direction
has no peak at all (see inset in fig. \ref{fig::S dipoles stripes}),
thus showing no signs of ordering in $x$-direction.
Figure~\ref{fig::S3D dipoles stripes} shows the full $S(k_\mathrm{x},k_\mathrm{y})$ at the critical density $\rho_c$.
Fig.~\ref{fig::g3D dipoles stripes} shows the corresponding pair distribution function $g(x,y)$
obtained as the Fourier transform of $S(k_x,k_y)$. As expected $g(x,y)$ shows a small peak
along the $x$ direction where the attractive well of the dipole-dipole interaction (fig. \ref{FIG:dipole}) is deepest.
Other than that, the pair distribution function is anisotropic but smooth.
In the small range depicted, $g(x,y)$ has little structure along the $y$-direction,
and the fluctuations in $g(0,y)$ that lead to the peak in $S(0,k_y)$ cannot be seen.
\begin{figure}[h!]
\begin{center}
\includegraphics*[width=0.48\textwidth]{stability_analysis.pdf}
\end{center}
\caption{Stability analysis for $C_\mathrm{h}=0.33 r_0$, $\theta=1.08/1.16$ and $C_\mathrm{h}=0.28 r_0$, $\theta=1.19$.
The left column shows the lowest eigenvalue $\lambda_0$
of the Hessian as a function of the density $\rho_0$, while the right colum shows
the eigenvector $f_0(\Vec{r})$ at $\rho_\mathrm{c}$.}
\label{fig:stability analysis}
\end{figure}
The emergence of a peak in the static structure factor along the
$y$-direction is already a strong indicator for the transition to a stripe phase.
The more rigorous stability analysis explained above leads to the same conclusion, as illustrated in
fig.~\ref{fig:stability analysis}. The top, middle and bottom panels correspond to
($C_\mathrm{h}=0.33 r_0$, $\theta=1.08$),
($C_\mathrm{h}=0.33 r_0$, $\theta=1.16$) and
($C_\mathrm{h}=0.28 r_0$, $\theta=1.19$), respectively. The left panels depict the density dependence of
the lowest eigenvalue $\lambda_0$ of Eq.(\ref{eq:ev2}).
As can be seen, $\lambda_0$ approaches zero as the density approaches $\rho_\mathrm{c}$, confirming
that the homogeneous phase becomes unstable in that limit. As a further confirmation
we show the lowest eigenvectors $f_0(\Vec{r})$ in the right column of of the same figure.
The shape of $f_0$ informs about the least stable fluctuation in
$g(x,y)$, which drives the transition to a stripe phase,
showing oscillations along the $y$-direction. These oscillations are most pronounced
at high densities, as seen in the bottom panel of the figure.
The wave length corresponds to the location of the peak in $S(0,k_y)$
at $\rho=\rho_c$, see Fig.~\ref{fig::S dipoles stripes}.
The wave number of this oscillation is the same as the wave number
of the peak in $S(0,k_y)$ for $\rho_\mathrm{c}$ (see fig. \ref{fig::S dipoles stripes}).
This behavior clearly signals the transition to a phase with long range order in the $y$-direction.
Note that in all cases, the oscillation is strongly damped, which
was not found for the stripe phase transition at high densities
in Ref~.\cite{maciaPRL12}, where the system is not self-bound.
When the tilting angle is increased beyond $\theta= 1.12$ while still keeping $C_\mathrm{h}=0.33 r_0$, the energy
decreases further due to the stronger attraction, but there is also a qualitative change of the shape of the equation
of state: $e(\rho_0)$ is not monotonous anymore, but reaches a minimum at
a homogeneous equilibrium density $\rho_\mathrm{eq}$ before reaching the critical density $\rho_c$ of the transition to the stripe phase.
This new minimum corresponds to a self-bound homogeneous liquid state. If the system were
finite, the dipoles would form a two-dimensional droplet adjusting its radius so
that the pressure inside the droplet is zero at $\rho_\mathrm{eq}$.
When the density is lowered further the energy per particle starts
increasing again, up to the point where the transition to the stripe phase takes place at $\rho_\mathrm{c}$,
as evidenced by the stability analysis shown in fig.~\ref{fig:stability analysis}).
The lower inset in fig. \ref{fig::energy cut constant C} shows the result for $\theta=1.14$,
where the energy minimum at $\rho_\mathrm{eq}$ is clearly visible.
\section{Phase Diagram}
\begin{figure}[h!]
\begin{center}
\includegraphics*[width=0.48\textwidth]{phase_diagram.pdf}
\includegraphics*[width=0.48\textwidth]{phase_diagram_scattering_length_with_different_potential.pdf}
\end{center}
\caption{
Phase diagram as a function of the
tilting angle $\theta$ and short-range repulsion $C_\mathrm{h}$ (upper panel) and
as a function of the scattering length $a_0$ (lower panel). The insets (a), (b), and (c)
show a representative $e(\rho_0)$ for each phase for $\theta=1.02$, $1.08$, and $1.16$, respectively,
all for the same value of $C_\mathrm{h}=0.33 r_0$ (white line).
In the lower panel the three black symbols show the result of a calculation with a $1/r^6$-repulsion, which gives
the same result as the $1/r^{12}$-repulsion used for all other calculations.}
\label{fig:Phase Diagram}
\end{figure}
In order to obtain the full phase diagram, we calculate $e(\rho_0)$ for a wide range of $\theta$ and $C_\mathrm{h}$
values following the protocol described above, i.e. lowering $\rho_0$ from a sufficiently large value down to zero or
until a homogeneous phase ceases to be stable at a finite density value $\rho_\mathrm{c}$.
We analyze the results in the same way as in the previous section
to classify the phase into gas, striped liquid, and homogeneous liquid.
In fig.~\ref{fig:Phase Diagram} we indicate the phase by different colors.
The lower panel of fig. \ref{fig:Phase Diagram} depicts the same phase diagram as the upper panel, but
in terms of the scattering length $a_0$ of the interaction (\ref{eq::dipole-dipole interaction}) instead of $C_\mathrm{h}$
to make it independent of the interaction~\cite{maciaPRA11}.
The scattering length $a_0$ depends on both $C_\mathrm{h}$ and $\theta$, and increasing $C_\mathrm{h}$ decreases $a_0$,
see fig.~\ref{fig:Scatt_Length}.
We also include three points, marked in black, corresponding to an
interaction $1/r^6$ but tuned to give the same scattering length of the compound system.
These calculations give the same phase as those with the $1/r^{12}$-repulsion,
which indicates a universal behavior that is independent of the
details of the short-range interaction beyond its scattering length.
Based on the shown grid of $\theta$ and $C_\mathrm{h}$ values,
the boundaries between the phases were calculated using a C-support vector machine \cite{scikit-learn}, and
an analytic expression for those phase boundaries is given in appendix \ref{sec::phase boundaries}.
In the upper panel of fig. \ref{fig:Phase Diagram} a white line is drawn for the fixed repulsion parameter
$C_\mathrm{h}=0.33 r_0$ that we used in the previous section. The three insets show three examples for the equation
of state $e(\rho_0)$ along this line, each representing one of the phases. The points along the white line are
the results of the energies $e(\rho_0)$ discussed earlier in fig.~\ref{fig::energy cut constant C}.
Note, that a dipolar system in 2D can still have a density oscillation at densities $\rho_0 \sim 10^2/r_0^2$
as shown in \cite{maciaPRL12}. However, here we focus on
the formation of a stripe phase at orders of magnitude lower densities.
\begin{figure}[h!]
\begin{center}
\includegraphics*[width=0.48\textwidth]{Scatt_Length.png}
\end{center}
\caption{
Scattering length $a_0$ of the interaction (\ref{eq::dipole-dipole interaction}) as function of the
repulsion parameter $C_h$ for 4 different tilting angles $\theta$.}
\label{fig:Scatt_Length}
\end{figure}
Fig.~\ref{fig:Phase Diagram} shows that,
for tilting angle $\theta>1.07$, upon increasing the strength $C_h$ of the repulsion
the system undergoes first a transition from a striped liquid to a homogeneous liquid, and then
a transition to a homogeneous gas. The reason is that the
increasing isotropic short-range repulsion becomes dominant compared to the anisotropic dipole-dipole interaction,
which is responsible for the formation of stripes.
More interestingly, an increase of the tilting angle $\theta$, and thus of the anisotropy,
for a fixed short-range interaction, also drives the system to the homogeneous liquid phase,
as we have discussed in Sec.~\ref{sec::Energy and Stability}.
This might seem counter-intuitive at first. However, it
is actually the repulsive part of the dipole-dipole interaction in
$y$-direction causing the formation of stripes and this part of the
interaction does not change with $\theta$.
At the same time the attraction in $x$-direction increases with increasing $\theta$ and leads to a more strongly bound system with a higher density, and the stripes merge as their wave length decreases. For example if $C_\mathrm{h}=0.33r_0$ is fixed and the tilting angle is increased from $\theta=1.08$ to $\theta=1.16$ the wave length decreases from $\lambda=11.70r_0$ to $\lambda=4.82r_0$.
In the lower panel of fig.~\ref{fig:Phase Diagram}
we characterize the phase diagram using the scattering length $a_0$ of the full interaction
instead of the short-range repulsion parameter $C_\mathrm{h}$. This way, the phase diagram
becomes more intuitive: for a given $a_0$, the system undergoes a transition from an homogeneous liquid to
to a striped liquid phase with increasing $\theta$.
\section{Critical Density}
\begin{figure}[h!]
\begin{center}
\includegraphics*[width=0.48\textwidth]{critical_density.pdf}
\end{center}
\caption{Critical density $\rho_\mathrm{c}$ for the transition to the stripe phase as a function of $\theta$ and $C_\mathrm{h}$.
The red and blue surfaces denote the regimes of striped liquid and homogeneous liquid, respectively, see fig.~\ref{fig:Phase Diagram}.
At the border to the gas phase, $\rho_\mathrm{c}$ falls to zero.}
\label{fig::critical density}
\end{figure}
Figure~\ref{fig::critical density} shows the critical density $\rho_\mathrm{c}$ for the transition to a
stripe phase as a function of tilting angle $\theta$ and repulsion strength $C_\mathrm{h}$, showing
that $\rho_\mathrm{c}$ increases with increasing $\theta$ and decreasing $C_\mathrm{h}$.
The blue and red surfaces represent the regions of the homogeneous liquid and
the striped liquid, respectively. Both
surfaces approach $\rho_\mathrm{c}=0$ as the system approaches the gas phase, seen as a sharp
drop on the logarithmic scale in fig.~\ref{fig::critical density}.
This shows that with a proper choice of
$\theta$ and $C_\mathrm{h}$ (or $a_0$), stripes can be observed in arbitrarily dilute systems.
This is markedly different
from previous theoretical studies of tilted dipoles in two dimensions \cite{maciaPRL12}, where the dipole orientation was not
tilted above $\theta=\arcsin(1/\sqrt{3})$ such that the dipole-dipole interaction stays purely repulsive.
In that case a stripe phase was predicted only above a critical density $\rho_\mathrm{c}=\mathrm{O}(10^2)$.
We also note that in the range of $\theta$ and $C_\mathrm{h}$ explored in this work,
$\rho_\mathrm{c}r_0^2<1$ .
Despite the variational nature of the calculations presented,
the contribution of elementary diagrams to the energy per particle
is less than 3\%\ (see appendix~\ref{sec::Elementary Diagrams} for details),
while three- or higher-order correlations are negligible in the
low density regime considered.
\section{Conclusion}
In this work we have explored the phase diagram of a dipolar gas in two dimensions as a function of the
dipole tilting angle and their short-range repulsion.
We have found that this system exhibits three different
phases: a gas phase, a self-bound stripe phase, and a homogeneous self-bound stripeless phase.
We detected the transition to a stripe phase by monitoring a peak that emerges in the structure factor
close to a critical density $\rho_c$ that marks the phase transition.
The critical density depends on the short range repulsion and the tilting angle, and
can be low enough so that it can be realized with current experimental setups.
To further confirm the emergence of the low density stripe phase we have also
conducted a stability analysis based on the Hessian of the energy.
Although our calculations, based on the variational HNC-EL/0 method, always assumed homogeneity, this
provides strong evidence that a self-bound and striped droplet can form in a
two-dimensional anisotropic dipolar Bose gas.
In order to calculate the properties of the stripe phase
one can use a variational model that allows for density modulations.
This can be achieved with the inhomogeneous generalization of the HNC-EL method~\cite{kroValencia98},
which has been used for example
to describe dipolar systems in quasi-2D geometries~\cite{hufnaglPRL11,hufnaglPRA13}. Alternatively
one can use exact diffusion Monte Carlo methods that are able to describe the system also in the stripe phase,
using the optimal HNC-EL solutions as an ingredient for the guiding wave function.
We are currently working along this line, in an attempt
to study the self-bound nature of two-dimensional striped dipolar systems.
\begin{acknowledgments}
We acknowledge discussions with Jürgen T. Drachta and Tilman Pfau.
F. Mazzanti acknowledges financial support by grant PID2020-113565GB-C21 funded by MCIN/AEI/10.13039/501100011033,
and from Secretaria d'Universitats i Recerca del Departament d'Empresa i Coneixement de la
Generalitat de Catalunya, co-funded by the European Union Regional Development
Fund within the ERDF Operational Program of Catalunya (project QuantumCat, ref.
001-P-001644).
\end{acknowledgments}
|
\section{Introduction}
\label{sec:intro}
With the rapid development of 3D sensing technology, point cloud based scene understanding has become a hot research topic in recent years. Instance segmentation is one of the most fundamental tasks here, with many applications in robotics, autonomous driving, AV/VR, etc. Given a reconstructed scene, this task requires predicting not only semantic category but also an instance id to differentiate objects at point-level. Many deep learning methods have been proposed with promising performance. However, most methods need full annotations for every single point to supervise the network training.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=1.0\linewidth]{figure/weak_annotations_v3.png}
\end{center}
\vspace{-8mm}
\caption{Weak Annotations for 3D Instance Segmentation. We use the annotation setting on the left to train our approach. Each object is annotated with a single point on it (enlarged for visualization). The right plot shows 100\% supervision at point-level.}
\label{fig:weak_annotations}
\end{figure}
Manually creating data annotations at point-level is very cumbersome and labor-intensive. Even some tools have been adopted to assist, the average time used to annotate one scene is about 22.3 minutes on ScanNet-v2 dataset \cite{dai2017scannet}. To alleviate this issue, several types of weak annotations have been proposed, including scene-level annotation, subcloud-level annotation \cite{Wei_2020_CVPR} and 2D image based annotation. However, not all weak label types are easy to generate in practice. In this work, we follow the annotation method used in SegGroup \cite{tao2020seggroup} and ``One Thing One Click" \cite{otoc_Liu_2021_CVPR}, which only requires one object to be annotated by one point on it. As shown in Figure \ref{fig:weak_annotations}, initial annotations are very sparse, with less than 0.02\% of total points. According to \cite{tao2020seggroup, otoc_Liu_2021_CVPR}, it takes less than 2 minutes to annotate a scene with this method, which dramatically reduces the human efforts. Existing weakly supervised methods in 3D instance segmentation task still have a large gap in performance with their fully supervised counterparts. Thus users have to make a trade-off between performance and annotation cost.
In this paper, we propose a novel weakly supervised learning approach \textbf{RWSeg} for 3D point cloud instance segmentation. To make it feasible to use only weak labels, our core idea is to jointly perform feature learning and feature propagation. Convolutional neural network (CNN) can extract good local feature. However, long-range dependencies can hardly be captured due to its relatively small receptive field. Its shortcomings are magnified during weakly supervised learning. To this end, a common practice is to use a separated graph model as \cite{otoc_Liu_2021_CVPR} for feature propagation. Instead of implementing another model, we introduce a self-attention module within the same framework, which can effectively propagate long-range information to unknown regions. To leverage the pairwise relationship of point clouds, we propose a random walk algorithm to generate high-quality pseudo labels. Additionally, we design a mechanism to further encourage competition among different instance graphs. To sum up, the key contributions of our work are following:
\begin{itemize}
\item We design a unified framework with strong feature propagation capability for weakly supervised 3D instance segmentation.
\item We propose a random walk based algorithm to generate qualitative instance pseudo labels. Our pseudo label results show dramatic improvement over previous methods.
\item We introduce cross-graph competition among instances into the proposed random walk algorithm to further boost the performance.
\item With significantly fewer annotations, our method bridges the gap between weakly supervised learning and fully supervised learning in 3D instance segmentation.
\end{itemize}
\section{Related Work}
\label{sec:rltd}
\paragraph{Segmentation on point cloud data}
Unlike 2D images on grids, point cloud data \cite{bello2020review} is usually a set of unstructured and unordered points in 3D space. To handle these data , existing proposed feature learning methods can be roughly classified into point-based methods \cite{PointCNN, PointNet, PointNet++, KPConv, PointConv, PointWeb, 9552005, Guo_2021} and voxel-based methods \cite{multiview2, Submanifold, MVPNet, multiview}. The latter transforms data into 3D volumetric grids.
\vspace{-4mm}
\paragraph{Fully supervised 3D instance segmentation}
Instance segmentation on point clouds can be seen as a joint task of segmentation and localization. Existing approaches can be classified as proposal-based or proposal-free methods. Proposal-base methods \cite{SGPN, 3DSIS, 3DBoNet, 3dmpa} explicitly detect object boundaries first and then perform a binary mask segmentation as final output. Whereas proposal-free methods \cite{SGPN, MASC, ASIS, JSIS3D, MTML, pointgroup, occuseg, Chen_2021_ICCV, Liang_2021_ICCV} do not perform detection task but directly regress instance centroids. Liu et al. \cite{MASC} proposed MASC, a submanifold sparse convolution \cite{Submanifold} based U-net architecture with affinity between neighbouring voxels at different scales. Wang et al. \cite{ASIS} introduce a network with semantic branch and instance branch that can mutually support each other. J Lahoud et al. \cite{MTML} adopt discriminative loss to group feature embeddings. Jiang et al. \cite{pointgroup} proposed to use a breadth-first search clustering algorithm on dual coordinate sets. Some recent works \cite{Chen_2021_ICCV, Liang_2021_ICCV} are built on top of PointGroup \cite{pointgroup} with further refinement.
\vspace{-4mm}
\paragraph{Weakly supervised segmentation}
Many weakly supervised methods \cite{ahn2018learning,huang2018weakly,pinheiro2015image,papandreou2015weakly,ahn2019weakly,zhou2018weakly,arun2020weakly} have been proposed for image segmentation. Wei et al. \cite{wei2020multi} proposed the first weakly approach for point cloud semantic segmentation, with subcloud-level labels. They utilized Class Activation Map (CAM) to generate point-level pseudo labels. Some following works \cite{xu2020weakly, wang2020weakly, otoc_Liu_2021_CVPR} are also proposed for weakly semantic segmentation on point clouds. Hou et al. \cite{hou2021exploring} designed a pre-training method to assist prediction. Tao et al. \cite{tao2020seggroup} proposed Seggroup with graph convolution network (GCN) for instance label propagation.
\vspace{-4mm}
\paragraph{Random walk}
As a well-known graphical model, random walk represents a type of iterative process based on transition probability between graph nodes. Its character has been proven to be effective for image segmentation. Some works \cite{vernaza2017learning,Bertasius_2017_CVPR} utilized random walk to refine semantic segmentation results. Li et al. \cite{Li_2021_CVPR} adopted random walk to smooth the pseudo labels in scene flow estimation. In this paper, we propose a novel random walk based method for generating instance clusters on point clouds.
\begin{figure*}[thb]
\vspace{-8mm}
\begin{center}
\includegraphics[width=1.0\linewidth]{figure/CGCRW_Pipeline_v5.png}
\end{center}
\vspace{-8mm}
\caption{Pipeline of our proposed weakly supervised method for 3D instance segmentation. For input point cloud, we first employ submanifold sparse convolution \cite{Submanifold} based 3D U-Net backbone to extract point features. Then, average pooling is applied to the points within the same supervoxel. Self-attention module is used for semantic feature propagation. Our proposed Cross-graph Competing Random Walks (CGCRW) module considers the input from two branches and generates qualitative pseudo labels for further network training. }
\label{fig:CGCRW_Pipeline}
\end{figure*}
\section{Method}
\label{sec:mthd}
In this section, we first introduce our data annotation setting for point cloud instance segmentation in Section \ref{sec:anno}. Then, Section \ref{sec:ntwk} to \ref{sec:psd} present our approach in detail, including network architecture, semantic branch, instance branch and proposed pseudo label generation algorithm. Lastly, Section \ref{sec:stgy} describes our training strategy.
\subsection{Weak Annotation}
\label{sec:anno}
Following SegGroup \cite{tao2020seggroup}, we adopt the annotation setting of one point per object, as shown in Figure \ref{fig:weak_annotations}. To create initial pseudo labels, we spread the labels from annotated points to other nearby points under the same supervoxel segment. These segments are generated by unsupervised over-segmentation method\cite{felzenszwalb2004efficient} based on the surface normals of points. Points within the same segment have high internal consistency. Thus we treat these high confidence points as initial ground-truth to supervise the network training.
\subsection{Network Architecture}
\label{sec:ntwk}
Our network uses a shared U-Net backbone and two separate branches for point-level semantic feature learning and instance segment's centroid regression. In the semantic branch, a self-attention based module is used to further enhance semantic features, especially for those regions without supervision. Following that, our proposed \textbf{Cross-Graph Competing Random Walks (CGCRW)} algorithm leverages learned feature and existing ground-truth weak labels to generate instance-level pseudo labels. With refined weak labels, the network can be further trained for predicting better features, and thus benefit RWSeg again. All proposed modules are within the unified framework, as shown in Figure \ref{fig:CGCRW_Pipeline}.
\subsection{Semantic segmentation branch}
\label{sec:sem_branch}
The submanifold sparse convolution \cite{Submanifold} based backbone network takes point cloud $\bm P \in \mathbb{R}^{N \times 3}$ as input and extracts point-wise feature. The convolution operations are good at extracting local features, but with limited receptive field. To facilitate the capture of long-range feature dependencies, we introduce a self-attention module to further refine semantic feature from the backbone.
To reduce the computational complexity in self-attention and improve local geometric consistency, we utilize a supervoxel generation method \cite{Lin2018Supervoxel}. For each supervoxel set $\mathcal{V}=\{p_1, p_2, ...,p_i\}$, average pooling is applied on both point coordinates and semantic features.
\paragraph{Global self-attention module}
Following \cite{NIPS2017_3f5ee243, zhao2021pointtransformer}, we first perform linear transformations of the input semantic features $S_{\mathcal{V}}$ to three matrices as query, key, and value ($Q, K, V$). Then, matrix $\bm A$ captures the similarity between queries and keys and also includes encoded positional information for adjustment. This can be written as
\begin{equation}\label{eqn:attn}
\bm Q = S_{\mathcal{V}}W_Q,\quad \bm K = S_{\mathcal{V}}W_K,\quad \bm V = S_{\mathcal{V}}W_V,\\
\end{equation}
\begin{equation}\label{eqn:attn2}
\bm A = \gamma (\frac{\bm Q \bm K^\top}{\sqrt{d}}) + \delta(X_{\mathcal{V}}, X_{\mathcal{V}}),
\end{equation}
where $d$ is the dimension of $Q$ and $V$, $X_{\mathcal{V}}$ the coordinates of supervoxels, $\gamma$ is a mapping function via MLP, $\delta$ is a positional similarity function via MLP. The output of self-attention can be formulated as
\begin{equation}\label{eqn:attn3}
\mathrm{Attn}(S_{\mathcal{V}}) = \sigma({\bm A}) (\bar{\bm V} + \delta (X_{\mathcal{V}}, X_{\mathcal{V}})),
\end{equation}
where $\sigma(\cdot)$ is a softmax activation function. $\bar{\bm V}$ denotes a symmetric matrix created by repeatedly expanding $\bm V$.
Lastly, refined semantic features are interpolated to the original size in point cloud. The training process is supervised by a conventional cross-entropy loss $H_{CE}$ with incomplete labels. We define the semantic loss as
\begin{equation}\label{eqn:ce}
L_{sem} = - \frac{1}{N} \sum_{i=1}^{N} H_{CE}(y_i, \widehat{c}_i).
\end{equation}
where $\widehat{c}_i$ is the weak semantic label. Unlabelled points are ignored here.
\subsection{Instance centroid offset branch}
\label{sec:ins_branch}
Paralleled with semantic branch, we apply 2-layers MLP upon point feature to predict point-wise centroid shift vector $d_i \in \mathbb{R}^{3}$. The instance
centroid $\widehat{q}$ is defined as the mean coordinates of all points with this instance label. We use following regression loss to train offset prediction,
\begin{multline}
L_{offset} = \frac{1}{\sum_i m_i}\sum_i ||d_i - (\hat{q}_i - p_i)|| \cdot m_i - \\ \frac{1}{\sum_i m_i} \sum_i \frac{d_i}{||d_i||_2} \cdot \frac{\hat{q}_i - p_i}{||\hat{q}_i - p_i||_2} \cdot m_i.
\end{multline}
where $m \in \{0,1\}$ is a binary mask telling if a point belongs to any object instance, which means to consider foreground points with weak labels only.
With initial weak labels, real centroids of instances can hardly be inferred. However, we found it is still beneficial to apply offset loss, as it can help to slightly shift points towards inner part of objects.
The final joint loss function can be written as
\begin{equation}
L_{joint} = L_{sem} + L_{offset}.
\end{equation}
\subsection{Pseudo label Generation}
\label{sec:psd}
Since only a small amount of points can be used to supervise network training at the beginning, the predicted feature to be utilized might not be very reliable during this time. Thus, robustly finding instance segments among ambiguous unseeded regions is both challenging and important. To deal with this situation, we propose a random walk based algorithm to infer new pseudo labels for unlabelled points.
\paragraph{Random walk based approach}
According to the semantic prediction $\mathbf{S} = \{s_1, s_2, ..., s_N\}\in \mathbb{R}^{N}$ on point clouds $\bm P$, we treat each foreground category as a target group. For each group, we view each point as a node and build $k$ numbers of instance graphs $G_s=\{G_1(V^1,E^1),G_2(V^2,E^2),... ,G_k(V^k,E^k)\}$ on the $\bm P_s$, where $k$ is the number of instances based on ground-truth weak labels, $V$ and $E$ denotes vertex and edge set respectively.
Instance graphs are constructed as fully connected with undirected edges. The random walk operation on each graph can be modeled with a~$n \times n$ transition matrix~$\bm A$. $\bm A_{ij} \in [0, 1]$ denotes the transition probability between $i$-th and $j$-th nodes. The higher value represent a higher transition probability.
To build transition matrix~$\bm A$, we first consider a pairwise kernel function to derive a symmetric affinity matrix $\bm W$, which helps to enhance local smoothness. For each edge from $i$-th node to $j$-th node, we define the corresponding weight as
\begin{equation}\label{eq:W1}
{\bm W}_{ij} = {\rm \exp} (- \frac{\| x_i - x_j \|^2}{2\sigma^2}),
\end{equation}
where $\sigma$ is a hyperparameter, $x_i$ and $x_j$ are point coordinates. Thus, node pairs with low euclidean distance in 3D space tends to have high similarity. In addition to Eq. (\ref{eq:W1}), we use predicted offset vector $d$ to shift points from original coordinates $x$ towards their instance centers,
\begin{equation}\label{eq:W2}
{\bm W}_{ij} = {\rm \exp} (- \frac{\| (x_i + d_i) - (x_j + d_j) \|^2}{2\sigma^2}).
\end{equation}
Then, we use the following rules to obtain transition matrix $\bm A$. For edges between nodes with different instance labels, their edge weights are set to be zero, in order to limit the direct interaction between them.
\begin{equation}\label{eq:A}
\bm A_{ij} =
\begin{cases}
$0$ & \widehat{\bm L}_{i} \neq \widehat{\bm L}_{j}\\
\bm W_{ij} & otherwise
\end{cases},
\end{equation}
where $\widehat{\bm L}_{i}$ and $\widehat{\bm L}_{j}$ are weak instance labels of two nodes. Lastly, transition matrix $\bm A$ needs to be normalized as:
\begin{equation}\label{eq:A_norm}
{\bm A}_{ij} = \frac{{\bm A}_{ij}}{\sum_{j \in n}{\bm A}_{ij}}.
\end{equation}
Transition matrix $\bm A$ is shared to use among each group of instance graphs. For individual $k$-th instance graph, the initial graph node score vector $b_0$ is defined by the corresponding binary weak instance label mask $m \in n$, where $m_i=1$ if $\widehat{\bm L}_{i}$ is with $k$-th instance id.
\begin{equation}\label{eq:b_init}
{b_0} =
\begin{cases}
\frac{1}{\sum_{i=0}^{n}{m}_{i}} & {m}_{i} = 1 \\
$0$ & otherwise
\end{cases}
\end{equation}
The initial scores of seeding points are normalized by the total number of weak labels with the same id. The purpose is to fairly balance the initial potential between instance graphs and avoid dominance by instances with more positive weak label numbers. For unlabelled points and points with other weak instance label id, their initial node values are set to zero.
Random walk algorithm is performed by repeatedly adjusting node vector $b$ via transition matrix $\bm A$. At $t$-th iteration, the adjustment can be expressed as
\begin{equation}\label{eq:rw_1}
{b}_{t+1} = \alpha \bm A b_{t} + (1-\alpha)b_{0},
\end{equation}
where $b_{t}$ is the existing node vector derived at the previous random walk step, $b_0$ is the initial node vector, $\alpha \in [0, 1]$ is a blending coefficient between propagated score and initial score.
For random walk with multiple steps, we use $t$ to represent the $t$-th iteration and Expand Eq. (\ref{eq:rw_1}) to
\begin{equation}\label{eq:iter_rw_ex}
{b}_{t+1} =(\alpha \bm A)^{t+1}b_{0} + (1 - \alpha)\sum_{i=0}^{t}(\alpha \bm A)^{i}b_{0}.
\end{equation}
Applying $t \rightarrow \infty$, since $\alpha \in [0,1]$, the first term in Eq. (\ref{eq:iter_rw_ex}) turns into
\begin{align}
\underset{t\rightarrow\infty}{\lim}(\alpha \bm A)^{t+1}b_0 = 0.
\end{align}
For the second term with matrices can be expanded as
\begin{equation}
\underset{t\rightarrow \infty}{\lim} \sum_{i=0}^{t}(\alpha \bm A)^{i} b_0 = (\bm I - \alpha \bm A)^{-1} b_0,
\end{equation}
where ${\bm I}$ is the identity matrix. Thus, the final steady-state of random walk algorithm can be written as
\begin{equation}
{b^{(\infty)}} = (1 - \alpha) ({\bm I} -\alpha {\bm A})^{-1} b_0.
\end{equation}
\begin{figure*}[thb]
\begin{center}
\includegraphics[width=1.0\linewidth]{figure/CGCRW_v7.png}
\end{center}
\vspace{-7mm}
\caption{Illustration on Cross-graph Competing Random Walks (CGCRW). Our algorithm takes a group of points from the same semantic category as input and constructs $k$ graphs according to the number of instances. Proposed method enables the interactions among the same positioned nodes on $k$ instance graphs. Point score can be suppressed or enhanced after cross-graph competition, thereby affecting the following seeding points update strategy. High score points enjoy the priority to be grouped first. After performing several iterations, instance graphs are merged to generate the final output prediction.}
\label{fig:CGCRW}
\end{figure*}
\paragraph{Cross-graph Competing Random Walks (CGCRW)}
On top of random walk algorithm, we design a mechanism to encourage competitive interactions among instance graphs. Intuitively the idea is to suppress a point's activation score in the current graph if its scores in other instance graphs are relatively high. However, the level of repulsive effect needs to be well controlled. Otherwise, too strong repulsive effects are likely to distort the results from the previous random walk.
Based on the random walk results, we apply softmax function to every node score to adjust the probability distribution over $k$ instance graphs. Elements in score vector $b$ is re-scaled within the range of $[0,1]$. And the score values of the same positioned node on $k$ instance graphs are summed to $1$.
\begin{equation}\label{eq:softmax}
b_{i} = \frac{exp(b_{i})}{\sum_{j \in k}{exp(b_{ij}})},
\end{equation}
where $b_{ij}$ denotes $i$-th node on $j$-th graph.
In this simple manner, we bring repulsive interaction among instance graphs. A point that receives less competition from other instance graphs will be adjusted to a relatively higher score, and vice versa.
Then, for each instance, we pick a fixed percentage $\theta$ (i.e. $50\%$) of newly predicted pseudo labels with high confidence to be updated as seeding points for the next iteration. The selection is based on the sorted node scores. Only unlabelled points can be considered as new seeding points.
Our approach gradually groups relatively confident points into seed and performs a random walk step at each iteration.
\begin{algorithm}[thb]
\caption{\small Cross-graph Competing Random Walks (CGCRW)}
\hspace*{0.02in} {\bf Input:} coordinates $\mathbf{X} = \{x_1, x_2, ..., x_N\}\in \mathbb{R}^{N \times 3}$; number of instances $\bm K = \{k_1, k_2, ..., k_{s}\}$; hyperparameter $\alpha, \theta$; max iteration number ${t_1}_{max}$ and ${t_2}_{max}$; instance weak labels $\widehat{\bm L}$; semantic prediction $\bm S$; offset prediction $\bm D$ \\
\hspace*{0.02in}{\bf Output:}
Instance pseudo label prediction $\bm P$ \\
\vspace*{-3mm}
\begin{algorithmic}[1]
\For {$id$ in foreground semantic IDs}
\For {$\bm S \in id$}
\State build $k$ instance graphs ;
\State construct affinity matrix $\bm W$ via Eq. (\ref{eq:W2});
\State construct transition matrix $\bm A$ via Eq. (\ref{eq:A});
\State normalize transition matrix $\bm A$ via Eq. (\ref{eq:A_norm});
\For {$m = 1$ to $k_{s}$}
\State initialize graph node vector $b_{m}$ via Eq. (\ref{eq:b_init});
\EndFor
\While{ $t_1 \leq {t_1}_{max}$ }
\For {$m = 1$ to $k_{s}$}
\State propagate one step via Eq. (\ref{eq:rw_1});
\EndFor
\State $t_1 \gets t_1 + 1$
\EndWhile
\While{ $t_2 \leq {t_2}_{max}$ }
\State adjust node vectors $b$ via Eq. (\ref{eq:softmax})
\For {$m = 1$ to $k_{s}$}
\State reinitialize vector $b_{m}$ via Eq. (\ref{eq:b_init});
\State update top $\theta$ as new seeding points
\State propagate one step via Eq. (\ref{eq:rw_1});
\EndFor
\State $t_2 \gets t_2 + 1$
\EndWhile
\State $p_{i} \gets \argmax{(b^k_i)}$
\State $p_{i} \gets \widehat{\bm L}_{k}$ if under the same mask
\EndFor
\EndFor
\State \Return $\mathbf{P}$
\end{algorithmic}
\end{algorithm}
\begin{figure*}[thb]
\begin{center}
\includegraphics[width=1.0\linewidth]{figure/scannet_results_medium.png}
\end{center}
\caption{The qualitative visualization results of generated pseudo labels on ScanNet-v2 dataset\cite{dai2017scannet}}
\label{fig:scannet_result}
\end{figure*}
\begin{figure*}[thb]
\begin{center}
\includegraphics[width=1.0\linewidth]{figure/s3dis_results_medium.png}
\end{center}
\vspace{-12mm}
\caption{The qualitative visualization results of generated pseudo labels on S3DIS dataset \cite{S3DIS}}
\label{fig:s3dis_result}
\end{figure*}
\begin{figure}[hbt]
\begin{center}
\includegraphics[width=0.9\linewidth]{figure/CGCRW_training_v3.png}
\end{center}
\vspace{-6mm}
\caption{Learning cycle of our proposed weakly supervised method for 3D Instance Segmentation.}
\vspace{-4mm}
\label{fig:CGCRW_training}
\end{figure}
\vspace{-2mm}
\subsection{Learning Strategy}
\label{sec:stgy}
Network training of our method consists of two stages. The first stage is supervised by the initial weak labels. Afterward, predictions with high confidence from our pseudo label generation algorithm are further updated as new ground-truth labels for next stage training. With this learning strategy, the quality of learned feature can be consistently improved.
\begin{table*}[tb]
\footnotesize
\begin{center}
\begin{tabular}{l|c|l|l}
\toprule
\textbf{Semantic mIoU} & Label& wall~\,floor~\,cab~~\,bed~~\,chair~\,sofa~\,table~\,door~\,wind~\,bkshf~~pic~~~cntr~~\,desk~~curt~\;fridg~\,showr~\,toil~~\,sink~~bath~\,ofurn & ~avg\\
\midrule
MPRM~\cite{wei2020multi} & Scene & 47.3~\,41.1~\,10.4~\,43.2~~25.2~~43.1~\;21.9~\;\;9.8\;\;\,12.3~~\,45.0~~\,\;9.0\;\;\,13.9~~21.1~~40.9~~\;1.8\;\;\,\,29.4~~\,14.3~\,\;9.2~\;\,39.9~~10.0&~24.4\\
MPRM~\cite{wei2020multi} & Subcloud & 58.0~\,57.3~\,33.2~\,71.8~~50.4~~69.8~\;47.9~\;42.1~~44.9~~\,73.8~~\,28.0~~21.5~~49.5~~72.0~~38.8~~\,44.1~~\,42.4~\,20.0~~48.7~~34.4&~47.4\\
\midrule
SegGroup \cite{tao2020seggroup} & 0.02\% & 71.0~\,82.5~\,63.0~\,52.3~~72.7~~61.2~\;65.1~\;66.7~~55.9~~\,46.3~~\,42.7~~50.9~~50.6~~67.9~~67.3~~\,70.3~~\,70.7~\,53.1~~54.5~~63.7&~61.4\\
\midrule
\midrule
\textbf{RWSeg (Ours)} & 0.02\% & \textbf{88.8}~\,\textbf{94.4}~\,\textbf{80.2}~\,\textbf{82.4~~}\textbf{85.9}~~\textbf{91.2}~\;\textbf{76.5}~\;\textbf{76.6}~~\textbf{78.2}~~\,\textbf{87.5}~~\,\textbf{66.3}~~\textbf{64.1}~~\textbf{67.7}~~\textbf{85.6}~~\textbf{86.9}~~\,\textbf{88.9}~~\,\textbf{92.4}~\,\textbf{71.5}~~\textbf{91.7}~~\textbf{75.3}&~\textbf{81.6}\\
\bottomrule
\end{tabular}
\end{center}
\vspace{-4mm}
\caption{Pseudo label quality of semantic segmentation (category-level) on ScanNet-2 \cite{dai2017scannet} training set. }
\label{sem_iou_train}
\end{table*}
\begin{table*}[tb]
\footnotesize
\begin{center}
\begin{tabular}{l|c|l|l}
\toprule
\textbf{Instance AP} & Metric& cab~~\,bed~~\,chair~\,sofa~\,table~\,door~\,wind~\,bkshf~~pic~~~cntr~~\,desk~~curt~\;fridg~\,showr~\,toil~~\,sink~~bath~\,ofurn & ~avg\\
\midrule
\midrule
\multirow{3}{*}{\textbf{RWSeg (Ours)}} & AP & 59.0~\,65.9~~70.3~~82.1~\;59.3~\;38.2~~54.0~~\,68.0~~\,54.7~~35.6~~35.8~~48.0~~73.9~~\,80.6~~\,88.4~\,44.6~~85.4~~54.3&~\textbf{61.0}\\
& AP{\tiny 50} & 85.7~\,94.2~~93.8~~90.3~\;87.1~\;60.8~~77.1~~\,84.5~~\,81.6~~79.9~~74.1~~69.7~~92.1~~\,92.4~~\,97.8~\,82.3~~97.4~~77.6&~\textbf{84.4}\\
& AP{\tiny 25} & 96.4~\,99.0~~98.3~~95.7~\;95.2~\;87.0~~91.6~~\,91.5~~\,92.8~~96.3~~93.9~~87.5~~99.2~~\,97.4~~\,99.3~\,95.8~100.0~92.8&~\textbf{95.0}\\
\bottomrule
\end{tabular}
\vspace{-4mm}
\end{center}
\caption{Pseudo label quality of instance segmentation on ScanNet-2 \cite{dai2017scannet} training set. }
\label{ins_ap_train}
\end{table*}
\vspace{-4mm}
\section{Experiments}
\label{sec:exp}
\vspace{-2mm}
In this section, we show our experimental results on two public datasets: ScanNet-v2 \cite{dai2017scannet} and S3DIS \cite{S3DIS} to show the effectiveness of our proposed method. We first evaluate the accuracy of the generated pseudo labels, in comparison with full ground-truth labels on the training set of ScanNet-v2 \cite{dai2017scannet}. Then, we compare our method with other existing methods on the validation set and test set. Comparisons are made with both weakly supervised methods and fully supervised methods.
\vspace{-5mm}
\paragraph{Implementation details}
We set the voxel size as $2 cm$ for submanifold sparse convolution \cite{Submanifold} based backbone, following \cite{pointgroup}. Our network is trained on a single GPU card. For each stage of training, the backbone network and self-attention module are trained sequentially, with a batch size of 4 and 2. We set $\gamma$ and $\delta$ in the self-attention module as two-layer MLPs with the hidden dimension of 64 and 32 respectively. For CGCRW algorithm, we set hyerparameters $\alpha = 0.2, {t_1}_{max} = 1, {t_2}_{max} = 5$ and $\theta$ as 50\%. Due to GPU memory limit, we subsample input points to CGCRW, if the number is above 25k. Last output remains at original resolution. For network training, we use Adam solver for optimization with an initial learning rate of 0.001.
\subsection{Evaluations on ScanNet-v2 Dataset}
ScanNet-v2 dataset \cite{dai2017scannet} is a popular 3D indoor dataset containing 2.5 million RGB-D views in 1513 real-world scenes. The 3D meshed data are annotated with semantic labels and instance labels for 18 class categories. Compared with other 3D indoor datasets \cite{S3DIS, sunrgbd}, ScanNet-v2 \cite{dai2017scannet} is more complete and well-recognized. Evaluation of 3D instance segmentation uses mean average precision at different overlap percentage as metric, which are [email protected], [email protected] and mAP respectively.
\paragraph{Pseudo label evaluation}
As shown in Table \ref{sem_iou_train} and Table \ref{ins_ap_train}, we present the quality of generated pseudo labels based on our proposed method. Reported final pseudo labels are created after two stages of network training. Our network is trained only on the training set of ScanNet-v2 \cite{dai2017scannet} with 1201 scenes, no extra data is needed.
In Table \ref{sem_iou_train}, the semantic quality of our pseudo labels largely outperforms previous methods by at least $20.2\%$. Besides, we also report the instance quality of pseudo labels in Table \ref{ins_ap_train}. However, no available data from other methods can be used for comparison at present. Our qualitative pseudo labels can be used by any fully supervised method to resolve their annotation cost issue.
\paragraph{Prediction evaluation}
Different from weakly supervised methods like SegGroup \cite{tao2020seggroup} that require training another a new network for prediction, we can directly adopt other methods on the same network for prediction without retraining. Here we employ a Breadth-First Search (BFS) clustering algorithm from PointGroup \cite{pointgroup} to our network.
In Table \ref{ins_val}, we compare the prediction results with fully supervised PointGroup \cite{pointgroup} and other weakly supervised methods on ScanNet-v2 \cite{dai2017scannet} validation set.
Our method significantly outperforms SegGroup \cite{tao2020seggroup} over all evaluation metrics, generally \textbf{ by an absolute margin of over 10 points}.
Remarkably, with only $0.02\%$ of annotated points, we \textbf{achieve comparable results with fully supervised method \cite{pointgroup}}.
We also report the instance segmentation results on ScanNet-v2 \cite{dai2017scannet} test set in Table \ref{ins_test}.
Our method again performs significantly better than the recent SegGroup \cite{tao2020seggroup} which uses the same amount of supervision as ours.
\begin{table}[tbh]
\footnotesize
\vspace{-7pt}
\begin{center}
\begin{tabular}{l|c|ccc}
\toprule
Method & Supervision & ~AP & \hspace{-3pt}AP{\tiny 50}\hspace{-3pt} & AP{\tiny 25}\\
\midrule
Full Supervision: &&&&\\
PointGroup~\cite{pointgroup} & 100\% & 34.8 & \hspace{-3pt}56.9\hspace{-3pt} & 71.3\\
\midrule
Init+Act. Point Supervision: &&&&\\
CSC-20 (PointGroup)~\cite{hou2021exploring} & 20 pts/scene & - & \hspace{-3pt}27.2\hspace{-3pt} & - \\
CSC-50 (PointGroup)~\cite{hou2021exploring} & 50 pts/scene & - & \hspace{-3pt}35.7\hspace{-3pt} & - \\
\midrule
One Obj One Pt Supervision: &&&&\\
SegGroup (PointGroup) \cite{tao2020seggroup} &0.02\% & ~23.4 & \hspace{-3pt}43.4\hspace{-3pt} & 62.9 \\
\midrule
\midrule
\textbf{RWSeg (Ours)} &0.02\% & ~34.7 & \hspace{-3pt}56.4\hspace{-3pt} & 71.2 \\
\bottomrule
\end{tabular}
\end{center}
\vspace{-4mm}
\caption{Instance segmentation results on ScanNet-v2 \cite{dai2017scannet} validation set. Methods marked with brackets in the names represents using generated pseudo labels to train another fully-supervised method (such as PointGroup) for evaluation.}
\label{ins_val}
\vspace{-10pt}
\end{table}
\begin{table}[htb]
\footnotesize
\begin{center}
\begin{tabular}{l|c|ccc}
\toprule
Method & Supervision & ~AP & AP{\tiny 50} & ~AP{\tiny 25}~\\
\midrule
Full Supervision:~~ &&&&\\
OccuSeg~\cite{occuseg} & 100\% & ~48.6 & 67.2 & 78.8~\\
PointGroup~\cite{pointgroup} & 100\% & ~40.7 & 63.6 & 77.8~\\
3D-MPA~\cite{3dmpa} & 100\% & ~35.5 & 61.1 & 73.7~\\
MTML~\cite{MTML} & 100\% & ~28.2 & 54.9 & 73.1~\\
3D-BoNet~\cite{3DBoNet} & 100\% & ~25.3 & 48.8 & 68.7~\\
3D-SIS~\cite{3DSIS} & 100\% & ~16.1 & 38.2 & 55.8~\\
GSPN~\cite{gspn} & 100\% & ~15.8 & 30.6 & 54.4~\\
SGPN~\cite{SGPN} & 100\% & ~4.9 & 14.3 & 39.0~\\
\midrule
One Obj One Pt Supervision: &&&&\\
SegGroup (PointGroup) \cite{tao2020seggroup} & 0.02\% & ~24.6 & 44.5 & 63.7~ \\
\midrule
\midrule
\textbf{RWSeg (Ours)} & 0.02\% & ~34.8 & 56.7 & 73.9~ \\
\bottomrule
\end{tabular}
\end{center}
\vspace{-4mm}
\caption{Instance segmentation results on ScanNet-v2 \cite{dai2017scannet} test set.}
\label{ins_test}
\end{table}
\subsection{Evaluations on S3DIS Dataset}
We also evaluate our proposed method on S3DIS dataset \cite{S3DIS}. The dataset has 272 scenes under six large-scale indoor areas. Unlike ScanNet \cite{dai2017scannet}, all 13 classes including background are annotated as instances and require prediction. We use the mean precision (mPre) and mean recall (mRec) with an IoU threshold of 0.5 as evaluation metric. Our results is $60.1/45.8$ on Area 5 and $68.9/56$ on 6-fold cross validation over six areas. Other weakly supervised methods have not reported their results on S3DIS dataset \cite{S3DIS}, thus cannot be compared at present.
\subsection{Ablation Study}
In this section, we proceed to study the impacts of different components of our proposed method.
Table \ref{tab:ablation_network} shows the network performance at different stages of training. We use ``XFMR" to represent the self-attention module in our network. In the setting of ``3D U-Net + XFMR", we freeze the backbone network and only train self-attention module, which shows the effectiveness of this component. Stage 1 training is supervised by initial weak labels. And Stage 2 training is supervised by the generated pseudo labels from our algorithm at the end of Stage 1. With our training strategy, the quality of semantic features can be steadily improved.
\begin{table} [h]
\centering
\scalebox{0.9}{
\begin{tabular}{c|l|cc}
\toprule
mIoU & ~~~~~~~~Method & train set & val set \\
\midrule
Stage 1 & 3D U-Net & 74.6 & 61.7 \\
Stage 1 & 3D U-Net + XFMR & 78.9 & 66 \\
Stage 2 & 3D U-Net & 80 & 68.4 \\
Stage 2 & 3D U-Net + XFMR & 81.6 & 70.3 \\
\bottomrule
\end{tabular}
}
\vspace{-2mm}
\caption{Ablation study for network components. ``3D U-Net'' indicates our backbone network, and ``XFMR'' means our proposed self-attention module for feature propagation. Evaluated on ScanNet-v2 \cite{dai2017scannet} validation set.}
\label{tab:ablation_network}
\end{table}
\vspace{-6mm}
\paragraph{Ablations on Cross-graph Competing Random Walks (CGCRW)}
To make fair comparisons on clustering algorithms for pseudo label generation, we train a PointGroup \cite{pointgroup} backbone network with initial weak labels. On top of the shared network, we evaluate the performance of our CGCRW and other baseline methods in Table \ref{tab:ablation_CGCRW}. ``PointGroup BFS" represents a popular Breadth-First Search algorithm used in fully supervised 3D instance segmentation. K-means \cite{Hartigan1979} is a simple yet powerful unsupervised clustering algorithm to separate samples in $k$ groups of equal variance. Its character suits our task very well by nature. However, we found K-means is very sensitive to noise. The performance highly depends on the quality of semantic predictions and shift vectors. In contrast, our CGCRW is more robust and works well in different situations.
\begin{table} [h]
\centering
\scalebox{0.9}{
\begin{tabular}{c|ccc}
\toprule
Baseline Methods & AP & AP{\tiny 50} & AP{\tiny 25} \\
\midrule
PointGroup BFS \cite{pointgroup} & 15.8 & 32.4 & 58.9 \\
K-means$^\dag$ \cite{Hartigan1979} & 14.5 & 28.5 & 66.9 \\
K-means$^\ddag$ \cite{Hartigan1979} & 23.5 & 44.1 & 72.5 \\
\midrule
\midrule
\textbf{CGCRW$^\dag$ (Ours)} & 53.2 & 80.6 & 95.2 \\
\textbf{CGCRW$^\ddag$ (Ours)} & \textbf{55} & \textbf{82} & \textbf{95.9} \\
\bottomrule
\end{tabular}
}
\vspace{-2mm}
\caption{Comparison with pseudo label generation baseline methods on ScanNet-v2 \cite{dai2017scannet} training set. Methods marked with $^\dag$ are based on original coordinates. Methods marked with $^\ddag$ are based on shifted coordinates. BFS uses both sets of coordinates.}
\label{tab:ablation_CGCRW}
\end{table}
\begin{figure}[htb]
\vspace{-2mm}
\begin{center}
\includegraphics[width=1.0\linewidth]{figure/CGCRW_abalation.png}
\end{center}
\vspace{-8mm}
\caption{Visualized example of CGCRW's seeding regions at different iterations.}
\label{fig:CGCRW_ablation_iter}
\end{figure}
Figure \ref{fig:CGCRW_ablation_iter} shows the change of seeding regions during the process of Cross-graph Competing Random Walks. At each step, the top $50\%$ of the new predictions on unlabelled points are added as seed. It can be seen that new seeding points tend to be distributed at those regions relatively far from other seeds, as a result of cross-graph competition.
\begin{table} [h]
\centering
\scalebox{0.9}{
\begin{tabular}{c|cccc}
\toprule
Iteration number (${t_2}_{max}$) & 0 & 1 & 5 & 10 \\
\midrule
~~~chair~~~~~~~~ AP $\uparrow$ & 64.2 & 66.3 & 67.4 & 67.4 \\
bookshelf~~~ AP $\uparrow$ & 48.1 & 51.0 & 52.3 & 52.3 \\
\bottomrule
\end{tabular}
}
\vspace{-2mm}
\caption{Impact of iteration number on CGCRW ($\theta=50\%$). Evaluated in AP for chair and bookshelf class on ScanNet-v2 \cite{dai2017scannet} training set.}
\label{tab:ablation_CGCRW_iter}
\end{table}
\vspace{-4mm}
The impact of using multiple random walk steps in Cross-graph Competing Random Walks (CGCRW) is shown in Table \ref{tab:ablation_CGCRW_iter}. ${t2}_{max}=0$ represents converged results from the basic RWSeg without competing mechanism. As we expected, cross-graph competition is useful to resolve those ambiguous cases in instance segmentation, where objects from same category are compactly placed. Meanwhile, for those sparsely placed object categories, such as bathtub and door, their instance segments can already be well defined by proposed basic random walk algorithm. Competitions usually not exist for such cases.
\vspace{-2mm}
\section{Conclusion}
\label{sec:ccl}
\vspace{-2mm}
In this paper, we propose a novel weakly supervised method for 3D instance segmentation on point clouds. With significantly fewer annotations, our network uses a self-attention module to propagate semantic feature and a random walk based algorithm with cross-graph competition to generate qualitative pseudo labels. Comprehensive experiments show that our method achieves solid improvements on performance.
\clearpage
\newpage
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction:}
Let $D\in {\mathbb Z}$ be squarefree and odd. Our convention is that $\infty |D$ if $D<0$.
\textcolor{black}{It is a standard result} that there exists a quadratic extension
$K/ {\mathbb Q}$ ramified at exactly the set $\{v : v| D \}$ if and only if $D\equiv 1$ mod $4$.
The key is how the Frobenius elements of the $v|D$ lie in the Galois group of the `governing extension' $ {\mathbb Q}(i)/ {\mathbb Q}$.
Let $\sigma_v$ denote Frobenius at $v$ in this extension with $\sigma_\infty$ being the nontrivial element of $Gal( {\mathbb Q}(i)/ {\mathbb Q})$.
We frame this as the following theorem:
\vskip1em\noindent
{\bf Theorem} {\it There exists a quadratic extension $K/ {\mathbb Q}$ ramified exactly at a tame (not containing $2$
\textcolor{black}{but allowing $\infty$}) set $S$ of places if and only if $\displaystyle \sum_{v\in S} \sigma_v $ is the trivial element in $Gal( {\mathbb Q}(i)/ {\mathbb Q})$.}
\noindent
In \cite{GM} this result was generalized to $ {\mathbb Z}/p{\mathbb Z} $-extensions of a general number field $K$.
For a fixed prime $p$ and set $S$ of tame places, set
$$V_S = \{ x \in K^\times \mid (x)=J^p; \,\,x \in K^{\times p}_v \,\,\forall \,\,v \in S\}.$$
Note $K^{\times p} \subset V_S$ for all $S$ and $S\subset T \implies V_T \subseteq V_S$.
Let ${\mathcal O}^{\times}_K$ and $Cl_K[p]$ be, respectively, the units of $K$ and the $p$-torsion in the class group of $K$.
\textcolor{black}{It is a standard result that } $V_\emptyset/K^{\times p}$ lies in the exact sequence (see Proposition 10.7.2 of \cite{NSW}):
$$0 \to {\mathcal O}^{\times}_K \otimes {\mathbb F} _p \to V_\emptyset / K^{\times p} \to Cl_K[p] \to 0.$$
Set $K'=K(\mu_p)$, $L=K'(\sqrt[p]{V_\emptyset})$ and let $r_1$ and $r_2$ be the number of real and pairs of complex embeddings of $K$. We call $L/K'$ the {\it governing extension} for $K$.
When $K= {\mathbb Q}$ and $p=2$ we see $L= {\mathbb Q}(i)$ and have recovered the field of the theorem above.
\begin{center}
\begin{tikzcd}
& L:=K'(\sqrt[p]{V_\emptyset}) \\
K':=K(\mu_p) \arrow[ur,dash]\\
K\arrow[u, dash]
\end{tikzcd}
\end{center}
As $L$ is obtained by adjoining to $K'$ the $p$th roots of elements of $K$ (not $K'$), one easily shows that places $v'_1,v'_2$ of $K'$ above a fixed place $v$ of $K$ have Frobenius elements in $ Gal(L/K')$ that differ by a nonzero scalar multiple
We abuse notation and for any $v'$ of $K'$ above $v$ in $K$ denote Frobenius at $v'$ by $\sigma_v$. The theorem of
\cite{GM} \textcolor{black}{(also see Chapter V of \cite{gras})}
below and Theorem~\ref{theorem:main} implicitly use this abuse of notation.
\vskip1em\noindent
{\bf Theorem} (Gras-Munnier) {\it Let $p$ be a prime and $S$ a finite set of {\it tame} places \textcolor{black}{ (allowing Archimedean places if $p=2$)} of $K$. There exists a ${\mathbb Z}/p{\mathbb Z}$-extension of $K$ ramified at {\it exactly} the places of $S$ if and only if there exists a dependence relation
$\displaystyle \sum_{v\in S} a_v \sigma_v =0$ in the $ {\mathbb F} _p$-vector space $Gal(L/K')$ with all $a_v \neq 0$.}
\noindent
The original proof uses class field theory in a fairly complicated way. Theorem~\ref{theorem:main}
is a generalization of the Gras-Munnier Theorem. We first give a short proof
that uses only one element of class field theory, \eqref{eq:cft} below,
and elementary linear algebra.
We easily prove Proposition~\ref{prop:dims} from \eqref{eq:cft}, after which one only needs a standard inclusion-exclusion argument.
\textcolor{black}{The cardinalities of the two sets of Theorem~\ref{theorem:main} being equal suggests a duality.
In the final section of this note we give an alternative proof of Proposition~\ref{prop:dims} using the Wiles-Greenberg formula whose proof requires the full strength of global duality.}
\textcolor{black}{Denote by $G_S$ the Galois group over $K$ of its maximal extension pro-$p$ unramified outside $S$.}
Our main result is:
\begin{theorem}\label{theorem:main}
Let $p$ be a prime and $S$ a finite set of {\it tame} places of a number field $K$ \textcolor{black}{ (allowing Archimedean places if $p=2$)}. The sets
$$\left\{ f\in \frac{H^1(G_S, {\mathbb Z}/p{\mathbb Z} )}{H^1(G_\emptyset, {\mathbb Z}/p{\mathbb Z} )} \,\mid \, \mbox{ the extension } K_f/K \mbox{ fixed by } Ker(f) \mbox{ is ramified exactly at the places of } S\right\}$$
and
$$\{\mbox{The dependence relations }\sum_{v\in S} a_v\sigma_v=0 \mbox{ in } Gal(L/K')\mbox{ with all } a_v\neq 0 \}$$
have the same cardinality.
\end{theorem}
It is an easy exercise to see both sets have cardinality at most one when $p=2$, so the bijection is natural in this case.
\textcolor{black}{We thank Brian Conrad for pointing out to us a proof of Lemma~\ref{lemma:XXX}(ii) and Peter Uttenthal for helpful suggestions.}
\section{Proof of the Gras-Munnier Theorem}
\textcolor{black}{For any field $E$ set $\delta(E) = \left\{\begin{array}{cc} 1 & \mu_p \subset E\\0 &\mu_p \not \subset E \end{array}\right. .$ Observe that by Dirichlet's unit theorem $Gal(L/K')$ is an $ {\mathbb F} _p$-vector space of dimension $r_1+r_2-1+\delta(K)+ \dim Cl_K[p]$. }
The standard fact from class field theory that we need (see \S 11.3 of \cite{K} or \S 10.7 of \cite{NSW})
\textcolor{black}{is a formula of Shafarevich and Koch} for
the dimension of the space of $ {\mathbb Z}/p{\mathbb Z} $-extensions of $K$ unramified outside a tame set $Z$:
\begin{equation}\label{eq:cft}
\dim H^1(G_Z, {\mathbb Z}/p{\mathbb Z} ) =-r_1-r_2+1-\delta(K) + \dim (V_Z/K^{\times p}) + \left(\sum_{v \in Z} \delta(K_v) \right) .
\end{equation}
Fix a tame set $S$ noting that $H^1(G_S, {\mathbb Z}/p{\mathbb Z} )$ may include cohomology classes that cut out
$ {\mathbb Z}/p{\mathbb Z} $-extensions of $K$ that could be ramified at proper subsets of $S$.
If $\delta(K_v) =0$ for $v \in S$, there are no ramified $ {\mathbb Z}/p{\mathbb Z} $-extensions of $K_v$ and thus no $ {\mathbb Z}/p{\mathbb Z} $-extensions of $K$ ramified at $v$, so we always assume $\delta(K_v)=1$. Then, as we vary $Z$ from $\emptyset$ to $S$ one place at a time, $\dim (V_Z/K^{\times p})$ may remain the same or decrease by $1$. In these cases $\dim H^1(G_Z, {\mathbb Z}/p{\mathbb Z} )$ increases by $1$ or remains the same respectively.
Let $W \subset Gal(L/K')$ be the $ {\mathbb F} _p$-subspace spanned by $\langle \sigma_v\rangle_{v\in S}$,
\textcolor{black}{the Frobenius elements of the places in $S$. We will show the set of dependence relations on these Frobenius elements all of whose coefficients are nonzero has the same cardinality as the set of ${\mathbb Z}/p{\mathbb Z}$-extensions of $K$ ramified exactly at the places of $S$.}
Let
$I:= \{u_1,u_2,\cdots,u_r\} \subset S$
be such that
$\{\sigma_{u_1},\sigma_{u_2},\cdots,\sigma_{u_r} \}$
form a basis of $W$ and let
$D:=\{ w_1,w_2,\cdots ,w_s\} \subset S$
be the remaining elements of $S$. We think of the $\sigma_{u_i}$ as independent elements
and the $\sigma_{w_j}$ as the dependent elements.
As we vary $X$ in \eqref{eq:cft} from $\emptyset$ to $I$ by adding in one $u_i$ at a time, we are adding $1$ through the $\delta(K_{u_i})$ term
to the right side of \eqref{eq:cft},
but $\dim V_X/K^{\times p}$ becomes
one dimension smaller, so both sides remain unchanged.
Then, as we add in the places $w_j$ of $D$ to get to $S=I\cup D$ we have $V_I/K^{\times p} = V_S/K^{\times p}$.
Thus
\begin{equation}\label{eq:h1eq}
H^1(G_\emptyset, {\mathbb Z}/p{\mathbb Z} ) = H^1(G_I, {\mathbb Z}/p{\mathbb Z} )=\dim H^1(G_S, {\mathbb Z}/p{\mathbb Z} ) - s \implies \dim \frac{H^1(G_S, {\mathbb Z}/p{\mathbb Z} )}{H^1(G_\emptyset, {\mathbb Z}/p{\mathbb Z} )}=s.
\end{equation}
We can write each $\sigma_{w_j}$ as a linear combination of the
$\sigma_{u_i}$
in a unique way:
$$R_j:\,\, \sigma_{w_j} - \sum^r_{i=1} F_{ji} \sigma_{u_i}=0. $$
For $X\subset S$ let $R_X$ be the $ {\mathbb F} _p$-vector space of all dependence relations on the elements
$\{\sigma_v\}_{v\in X} \subset Gal(L/K')$.
We prove a preliminary result in the spirit of Theorem~\ref{theorem:main}.
\begin{lemma}\label{lemma:dep}
The set $\{R_1,R_2,\cdots,R_s\}$ forms a basis of the \textcolor{black}{$ {\mathbb F} _p$-vector space of} dependence relations on the $\sigma_{u_i}$ and $\sigma_{w_j}$.\\
\end{lemma}
\begin{proof}
Consider {\it any} dependence relation $R$ among the $\sigma_{u_i}$ and $\sigma_{w_j}$. We can eliminate each $\sigma_{w_j}$ by adding to $R$ a suitable multiple of $R_j$. We are then left with a dependence relation on the $\sigma_{u_i}$, which are independent, so it is trivial, proving the lemma.
\end{proof}
\begin{prop} \label{prop:dims} For any $X \subseteq S$,
$\displaystyle\dim R_X = \dim \left( \frac{H^1(G_X, {\mathbb Z}/p{\mathbb Z} )}{H^1(G_\emptyset, {\mathbb Z}/p{\mathbb Z} )}\right).$
\end{prop}
\begin{proof} Lemma~\ref{lemma:dep} and ~\eqref{eq:h1eq} prove this for $X=S$. Apply the same proof to $X\subset S$.
\end{proof}
\vskip1em\noindent
\textcolor{black}{\it Proof of Theorem~\ref{theorem:main}}:
Proposition~\ref{prop:dims} does {\it not} complete the proof as $R_S$ may contain dependence relations with support properly contained in $S$ and
$\displaystyle \frac{H^1(G_S, {\mathbb Z}/p{\mathbb Z} )}{H^1(G_\emptyset, {\mathbb Z}/p{\mathbb Z} )}$ may contain elements giving rise to extensions of $K$ ramified at proper subsets of $S$.
\vskip1em\noindent
{\it Proof Theorem~\ref{theorem:main}:}
The set of dependence relations with support {\it exactly} in $S$ is
\begin{equation}\label{eq:rs}
R_S \backslash \bigcup_{v \in S} R_{S \backslash \{v\}},
\end{equation}
those with support contained in $S$ less \textcolor{black}{the union of} those with proper maximal support in $S$.
For any sets $A_i \subset S$ \textcolor{black}{it is clear that}
$$\bigcap^k_{i=1} R_{A_i} = R_{\bigcap^k_{i=1} A_i},$$
so
by inclusion-exclusion
\begin{equation}\label{eq:rai2} \# \bigcup_{v \in S} R_{S \backslash \{v\}} =\sum_{v\in S} \# R_{S \backslash \{v\}} -
\sum_{v\neq w \in S} \# R_{S \backslash \{v,w\}} +\cdots
\end{equation}
Similarly the set of cohomology classes giving rise to ${\mathbb Z}/p{\mathbb Z}$-extensions ramified exactly at the places of $S$ (up to unramified extensions) is
\begin{equation}\label{eq:aai2}
\frac{H^1(G_{S}, {\mathbb Z}/p{\mathbb Z} )}{H^1(G_\emptyset, {\mathbb Z}/p{\mathbb Z} )} \backslash \bigcup_{v\in S} \frac{ H^1(G_{ S\backslash \{v\} }, {\mathbb Z}/p{\mathbb Z} )}{H^1(G_\emptyset, {\mathbb Z}/p{\mathbb Z} )}.
\end{equation}
Since for any sets $A_i \subset S$ we have
$$\bigcap^k_{i=1} \frac{H^1(G_{A_i}, {\mathbb Z}/p{\mathbb Z} )}{H^1(G_\emptyset, {\mathbb Z}/p{\mathbb Z} )} =\frac{ H^1(G_{ \cap^k_{i=1} A_i}, {\mathbb Z}/p{\mathbb Z} )}{H^1(G_\emptyset, {\mathbb Z}/p{\mathbb Z} )},$$
we see
\begin{equation}\label{eq:aai3}
\# \bigcup_{v\in S} \frac{ H^1(G_{ S\backslash \{v\} }, {\mathbb Z}/p{\mathbb Z} )}{H^1(G_\emptyset, {\mathbb Z}/p{\mathbb Z} )} = \sum_{v\in S}
\# \frac{ H^1(G_{ S\backslash \{v\} }, {\mathbb Z}/p{\mathbb Z} )}{H^1(G_\emptyset, {\mathbb Z}/p{\mathbb Z} )}
- \sum_{v\neq w \in S} \# \frac{ H^1(G_{ S\backslash \{v,w\} }, {\mathbb Z}/p{\mathbb Z} )}{H^1(G_\emptyset, {\mathbb Z}/p{\mathbb Z} )} + \cdots
\end{equation}
Proposition~\ref{prop:dims} implies the terms on the right sides of \eqref{eq:rai2} and \eqref{eq:aai3} are equal so the left sides are equal as well. The result follows from~\eqref{eq:rs}, ~\eqref{eq:aai2} and applying Proposition~\ref{prop:dims} with $X=S$.
\hfill $\square$
\section{A proof via the Wiles-Greenberg formula}
As the association of dependence relations and cohomology classes in Theorem~\ref{theorem:main} resembles a duality result, we now prove Proposition~\ref{prop:dims} using the Wiles-Greenberg formula, which follows from global duality.
We assume familiarity with local and global Galois cohomology and their duality theories.
\textcolor{black}{As we will need to apply the Wiles-Greenberg formula, we henceforth assume its hypothesis that
$Z$} is a set of places of $K$ containing all those above infinity and $p$.
For each $v\in Z$, let $G_v:=Gal(\bar{K}_v/K_v)$ and consider a subspace $L_v \subset H^1(G_v,{\mathbb Z}/p{\mathbb Z})$. Under the local duality pairing (see Chapter 7, \S 2 of \cite{NSW})
$$H^1(G_v,{\mathbb Z}/p{\mathbb Z}) \times H^1(G_v,\mu_p) \to H^2(G_v,\mu_p) \simeq \frac1p{\mathbb Z}/{\mathbb Z}$$
$L_v$ has an annihilator $L^\perp_v \subset H^1(G_v,\mu_p)$.
Set $$H^1_{\mathcal L}(G_Z,{\mathbb Z} / p{\mathbb Z}) := \mbox{Ker} \left( H^1(G_Z,{\mathbb Z} / p{\mathbb Z}) \to \oplus_{v\in Z}
\frac{H^1(G_v, { {\mathbb Z}/p{\mathbb Z}})}{L_v}\right)$$
and
$$H^1_{{\mathcal L}^\perp}(G_Z,\mu_p) := \mbox{Ker} \left( H^1(G_Z,\mu_p) \to \oplus_{v\in Z}
\frac{H^1(G_v,\mu_p)}{L^\perp_v}\right).$$
We call $\{L_v\}_{v \in Z}$ and $\{ L^\perp_v \}_{v\in Z}$ the Selmer and dual Selmer conditions and
$H^1_{\mathcal L}(G_Z,{\mathbb Z} / p{\mathbb Z})$ and $H^1_{{\mathcal L}^\perp}(G_Z,\mu_p)$ the Selmer and dual Selmer groups.
We state two results that we need for our second proof of Proposition~\ref{prop:dims}. As Lemma~\ref{lemma:XXX} (ii) is perhaps not so well-known, we include \textcolor{black}{ a sketch of its proof.}
\begin{lemma} \label{lemma:XXX}
(i) Suppose $v\nmid p$. Then $H_{nr}^1(G_v,{\mathbb Z}/p{\mathbb Z})$ and $H_{nr}^1(G_v,\mu_p)$, the unramified cohomology classes, are exact annihilators of one another under the local duality pairing.\\
(ii) Suppose $v|p$. \textcolor{black}{Recall $K'_v=K_v(\mu_p)$.} The annihilator in
$H^1(G_v,\mu_p)$ of $H_{nr}^1(G_v,{\mathbb Z}/p{\mathbb Z}) \subset H^1(G_v,{\mathbb Z}/p{\mathbb Z})$ is $H^1_f(G_v,\mu_p)$, the peu ramifi\'{e}e classes, namely those $f\in H^1_f(G_v,\mu_p)$ whose fixed field $L_{v,f}$ of Kernel$(f |_{G_{K'_v}})$ arises from adjoining the $p$th root of a unit $u_f \in K_v$.
\end{lemma}
\begin{proof}
(i) This is standard - see 7.2.15 of \cite{NSW}. \newline\noindent
(ii)
\textcolor{black}{Cohomology taken over $Spec({\mathcal O}_{K_v})${ in what follows is flat.}
Here $$H^1_{f}(G_v,\mu_p) =
H^1(Spec({\mathcal O}_{K_v}),\mu_p) = { {\mathcal O}^\times_{K_v}}/ { {\mathcal O}^{\times p}_{K_v}}
\subset K^{\times}_v/K^{\times p}_v$$
where the containment is codimension one as $ {\mathbb F}_p$-vector spaces. Recall $$ { {\mathbb Z}/p{\mathbb Z}} \simeq H^1_{nr}(G_v, { {\mathbb Z}/p{\mathbb Z}})
= H^1(Spec({\mathcal O}_{K_v}), { {\mathbb Z}/p{\mathbb Z}})$$ and by Lemma 1.1 of Chapter III of \cite{Milne} we have the injections
$$H^1(Spec({\mathcal O}_{K_v}, { {\mathbb Z}/p{\mathbb Z}})) \hookrightarrow H^1(G_v, { {\mathbb Z}/p{\mathbb Z}}) \mbox{ and } H^1(Spec({\mathcal O}_{K_v},\mu_p)) \hookrightarrow H^1(G_v,\mu_p)$$
and the pairing
$$H^1(Spec({\mathcal O}_{K_v}), { {\mathbb Z}/p{\mathbb Z}}) \times H^1(Spec({\mathcal O}_{K_v}),\mu_p) \to H^2(Spec({\mathcal O}_{K_v}),\mu_p)=0$$
which is consistent with the local duality pairing
$$H^1(G_v, { {\mathbb Z}/p{\mathbb Z}})\times H^1(G_v,\mu_p) \to H^2(G_v,\mu_p)=\frac1{p}{\mathbb Z} /{\mathbb Z}.$$
As $H^1(Spec({\mathcal O}_{K_v}), { {\mathbb Z}/p{\mathbb Z}})= H^1_{nr}(G_v,{\mathbb Z}/p{\mathbb Z})$ and $H^1(Spec({\mathcal O}_{K_v}),\mu_p) =H^1_f(G_v,\mu_p)$ are, respectively, dimension $1$ and codimension $1$ in
$H^1(G_v, { {\mathbb Z}/p{\mathbb Z}})$ and $H^1(G_v,\mu_p)$, they are exact annihilators of one another,
proving (ii).}
\end{proof}
\noindent
{\bf Theorem} {\it (Wiles-Greenberg) \textcolor{black}{Assume $Z$ contains all places above $\{p,\infty\}$. Then}
\begin{eqnarray*}
\begin{tabular}{ll}
$\dim H^1_{\mathcal L}(G_Z,{\mathbb Z} / p{\mathbb Z}) - \dim H^1_{{\mathcal L}^\perp}(G_Z,\mu_p)$ & \\
$=\dim H^0(G_Z,{\mathbb Z}/p{\mathbb Z}) - \dim H^0(G_Z,\mu_p) + \sum_{v\in Z} \left(\dim L_v - \dim H^0(G_v,{\mathbb Z}/p{\mathbb Z}) \right).$
&
\end{tabular}
\end{eqnarray*}}
See $8.7.9$ of \cite{NSW} for details of this result.
\vskip1em
\noindent
{\it Second proof of Proposition~\ref{prop:dims}.}
Recall $X$ is tame and write $X:=X_{<\infty} \cup X_\infty$.
Set $Z:= Z_p \cup X_{<\infty} \cup Z_\infty$ where
$Z_p:= \{v : v|p\}$ and $Z_\infty$ is the set of all real Archimedean places of $K$ (so $X_\infty \subseteq Z_\infty$).
We assume for all $v\in X_{<\infty}$ that $N(v) \equiv 1\mod p$.
Recall that for a complex Archimedean place $v$ of $K$ we have $G_v=\{e\}$ so the Selmer and dual Selmer conditions are trivial. For a real Archimedean place $v$, $\dim H^1(G_v,{\mathbb Z}/2{\mathbb Z}) = \dim H^1(G_v,\mu_2)=1$ and the pairing between them is perfect - see Chapter I, Theorem $2.13$ of \cite{Milne}. It is easy to see in this case that the unramified cohomology groups are trivial.
In the table below we choose $\{M_v\}_{v\in Z}$ and $\{N_v\}_{v\in Z}$ so that
$H^1_{\mathcal M}(G_{Z}, { {\mathbb Z}/p{\mathbb Z}}) =H^1(G_\emptyset, { {\mathbb Z}/p{\mathbb Z}})$ and
$H^1_{\mathcal N}(G_{Z}, { {\mathbb Z}/p{\mathbb Z}}) =H^1(G_{X}, { {\mathbb Z}/p{\mathbb Z}})$. \textcolor{black}{The previous paragraph and Lemma~\ref{lemma:XXX}
justify
the stated dual Selmer conditions of the table.}
\begin{eqnarray*}
\begin{tabular}{l|cccc}
& $M_v$ & $M^\perp_v$ & $N_v$ & $N^\perp_v$ \\
\hline
$v\in Z_p$ & $H^1_{nr}(G_v,{\mathbb Z}/p{\mathbb Z})$ & $H^1_{f}(G_v,\mu_p)$ & $H^1_{nr}(G_v,{\mathbb Z}/p{\mathbb Z})$ & $H^1_{f}(G_v,\mu_p)$\\
$v\in X_\infty$ & $H^1_{nr}(G_v,{\mathbb Z}/2{\mathbb Z})=0$ & $H^1(G_v,\mu_2)$ & $H^1(G_v,{\mathbb Z}/2{\mathbb Z})$ & $0$ \\
$v\in Z_\infty \backslash X_\infty$ & $H^1_{nr}(G_v,{\mathbb Z}/2{\mathbb Z}) =0$ & $H^1(G_v,\mu_2)$ & $0$ & $H^1(G_v,\mu_2)$ \\
$v\in X_{<\infty}$ & $H^1_{nr}(G_v,{\mathbb Z}/p{\mathbb Z})$ & $ H^1_{nr}(G_v,\mu_p)$ & $H^1(G_v,{\mathbb Z}/p{\mathbb Z})$ & $0$\\
\end{tabular}
\end{eqnarray*}
Applying the Wiles-Greenberg formula for $\{M_v\}_{v\in Z}$ and $\{N_v\}_{v\in Z}$ and subtracting the first equation from the second:
\begin{eqnarray*}
\begin{tabular}{l}
$\dim H^1(G_X, { {\mathbb Z}/p{\mathbb Z}}) -\dim H^1(G_\emptyset, { {\mathbb Z}/p{\mathbb Z}})= $\\
$\dim H^1_{\mathcal N}(G_{Z}, { {\mathbb Z}/p{\mathbb Z}}) - \dim H^1_{\mathcal M}(G_{Z}, { {\mathbb Z}/p{\mathbb Z}}) =$\\
$\dim H^1_{{\mathcal N}^\perp}(G_{Z},\mu_p) - \dim H^1_{{\mathcal M}^\perp}(G_{Z},\mu_p)
+\sum_{v\in Z} (\dim N_v - \dim M_v)$
\end{tabular}
\end{eqnarray*}
For $v \in X_{<\infty}$ local class field theory implies $\dim H^1_{nr}(G_v, { {\mathbb Z}/p{\mathbb Z}}) =1$ and $\dim H^1(G_v, { {\mathbb Z}/p{\mathbb Z}}) =2$ so
$$\dim N_v - \dim M_v =\left\{
\begin{array}{ll}
0 & v\in Z_p \\
1 & v \in X_\infty, \mbox{ }p=2\\
0 & v\in Z_\infty \backslash X_\infty \\
1 & v\in X_{<\infty}
\end{array} \right. ,$$
and we have
\begin{equation}\label{eq:prop32}
\dim \left(\frac{H^1(G_X, { {\mathbb Z}/p{\mathbb Z}})}{H^1(G_\emptyset, { {\mathbb Z}/p{\mathbb Z}})} \right)= \dim H^1_{{\mathcal N}^\perp}(G_{Z},\mu_p) - \dim H^1_{{\mathcal M}^\perp}(G_{Z},\mu_p) +\#X .
\end{equation}
To prove Proposition~\ref{prop:dims} we need to show this last quantity is $\dim R_X=s$, the dimension of the space of dependence relations on the set $\{\sigma_v\}_{v\in X} \subset W =
Gal(K'(\sqrt[p]{V_\emptyset})/K')$.
An element $f \in H^1_{{\mathcal M}^\perp}(G_{Z},\mu_p)$ gives rise to the field diagram below
where $L_f/K'$ is a $ { {\mathbb Z}/p{\mathbb Z}}$ extension peu ramifi\'{e}e at $v\in Z_p$, with no condition on $v\in Z_\infty$ and unramified at $v\in X_{<\infty}$.
We show the composite of all such $L_f$ is $K'\left(\sqrt[p]{V_\emptyset}\right)$.
\vskip1em
\hspace{2.3in}
\begin{tikzcd}
& L_{f}:=K'\left(\sqrt[p]{\alpha_f}\right) \\
K':=K(\mu_p) \arrow[ur,dash]\\
K\arrow[u, dash]
\end{tikzcd}
\vskip1em\noindent
\textcolor{black}{Kummer Theory implies $\alpha_f \in K'/{K'}^{\times p}$, which decomposes into eigenspaces under the action of
$Gal(K'/K)$. If it is not in the trivial eigenspace,
then $Gal(L_f/K')$ is not acted on by $Gal(K'/K)$ via the cyclotomic character, a contradiction, so
we may assume (up to $p$th powers) $\alpha_f \in K$.}
Since $L_f/K'$ is unramified at $v\in X_{<\infty}$, we see that at all such $v$ that $\alpha_f = u \pi_v^{pr}$ where $u\in K_v$ is a unit. At $v\in Z_p$ being peu ramifi\'{e}e implies that locally at
$v\in X_p$ we have $\alpha_f = u \pi_v^{pr}$ where $u\in K_v$ is a unit. Together, these mean that the fractional ideal $(\alpha_f)$ of $K$ is a $p$th power, which implies that $\alpha_f \in V_\emptyset$. Conversely, if $\alpha \in V_\emptyset$, then, recalling that
$(\alpha)=J^p$ for some ideal of $K$, we have that
$K'\left(\sqrt[p]{\alpha}\right)/K'$ is a $ { {\mathbb Z}/p{\mathbb Z}}$-extension peu ramifi\'{e}e at $v\in Z_p$, with no condition at $v\in Z_\infty$.
Thus $\alpha $ gives rise to an element $f_\alpha \in H^1_{{\mathcal M}^\perp}(G_Z,\mu_p)$
so $L:=K'\left( \sqrt[p]{V_\emptyset}\right)$ is the composite of all $L_f$ for
$f \in H^1_{{\mathcal M}^\perp}(G_Z,\mu_p)$.
An element $f \in H^1_{{\mathcal N}^\perp}(G_{Z},\mu_p)$ gives rise to a $ { {\mathbb Z}/p{\mathbb Z}}$-extension of $K'$
peu ramifi\'{e}e at $v\in Z_p$ and split completely at $v\in X$. We denote the composite of all these fields by
$D \subset K'\left( \sqrt[p]{V_\emptyset}\right)$.
\vskip1em
\hspace{2.3in}
\begin{tikzcd}
& & L:=K'\left( \sqrt[p]{V_\emptyset}\right) \\
& D\arrow[ur,dash] \\
K':=K(\mu_p) \arrow[ur,dash]\\
K\arrow[u, dash]
\end{tikzcd}
\vskip1em
\noindent
Recall that $r$ is the dimension of the space $\langle \sigma_v \rangle_{v\in X} \subset Gal(L/K')$.
Clearly $D$ is the field fixed of $\langle \sigma_v\rangle_{v\in X}$ so
$\dim_{{\mathbb F}_p} Gal\left( K'\left( \sqrt[p]{V_\emptyset}\right)/ D \right) =r =\# I$ from the first section of this note.
Thus $$\dim H^1_{{\mathcal N}^\perp}(G_{Z},\mu_p)=\dim (V_\emptyset/K^{\times p}) -r $$
so
\begin{eqnarray*}
\dim H^1_{{\mathcal N}^\perp}(G_{Z},\mu_p) - \dim H^1_{{\mathcal M}^\perp}(G_{Z},\mu_p)
+\#X=\\
\left(\dim (V_\emptyset/K^{\times p}) -r \right) - \dim (V_\emptyset/K^{\times p}) +(r+s) =s = \dim R_X
\end{eqnarray*}
and we have shown the the left hand side of~\eqref{eq:prop32} is $\dim R_X$ proving Proposition~\ref{prop:dims}.
\hfill$\square$
\vskip1em
We have now proven Proposition~\ref{prop:dims} using the Wiles-Greenberg formula. The rest of the proof of Theorem~\ref{theorem:main} follows as in the previous section.
|
\section{Introduction}
\label{sec:introduction}
Reinforcement learning (RL) algorithms often require a large number of data samples to learn a control policy. As a result, training them directly on the real-world systems is expensive and potentially dangerous. This problem is typically overcome by training them on a simulator (online RL) or using a pre-collected offline dataset (offline RL). The offline dataset is usually collected either from a sophisticated simulator of the real-world system or from the historical measurements. The trained RL policy is then deployed assuming that the training environment, the simulator or the offline data, faithfully represents the model of the real-world system. This assumption is often incorrect due to multiple factors such as the approximation errors incurred while modeling, changes in the real-world parameters over time and possible adversarial disturbances in the real-world. For example, the standard simulator settings of the sensor noise, action delay, friction, and mass of a mobile robot can be different from that of the actual real-world robot, in addition to changes in the terrain, weather conditions, lighting, and obstacle densities of the testing environment.
Unfortunately, the current RL control policies can fail dramatically when faced with even mild changes in the training and testing environments \citep{sunderhauf2018limits, tobin2017domain, peng2018sim}.
The goal in robust RL is to learn a policy that is robust against the model parameter mismatches between the training and testing environments. The robust planning problem is formalized using the framework of Robust Markov Decision Process (RMDP) \citep{iyengar2005robust,nilim2005robust}. Unlike the standard MDP which considers a single model (transition probability function), the RMDP formulation considers a set of models which is called the \textit{uncertainty set}. The goal is to find an optimal robust policy that performs the best under the worst possible model in this uncertainty set.
The minimization over the uncertainty set makes the robust MDP and robust RL problems significantly more challenging than their non-robust counterparts.
In this work, we study the problem of developing a robust RL algorithm with provably optimal performance for an RMDP with arbitrarily large state spaces, using only offline data with function approximation. Before stating the contributions of our work, we provide a brief overview of the results in offline and robust RL that are directly related to ours. We leave a more thorough discussion on related works to Appendix~\ref{sec:app:related-works}.
\looseness=-1\textbf{Offline RL:} Offline RL considers the problem of learning the optimal policy only using a pre-collected (offline) dataset. Offline RL problem has been addressed extensively in the literature \citep{antos2008learning, bertsekas2011approximate, lange2012batch, chen2019information, xie2020q, levine2020offline, xie2021bellman}. Many recent works develop deep RL algorithms and heuristics for the offline RL problem, focusing on the algorithmic and empirical aspects \citep{fujimoto2019off, kumar2019stabilizing, kumar2020conservative, yu2020mopo, zhang2021towards}. A number of theoretical work focus on analyzing the variations of Fitted Q-Iteration (FQI) algorithm \citep{gordon1995stable, ernst2005tree}, by identifying the necessary and sufficient conditions for the learned policy to be approximately optimal and characterizing the performance in terms of sample complexity \citep{munos08a, farahmand2010error, lazaric2012finite, chen2019information, liu2020provably, xie2021bellman}. All these works assume that the offline data is generated according to a single model and the goal is to find the optimal policy for the MDP with the same model. In particular, none of these works consider the \textit{offline robust RL problem} where the offline data is generated according to a (training) model which can be different from the one in testing, and the goal is to learn a policy that is robust w.r.t.~an uncertainty set.
\textbf{Robust RL:} The RMDP framework was first introduced in~\citet{iyengar2005robust, nilim2005robust}. The RMDP problem has been analyzed extensively in the literature \citep{xu2010distributionally, wiesemann2013robust, yu2015distributionally, mannor2016robust, russel2019beyond} providing computationally efficient algorithms, but these works are limited to the {planning} problem. Robust RL algorithms with provable guarantees have also been proposed \citep{lim2013reinforcement,tamar2014scaling,roy2017reinforcement,panaganti2020robust, wang2021online}, but they are limited to tabular or linear function approximation settings and only provide asymptotic convergence guarantees. Robust RL problem has also been addressed using deep RL methods \citep{pinto2017robust, derman2018soft, derman2020bayesian, Mankowitz2020Robust, zhang2020robust}. However, these works do not provide any theoretical guarantees on the performance of the learned policies.
The works that are closest to ours are by \citet{zhou2021finite,yang2021towards,panaganti22a} that address the robust RL problem in a tabular setting under the generative model assumption. Due to the generative model assumption, the offline data has the same uniform number of samples corresponding to each and every state-action pair, and tabular setting allows the estimation of the uncertainty set followed by solving the planning problem. Our work is significantly different from these in the following way: $(i)$ we consider a robust RL problem with arbitrary large state space, instead of the small tabular setting, $(ii)$ we consider a true offline RL setting where the state-action pairs are sampled according to an arbitrary distribution, instead of using the generative model assumption, $(iii)$ we focus on a function approximation approach where the goal is to directly learn optimal robust value/policy using function approximation techniques, instead of solving the tabular planning problem with the estimated model. \textit{To the best of our knowledge, this is the first work that addresses the offline robust RL problem with arbitrary large state space using function approximation, with provable guarantees on the performance of the learned policy.}
\textbf{Offline Robust RL: Challenges and Our Contributions:} Offline robust RL is significantly more challenging than its non-robust counterpart mainly because of the following key difficulties. \\
$(i)$ Data generation: The optimal robust policy is computed by taking the infimum over all models in the uncertainty set $\mathcal{P}$. However, generating data according to all models in $\mathcal{P}$ is clearly infeasible. It may only be possible to get the data from a nominal (training) model $P^{o}$. \textit{How do we use the data from a nominal model to account for the behavior of all the models in the uncertainty set $\mathcal{P}$?} \\
$(ii)$ Optimization over the uncertainty set $\mathcal{P}$: The robust Bellman operator (defined in \eqref{eq:robust-bellman-primal}) involves a minimization over $\mathcal{P}$, which is a significant computational challenge. Moreover, the uncertainty set $\mathcal{P}$ itself is unknown in the RL setting. \textit{How do we solve the optimization over $\mathcal{P}$?} \\
$(iii)$ Function approximation: Approximation of the robust Bellman update requires a modified target function which also depends on the approximate solution of the optimization over the uncertainty set. \textit{How do we perform the offline RL update accounting for both approximations? }
\looseness=-1 As the \textit{key technical contributions} of this work, we first derive a dual reformulation of the robust Bellman operator which replaces the expectation w.r.t.~all models in the uncertainty set $\mathcal{P}$ with an expectation only w.r.t. the nominal (training) model $P^o$. This enables using the offline data generated by $P^o$ for learning, without relying on high variance importance sampling techniques to account for all models in $\mathcal{P}$. Following the same reformulation, we then show that the optimization problem over $\mathcal{P}$ can be further reformulated as functional optimization. We solve this functional optimization problem using empirical risk minimization and obtain performance guarantees using the Rademacher complexity based bounds. We then use the approximate solution obtained from the empirical risk minimization to generate modified target samples that are then used to approximate robust Bellman update through a generalized least squares approach with provably bounded errors. Performing these operations iteratively results in our proposed Robust Fitted Q-Iteration (RFQI) algorithm, for which we prove that its learned policy achieves non-asymptotic and approximately optimal performance guarantees.
\textbf{Notations:} For a set $\mathcal{X}$, we denote its cardinality as $|\mathcal{X}|$. The set of probability distribution over $\mathcal{X}$ is denoted as $\Delta(\mathcal{X})$, and its power set sigma algebra as $\Sigma(\mathcal{X})$. For any $x \in \mathbb{R}$, we denote $\max\{x, 0\}$ as $(x)_{+}$.
For any function $f:\cS\times\cA\rightarrow \mathbb{R}$, state-action distribution $\nu\in\Delta({\cS\times\cA})$, and real number $p\geq 1$, the $\nu$-weighted $p$-norm of $f$ is defined as $\|f\|_{p,\nu} = \mathbb{E}_{s,a\sim\nu}[|f(s,a)|^p]^{1/p}$.
\section{Preliminaries}
\label{sec:preliminaries}
A Markov Decision Process (MDP) is a tuple $(\cS, \mathcal{A}, r, P, \gamma, d_0)$, where $\cS$ is the state space, $\mathcal{A}$ is the action space, $r: \cS\times \mathcal{A}\rightarrow \mathbb{R}$ is the reward function, $\gamma \in (0, 1)$ is the discount factor, and $d_0\in \Delta({\cS})$ is the initial state distribution. The transition probability function $P_{s,a}(s')$ is the probability of transitioning to state $s'$ when action $a$ is taken at state $s$. In the literature, $P$ is also called the \textit{model} of the MDP. We consider a setting where $|\cS|$ and $|\mathcal{A}|$ are finite but can be arbitrarily large. We will also assume that $r(s,a) \in [0,1]$, for all $(s,a) \in \cS\times\cA$, without loss of generality. A policy $\pi : \cS \rightarrow \Delta({\mathcal{A}})$ is a conditional distribution over actions given a state. The value function $V_{\pi, P}$ and the state-action value function $Q_{\pi,P}$ of a policy $\pi$ for an MDP with model $P$ are defined as
\vspace{-0.1cm}
\begin{align*}
V_{\pi, P}(s) = \mathbb{E}_{\pi, P}[\sum^{\infty}_{t=0} \gamma^{t} r(s_{t}, a_{t}) ~|~ s_{0} = s], \quad Q_{\pi, P}(s, a) = \mathbb{E}_{\pi, P}[\sum^{\infty}_{t=0} \gamma^{t} r(s_{t}, a_{t}) ~|~ s_{0} = s, a_{0} = a],
\end{align*}
where the expectation is over the randomness induced by the policy $\pi$ and model $P$.
The optimal value function $V^{*}_{P}$ and the optimal policy $\pi^{*}_{P}$ of an MDP with the model $P$ are defined as $V^{*}_{P} = \max_{\pi} V_{\pi, P}$ and $\pi^{*}_{P} = \argmax_{\pi} V_{\pi, P}$.
The optimal state-action value function is given by $Q^{*}_{P} = \max_{\pi} Q_{\pi, P}$. The optimal policy can be obtained as $\pi^{*}_{P}(s) = \argmax_{a} Q_P^{*}(s, a)$. The discounted state-action occupancy of a policy $\pi$ for an MDP with model $P$, denoted as $d_{\pi, P} \in \Delta({\cS\times\cA})$, is defined as $d_{\pi,P}(s,a)=(1-\gamma) \mathbb{E}_{\pi,P}[\sum_{t=0}^\infty \gamma^{t} \mathbbm{1}(s_t=s,a_t=a)]$.
\textbf{Robust Markov Decision Process (RMDP):} Unlike the standard MDP which considers a single model (transition probability function), the RMDP formulation considers a set of models. We refer to this set as the \textit{uncertainty set} and denote it as $\mathcal{P}$. We consider $\mathcal{P}$ that satisfies the standard \textit{$(s,a)$-rectangularity condition} \citep{iyengar2005robust}. We note that a similar {uncertainty set} can be considered for the reward function at the expense of additional notations. However, since the analysis will be similar and the sample complexity guarantee will be identical up to a constant factor, without loss of generality, we assume that the reward function is known and deterministic.
We specify an RMDP as $M = (\cS, \mathcal{A}, r, \mathcal{P}, \gamma, d_0)$, where the uncertainty set $\mathcal{P}$ is typically defined as
\begin{align}
\label{eq:uncertainty-set}
\mathcal{P} = \otimes_{(s,a) \in \cS\times\cA }\, \mathcal{P}_{s,a}, \quad \text{where} \;\;
\mathcal{P}_{s,a} = \{ P_{s,a} \in \Delta(\cS)~:~ D(P_{s,a}, P^o_{s,a}) \leq \rho \},
\end{align}
$P^{o} = (P^o_{s,a}, (s, a) \in \cS\times\cA)$ is the \textit{nominal model}, $D(\cdot, \cdot)$ is a distance metric between two probability distributions, and $\rho>0$ is the radius of the uncertainty set that indicates the level of robustness. The nominal model $P^{o}$ can be thought as the model of the training environment. It is either the model of the simulator on which the (online) RL algorithm is trained, or in our setting, it is the model according to which the offline data is generated. The uncertainty set $\mathcal{P}$ \eqref{eq:uncertainty-set} is the set of all valid transition probability functions (valid testing models) in the neighborhood of the nominal model $P^{o}$, which by definition satisfies $(s,a)$-rectangularity condition \citep{iyengar2005robust}, where the neighborhood is defined using the distance metric $D(\cdot, \cdot)$ and radius $\rho$. In this work, we consider the \textit{Total Variation (TV) uncertainty set} defined using the TV distance, i.e., $D(P_{s,a}, P^{o}_{s,a}) = (1/2) \|P_{s,a} - P^{o}_{s,a} \|_{1}$.
The RMDP problem is to find the optimal robust policy which maximizes the value against the worst possible model in the uncertainty set $\mathcal P$. The \textit{robust value function} $V^{\pi}$ corresponding to a policy $\pi$ and the \textit{optimal robust value function} $V^{*}$ are defined as \citep{iyengar2005robust,nilim2005robust}
\begin{align}
\label{eq:robust-value-function}
V^{\pi} = \inf_{P \in \mathcal{P}} ~V_{\pi, P},\qquad V^{*} = \sup_{\pi} \inf_{P \in \mathcal{P}} ~V_{\pi, P} .
\end{align}
The \textit{optimal robust policy} $\pi^{*}$ is such that the robust value function corresponding to it matches the optimal robust value function, i.e., $V^{\pi^*} =V^{*} $. It is known that there exists a deterministic optimal policy \citep{iyengar2005robust} for the RMDP. The \textit{robust Bellman operator} is defined as \citep{iyengar2005robust}
\begin{align}
\label{eq:robust-bellman-primal}
(T Q)(s, a) = r(s, a) + \gamma \inf_{P_{s,a} \in \mathcal{P}_{s,a}} \mathbb{E}_{s' \sim P_{s,a}} [ \max_{b} Q(s', b)].
\end{align}
It is known that $T$ is a contraction mapping in the infinity norm and hence it has a unique fixed point $Q^{*}$ with $V^{*}(s) = \max_{a} Q^{*}(s,a)$ and $\pi^{*}(s) = \argmax_{a} Q^{*}(s,a)$ \citep{iyengar2005robust}. The \textit{Robust Q-Iteration (RQI)} can now be defined using the robust Bellman operator as $Q_{k+1} = T Q_{k}$. Since $T$ is a contraction, it follows that $Q_{k} \rightarrow Q^{*}$. So, RQI can be used to compute (solving the planning problem) $Q^{*}$ and $\pi^{*}$ in the tabular setting with a known $\mathcal{P}$. Due to the optimization over the uncertainty set $\mathcal{P}_{s,a}$ for each $(s,a)$ pair, solving the planning problem in RMDP using RQI is much more computationally intensive than solving it in MDP using Q-Iteration.
\textbf{Offline RL:} Offline RL considers the problem of learning the optimal policy of an MDP when the algorithm does not have direct access to the environment and cannot generate data samples in an online manner. For learning the optimal policy $\pi^{*}_{P}$ of an MDP with model $P$, the algorithm will only have access to an offline dataset $\mathcal{D}_{P}=\{(s_i,a_i,r_{i}, s'_{i})\}_{i=1}^N$, where $(s_{i}, a_{i}) \sim \mu$, $\mu \in \Delta (\cS\times\cA)$ is some distribution, and $s'_{i} \sim P_{s_{i},a_{i}}$. \textit{Fitted Q-Iteration (FQI)} is a popular offline RL approach which is amenable to theoretical analysis while achieving impressive empirical performance. In addition to the dataset $\mathcal{D}_{P}$, FQI uses a function class $\mathcal{F}=\{f:\cS\times\cA\to [0, 1/(1-\gamma)] \}$ to approximate $Q^{*}_{P}$. The typical FQI update is given by $ f_{k+1} = \argmin_{f \in \mathcal{F}} \sum^{N}_{i=1} (r(s_{i}, a_{i}) + \gamma \max_{b} f_{k}(s'_{i}, b) - f(s_{i}, a_{i}) )^{2}$, which aims to approximate the non-robust Bellman update using offline data with function approximation. Under suitable assumptions, it is possible to obtain provable performance guarantees for FQI \citep{szepesvari2005finite, chen2019information, liu2020provably}.
\section{Offline Robust Reinforcement Learning}
\label{sec:offline-robust-RL-basics}
The goal of an offline robust RL algorithm is to learn the optimal robust policy $\pi^{*}$ using a pre-collected offline dataset $\mathcal{D}$. The data is typically generated according to a nominal (training) model $P^{o}$, i.e., $\mathcal{D} =\{(s_i,a_i,r_{i}, s'_{i})\}_{i=1}^N$, where $(s_{i}, a_{i}) \sim \mu, \mu \in \Delta (\cS\times\cA)$ is some data generating distribution, and $s'_{i} \sim P^{o}_{s_{i},a_{i}}$. The uncertainty set $\mathcal{P}$ is defined around this nominal model $P^{o}$ as given in \eqref{eq:uncertainty-set} w.r.t. the total variation distance metric. We emphasize that the learning algorithm does not know the nominal model $P^{o}$ as it has only access to $\mathcal{D}$, and hence it also does not know $\mathcal{P}$. Moreover, the learning algorithm does not have data generated according to any other models in $\mathcal{P}$ and has to rely only on $\mathcal{D}$ to account for the behavior w.r.t. all models in $\mathcal{P}$.
Learning policies for RL problems with large state-action spaces is computationally intractable. RL algorithms typically overcome this issue by using function approximation. In this paper, we consider two function classes $\mathcal{F}=\{f:\cS\times\cA\to [0, 1/(1-\gamma)] \}$ and $\mathcal{G}=\{g :\cS\times\cA\to [0, 2/(\rho(1-\gamma))]\}$. We use $\mathcal{F}$ to approximate $Q^{*}$ and $\mathcal{G}$ to approximate the dual variable functions which we will introduce in the next section. For simplicity, we will first assume that these function classes are finite but exponentially large, and we will use the standard log-cardinality to characterize the sample complexity results, as given in Theorem \ref{thm:tv-guarantee}. We note that, at the cost of additional notations and analysis, infinite function classes can also be considered where the log-cardinalities are replaced by the appropriate notions of covering number.
Similar to the non-robust offline RL, we make the following standard assumptions about the data generating distribution $\mu$ and the representation power of $\mathcal{F}$.
\begin{assumption}[Concentratability]
\label{assum-concentra-condition}
There exists a finite constant $C>0$ such that for any $\nu\in \{ d_{\pi,P^o} ~|\text{ any policy}~ \pi\} \subseteq \Delta(\cS\times\cA)$, we have $\norm{\nu/\mu}_{\infty} \leq \sqrt{C}$.
\end{assumption}
Assumption \ref{assum-concentra-condition} states that the ratio of the distribution $\nu$ and the data generating distribution $\mu$, $\nu(s,a)/\mu(s,a)$, is uniformly bounded. This assumption is widely used in the offline RL literature \citep{munos2003error,agarwal2019reinforcement,chen2019information,wang2021what,xie2021bellman} in many different forms. We borrow this assumption from \citet{chen2019information}, where they used it for non-robust offline RL. In particular, we note that the distribution $\nu$ is in the collection of discounted state-action occupancies on model $P^o$ alone for the robust RL.
\begin{assumption}[Approximate completeness]
\label{assum-bellman-completion}
Let $\mu\in\Delta(\cS\times\cA)$ be the data distribution. Then, $\sup_{f\in\mathcal{F}} \inf_{f'\in\mathcal{F}} \| f' - T f \|_{2,\mu}^2 \leq \epsilon_{\mathrm{c}}.$
\end{assumption}
Assumption \ref{assum-bellman-completion} states that the function class $\mathcal{F}$ is approximately closed under the robust Bellman operator $T$. This assumption has also been widely used in the offline RL literature~\citep{agarwal2019reinforcement,chen2019information,wang2021what,xie2021bellman}.
One of the most important properties that the function class $\mathcal{F}$ should have is that there must exist a function $f' \in \mathcal{F}$ which well-approximates $Q^{*}$. This assumption is typically called \textit{approximate realizability} in the offline RL literature. This is typically formalized by assuming $\inf_{f \in \mathcal{F}} \norm{f - Tf}^{2}_{2, \mu} \leq \epsilon_{\mathrm{r}}$ \citep{chen2019information}. It is known that the approximate completeness assumption and the concentratability assumption imply the realizability assumption \citep{chen2019information,xie2021bellman}.
\section{Robust Fitted Q-Iteration: Algorithm and Main Results}
\label{sec:algorithm-results}
In this section, we give a step-by-step approach to overcome the challenges of the offline robust RL outlined in Section \ref{sec:introduction}. We then combine these intermediate steps to obtain our proposed RFQI algorithm. We then present our main result about the performance guarantee of the RFQI algorithm, followed by a brief description about the proof approach.
\subsection{Dual Reformulation of Robust Bellman Operator}
\label{sec:rob-Bellman-tv}
One key challenge in directly using the standard definition of the optimal robust value function given in \eqref{eq:robust-value-function} or of the robust Bellman operator given in \eqref{eq:robust-bellman-primal} for developing and analyzing robust RL algorithms is that both involve computing an expectation w.r.t. each model $P \in \mathcal{P}$. Given that the data is generated only according to the nominal model $P^{o}$, estimating these expectation values is really challenging. We show that we can overcome this difficulty through the dual reformulation of the robust Bellman operator, as given below.
\begin{proposition}
\label{prop:robust-bellman-dual}
Let $M$ be an RMDP with the uncertainty set $\mathcal{P}$ specified by~\eqref{eq:uncertainty-set} using the total variation distance $D(P_{s,a}, P^{o}_{s,a}) = (1/2) \|P_{s,a} - P^{o}_{s,a} \|_{1}$. Then, for any $Q: \cS\times\cA\to [0, 1/(1-\gamma)]$, the robust Bellman operator $T$ given in \eqref{eq:robust-bellman-primal} can be equivalently written as
\begin{align}
\label{eq:robust-bellman-dual}
(T Q)(s, a) = r(s, a) - \gamma ~ \inf_{\eta\in[0, \frac{2}{\rho(1-\gamma)}]} (\mathbb{E}_{s' \sim P^{o}_{s,a}}[(\eta -V(s'))_+] - \eta + \rho (\eta - \inf_{s''} V(s''))_+ ),
\end{align}
where $V(s)=\max_{a\in\mathcal{A}} Q(s,a)$. Moreover, the inner optimization problem in \eqref{eq:robust-bellman-dual} is convex in $\eta$.
\end{proposition}
This result mainly relies on \citet[Section 3.2]{shapiro2017distributionally} and~\citet[Proposition 1]{duchi2018learning}.
Note that in \eqref{eq:robust-bellman-dual}, the expectation is now only w.r.t.~the nominal model $P^{o}$, which opens up the possibility of using empirical estimates obtained from the data generated according to $P^{o}$. This avoids the need to use importance sampling based techniques to account for all models in $\mathcal{P}$, which often have high variance, and thus, are not desirable.
While \eqref{eq:robust-bellman-dual} provides a form that is amenable to estimation using offline data, it involves finding $ \inf_{s''} V(s'')$. Though this computation is straightforward in a tabular setting, it is infeasible in a function approximation setting. In order to overcome this issue, we make the following assumption.
\begin{assumption}[Fail-state]
\label{assum-fail-state}
The RMDP $M$ has a `fail-state' $s_{f}$, such that $r(s_{f}, a) = 0$ and $P_{s_{f}, a}(s_{f}) = 1,\;\forall a \in \mathcal{A},\;\forall P \in \mathcal{P}$.
\end{assumption}
We note that this is not a very restrictive assumption because such a `fail-state' is quite natural in most simulated or real-world systems. For example, a state where a robot collapses and is not able to get up, either in a simulation environment like MuJoCo or in real-world setting, is such a fail state.
Assumption~\ref{assum-fail-state} immediately implies that $V_{\pi, P}(s_{f}) = 0,\;\forall P \in \mathcal{P}$, and hence $V^{*}(s_{f}) = 0$ and $Q^{*}(s_{f}, a) = 0,\;\forall a \in \mathcal{A}$. It is also straightforward to see that $Q_{k+1}(s_{f}, a) = 0,\;\forall a \in \mathcal{A}$, where $Q_{k}$'s are the RQI iterates given by the robust Bellman update $Q_{k+1} = T Q_{k}$ with the initialization $Q_{0}=0$. By the contraction property of $T$, we have $Q_{k} \rightarrow Q^{*}$. So, under Assumption \ref{assum-fail-state}, without loss of generality, we can always keep $Q_{k}(s_{f}, a) = 0,\;\forall a \in \mathcal{A}$ and for all $k$ in RQI (and later in RFQI). So, in the light of the above description, for the rest of the paper we will use the robust Bellman operator $T$ by setting $ \inf_{s''} V(s'') = 0$. In particular, for any function $f : \cS\times\cA \rightarrow [0,1/(1-\gamma)]$ with $f(s_{f}, a) = 0$, the robust Bellman operator $T$ is now given by
\begin{align}
\label{eq:robust-bellman-dual-2}
(T f) (s, a) = r(s, a) - \gamma \inf_{\eta\in[0, \frac{2}{(\rho(1-\gamma))}]} ~ (\mathbb{E}_{s' \sim P^{o}_{s,a}}[(\eta -\max_{a'}f(s',a'))_+] - (1-\rho)\eta ).
\end{align}
\subsection{Approximately Solving the Dual Optimization using Empirical Risk Minimization}
Another key challenge in directly using the standard definition of the optimal robust value function given in \eqref{eq:robust-value-function} or of the robust Bellman operator given in \eqref{eq:robust-bellman-primal} for developing and analyzing robust RL algorithms is that both involve an optimization over $\mathcal{P}$. The dual reformulation given in \eqref{eq:robust-bellman-dual-2} partially overcomes this challenge also, as the optimization over $\mathcal{P}$ is now replaced by a {convex} optimization over a scalar $\eta\in[0, 2/(\rho (1-\gamma))]$. However, this still requires solving an optimization for each $(s, a)\in\cS\times\cA$, which is clearly infeasible even for moderately sized state-action spaces, not to mention the function approximation setting. Our key idea to overcome this difficulty is to reformulate this as a functional optimization problem instead of solving it as multiple scalar optimization problems. This functional optimization method will make it amenable to approximately solving the dual problem using an empirical risk minimization approach with offline data.
Consider the probability (measure) space $(\cS\times\cA,\Sigma(\cS\times\cA),\mu)$ and let $L^{1}(\cS\times\cA,\Sigma(\cS\times\cA),\mu)$ be the set of all absolutely integrable functions defined on this space.\footnote{In the following, we will simply denote $L^{1}(\cS\times\cA,\Sigma(\cS\times\cA),\mu)$ as $L^{1}$ for conciseness. } In other words, $L^{1}$ is the set of all functions $g : \cS\times\cA \rightarrow \mathcal{C} \subset \mathbb{R}$, such that $\norm{g}_{1,\mu}$ is finite. We set $\mathcal{C} = [0, 2/\rho (1 - \gamma)]$, anticipating the solution of the dual optimization problem~\eqref{eq:robust-bellman-dual-2}. We also note $\mu$ is the data generating distribution which is a $\sigma$-finite measure.
For any given function $f : \cS\times\cA \rightarrow [0,1/(1-\gamma)]$, we define the loss function $L_{\mathrm{dual}}(\cdot; f)$ as
\begin{align}
\label{L-f-eta-loss}
L_{\mathrm{dual}}(g; f) = \mathbb{E}_{s,a\sim\mu} [ \mathbb{E}_{s'\sim P^o_{s,a}} [(g(s,a) - \max_{a'} f(s',a'))_+] - (1-\rho) g(s,a)], \quad \forall g\in L^1.
\end{align}
In the following lemma, we show that the scalar optimization over $\eta$ for each $(s,a)$ pair in \eqref{eq:robust-bellman-dual-2} can be replaced by a single functional optimization w.r.t.~the loss function $L_{\mathrm{dual}}$.
\begin{lemma}
\label{lem:interchanging-lemma-1}
Let $L_{\mathrm{dual}}$ be the loss function defined in \eqref{L-f-eta-loss}. Then, for any function $f : \cS\times\cA \rightarrow [0,1/(1-\gamma)]$, we have
\begin{align}
\label{eq:interchanging-lemma-1}
\inf_{g \in L^{1}} L_{\mathrm{dual}}(g; f) = \mathbb{E}_{s,a\sim\mu} \Big[ \inf_{\eta\in[0, \frac{2}{(\rho(1-\gamma))}]} \Big(\mathbb{E}_{s'\sim P^o_{s,a}} \big[\big(\eta - \max_{a'} f(s',a')\big)_+\big] - (1-\rho)\eta\Big)\Big].
\end{align}
\end{lemma}
Note that the RHS of \eqref{eq:interchanging-lemma-1} has minimization over $\eta$ for each $(s, a)$ pair and minimization is inside the expectation $\mathbb{E}_{s,a\sim\mu}[\cdot]$. However, the LHS of \eqref{eq:interchanging-lemma-1} has a single functional minimization over $g \in L^{1}$ and this minimization is outside the expectation. For interchanging the expectation and minimization, and for moving from point-wise optimization to functional optimization, we use the result from \citet[Theorem 14.60]{rockafellar2009variational}, along with the fact that $L^{1}$ is a decomposable space. We also note that this result has been used in many recent works on distributionally robust optimization \citep{shapiro2017distributionally,duchi2018learning} (see Appendix \ref{appendix:useful-tech-results} for more details).
We can now define the empirical loss function $\widehat{L}_{\mathrm{dual}}$ corresponding to the true loss $L_{\mathrm{dual}}$ as
\begin{align} \label{eq:L-dual-empirical-loss}
\widehat{L}_{\mathrm{dual}}(g;f) = \frac{1}{N} \sum^{N}_{i=1} (g(s_{i},a_{i}) - \max_{a'} f(s'_{i},a'))_+ - (1-\rho) g(s_{i},a_{i}).
\end{align}
Now, for any given $f$, we can find an approximately optimal dual function through the \textit{empirical risk minimization} approach as $\inf_{g \in L^{1}} \widehat{L}_{\mathrm{dual}}(g;f)$.
As we mentioned in Section \ref{sec:offline-robust-RL-basics}, our offline robust RL algorithm is given an input function class $\mathcal{G}=\{g :\cS\times\cA\to [0, 2/(\rho(1-\gamma))]\}$ to approximate the dual variable functions. So, in the empirical risk minimization, instead of taking the infimum over all the functions in $L^{1}$, we can only take the infimum over all the functions in $\mathcal{G}$. For this to be meaningful, $\mathcal{G}$ should have sufficient representation power. In particular, the result in Lemma \ref{lem:interchanging-lemma-1} should hold approximately even if we replace the infimum over $L^{1}$ with infimum over $\mathcal{G}$.
One can see that this is similar to the realizability requirement for the function class $\mathcal{F}$ as described in Section \ref{sec:offline-robust-RL-basics}. We formalize the representation power of $\mathcal{G}$ in the following assumption.
\begin{assumption}[Approximate dual realizability] \label{assum-dual-realizability}
For all $f \in \mathcal{F}$, there exists a uniform constant $\epsilon_{\textrm{dual}}$ such that $ \inf_{g \in \mathcal{G}} L_{\mathrm{dual}}(g; f) - \inf_{g \in L^{1}} L_{\mathrm{dual}}(g; f) \leq \epsilon_{\textrm{dual}}$
\end{assumption}
Using the above assumption, for any given $f\in\mathcal F$, we can find an approximately optimal dual function $\widehat{g}_{f} \in \mathcal{G}$ through the \textit{empirical risk minimization} approach as $\widehat{g}_{f} = \argmin_{g \in \mathcal{G}} \widehat{L}_{\mathrm{dual}}(g;f)$.
In order to characterize the performance of this approach, consider the operator $T_g$ for any $g\in\mathcal{G}$ as
\begin{align}
\label{eq:Tg}
(T_g f) (s, a) = r(s, a) - \gamma (\mathbb{E}_{s' \sim P^{o}_{s,a}}[(g(s,a) -\max_{a'}f(s',a'))_+] - (1-\rho)g(s,a) ),
\end{align} for all $f\in\mathcal{F}$ and $(s,a)\in\cS\times\cA$. We will show in Lemma \ref{lem:erm-high-prob-bound} in Appendix \ref{appendix:thm:tv-guarantee} that the error $\sup_{f\in\mathcal{F}} \|T f - T_{\widehat{g}_f} f\|_{1,\mu}$ is $\mathcal{O}(\log(|\mathcal{F}|/\delta)/\sqrt{N})$ with probability at least $1-\delta$.
\subsection{Robust Fitted Q-iteration}
\label{sec:rofqi-algorithm}
The intuitive idea behind our robust fitted Q-iteration (RFQI) algorithm is to approximate the exact RQI update step $Q_{k+1} = T Q_{k}$ with function approximation using offline data. The exact RQI step requires updating each $(s,a)$-pair separately, which is not scalable to large state-action spaces. So, this is replaced by the function approximation as $Q_{k+1} = \argmin_{f \in \mathcal{F}} \norm{T Q_{k} - f}^{2}_{2,\mu}$. It is still infeasible to perform this update as it requires to exactly compute the expectation (w.r.t. $P^{o}$ and $\mu$) and to solve the dual problem accurately. We overcome these issues by replacing both these exact computations with empirical estimates using the offline data. We note that this intuitive idea is similar to that of the FQI algorithm in the non-robust case. However, RFQI has unique challenges due to the nature of the robust Bellman operator $T$ and the presence of the dual optimization problem within $T$.
Given a dataset $\mathcal{D}$, we also follow the standard non-robust offline RL choice of least-squares residual minimization \citep{chen2019information,xie2021bellman,wang2021what}. Define the empirical loss of $f$ given $f'$ (which represents the $Q$-function from the last iteration) and dual variable function $g$ as
\begin{align}
\label{eq:loss-rfqi-1}
\widehat{L}_{\mathrm{RFQI}}(f;f', g) = \frac{1}{N} \sum^{N}_{i=1} \bigg( \begin{array}{ll} r(s_i, a_i) + \gamma \big(-(g(s_{i},a_{i}) -\max_{a'}f'(s'_{i},a'))_+ \\ \hspace{2cm}+ \;(1-\rho) g(s_{i},a_{i})\big) - f(s_{i}, a_{i}) \end{array} \bigg)^{2}.
\end{align}
The correct dual variable function to be used in \eqref{eq:loss-rfqi-1} is the optimal dual variable $g^{*}_{f'} = \argmin_{g \in \mathcal{G}} L_{\mathrm{dual}}(g; f')$ corresponding to the last iterate $f'$, which we will approximate it by $\widehat{g}_{f'} = \argmin_{g \in \mathcal{G}} \widehat{L}_{\mathrm{dual}}(g; f')$. The RFQI update is then obtained as $\argmin_{f \in \mathcal{F}} \widehat{L}_{\mathrm{RFQI}}(f;f', \widehat{g}_{f'})$.
Summarizing the individual steps described above, we formally give our RFQI algorithm below.
\begin{algorithm}[H]
\caption{Robust Fitted Q-Iteration (RFQI) Algorithm}
\label{alg:RFQI-Algorithm}
\begin{algorithmic}[1]
\STATE \textbf{Input:} Offline dataset $\mathcal{D} =(s_i,a_i,r_{i}, s'_{i})_{i=1}^N$, function classes $\mathcal{F}$ and $\mathcal{G}$.
\STATE \textbf{Initialize:} $Q_{0}\equiv 0 \in \mathcal{F}$.
\FOR {$k=0,\cdots,K-1$ }
\STATE \textbf{Dual variable function optimization:} Compute the dual variable function corresponding to $Q_{k}$ through empirical risk minimization as $g_{k} = \widehat{g}_{Q_{k}} = \argmin_{g \in \mathcal{G}} \widehat{L}_{\mathrm{dual}}(g; Q_{k})\;\;$ (see \eqref{eq:L-dual-empirical-loss}).
\STATE \textbf{Robust Q-update:} Compute the next iterate $Q_{k+1}$ through least-squares regression as
$\quad Q_{k+1}=\argmin_{Q\in\mathcal{F}}\widehat{L}_{\mathrm{RFQI}}(Q;Q_{k}, g_{k})\;\;$ (see \eqref{eq:loss-rfqi-1}).
\ENDFOR
\STATE \textbf{Output:} $\pi_{K} = \argmax_a Q_{K}(s,a)$
\end{algorithmic}
\end{algorithm}
Now we state our main theoretical result on the performance of the RFQI algorithm.
\begin{theorem}
\label{thm:tv-guarantee}
Let Assumptions \ref{assum-concentra-condition}-\ref{assum-dual-realizability} hold. Let $\pi_{K}$ be the output of the RFQI algorithm after $K$ iterations. Denote $J^{\pi}=\mathbb{E}_{s\sim d_0}[V^{\pi}(s)]$ where $d_0$ is initial state distribution. Then, for any $\delta\in(0,1)$, with probability at least $1 - 2 \delta$, we have
\begin{align*}
J^{\pi^*} - J^{\pi_{K}} \leq \frac{\gamma^K}{(1-\gamma)^2} &+ \frac{\sqrt{C}(\sqrt{6\epsilon_{\textrm{c}}} + \gamma \epsilon_{\textrm{dual}}) }{(1-\gamma)^2} + \frac{16}{\rho(1-\gamma)^3} \sqrt{\frac{18 C \log(2|\mathcal{F}||\mathcal{G}|/\delta)}{N}}. \end{align*}
\end{theorem}
\begin{remark}
Theorem \ref{thm:tv-guarantee} states that the RFQI algorithm can achieve approximate optimality. To see this, note that with $K\geq\mathcal{O}(\frac{1}{ \log(1/\gamma)} \log(\frac{1}{\epsilon (1-\gamma)}))$, and neglecting the second term corresponding to (inevitable) approximation errors $\epsilon_{\textrm{c}}$ and $\epsilon_{\textrm{dual}}$, we get $J^{\pi^*} - J^{\pi_{K}} \leq {\epsilon}/{(1-\gamma)}$ with probability greater than $1-2\delta$ for any $\epsilon, \delta \in (0,1)$, as long as the number of samples $N\geq\mathcal{O}(\frac{1}{(\rho\epsilon)^{2} (1-\gamma)^{4}} \log \frac{|\mathcal{F}| |\mathcal{G}|}{\delta })$. So, the above theorem can also be interpreted as a \textbf{sample complexity} result.
\end{remark}
\begin{remark}
The known sample complexity of robust-RL in the tabular setting is $\widetilde{\mathcal{O}}(\frac{|\cS|^2 |\mathcal{A}|}{(\rho\epsilon)^{2} (1-\gamma)^{4}})$ \citep{yang2021towards,panaganti22a}. Considering $\widetilde{\mathcal{O}}(\log(|\mathcal{F}| | \mathcal{G}|))$ to be $\widetilde{\mathcal{O}}(|\cS| |\mathcal{A}|)$, we can recover the same bound as in the tabular setting (we save $|\cS|$ due to the use of Bernstein inequality).
\end{remark}
\begin{remark}
Under similar Bellman completeness and concentratability assumptions, RFQI sample complexity is comparable to that of a non-robust offline RL algorithm, i.e.,~$\mathcal{O}(\frac{1}{\epsilon^{2} (1-\gamma)^{4}} \log \frac{|\mathcal{F}|}{\delta })$~\citep{chen2019information}. As a consequence of robustness, we have $\rho^{-2}$ and $\log(|\mathcal{G}|)$ factors in our bound.
\end{remark}
\subsection{Proof Sketch} \label{sec:proof-sketch}
Here we briefly explain the key ideas used in the analysis of RFQI for obtaining the optimality gap bound in Theorem \ref{thm:tv-guarantee}. The complete proof is provided in Appendix \ref{appendix:thm:tv-guarantee}.
\textit{Step 1:} To bound $J^{\pi^*} - J^{\pi_K}$, we connect it to the error $\|Q^{\pi^*} - Q_{K}\|_{1,\nu }$ for any state-action distribution $\nu$. While the similar step follows almost immediately using the well-known performance lemma in the analysis of non-robust FQI, such a result is not known in the robust RL setting. So, we derive the basic inequalities to get a recursive form and to obtain the bound $J^{\pi^*} - J^{\pi_K} \leq 2\|Q^{\pi^*} - Q_{K}\|_{1,\nu }/(1-\gamma)$ (see~\eqref{eq:thm-bound-part-1} and the steps before in Appendix~\ref{appendix:thm:tv-guarantee}). \\
\textit{Step 2:} To bound $\|Q^{\pi^*} - Q_{K}\|_{1,\nu }$ for any state-action distribution $\nu$ such that $\norm{\nu/\mu}_{\infty}\leq\sqrt{C}$, we decompose it to get a recursion, with approximation terms based on the least-squares regression and empirical risk minimization. Recall that $\widehat{g}_f$ is the dual variable function from the algorithm for state-action value function $f\in\mathcal{F}$. Denote $\widehat{f}_{g}$ as the least squares solution from the algorithm for the state-action value function $f\in\mathcal{F}$ and dual variable function $g\in\mathcal{G}$, i.e., $\widehat{f}_{g}=\argmin_{Q\in\mathcal{F}}\widehat{L}_{\mathrm{RFQI}}(Q;f, g)$. By recursive use of the obtained inequality \eqref{eq:pf-sketech-stp4-1} (see Appendix \ref{appendix:thm:tv-guarantee}) and using uniform bound, we get
\begin{align*}
\|Q^{\pi^*} - Q_{K}\|_{1,\nu } \leq \frac{\gamma^K }{1-\gamma} + \frac{\sqrt{C} }{1-\gamma} \sup_{f\in\mathcal{F}} \|T f - T_{\widehat{g}_f} f\|_{1,\mu}
+ \frac{\sqrt{C} }{1-\gamma} \sup_{f\in\mathcal{F}} \sup_{g\in\mathcal{G}} \| T_{g} f - \widehat{f}_{g}\|_{2,\mu}.
\end{align*}
\textit{Step 3:} We recognize that $\sup_{f\in\mathcal{F}} \|T f - T_{\widehat{g}_f} f\|_{1,\mu}$ is an empirical risk minimization error term. Using Rademacher
complexity based bounds, we show in Lemma~\ref{lem:erm-high-prob-bound} that this error is $\mathcal{O}(\log(|\mathcal{F}|/\delta)/\sqrt{N})$ with high probability. \\
\textit{Step 4:} Similarly, we also recognize that $ \sup_{f\in\mathcal{F}} \sup_{g\in\mathcal{G}} \| T_{g} f - \widehat{f}_{g}\|_{2,\mu}$ is a least-squares regression error term. We also show that this error is $\mathcal{O}(\log(|\mathcal{F}||\mathcal{G}|/\delta)/\sqrt{N})$ with high probability. We adapt the generalized least squares regression result to accommodate the modified target functions resulting from the robust Bellman operator to obtain this bound (see Lemma~\ref{lem:least-squares-generalization-bound}). \\
The proof is complete after combining steps 1-4 above.
\section{Experiments}
\label{sec:main-experiments}
\begin{figure*}[h]
\centering
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/CartPolePerturbed-v0_force_mag.pdf}
\captionof{figure}{CartPole}
\label{fig:cart_force-in}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/CartPolePerturbed-v0_action.pdf}
\captionof{figure}{CartPole}
\label{fig:cart_action-in}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/Hopper_leg_joint.pdf}
\captionof{figure}{Hopper}
\label{fig:hopper_leg}
\end{minipage}
\end{figure*}
\looseness=-1 Here, we demonstrate the robust performance of our RFQI algorithm by evaluating it on \textit{Cartpole} and \textit{Hopper} environments in OpenAI Gym \citep{brockman2016openai}. In all the figures shown, the quantity in the vertical axis is averaged over $20$ different seeded runs depicted by the thick line and the band around it is the $\pm 0.5$ standard deviation. \textit{A more detailed description of the experiments, and results on additional experiments, are deferred to Appendix \ref{appendix:simulations}.} We provide our code in \textbf{github webpage} \url{https://github.com/zaiyan-x/RFQI} containing instructions to reproduce all results in this paper.
\looseness=-1 For the \textit{Cartpole}, we compare RFQI algorithm against the non-robust RL algorithms FQI and DQN, and the soft-robust RL algorithm proposed in \cite{derman2018soft}. We test the robustness of the algorithms by changing the parameter \textit{force\_mag} (to model external force disturbance), and also by introducing action perturbations (to model actuator noise). Fig. \ref{fig:cart_force-in} and Fig. \ref{fig:cart_action-in} shows superior robust performance of RFQI compared to the non-robust FQI and DQN. The RFQI performance is similar to that of soft-robust DQN. We note that soft-robust RL algorithm (here soft-robust DQN) is an online deep RL algorithm (and not an offline RL algorithm) and has no provable performance guarantee. Moreover, soft-robust RL algorithm requires generating online data according a number of models in the uncertainty set, whereas RFQI only requires offline data according to a single nominal training model.
For the \textit{Hopper}, we compare RFQI algorithm against the non-robust RL algorithms FQI and TD3 \citep{fujimoto2018addressing}, and the soft-robust RL (here soft-robust DDPG) algorithm proposed in \cite{derman2018soft}. We test the robustness of the algorithms by changing the parameter \textit{leg\_joint\_stiffness}. Fig. \ref{fig:hopper_leg} shows the superior performance of our RFQI algorithm against the non-robust algorithms and soft-robust DDPG algorithm. The average episodic reward of RFQI remains almost the same initially, and later decays much less and gracefully when compared to the non-robust FQI and TD3.
\section{Conclusion}
In this work, we presented a novel robust RL algorithm called Robust Fitted Q-Iteration algorithm with provably optimal performance for an RMDP with arbitrarily large state space, using only offline data with function approximation. We also demonstrated the superior performance of the proposed algorithm on standard benchmark problems.
\looseness=-1 One limitation of our present work is that, we considered only the uncertainty set defined with respect to the total variation distance. In future work, we will consider uncertainty sets defined with respect to other $f$-divergences such as KL-divergence and Chi-square divergence. Finding a lower bound for the sample complexity and relaxing the assumptions used are also important and challenging problems.
\section{Acknowledgements}
This work was supported in part by the National Science Foundation (NSF) grants NSF-CAREER- EPCN-2045783 and NSF ECCS 2038963. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsoring agencies.
\section*{Checklist}
\begin{enumerate}
\item For all authors...
\begin{enumerate}
\item Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
\answerYes{See \textbf{contributions} in the Introduction.}
\item Did you describe the limitations of your work?
\answerYes{The discussions on the assumptions describes the limitations.}
\item Did you discuss any potential negative societal impacts of your work?
\answerNA{}
\item Have you read the ethics review guidelines and ensured that your paper conforms to them?
\answerYes{}
\end{enumerate}
\item If you are including theoretical results...
\begin{enumerate}
\item Did you state the full set of assumptions of all theoretical results?
\answerYes{See Sections \ref{sec:offline-robust-RL-basics}-\ref{sec:rofqi-algorithm} }
\item Did you include complete proofs of all theoretical results?
\answerYes{We provide proof sketch \ref{sec:proof-sketch} in main paper and the complete proof in Appendix with self-contained material.}
\end{enumerate}
\item If you ran experiments...
\begin{enumerate}
\item Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)?
\answerYes{}
\item Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)?
\answerYes{Described in the Appendix.}
\item Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)?
\answerYes{Described in the main paper and the Appendix.}
\item Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)?
\answerYes{Mentioned in the Appendix.}
\end{enumerate}
\item If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
\begin{enumerate}
\item If your work uses existing assets, did you cite the creators?
\answerYes{}
\item Did you mention the license of the assets?
\answerYes{}
\item Did you include any new assets either in the supplemental material or as a URL?
\answerNA{}
\item Did you discuss whether and how consent was obtained from people whose data you're using/curating?
\answerNA{}
\item Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content?
\answerNA{}
\end{enumerate}
\item If you used crowdsourcing or conducted research with human subjects...
\begin{enumerate}
\item Did you include the full text of instructions given to participants and screenshots, if applicable?
\answerNA{}
\item Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable?
\answerNA{}
\item Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation?
\answerNA{}
\end{enumerate}
\end{enumerate}
\section*{Appendix}
\section{Useful Technical Results}
\label{appendix:useful-tech-results}
In this section, we state some existing results from concentration inequalities, generalization bounds, and optimization theory that we will use later in our analysis. We first state the Berstein's inequality that utilizes second-moment to get a tighter concentration inequality.
\begin{lemma}[Bernstein's inequality \hspace{-0.1cm} \text{\citep[Theorem 2.8.4]{vershynin2018high}}] \label{lem:bernstein-ineq}
Let $X_1,\cdots, X_T$ be independent random variables. Assume that $|X_t - \mathbb{E}[X_t]| \leq M$, for all $t$. Then, for any $\epsilon > 0$, we have
\[
\mathbb{P} \left( \Big|\frac{1}{T} \sum_{t=1}^T (X_t - \mathbb{E}[X_t])\Big| \geq \epsilon \right) \leq 2\exp\left( -\frac{T^2\epsilon^2}{2\sigma^2 + \frac{2MT\epsilon}{3}} \right),
\]
where $\sigma^2 = \sum_{t=1}^T \mathbb{E}[X_t^2]$. Furthermore, if $X_1,\cdots, X_T$ are independent and identically distributed random variables, then for any $\delta\in(0,1)$, we have
\[
\Big|\mathbb{E}[X_1] - \frac{1}{T} \sum_{t=1}^T X_t \Big| \leq \sqrt{\frac{2 \mathbb{E}[X_1^2] \log(2/\delta)}{T}} +\frac{M\log(2/\delta)}{3T},
\]
with probability at least $1-\delta$.
\end{lemma}
We now state a result
for the generalization bounds on empirical risk minimization (ERM) problems. This result is adapted from \citet[Theorem 26.5, Lemma 26.8, Lemma 26.9]{shalev2014understanding}.
\begin{lemma}[ERM generalization bound] \label{lem:erm-generalization-bound}
Let $P$ be the data generating distribution on the space $\mathcal{X}$ and let $\mathcal{H}$ be a given hypothesis class of functions. Assume that for all $x\in\mathcal{X}$ and $h \in \mathcal{H}$ we have that $|l(h,x)| \leq c_1$ for some positive constant $c_1>0$. Given a dataset $\mathcal{D}=\{X_i\}_{i=1}^N$, generated independently from $P$, denote $\hat{h}$ as the ERM solution, i.e. $\hat{h}=\argmin_{h\in\mathcal{H}} (1/N) \sum_{i=1}^N l(h, X_i)$. For any fixed $\delta\in(0,1)$ and $h^*\in\argmin_{h\in\mathcal{H}} \mathbb{E}_{X\sim P}[l(h,X)]$, we have
\begin{align} \label{eq:erm-generalization-bound-part-1} \mathbb{E}_{X\sim P}[l(\hat{h},X)] - \mathbb{E}_{X\sim P}[l(h^*,X)] \leq 2 R(l\circ \mathcal{H} \circ \mathcal{D}) + 5 c_1 \sqrt{\frac{2\log(8/\delta)}{N}},
\end{align}
with probability at least $1-\delta$, where $R(\cdot)$ is the Rademacher complexity of $l\circ \mathcal{H}$ given by
\[
R(l\circ \mathcal{H} \circ \mathcal{D}) = \frac{1}{N} \mathbb{E}_{\{\sigma_i\}_{i=1}^N} \left( \sup_{g\,\in\, l\circ \mathcal{H}}~ \sum_{i=1}^N \sigma_i g(X_i) \right),
\]
in which $\sigma_i$'s are independent from $X_i$'s and are independently and identically distributed according to the Rademacher random variable $\sigma$, i.e. $\mathbb{P}(\sigma=1) = 0.5 = \mathbb{P}(\sigma=-1)$.
Furthermore, if $\mathcal{H}$ is a finite hypothesis class, i.e. $|\mathcal{H}|<\infty$, with $|h\circ x| \leq c_2$ for all $h\in\mathcal{H}$ and $x\in\mathcal{X}$, and $l(h,x)$ is $c_3$-Lipschitz in $h$, then we have
\begin{align} \label{eq:erm-generalization-bound-part-2}
\mathbb{E}_{X\sim P}[l(\hat{h},X)] - \mathbb{E}_{X\sim P}[l(h^*,X)] \leq 2 c_2 c_3 \sqrt{\frac{2\log(|\mathcal{H}|)}{N}} + 5 c_1 \sqrt{\frac{2\log(8/\delta)}{N}},
\end{align}
with probability at least $1-\delta$.
\end{lemma}
We now mention two important concepts from variational analysis \citep{rockafellar2009variational} literature that is useful to relate minimization of integrals and the integrals of pointwise minimization under special class of functions.
\begin{definition}[Decomposable spaces and Normal integrands \text{\citep[Definition 14.59, Example 14.29]{rockafellar2009variational}}] \label{def:decomposable-normal-integrand}
A space $\mathcal{X}$ of measurable functions is a {decomposable space} relative to an underlying measure space $(\Omega, \mathcal{A}, \mu)$, if for every function $x_0 \in \mathcal{X}$, every set $A \in \mathcal{A}$ with $\mu(A) < \infty$, and any bounded measurable function $x_1\,:\,A\to\mathbb{R}$, the function $x(\omega) = x_0(\omega)\mathbbm{1}(\omega \notin A) + x_1(\omega)\mathbbm{1}(\omega \in A)$ belongs to $\mathcal{X}$. A function $f : \Omega \times \mathbb{R} \to \mathbb{R}$ (finite-valued) is a {normal integrand}, if and only if $f(\omega,x)$ is $\mathcal{A}$-measurable in $\omega$ for each $x$ and is continuous in $x$ for each $\omega$.
\end{definition}
\begin{remark} \label{remark:decomposable}
A few examples of decomposable spaces are $L^p(\cS\times\cA,\Sigma(\cS\times\cA),\mu)$ for any $p\geq 1$ and $\mathcal{M}(\cS\times\cA,\Sigma(\cS\times\cA))$, the space of all $\Sigma(\cS\times\cA)$-measurable functions. \end{remark}
\begin{lemma}[\text{\citealp[Theorem 14.60]{rockafellar2009variational}}] \label{lem:integral-min-exchange-rockafellar}
Let $\mathcal{X}$ be a space of measurable functions from $\Omega$ to $\mathbb{R}$ that is decomposable relative to a $\sigma$-finite measure $\mu$ on the $\sigma$-algebra $\mathcal{A}$. Let $f : \Omega \times \mathbb{R} \to \mathbb{R}$ (finite-valued) be a normal integrand. Then, we have
\[
\inf_{x\in\mathcal{X}} \int_{\omega\in\Omega} f(\omega,x(\omega)) \mu(\mathrm{d}\,\omega) = \int_{\omega\in\Omega} \left( \inf_{x\in\mathbb{R}} f(\omega,x) \right) \mu(\mathrm{d}\,\omega).
\]
Moreover, as long as the above infimum is not $-\infty$, we have that
\begin{align*}
x' \in \argmin_{x\in\mathcal{X}} \int_{\omega\in\Omega} f(\omega,x(\omega)) \mu(\mathrm{d}\,\omega),
\end{align*}
if and only if $\;x'(\omega) \in \argmin_{x\in\mathbb{R}} f(\omega,x)\cdot\mu\;$ almost surely.
\end{lemma}
We now give one result from distributioanlly robust optimization. The $f$-divergence between the distributions $P$ and $P^o$ is defined as
\begin{align}
\label{eq:f-divergence-defn}
D_{f}(P \| P^o) = \int f(\frac{\mathrm{d}\, P}{\mathrm{d}\, P^o})\mathrm{d}\, P^o,
\end{align}
where $f$ is a convex function \citep{csiszar1967information,moses2011further}. We obtain different divergences for different forms of the function $f$, including some well-known divergences. For example,
$f(t) = |t-1|/2$ gives Total Variation (TV), $f(t) = t\log t$ gives Kullback-Liebler (KL), $f(t) = (t-1)^2$ gives Chi-square, and $f(t) = (\sqrt{t}-1)^2$ gives squared Hellinger divergences.
Let $P^o$ be a distribution on the space $\mathcal{X}$ and let $l: \mathcal{X} \to \mathbb{R}$ be a loss function. We have the following result from the \textit{distributionally robust optimization} literature, see e.g.,~\citet[Section 3.2]{shapiro2017distributionally} and~\citet[Proposition 1]{duchi2018learning}.
\begin{proposition}
\label{prop:dro-inner-problem-solution}
Let $D_{f}$ be the $f$-divergence as defined in \eqref{eq:f-divergence-defn}. Then,
\begin{align}
\label{eq:dro-inner-problem-solution}
\sup_{D_f(P\|P^o) \leq \rho} \mathbb{E}_{P}[l(X)] = \inf_{\lambda>0, \eta\in\mathbb{R}} ~~ \mathbb{E}_{P^o} \left[ \lambda f^* \left(\frac{l(X)-\eta}{\lambda}\right)\right] + \lambda\rho + \eta,
\end{align}
where $f^*(s) = \sup_{t\geq 0} \{st - f(t)\}$ is the Fenchel conjugate.
\end{proposition}
Note that on the right hand side of \eqref{eq:dro-inner-problem-solution}, the expectation is taken only with respect to $P^{o}$. We will use the above result to derive the dual reformulation of the robust Bellman operator.
\section{Proof of the Proposition \ref{prop:robust-bellman-dual}}
\label{appendix:prop:robust-bellman-dual}
As the first step, we adapt the result given in Proposition~\ref{prop:dro-inner-problem-solution} in two ways: $(i)$ Since Proposition~\ref{prop:robust-bellman-dual} considers the TV uncertainty set, we will derive the specific form of this result for the TV uncertainty set, $(ii)$ Since Proposition~\ref{prop:robust-bellman-dual} considers the minimization problem instead of the maximization problem, unlike in Proposition~\ref{prop:dro-inner-problem-solution}, we will derive the specific form of this result for minimization.
\begin{lemma}
\label{lem:TV-inner-problem}
Let $D_{f}$ be as defined in \eqref{eq:f-divergence-defn} with $f(t)=|t-1|/2$ corresponding to the TV uncertainty set. Then,
\begin{align*}
\inf_{D_f(P\|P^o) \leq \rho} \mathbb{E}_{P}[l(X)] = -\inf_{\eta \in\mathbb{R}}\;\mathbb{E}_{P^o}[(\eta-l(X))_+] + (\eta - \inf_{x\in\mathcal{X}} l(x))_+ \times \rho - \eta,
\end{align*}
\end{lemma}
\begin{proof}
First, we will compute the Fenchel conjugate of $f(t)=|t-1|/2$. We have
\begin{align*}
f^*(s) &= \sup_{t\geq 0} ~ \{st - \frac{1}{2}|t-1|\}
= \max \big\{\!\!\sup_{t\in[0,1]} \!\{(s+\frac{1}{2})t - \frac{1}{2}\} \;,\; \sup_{t> 1} \!\{(s-\frac{1}{2}) t +\frac{1}{2}\} \big\}.
\end{align*}
It is easy to see that for $s>1/2$, we have $f^*(s)=+\infty$, and for $s\leq -1/2$, we have $f^*(s)=-1/2$. For $s \in [-1/2,1/2]$, we have
\begin{align*}
f^*(s) &= \max \big\{ \sup_{t\in[0,1]}\{(s+\frac{1}{2})t - \frac{1}{2}\} \;,\; \sup_{t> 1} (\{(s-\frac{1}{2}) t +\frac{1}{2}\} \big\} \\
&= \max \big\{ ((s+\frac{1}{2})\cdot 1 - \frac{1}{2}),~ ((s-\frac{1}{2})\cdot 1 +\frac{1}{2}) \big\} = s.
\end{align*}
Thus, we have
\begin{equation*}
f^*(s) =
\begin{cases}
-\frac{1}{2} & s\leq -\frac{1}{2},\\
s & s\in [-\frac{1}{2},\frac{1}{2}]\\
+\infty & s > \frac{1}{2}.
\end{cases} .
\end{equation*}
From Proposition \ref{prop:dro-inner-problem-solution}, we obtain
\begin{align*}
\sup_{D_f(P\|P^o) \leq \rho} \mathbb{E}_{P}[l(X)] &= \inf_{\lambda>0, \eta\in\mathbb{R}} ~ \mathbb{E}_{P^o} [\lambda f^*(\frac{l(X)-\eta}{\lambda})] + \lambda\rho + \eta \\
&= \inf_{\lambda,\eta: \lambda>0, \eta\in\mathbb{R}, \frac{\sup_{x\in\mathcal{X}}l(x)-\eta}{\lambda} \leq \frac{1}{2}} ~ \mathbb{E}_{P^o} [\lambda \max\{ \frac{l(X)-\eta}{\lambda}, -\frac{1}{2} \}] + \lambda\rho + \eta \\
&= \inf_{\lambda,\eta: \lambda>0, \eta\in\mathbb{R}, \frac{\sup_{x\in\mathcal{X}}l(x)-\eta}{\lambda} \leq \frac{1}{2}} ~ \mathbb{E}_{P^o} [\max\{ l(X)-\eta, -\lambda/2 \}] + \lambda\rho + \eta \\
&= \inf_{\lambda,\eta: \lambda>0, \eta\in\mathbb{R}, \frac{\sup_{x\in\mathcal{X}}l(x)-\eta}{\lambda} \leq \frac{1}{2}} ~ \mathbb{E}_{P^o} [ (l(X)-\eta +\lambda/2)_+ ] - \lambda/2 + \lambda\rho + \eta \\
&= \inf_{\lambda,\eta': \lambda>0, \eta'\in\mathbb{R}, \frac{\sup_{x\in\mathcal{X}}l(x)-\eta'}{\lambda} \leq 1} ~ \mathbb{E}_{P^o} [ (l(X)-\eta')_+] + \lambda\rho + \eta'.
\end{align*}
The second equality follows since $f^*(\frac{l(X)-\eta}{\lambda})=+\infty$ whenever $\frac{l(X)-\eta}{\lambda}>\frac{1}{2}$, which can be ignored as we are minimizing over $\lambda$ and $\eta$. The fourth equality follows form the fact that $\max\{x,y\} = (x-y)_+ + y$ for any $x,y\in\mathbb{R}$. Finally, the last equality follows by making the substitution $\eta'=\eta-\lambda/2$. Taking the optimal value of $\lambda$, i.e.,~$\lambda=(\sup_{x\in\mathcal{X}}l(x)-\eta')_+$, we get
\begin{align*}
\sup_{D_f(P\|P^o) \leq \rho} \mathbb{E}_{P}[l(X)] = \inf_{\eta\in\mathbb{R}} \mathbb{E}_{P^o}[(l(X)-\eta)_+] + (\sup_{x\in\mathcal{X}}l(x) - \eta)_+ \rho + \eta.
\end{align*}
Now,
\begin{align*}
\inf_{D_f(P\|P^o) \leq \rho} \mathbb{E}_{P}[l(X)] &= -\sup_{D_f(P\|P^o) \leq \rho} \mathbb{E}_{P}[-l(X)] \\
&= -\inf_{\eta\in\mathbb{R}} ~~ \mathbb{E}_{P^o}[(-l(X)-\eta)_+] + (\sup_{x\in\mathcal{X}} -l(x) - \eta)_+ \rho + \eta \\
&= -\inf_{\eta' \in\mathbb{R}} ~~ \mathbb{E}_{P^o}[(\eta'-l(X))_+] + (\eta' - \inf_{x\in\mathcal{X}} l(x))_+ \rho - \eta',
\end{align*}
which completes the proof.
\end{proof}
We are now ready to prove Proposition \ref{prop:robust-bellman-dual}.
\begin{proof}[\textbf{Proof of Proposition~\ref{prop:robust-bellman-dual}}]
For each $(s, a)$, the optimization problem in~\eqref{eq:robust-bellman-primal} is given by $\min_{P_{s,a} \in \mathcal{P}_{s,a}} \mathbb{E}_{s' \sim P_{s,a}} [ V(s')]$, and our focus is on the setting where $\mathcal{P}_{s,a}$ is given by the TV uncertainty set. So, $\mathcal{P}_{s,a}$ can be equivalently defined using the $f$-divergence with $f(t) = |t-1|/2$ as $\mathcal{P}_{s,a} = \{P_{s,a} : D_{f}(P_{s,a}||P^{o}_{s,a}) \leq \rho\}$. We can now use the result of Lemma~\ref{lem:TV-inner-problem} to get
\begin{align*}
\inf_{P_{s,a} \in \mathcal{P}_{s,a}} \mathbb{E}_{s' \sim P_{s,a}} [ V(s')] = -\inf_{\eta \in\mathbb{R}} ~~ \mathbb{E}_{s' \sim P^o_{s,a}}[(\eta-V(s'))_+] + (\eta - \inf_{s'' \in \cS} V(s''))_+ \rho - \eta.
\end{align*}
From Proposition~\ref{prop:dro-inner-problem-solution}, the function $h(\eta)=\mathbb{E}_{s' \sim P^{o}_{s,a}}[(\eta -V(s'))_+] + \rho (\eta- \inf_{s''} V(s''))_+ - \eta $ is convex in $\eta$. Since $V(s') \geq 0,$ $h(\eta)=-\eta \geq 0$ when $\eta \leq 0$. So, $\inf_{\eta \in (-\infty, 0]} h(\eta)$, achieved at $\eta = 0$. Also, since $V(s)\leq 1/(1-\gamma)$, we have
\begin{align*}
h(\frac{2}{\rho(1-\gamma)}) &= \mathbb{E}_{s' \sim P^{o}_{s,a}}[\frac{2}{\rho(1-\gamma)} -V(s')] + \rho (\frac{2}{\rho(1-\gamma)}- \inf_{s''} V(s'')) - \frac{2}{\rho(1-\gamma)} \\
&\geq -\frac{1}{(1-\gamma)} + \rho (\frac{2}{\rho(1-\gamma)}- \frac{1}{(1-\gamma)} ) = \frac{2}{(1-\gamma)} - \frac{(1+\rho)}{(1-\gamma)} \geq 0.
\end{align*}
So, it is sufficient to consider $\eta \in [0, \frac{2}{\rho (1-\gamma)}]$ for the above optimization problem.
Using these, we get
\begin{align*}
(T Q)&(s, a) = r(s, a) + \gamma \inf_{P_{s,a} \in \mathcal{P}_{s,a}} \mathbb{E}_{s' \sim P_{s,a}} [ V(s')] \\
&= r(s, a) + \gamma \cdot -1 \cdot \inf_{\eta \in \eta \in [0, \frac{2}{\rho (1-\gamma)}]} ~~ \mathbb{E}_{s' \sim P^o_{s,a}}[(\eta-V(s'))_+] + (\eta - \inf_{s'' \in \cS} V(s''))_+ \rho - \eta.
\end{align*}
This completes the proof of Proposition~\ref{prop:robust-bellman-dual}.
\end{proof}
\section{Proof of Theorem \ref{thm:tv-guarantee}}
\label{appendix:thm:tv-guarantee}
We start by proving Lemma \ref{lem:interchanging-lemma-1} which mainly follows from Lemma~\ref{lem:integral-min-exchange-rockafellar} in Appendix~\ref{appendix:useful-tech-results}.
\begin{proof}[Proof of Lemma \ref{lem:interchanging-lemma-1}]
Let $h((s,a),\eta)=\mathbb{E}_{s'\sim P^o_{s,a}} ((\eta - \max_{a'} f(s',a'))_+ - (1-\rho)\eta)$. We note that $h((s,a),\eta)$ is $\Sigma(\cS\times\cA)$-measurable in $(s,a)\in\cS\times\cA$ for each $\eta\in[0,{1}/{(\rho(1-\gamma))}]$ and is continuous in $\eta$ for each $(s,a)\in\cS\times\cA$. Now it follows that $h((s,a),\eta)$ is a normal integrand (see Definition~\ref{def:decomposable-normal-integrand} in Appendix~\ref{appendix:useful-tech-results}). We now note that $L^{1}(\cS\times\cA,\Sigma(\cS\times\cA),\mu)$ is a decomposable space (Remark \ref{remark:decomposable} in Appendix~\ref{appendix:useful-tech-results}). Thus, this lemma now directly follows from Lemma~\ref{lem:integral-min-exchange-rockafellar}.
\end{proof}
Now we state a result and provide its proof for the empirical risk minimization on the dual parameter.
\begin{lemma}[Dual Optimization Error Bound] \label{lem:erm-high-prob-bound}
Let $\widehat{g}_{f}$ be the dual optimization parameter from the algorithm (Step 4) for the state-action value function $f$ and let $T_{g}$ be as defined in \eqref{eq:Tg}. With probability at least $1-\delta$, we have
\[
\sup_{f\in\mathcal{F}} \|T f - T_{\widehat{g}_{f}} f\|_{1,\mu} \leq \frac{4\gamma(2-\rho)}{\rho(1-\gamma)}\sqrt{\frac{2\log(|\mathcal{G}|)}{N}} + \frac{25\gamma}{\rho(1-\gamma)} \sqrt{\frac{2\log(8|\mathcal{F}|/\delta)}{N}} + \gamma \epsilon_{\textrm{dual}}.
\]
\end{lemma}
\begin{proof}
Fix an $f\in\mathcal{F}$. We will also invoke union bound for the supremum here. We recall from~\eqref{eq:L-dual-empirical-loss} that $\widehat{g}_f = \argmin_{g \in \mathcal{G}} \widehat{L}_{\mathrm{dual}}(g; f)$. From the robust Bellman equation, we directly obtain
\begingroup
\allowdisplaybreaks
\begin{align*}
&\|T_{\widehat{g}_{f}} f - T f \|_{1,\mu} = \gamma (\mathbb{E}_{s,a\sim\mu} | \mathbb{E}_{s'\sim P^o_{s,a}} ((\widehat{g}_{f}(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)\widehat{g}_{f}(s,a)) \\&\hspace{2cm}- \inf_{\eta \in[0, {2}/{(\rho(1-\gamma))}]} \mathbb{E}_{s'\sim P^o_{s,a}} ((\eta - \max_{a'} f(s',a'))_+ - (1-\rho) \eta) | ) \\
&\stackrel{(a)}{=} \gamma (\mathbb{E}_{s,a\sim\mu}\mathbb{E}_{s'\sim P^o_{s,a}} ((\widehat{g}_{f}(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)\widehat{g}_{f}(s,a)) \\&\hspace{1cm}- \mathbb{E}_{s,a\sim\mu} [ \inf_{\eta \in[0, {2}/{(\rho(1-\gamma))}]} \mathbb{E}_{s'\sim P^o_{s,a}} ((\eta - \max_{a'} f(s',a'))_+ - (1-\rho) \eta) ] ) \\
&\stackrel{(b)}{=} \gamma (\mathbb{E}_{s,a\sim\mu,s'\sim P^o_{s,a}} ((\widehat{g}_{f}(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)\widehat{g}_{f}(s,a)) \\&\hspace{1cm}- \inf_{g \in L^{1}} \mathbb{E}_{s,a\sim\mu,s'\sim P^o_{s,a}} ((g(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)g(s,a)) )\\
&= \gamma (\mathbb{E}_{s,a\sim\mu,s'\sim P^o_{s,a}} ((\widehat{g}_{f}(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)\widehat{g}_{f}(s,a)) \\&\hspace{1cm}- \inf_{g\in \mathcal{G}} \mathbb{E}_{s,a\sim\mu,s'\sim P^o_{s,a}} ((g(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)g(s,a)) )
\\ &\hspace{1cm} + \gamma (\inf_{g\in \mathcal{G}} \mathbb{E}_{s,a\sim\mu,s'\sim P^o_{s,a}} ((g(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)g(s,a))
\\&\hspace{1.5cm}- \inf_{g \in L^{1}} \mathbb{E}_{s,a\sim\mu,s'\sim P^o_{s,a}} ((g(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)g(s,a)) )\\
&\stackrel{(c)}{\leq} \gamma (\mathbb{E}_{s,a\sim\mu,s'\sim P^o_{s,a}} ((\widehat{g}_{f}(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)\widehat{g}_{f}(s,a)) \\&\hspace{1cm}- \inf_{g\in \mathcal{G}} \mathbb{E}_{s,a\sim\mu,s'\sim P^o_{s,a}} ((g(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)g(s,a)) ) + \gamma \epsilon_{\textrm{dual}} \\
&\stackrel{(d)}{\leq} 2\gamma R(l\circ \mathcal{G} \circ \mathcal{D}) + \frac{25\gamma}{\rho(1-\gamma)} \sqrt{\frac{2\log(8/\delta)}{N}} + \gamma \epsilon_{\textrm{dual}}
\\&\stackrel{(e)}{\leq} \frac{4\gamma(2-\rho)}{\rho(1-\gamma)}\sqrt{\frac{2\log(|\mathcal{G}|)}{N}} + \frac{25\gamma}{\rho(1-\gamma)} \sqrt{\frac{2\log(8/\delta)}{N}} + \gamma \epsilon_{\textrm{dual}}.
\end{align*}
\endgroup
$(a)$ follows since $\inf_{g} h(g) \leq h(\widehat{g}_{f})$. $(b)$ follows from Lemma \ref{lem:interchanging-lemma-1}. $(c)$ follows from the approximate dual realizability assumption (Assumption \ref{assum-dual-realizability}).
For $(d)$, we consider the loss function $l(g,(s,a,s'))=(g(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)g(s,a)$ and dataset $\mathcal{D}=\{s_i,a_i,s_i'\}_{i=1}^N$. Note that $|l(g,(s,a,s'))|\leq 5/(\rho(1-\gamma))$ (since $f\in\mathcal{F}$ and $g\in\mathcal{G}$). Now, we can apply the empirical risk minimization result \eqref{eq:erm-generalization-bound-part-1} in Lemma \ref{lem:erm-generalization-bound} to get $(d)$, where $R(\cdot)$ is the Rademacher complexity.
Finally, $(e)$ follows from~\eqref{eq:erm-generalization-bound-part-2} in Lemma \ref{lem:erm-generalization-bound} when combined with the facts that $l(g,(s,a,s'))$ is $(2-\rho)$-Lipschitz in $g$ and $g(s,a) \leq 2/(\rho(1-\gamma))$, since $g\in\mathcal{G}$.
With union bound, with probability at least $1-\delta$, we finally get \begin{align*}
&\sup_{f\in\mathcal{F}}\|T f - T_{\widehat{g}_{f}} f\|_{1,\mu} {\leq} \frac{4\gamma(2-\rho)}{\rho(1-\gamma)}\sqrt{\frac{2\log(|\mathcal{G}|)}{N}} + \frac{25\gamma}{\rho(1-\gamma)} \sqrt{\frac{2\log(8|\mathcal{F}|/\delta)}{N}} + \gamma \epsilon_{\textrm{dual}},
\end{align*}
which concludes the proof.
\end{proof}
We next prove the least-squares generalization bound for the RFQI algorithm.
\begin{lemma}[Least squares generalization bound] \label{lem:least-squares-generalization-bound}
Let $\widehat{f}_{g}$ be the least-squares solution from the algorithm (Step 5) for the state-action value function $f$ and dual variable function $g$. Let $T_{g}$ be as defined in \eqref{eq:Tg}. Then, with probability at least $1-\delta$, we have
\begin{align*}
\sup_{f\in\mathcal{F}} \sup_{g\in\mathcal{G}} \| T_{g} f - \widehat{f}_{g} \|_{2,\mu}
&\leq \sqrt{6 \epsilon_{\textrm{c}}} + \frac{16}{\rho(1-\gamma)} \sqrt{\frac{18 \log(2|\mathcal{F}||\mathcal{G}|/\delta)}{N}}.
\end{align*}
\end{lemma}
\begin{proof}
We adapt the least-squares generalization bound given in~\citet[Lemma A.11]{agarwal2019reinforcement} to our setting. We recall from~\eqref{eq:loss-rfqi-1} that $\widehat{f}_{g}=\argmin_{Q\in\mathcal{F}}\widehat{L}_{\mathrm{RFQI}}(Q;f, g)$. We first fix functions $f\in\mathcal{F}$ and $g\in\mathcal{G}$. For any function $f'\in\mathcal{F}$, we define random variables $z_i^{f'}$ as
\begin{align*}
z_i^{f'} = \left( f'(s_i,a_i) - y_{i} \right)^2 - \left( (T_{g} f)(s_i,a_i) - y_{i} \right)^2,
\end{align*}
where $y_{i} = r_i - \gamma(g(s_i,a_i) - \max_{a'} f(s_i',a'))_+ + \gamma (1-\rho)g(s_i,a_i)$, and $(s_{i}, a_{i}, s'_{i}) \in \mathcal{D}$ with $(s_{i}, a_{i}) \sim \mu, s'_{i} \sim P^{o}_{s_{i}, a_{i}}$. It is straightforward to note that for a given $(s_{i}, a_{i})$, we have $ \mathbb{E}_{s'_i \sim P^o_{s_i,a_i}} [y_{i}] = (T_{g} f)(s_{i}, a_{i})$.
Also, since $g(s_i,a_i)\leq 2/(\rho(1-\gamma))$ (because $g\in\mathcal{G}$) and $f(s_i,a_i), f'(s_i,a_i)\leq 1/(1-\gamma)$ (because $f, f'\in\mathcal{F})$, we have $(T_{g}f)(s_i,a_i) \leq 5/(\rho(1-\gamma))$. This also gives us that $y_{i} \leq 5/(\rho(1-\gamma))$.
Using this, we obtain the first moment and an upper-bound for the second moment of $z_i^{f'}$ as follows:
\begin{align*}
\mathbb{E}_{s'_i \sim P^o_{s_i,a_i}} [z_i^{f'}] &= \mathbb{E}_{s'_i \sim P^o_{s_i,a_i}} [( f'(s_i,a_i) - (T_{g} f)(s_i,a_i)) \cdot (f'(s_i,a_i) + (T_{g} f)(s_i,a_i) -2 y_{i} ) ] \\
&= ( f'(s_i,a_i) - (T_{g} f)(s_i,a_i))^{2}, \\
\mathbb{E}_{s'_i \sim P^o_{s_i,a_i}} [(z_i^{f'})^2] &= \mathbb{E}_{s'_i \sim P^o_{s_i,a_i}} [( f'(s_i,a_i) - (T_{g} f)(s_i,a_i))^{2} \cdot (f'(s_i,a_i) + (T_{g} f)(s_i,a_i) -2 y_{i} )^{2} ] \\
&= ( f'(s_i,a_i) - (T_{g} f)(s_i,a_i))^{2} \cdot \mathbb{E}_{s'_i \sim P^o_{s_i,a_i}} [ (f'(s_i,a_i) + (T_{g} f)(s_i,a_i) -2 y_{i} )^{2} ] \\
&\leq C_{1} ( f'(s_i,a_i) - (T_{g} f)(s_i,a_i))^{2} ,
\end{align*}
where $C_{1}={16^2}/{(\rho^2(1-\gamma)^2)}$. This immediately implies that
\begin{align*}
\mathbb{E}_{s_i,a_i\sim \mu, s'_i \sim P^o_{s_i,a_i}} [z_i^{f'}] &= \norm{T_{g}f -f'}^{2}_{2,\mu}, \\
\mathbb{E}_{s_i,a_i\sim \mu, s'_i \sim P^o_{s_i,a_i}} [(z_i^{f'})^2] &\leq C_{1} \norm{T_{g}f -f'}^{2}_{2,\mu}.
\end{align*}
From these calculations, it is also straightforward to see that $|z_i^{f'}-\mathbb{E}_{s_i,a_i\sim \mu, s'_i \sim P^o_{s_i,a_i}} [z_i^{f'}]| \leq 2 C_{1}$ almost surely.
Now, using the Bernstein's inequality (Lemma \ref{lem:bernstein-ineq}), together with a union bound over all $f'\in \mathcal{F}$, with probability at least $1-\delta$, we have
\begin{align}
\label{eq:least-squares-bern-main}
|\| T_{g} f - f' \|_{2,\mu}^2 - \frac{1}{N} \sum_{i=1}^N z_i^{f'}| \leq \sqrt{\frac{2 C_{1} \| T_{g} f - f' \|_{2,\mu}^2 \log(2|\mathcal{F}|/\delta)}{N}} + \frac{2C_{1} \log(2|\mathcal{F}|/\delta)}{3N},
\end{align}
for all $f'\in \mathcal{F}$. Setting $f'=\widehat{f}_{g}$, with probability at least $1-\delta/2$, we have
\begin{align}
\label{eq:least-squares-bern-1}
\| T_{g} f - \widehat{f}_{g} \|_{2,\mu}^2 \leq \frac{1}{N} \sum_{i=1}^N z_i^{\widehat{f}_{g}}+ \sqrt{\frac{2 C_{1} \| T_{g} f - \widehat{f}_{g} \|_{2,\mu}^2 \log(4|\mathcal{F}|/\delta)}{N}} + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{3N}.
\end{align}
Now we upper-bound $(1/N) \sum_{i=1}^N z_i^{\widehat{f}_{g}}$ in the following. Consider a function $\widetilde{f} \in \argmin_{h\in\mathcal{F}} \| h - T_{g} f \|_{2,\mu}^2$. Note that $\widetilde{f}$ is independent of the dataset. We note that our earlier first and second moment calculations hold true for $\widetilde{f}$, replacing $f'$, as well. Now, from \eqref{eq:least-squares-bern-main} setting $f'=\widetilde{f}$, with probability at least $1-\delta/2$ we have
\begin{align} \label{eq:least-squares-z-main}
\frac{1}{N} \sum_{i=1}^N z_i^{\widetilde{f}} -\| T_{g} f - \widetilde{f} \|_{2,\mu}^2 \leq \sqrt{\frac{2 C_{1} \| T_{g} f - \widetilde{f} \|_{2,\mu}^2 \log(4|\mathcal{F}|/\delta)}{N}} + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{3N}.
\end{align}
Suppose $(1/N) \sum_{i=1}^N z_i^{\widetilde{f}} \geq 2C_{1}\log(4|\mathcal{F}|/\delta)/N$ holds, then from \eqref{eq:least-squares-z-main} we get
\begin{align} \label{eq:least-squares-z-main-2}
\frac{1}{N} \sum_{i=1}^N z_i^{\widetilde{f}} -\| T_{g} f - \widetilde{f} \|_{2,\mu}^2 \leq \sqrt{ \| T_{g} f - \widetilde{f} \|_{2,\mu}^2 \cdot \frac{1}{N} \sum_{i=1}^N z_i^{\widetilde{f}}} + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{N}.
\end{align}
We note the following algebra fact: Suppose $x^2 - ax + b \leq 0$ with $b>0$ and $a^2\geq 4b$, then we have $x\leq a$. Taking $x=(1/N) \sum_{i=1}^N z_i^{\widetilde{f}}$ in this fact, from~\eqref{eq:least-squares-z-main-2} we get
\begin{align} \label{eq:least-squares-z-part-1}
\frac{1}{N} \sum_{i=1}^N z_i^{\widetilde{f}} \leq 3\| T_{g} f - \widetilde{f} \|_{2,\mu}^2 + \frac{4C_{1} \log(4|\mathcal{F}|/\delta)}{3N} \leq 3\| T_{g} f - \widetilde{f} \|_{2,\mu}^2 + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{N}.
\end{align}
Now suppose $(1/N) \sum_{i=1}^N z_i^{\widetilde{f}} \leq 2 C_{1}\log(4|\mathcal{F}|/\delta)/N$, then \eqref{eq:least-squares-z-part-1} holds immediately. Thus,~\eqref{eq:least-squares-z-part-1} always holds with probability at least $1-\delta/2$. Furthermore, recall $\widetilde{f} \in \argmin_{h\in\mathcal{F}} \| h - T_{g} f \|_{2,\mu}^2$, we have
\begin{align}
\frac{1}{N} \sum_{i=1}^N z_i^{\widetilde{f}} &\leq 3\| T_{g} f - \widetilde{f} \|_{2,\mu}^2 + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{N} \nonumber \\
&= 3\min_{h\in\mathcal{F}} \| h - T_{g} f \|_{2,\mu}^2 + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{N} \leq 3 \epsilon_{\textrm{c}} + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{N}
\label{eq:least-squares-z-part-2},
\end{align}
where the last inequality follows from the approximate robust Bellman completion assumption (Assumption \ref{assum-bellman-completion}).
We note that since $\widehat{f}_{g}$ is the least-squares regression solution, we know that $(1/N) \sum_{i=1}^N z_i^{\widehat{f}_{g}} \leq (1/N) \sum_{i=1}^N z_i^{\widetilde{f}}$. With this note in \eqref{eq:least-squares-z-part-2}, from \eqref{eq:least-squares-bern-1}, with probability at least $1-\delta$, we have
\begin{align*}
\| T_{g} f - \widehat{f}_{g} \|_{2,\mu}^2 &\leq 3 \epsilon_{\textrm{c}} + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{N}
\\&+ \sqrt{\frac{2 C_{1} \| T_{g} f - \widehat{f}_{g} \|_{2,\mu}^2 \log(4|\mathcal{F}|/\delta)}{N}} + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{3N} \\
&\leq 3 \epsilon_{\textrm{c}} + \frac{3C_{1} \log(4|\mathcal{F}|/\delta)}{N} + \sqrt{\frac{3 C_{1} \| T_{g} f - \widehat{f}_{g} \|_{2,\mu}^2 \log(4|\mathcal{F}|/\delta)}{N}}.
\end{align*}
From the earlier algebra fact, taking $x=\| T_{g} f - \widehat{f}_{g} \|_{2,\mu}^2$, with probability at least $1-\delta$, we have
\begin{align*}
\| T_{g} f - \widehat{f}_{g} \|_{2,\mu}^2
&\leq 6 \epsilon_{\textrm{c}} + \frac{9C_{1} \log(4|\mathcal{F}|/\delta)}{N}.
\end{align*}
From the fact $\sqrt{x+y}\leq \sqrt{x}+\sqrt{y}$, with probability at least $1-\delta$, we get
\begin{align*}
\| T_{g} f - \widehat{f}_{g} \|_{2,\mu}
&\leq \sqrt{6 \epsilon_{\textrm{c}}} + \sqrt{\frac{9C_{1} \log(4|\mathcal{F}|/\delta)}{N}}.
\end{align*}
Using union bound for $f\in\mathcal{F}$ and $g\in\mathcal{G}$, with probability at least $1-\delta$, we finally obtain
\begin{align*}
\sup_{f\in\mathcal{F}} \sup_{g\in\mathcal{G}} \| T_{g} f - \widehat{f}_{g} \|_{2,\mu}
&\leq \sqrt{6 \epsilon_{\textrm{c}}} + \sqrt{\frac{18 C_{1} \log(2|\mathcal{F}||\mathcal{G}|/\delta)}{N}},
\end{align*}
which completes the least-squares generalization bound analysis.
\end{proof}
We are now ready to prove the main theorem.
\begin{proof}[\textbf{Proof of Theorem \ref{thm:tv-guarantee}}]
We let $V_k(s) = Q_k(s,\pi_k(s))$ for every $s\in\cS$. Since $\pi_k$ is the greedy policy w.r.t $Q_k$, we also have $V_k(s) = Q_k(s,\pi_k(s)) = \max_a Q_k(s,a)$. We recall that $V^*=V^{\pi^*}$ and $Q^*=Q^{\pi^*}$. We also recall from Section~\ref{sec:preliminaries} that $Q^{\pi^*}$ is a fixed-point of the robust Bellman operator $T$ defined in~\eqref{eq:robust-bellman-primal}. We also note that the same holds true for any stationary deterministic policy $\pi$ from~\citet{iyengar2005robust} that $Q^{\pi}$ satisfies $Q^{\pi}(s, a) = r(s, a) + \gamma \min_{P_{s,a} \in \mathcal{P}_{s,a}} \mathbb{E}_{s' \sim P_{s,a}} [ V^{\pi}(s')].$ We can now further use the dual form \eqref{eq:robust-bellman-dual-2} under Assumption \ref{assum-fail-state}. We first characterize the performance decomposition between $V^{\pi^*}$ and ${V}^{\pi_K}$. For a given $s_0\in\cS$, we observe that
\begin{align*}
&V^{\pi^*}(s_0) - {V}^{\pi_K}(s_0) = (V^{\pi^*}(s_0) - V_K(s_0) ) - (V^{\pi_K}(s_0) - V_K(s_0) )\\
&= (Q^{\pi^*}(s_0,\pi^*(s_0)) - Q_K(s_0,\pi_K(s_0)) ) - (Q^{\pi_K}(s_0,\pi_K(s_0)) - Q_K(s_0,\pi_K(s_0))) \\
&\stackrel{(a)}{\leq} Q^{\pi^*}(s_0,\pi^*(s_0)) - Q_K(s_0,\pi^*(s_0)) + Q_K(s_0,\pi_K(s_0)) - Q^{\pi_K}(s_0,\pi_K(s_0)) \\
&= Q^{\pi^*}(s_0,\pi^*(s_0)) - Q_K(s_0,\pi^*(s_0)) + Q_K(s_0,\pi_K(s_0)) - Q^{\pi^*}(s_0,\pi_K(s_0))
\\&\hspace{5cm}+ Q^{\pi^*}(s_0,\pi_K(s_0)) - Q^{\pi_K}(s_0,\pi_K(s_0))
\\
&\stackrel{(b)}{\leq} Q^{\pi^*}(s_0,\pi^*(s_0)) - Q_K(s_0,\pi^*(s_0)) + Q_K(s_0,\pi_K(s_0)) - Q^{\pi^*}(s_0,\pi_K(s_0)) \\&\hspace{2cm}+ \gamma \sup_{\eta} (\mathbb{E}_{s_1\sim P^o_{s_0,\pi_{K}(s_0)}}( (\eta- V^{\pi_K}(s_1))_+ - (\eta-V^{\pi^*}(s_1))_+)) \\
&\stackrel{(c)}{\leq} |Q^{\pi^*}(s_0,\pi^*(s_0)) - Q_K(s_0,\pi^*(s_0))| + |Q^{\pi^*}(s_0,\pi_K(s_0)) - Q_K(s_0,\pi_K(s_0))|
\\&\hspace{4cm}+ \gamma \mathbb{E}_{s_1\sim P^o_{s_0,\pi_{K}(s_0)}}( |V^{\pi^*}(s_1)- V^{\pi_K}(s_1)| ).
\end{align*}
$(a)$ follows from the fact that $\pi_{K}$ is the greedy policy with respect to $Q_{K}$. $(b)$ follows from the Bellman optimality equations and the fact $|\sup_{x} f(x)-\sup_{x} g(x)| \leq \sup_{x} |f(x) - g(x)|$. Finally, $(c)$ follows from the facts $(x)_+ - (y)_+ \leq (x-y)_+$ and $(x)_+ \leq |x|$ for any $x,y\in\mathbb{R}$.
We now recall the initial state distribution $d_0$. Thus, we have
\begin{align*} &\mathbb{E}_{s_0\sim d_0}[{V}^{\pi^*}] - \mathbb{E}_{s_0\sim d_0}[V^{\pi_K}] \leq
\\&\hspace{0.5cm} \mathbb{E}_{s_0\sim d_0} \bigg[ |Q^{\pi^*}(s_0,\pi^*(s_0)) - Q_{K}(s_0,\pi^*(s_0))| + |Q^{\pi^*}(s_0,\pi_K(s_0)) - Q_{K}(s_0,\pi_K(s_0))|
\\&\hspace{3cm}+ \gamma \mathbb{E}_{s_1\sim P^{o}_{s_0,\pi_{K}(s_0)}} (|V^{\pi^*}(s_1) - {V}^{\pi_K}(s_1) |) \bigg].
\end{align*}
Since $V^{\pi^*}(s) \geq V^{\pi_K}(s)$ for any $s\in\cS$, by telescoping we get
\begin{align}
&\mathbb{E}_{s_0\sim d_0}[{V}^{\pi^*}] - \mathbb{E}_{s_0\sim d_0}[V^{\pi_K}] \leq \sum_{h=0}^\infty \gamma^h \times \nonumber \\
&\hspace{1cm} \bigg(\mathbb{E}_{s\sim d_{h,\pi_{K}}} [ |Q^{\pi^*}(s,\pi^*(s)) - Q_{K}(s,\pi^*(s))| + |Q^{\pi^*}(s,\pi_K(s)) - Q_{K}(s,\pi_K(s))| ] \bigg), \label{eq:perf-diff-1}
\end{align}
where $d_{h,\pi_K}\in \Delta(\cS)$ for all natural numbers $h\geq 0$ is defined as
\begin{equation*}
d_{h,\pi_{K}} =
\begin{cases}
d_0 & \text{if $h=0$},\\
P^{o}_{s',\pi_{K}(s')} & \text{otherwise, with } s'\sim d_{h-1,\pi_{K}}.
\end{cases}
\end{equation*}
We emphasize that the state distribution $d_{h,\pi_K}$'s are different from the discounted state-action occupancy distributions. We note that a similar state distribution proof idea is used in \citet{agarwal2019reinforcement}.
Recall $\|f\|_{p,\nu}^2 = (\mathbb{E}_{s,a\sim\nu} |f(s,a)|^p)^{1/p}$, where $\nu\in\Delta(\cS\times\cA)$. With this we have
\begin{align}
\label{eq:thm-bound-part-1}
&\mathbb{E}_{s_0\sim d_0}[{V}^{\pi^*}] - \mathbb{E}_{s_0\sim d_0}[V^{\pi_K}] \leq \sum_{h=0}^\infty \gamma^h \bigg( \|Q^{\pi^*} - Q_{K}\|_{1,d_{h,\pi_{K}}\circ \pi^* } + \|Q^{\pi^*} - Q_{K}\|_{1,d_{h,\pi_{K}}\circ \pi_K } \bigg),
\end{align}
where the state-action distributions $d_{h,\pi_{K}}\circ \pi^*(s,a)\propto d_{h,\pi_{K}}(s) \mathbbm{1}\{ a=\pi^*(s)\}$ and $d_{h,\pi_{K}}\circ \pi_K(s,a)\propto d_{h,\pi_{K}}(s) \mathbbm{1}\{ a=\pi_K(s)\}$ directly follows by comparing with \eqref{eq:perf-diff-1}.
We now bound one of the RHS terms above by bounding for any state-action distribution $\nu$ satisfying Assumption \ref{assum-concentra-condition} (in particular the following bound is true for $d_{h,\pi_{K}}\circ \pi^*$ or $d_{h,\pi_{K}}\circ \pi_K$ in \eqref{eq:perf-diff-1}):
\begin{align}
&\|Q^{\pi^*} - Q_{K}\|_{1,\nu} \leq \|Q^{\pi^*} - T Q_{K-1}\|_{1,\nu} + \|T Q_{K-1} - Q_{K}\|_{1,\nu} \nonumber \\
&\stackrel{(a)}{\leq} \|Q^{\pi^*} - T Q_{K-1}\|_{1,\nu} + \sqrt{C}\|T Q_{K-1} - Q_{K}\|_{1,\mu} \nonumber \\
&= (\mathbb{E}_{s,a\sim\nu} |Q^{\pi^*}(s,a) - T Q_{K-1}(s,a)|) + \sqrt{C}\|T Q_{K-1} - Q_{K}\|_{1,\mu} \nonumber \\
&\stackrel{(b)}{\leq} (\mathbb{E}_{s,a\sim\nu} \gamma \sup_{\eta}|
\mathbb{E}_{s'\sim P^o_{s,a}} ( (\eta-\max_{a'} Q_{K-1}(s',a'))_+ - (\eta-\max_{a'} Q^{\pi^*}(s',a'))_+)
|)
\nonumber \\&\hspace{6cm}+ \sqrt{C}\|T Q_{K-1} - Q_{K}\|_{1,\mu} \nonumber \\
&\stackrel{(c)}{\leq} (\mathbb{E}_{s,a\sim\nu} |
\mathbb{E}_{s'\sim P^o_{s,a}} (\max_{a'} Q^{\pi^*}(s',a')-\max_{a'} Q_{K-1}(s',a'))_+|) + \sqrt{C}\|T Q_{K-1} - Q_{K}\|_{1,\mu} \nonumber \\
&\stackrel{(d)}{\leq} \gamma (\mathbb{E}_{s,a\sim\nu} \mathbb{E}_{s'\sim {P}^o_{s,a}} \max_{a'} |Q^{\pi^*}(s',a') - Q_{K-1}(s',a')| ) + \sqrt{C}\|T Q_{K-1} - Q_{K}\|_{1,\mu} \nonumber\\
&\stackrel{(e)}{\leq} \gamma \|Q^{\pi^*} - Q_{K-1}\|_{1,\nu'} + \sqrt{C}\|T Q_{K-1} - Q_{K}\|_{1,\mu} \nonumber \\
\label{eq:pf-sketech-stp4-1}
&\stackrel{(f)}{\leq} \gamma \|Q^{\pi^*} - Q_{K-1}\|_{1,\nu'} + \sqrt{C}\|T_{g_{K-1}} Q_{K-1} - Q_{K}\|_{2,\mu}+ \sqrt{C}\|T Q_{K-1} - T_{g_{K-1}} Q_{K-1}\|_{1,\mu},
\end{align}
where $(a)$ follows by the concentratability assumption (Assumption \ref{assum-concentra-condition}), $(b)$ from Bellman equation, operator $T$, and the fact $|\sup_{x} p(x)-\sup_{x} q(x)| \leq \sup_{x} |p(x) - q(x)|$,
$(c)$ from the fact $|(x)_+ - (y)_+| \leq |(x-y)_+|$ for any $x,y\in\mathbb{R}$,
$(d)$ follows by Jensen's inequality and by the facts $|\sup_{x} p(x)-\sup_{x} q(x)| \leq \sup_{x} |p(x) - q(x)|$ and $(x)_+ \leq |x|$ for any $x,y\in\mathbb{R}$, and $(e)$ by defining the distribution $\nu'$ as $\nu'(s',a')=\sum_{s,a} \nu(s,a) {P}^o_{s,a}(s') \mathbbm{1} \{ a'= \argmax_b |Q^{\pi^*}(s',b)- Q_{K-1}(s',b)| \}$, and $(f)$ using the fact that $\|\cdot\|_{1,\mu}\leq \|\cdot\|_{2,\mu}$.
Now, by recursion until iteration 0, we get \begin{align} \|&Q^{\pi^*} - Q_{K}\|_{1,\nu} \leq \gamma^K \sup_{\bar{\nu}} \|Q^{\pi^*} - Q_{0}\|_{1,\bar{\nu}} + \sqrt{C} \sum_{t=0}^{K-1} \gamma^t \|T Q_{K-1-t} - T_{g_{K-1-t}} Q_{K-1-t}\|_{1,\mu} \nonumber
\\&\hspace{4cm}+ \sqrt{C} \sum_{t=0}^{K-1} \gamma^t \| T_{g_{K-1-t}} Q_{K-1-t} - Q_{K-t}\|_{2,\mu}
\nonumber\\
&\stackrel{(a)}{\leq} \frac{\gamma^K }{1-\gamma} + \sqrt{C} \sum_{t=0}^{K-1} \gamma^t \|T Q_{K-1-t} - T_{g_{K-1-t}} Q_{K-1-t}\|_{1,\mu}
\nonumber\\&\hspace{4cm}+ \sqrt{C} \sum_{t=0}^{K-1} \gamma^t \| T_{g_{K-1-t}} Q_{K-1-t} - Q_{K-t}\|_{2,\mu}
\nonumber\\&\stackrel{(b)}{\leq} \frac{\gamma^K }{1-\gamma} + \frac{\sqrt{C} }{1-\gamma} \sup_{f\in\mathcal{F}} \|T f - T_{\widehat{g}_{f}} f\|_{1,\mu}
+ \frac{\sqrt{C} }{1-\gamma} \sup_{f\in\mathcal{F}} \| T_{\widehat{g}_{f}} f - \widehat{f}_{\widehat{g}_{f}}\|_{2,\mu}
\nonumber\\&\leq \frac{\gamma^K }{1-\gamma} + \frac{\sqrt{C} }{1-\gamma} \sup_{f\in\mathcal{F}} \|T f - T_{\widehat{g}_{f}} f\|_{1,\mu}
+ \frac{\sqrt{C} }{1-\gamma} \sup_{f\in\mathcal{F}} \sup_{g\in\mathcal{G}} \| T_{g} f - \widehat{f}_{g}\|_{2,\mu}. \label{eq:thm-bound-part-2} \end{align} where $(a)$ follows since $|Q^{\pi^*}(s,a)|\leq 1/(1-\gamma), Q_{0}(s,a)= 0$, and $(b)$ follows since $\widehat{g}_{f}$ is the dual variable function from the algorithm for the state-action value function $f$ and $\widehat{f}_{g}$ as the least squares solution from the algorithm for the state-action value function $f$ and dual variable function $g$ pair.
The proof is now complete combining \eqref{eq:thm-bound-part-1} and \eqref{eq:thm-bound-part-2} with Lemma \ref{lem:erm-high-prob-bound} and Lemma \ref{lem:least-squares-generalization-bound}.
\end{proof}
\section{Related Works}
\label{sec:app:related-works}
Here we provide a more detailed description of the related work to complement what we listed in the introduction (Section~\ref{sec:introduction}).
\textbf{Offline RL:} The problem of learning the optimal policy only using an offline dataset is first addressed under the generative model assumption \citep{singh1994upper, AzarMK13, haskell2016empirical, sidford2018near, agarwal2020model,li2020breaking, kalathil2021empirical}. This assumption requires generating the same uniform number of next-state samples for each and every state-action pairs.
To account for large state spaces, there are number of works \citep{antos2008learning, bertsekas2011approximate, lange2012batch, chen2019information, xie2020q, levine2020offline, xie2021bellman} that utilize function approximation under similar assumption, concentratability assumption \citep{chen2019information} in which the data distribution $\mu$ sufficiently covers the discounted state-action occupancy. There is rich literature \citep{munos08a, farahmand2010error, lazaric2012finite, chen2019information, liu2020provably, xie2021bellman} in the conquest of identifying and improving these necessary and sufficient assumptions for offline RL that use variations of Fitted Q-Iteration (FQI) algorithm \citep{gordon1995stable, ernst2005tree}.
There is also rich literature \citep{fujimoto2019off, kumar2019stabilizing, kumar2020conservative, yu2020mopo, zhang2021towards} that develop offline deep RL algorithms focusing on the algorithmic and empirical aspects and propose multitude heuristic approaches to advance the field. All these results assume that the offline data is generated according to a single model and the goal is to find the optimal policy for the MDP with the same model. In particular, none of these works consider the \textit{offline robust RL problem} where the offline data is generated according to a (training) model which can be different from the one in testing, and the goal is to learn a policy that is robust w.r.t.~an uncertainty set.
\textbf{Robust RL:} To address the parameter uncertainty problem,~\citet{iyengar2005robust} and~\citet{nilim2005robust} introduced the RMDP framework.~\citet{iyengar2005robust} showed that the optimal robust value function and policy can be computed using the robust counterparts of the standard value iteration and policy iteration algorithms. To tackle the parameter uncertainty problem, other works considered distributionally robust setting \citep{xu2010distributionally}, modified policy iteration \citep{kaufman2013robust}, and {more general uncertainty set} \citep{wiesemann2013robust}. These initial works mainly focused on the planning problem (known transition probability dynamics) in the tabular setting.~\citet{tamar2014scaling} proposed linear function approximation method to solve large RMDPs. Though this work suggests a sampling based approach, a general model-free learning algorithm and analysis was not included.~\citet{roy2017reinforcement} proposed the robust versions of the classical model-free reinforcement learning algorithms, such as Q-learning, SARSA, and TD-learning in the tabular setting. They also proposed function approximation based algorithms for the policy evaluation. However, this work does not have a policy iteration algorithm with provable guarantees for learning the optimal robust policy.~\citet{derman2018soft} introduced soft-robust actor-critic algorithms using neural networks, but does not provide any global convergence guarantees for the learned policy.~\citet{tessler2019action} proposed a min-max game framework to address the robust learning problem focusing on the tabular setting.~\citet{lim2019kernel} proposed a kernel-based RL algorithm for finding the robust value function in a batch learning setting.~\citet{Mankowitz2020Robust} employed an entropy-regularized policy optimization algorithm for continuous control using neural network, but does not provide any provable guarantees for the learned policy.~\citet{panaganti2020robust} proposed least-squares policy iteration method to handle large state-action space in robust RL, but only provide asymptotic policy evaluation convergence guarantees whereas \citet{panaganti2020robust} provide finite time convergence for the policy iteration to optimal robust value.
\textbf{Other robust RL related works:} Robust control is a well-studied area in the classical control theory~\citep{zhou1996robust, dullerud2013course}. Recently, there are some interesting works that address the robust RL problem using this framework, especially focusing on the linear quadratic regulator setting \citep{zhang2020policy}. Risk sensitive RL algorithms \citep{borkar2002q,prashanth2016variance,fei2021exponential} and adversarial RL algorithms \citep{pinto2017robust,zhang2021provably,huang2022robust} also address the robustness problem implicitly under different frameworks which are independent from RMDPs. \citep{zhang2022corruption} addresses the problem of \textit{corruption-robust} offline RL problem, where an adversary is allowed to change a fraction of the samples of an offline RL dataset and the goal is to find the optimal policy for the nominal linear MDP model (according to which the offline data is generated).
Our framework and approach of robust MDP is significantly different from these line of works.
\textbf{This work:} The works that are closest to ours are by \citet{zhou2021finite,yang2021towards,panaganti22a} that address the robust RL problem in a tabular setting under the generative model assumption. Due to the generative model assumption, the offline data has the same uniform number of samples corresponding to each and every state-action pair, and tabular setting allows the estimation of the uncertainty set followed by solving the planning problem. Our work is significantly different from these in the following way: $(i)$ we consider a robust RL problem with arbitrary large state space, instead of the small tabular setting, $(ii)$ we consider a true offline RL setting where the state-action pairs are sampled according to an arbitrary distribution, instead of using the generative model assumption, $(iii)$ we focus on a function approximation approach where the goal is to directly learn optimal robust value/policy using function approximation techniques, instead of solving the tabular planning problem with the estimated model. \textit{To the best of our knowledge, this is the first work that addresses the offline robust RL problem with arbitrary large state space using function approximation, with provable guarantees on the performance of the learned policy.}
\section{Experiment Details}
\label{appendix:simulations}
We provide more detailed and practical version of our RFQI algorithm (Algorithm \ref{alg:RFQI-Algorithm}) in this section. We also provide more experimental results evaluated on \textit{Cartpole}, \textit{Hopper}, and \textit{Half-Cheetah} OpenAI Gym Mujoco \citep{brockman2016openai} environments.
We provide our code in \textbf{github webpage} \url{https://github.com/zaiyan-x/RFQI} containing instructions to reproduce all results in this paper.
We implemented our RFQI algorithm based on the architecture of Batch Constrained deep Q-learning (BCQ) algorithm \citep{fujimoto2019off} \footnote{Available at \url{https://github.com/sfujim/BCQ}} and Pessimistic Q-learning (PQL) algorithm \citep{liu2020provably} \footnote{Available at \url{https://github.com/yaoliucs/PQL}}. We note that PQL algorithm (with $b=0$ filtration thresholding \citep{liu2020provably}) and BCQ algorithm are the practical versions of FQI algorithm with neural network architecture.
\subsection{RFQI Practical Algorithm} \label{appendix-sub:rfqi-prac-implementation}
We provide the practical version of our RFQI algorithm in Algorithm \ref{alg:prac-RFQI-Algorithm} and highlight the difference with BCQ and PQL algorithms in \bluetext{blue} (steps 8 and 9).
\begin{algorithm}[t]
\caption{RFQI Practical Algorithm}
\label{alg:prac-RFQI-Algorithm}
\begin{algorithmic}[1]
\STATE \textbf{Input:} Offline dataset $\mathcal{D}$, radius of robustness $\rho$, maximum perturbation $\Phi$, target update rate $\tau$, mini-batch size $N$, maximum number of iterations $K$, number of actions $u$.
\STATE \textbf{Initialize:} Two state-action neural networks $Q_{\theta_1}$ and $Q_{\theta_2}$, \bluetext{one dual neural network $g_{\theta_3}$} \\policy (perturbation) model: $\xi_{\phi} \in [-\Phi, \Phi]),$ and action VAE $G^a_{\omega}$, with random parameters $\theta_1$, $\theta_2$, $\phi$, $\omega$, and target networks $Q_{\theta'_1}, Q_{\theta'_2}$, $\xi_{\phi'}$ with $\theta'_1 \leftarrow \theta_1, \theta'_2 \leftarrow \theta_2$, $\phi' \leftarrow \phi$
\FOR {$k=1,\cdots,K$ }
\STATE Sample a minibatch $B$ with $N$ samples from $\mathcal{D}$.
\STATE Train $\omega \leftarrow\argmin_{\omega} ELBO(B;G^a_{\omega}).$ Sample $u$ actions $a_i'$ from $G^a_{\omega} (s')$ for each $s'$.
\STATE Perturb $u$ actions $a_i'=a_i' + \xi_{\phi}(s',a_i')$.
\STATE Compute next-state value target for each $s'$ in $B$: $$V_t = \max_{a'_i} (0.75\cdot \min\{ Q_{\theta'_1}, Q_{\theta'_2} \} + 0.25\cdot \max\{ Q_{\theta'_1}, Q_{\theta'_2} \}).$$
\STATE \vspace{-0.3cm} \bluetext{$\theta_3 \leftarrow \argmin_{\theta} \sum [\max\{g_{\theta}(s,a) - V_t(s'), 0\} - (1-\rho) g_{\theta}(s,a)].$}
\STATE \bluetext{Compute next-state Q target for each $(s, a, r, s' )$ pair in $B$: $$ Q_t(s,a) = r - \gamma \cdot \max\{g_{\theta_3}(s,a) - V_t(s'), 0\} + \gamma (1-\rho) g_{\theta_3}(s,a). $$}
\STATE \vspace{-0.3cm} $\theta \leftarrow \argmin_{\theta} \sum (Q_t(s,a) - Q_{\theta}(s,a))^2.$
\STATE Sample $u$ actions $a_i$ from $G^a_{\omega} (s)$ for each $s$.
\STATE $\phi \leftarrow \argmax_{\phi} \sum \max_{a_i} Q_{\theta_1}(s,a_i+\xi_{\phi}(s,a_i))$.
\STATE Update target network: $\theta'=(1-\tau)\theta' + \tau\theta, \phi'=(1-\tau)\phi' + \tau\phi$.
\ENDFOR
\STATE \textbf{Output policy:} Given $s$, sample $u$ actions $a_i$ from $G^a_{\omega} (s)$. Select action $a = \argmax_{a_i} Q_{\theta_1}(s,a_i+\xi_{\phi}(s,a_i))$.
\end{algorithmic}
\end{algorithm}
\textbf{RFQI algorithm implementation details}: The Variational Auto-Encoder (VAE) $G^a_\omega$ \citep{kingma2013auto} is defined by two networks, an encoder $E_{\omega_1}(s,a)$ and decoder $D_{\omega_2}(s, z)$, where $\omega = \{\omega_1, \omega_2\}$. The encoder outputs mean and standard deviation, $(\mu, \sigma) = E_{\omega_1}(s,a)$, of a normal distribution.
A latent vector $z$ is sampled from the standard normal distribution and for a state $s$, the decoder maps them to an action $D_{\omega_2}:(s, z)\mapsto\tilde{a}$. Then the evidence lower bound ($ELBO$) of VAE is given by $ELBO(B;G^a_{\omega})=\sum_{B} (a - \tilde{a})^2 + D_{\text{KL}}(\mathcal{N}(\mu, \sigma),\mathcal{N}(0,1)),$ where $\mathcal{N}$ is the normal distribution with mean and standard deviation parameters. We refer to \citep{fujimoto2019off} for more details on VAE. We also use the default VAE architecture from BCQ algorithm \citep{fujimoto2019off} and PQL algorithm \citep{liu2020provably} in our RFQI algorithm.
We now focus on the additions described in blue (steps 8 and 9) in Algorithm \ref{alg:prac-RFQI-Algorithm}. For all the other networks we use default architecture from BCQ algorithm \citep{fujimoto2019off} and PQL algorithm \citep{liu2020provably} in our RFQI algorithm.\\
(1) In each iteration $k$, we solve the dual variable function $g_\theta$ optimization problem (step 4 in Algorithm \ref{alg:RFQI-Algorithm}, step 8 in Algorithm \ref{alg:prac-RFQI-Algorithm}) implemented by ADAM \citep{kingma2014adam} on the minibatch $B$ with the learning rate $l_1$ mentioned in Table \ref{tab:hp}.\\
(2) Our state-action value target function corresponds to the robust state-action value target function described in \eqref{eq:loss-rfqi-1}. This is reflected in step 9 of Algorithm \ref{alg:prac-RFQI-Algorithm}. The state-action value function $Q_\theta$ optimization problem (step 5 in Algorithm \ref{alg:RFQI-Algorithm}, step 9 in Algorithm \ref{alg:prac-RFQI-Algorithm}) is implemented by ADAM \citep{kingma2014adam} on the minibatch $B$ with the learning rate $l_2$ mentioned in Table \ref{tab:hp}.
\begin{table}[H]
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
Environment & Discount & Learning rates & Q Neural nets & Dual Neural nets \\
& $\gamma$ & $[l_1,l_2]$ & $\theta_1=\theta_2=[h_1,h_2]$ & $\theta_3=[h_1,h_2]$ \\
\hline & &&& \vspace{-0.3cm}\\
CartPole & $0.99$ & $[10^{-3},10^{-3}]$ & $[400, 300]$ & $[64, 64]$\\
Hopper & $0.99$ & $[10^{-3},8\times 10^{-4}]$ & $[400, 300]$ & $[64, 64]$\\
& & $[3\times 10^{-4},6\times10^{-4}]$ & & \\
Half-Cheetah & $0.99$ & $[10^{-3},8\times 10^{-4}]$ & $[400, 300]$ & $[64, 64]$\\
& & $[3\times 10^{-4},6\times10^{-4}]$ & & \\
\hline
\end{tabular}
\end{center}\caption{Details of hyper-parameters in FQI and RFQI algorithms experiments.} \label{tab:hp}
\end{table}
\textbf{Hyper-parameters details}: We now give the description of hyper-parameters used in our codebase in Table \ref{tab:hp}. We use same hyper-parameters across different algorithms. Across all learning algorithms we use $\tau=0.005$ for the target network update, $K=5\times10^5$ for the maximum iterations, $|\mathcal{D}|=10^6$ for the offline dataset, $|B|=1000$ for the minibatch size. We used grid-search for $\rho$ in $\{0.2,0.3,\cdots,0.6\}$. We also picked best of the two sets of learning rates mentioned in Table \ref{tab:hp}. For all the other hyper-parameters we use default values from BCQ algorithm \citep{fujimoto2019off} and PQL algorithm \citep{liu2020provably} in our RFQI algorithm that can be found in our code.
\textbf{Offline datasets}: Now we discuss the offline dataset used in the our training of FQI and RFQI algorithms. For the fair comparison in every plot, we train both FQI and RFQI algorithms on same offline datasets.
\textit{Cartpole dataset} $\mathcal{D}_{\textrm{c}}$: We first train proximal policy optimization (PPO) \citep{schulman2017proximal} algorithm, under default RL baseline zoo \citep{rl-zoo3} parameters. We then generate the Cartpole dataset $\mathcal{D}_{\textrm{c}}$ with $10^5$ samples using an $\epsilon$-greedy ($\epsilon=0.3$) version of this PPO trained policy. We note that this offline dataset contains non-expert behavior meeting the richness of the data-generating distribution assumption in practice.
\textit{Mixed dataset} $\mathcal{D}_{\textrm{m}}$: For the MuJoCo environments, \textit{Hopper} and \textit{Half-Cheetah}, we increase the richness of the dataset since these are high dimensional problems. We first train soft actor-critic (SAC) \citep{haarnoja2018soft} algorithm, under default RL baseline zoo \citep{rl-zoo3} parameters, with replay buffer updated by a fixed $\epsilon$-greedy ($\epsilon=0.1$) policy with the model parameter \textit{actuator\_ctrlrange} set to $[-0.85,0.85]$. We then generate the mixed dataset $\mathcal{D}_{\textrm{m}}$ with $10^6$ samples from this $\epsilon$-greedy ($\epsilon=0.3$) SAC trained policy. We note that such a dataset generation gives more diverse set of observations than the process of $\mathcal{D}_{\textrm{c}}$ generation for fair comparison between FQI and RFQI algorithms.
\textit{D4RL dataset} $\mathcal{D}_{\textrm{d}}$: We consider the \textit{hopper-medium} and \textit{halfcheetah-medium} offline datasets in \citep{fu2020d4rl} which are benchmark datasets in offline RL literature \citep{fu2020d4rl,levine2020offline,liu2020provably}. These `medium' datasets are generated by first training a policy online using Soft Actor-Critic \citep{haarnoja2018soft}, early-stopping the training, and collecting $10^6$ samples from this partially-trained policy. We refer to \citep{fu2020d4rl} for more details.
We end this section by mentioning the software and hardware configurations used. The training and evaluation is done using three computers with the following configuration. Operating system is Ubuntu 18.04 and Lambda Stack; main softwares are PyTorch, Caffe, CUDA, cuDNN, Numpy, Matplotlib; processor is AMD Threadripper 3960X (24 Cores, 3.80 GHz); GPUs are 2x RTX 2080 Ti; memory is 128GB RAM; Operating System Drive is 1 TB SSD (NVMe); and Data Drive is 4TB HDD.
\subsection{More Experimental Results}
\begin{figure*}[ht]
\vspace{-0.5cm}
\centering
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/CartPolePerturbed-v0_action.pdf}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/CartPolePerturbed-v0_force_mag.pdf}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/CartPolePerturbed-v0_length.pdf}
\end{minipage}
\captionof{figure}{ \textit{Cartpole simulation results on offline dataset $\mathcal{D}_{\textrm{c}}$.} Average cumulative reward in $20$ episodes versus different model parameter perturbations mentioned in the respective titles.}
\label{fig:cartpole}
\centering
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/hopper_action.pdf}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/hopper_actuator_ctrlrange.pdf}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/hopper_foot_joint_stiffness.pdf}
\end{minipage}
\captionof{figure}{ \textit{Hopper simulation results on offline dataset $\mathcal{D}_{\textrm{m}}$.} Average cumulative reward in $20$ episodes versus different model parameter perturbations mentioned in the respective titles.}
\label{fig:hopper-mixed}
\end{figure*}
Here we provide more experimental results and details in addition to Fig. \ref{fig:cart_force-in}-\ref{fig:hopper_leg} in Section \ref{sec:main-experiments}.
For the \textit{Cartpole}, we compare RFQI algorithm against the non-robust RL algorithms FQI and DQN, and the soft-robust RL algorithm proposed in \cite{derman2018soft}. We trained FQI and RFQI algorithms on the dataset $\mathcal{D}_{\textrm{c}}$ (a detailed description of data set is provided in Appendix \ref{appendix-sub:rfqi-prac-implementation}).
We test the robustness of the algorithms by changing the parameters \textit{force\_mag} (to model external force disturbance), \textit{length} (to model change in pole length), and also by introducing action perturbations (to model actuator noise). The nominal value of \textit{force\_mag} and \textit{length} parameters are $10$ and $0.5$ respectively. Fig. \ref{fig:cartpole} shows superior robust performance of RFQI compared to the non-robust FQI and DQN.
For example, consider the action perturbation performance plot in Fig. \ref{fig:cartpole} where RFQI algorithm improves by $75\%$ compared to FQI algorithm in average cumulative reward for a $40\%$ chance of action perturbation.
We note that we found $\rho=0.5$ is the best from grid-search for RFQI algorithm. The RFQI performance is similar to that of soft-robust DQN. We note that soft-robust DQN algorithm is an online deep RL algorithm (and not an offline RL algorithm) and has no provable performance guarantee. Moreover, soft-robust DQN algorithm requires generating online data according a number of models in the uncertainty set, whereas RFQI only requires offline data according to a single nominal training model.
Before we proceed to describe our results on the OpenAI Gym MuJoCo \citep{brockman2016openai} environments \textit{Hopper} and \textit{Half-Cheetah}, we first mention their model parameters and its corresponding nominal values in Table \ref{tab:model-parameters}. The model parameter names are self-explanatory, for example, stiffness control on the leg joint is the \textit{leg\_joint\_stiffness}, range of actuator values is the \textit{actuator\_ctrlrange}. The front and back parameters in Half-Cheetah are for the front and back legs. We refer to the perturbed environments provided in our code and the \textit{hopper.xml, halfcheetah.xml} files in the environment assets of OpenAI Gym MuJoCo \citep{brockman2016openai} for more information regarding these model parameters.
\begin{table}[H]
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
Environment & Model parameter & Nominal range/value \\
\hline & & \vspace{-0.3cm}\\
Hopper & \textit{actuator\_ctrlrange} & $[-1,1]$ \\
& \textit{foot\_joint\_stiffness} & $0$ \\
& \textit{leg\_joint\_stiffness} & $0$ \\
& \textit{thigh\_joint\_stiffness} & $0$ \\
& \textit{joint\_damping} & $1$ \\
& \textit{joint\_frictionloss} & $0$ \\
\hline & & \vspace{-0.3cm}\\
& \textit{joint\_frictionloss} & $0$ \\
Half-Cheetah & front \textit{actuator\_ctrlrange} & $[-1,1]$ \\
& back \textit{actuator\_ctrlrange} & $[-1,1]$ \\
& front \textit{joint\_stiffness} = (\textit{thigh\_joint\_stiffness}, & \\
& \textit{shin\_joint\_stiffness}, \textit{foot\_joint\_stiffness}) & $(180,120,60)$ \\
& back \textit{joint\_stiffness} = (\textit{thigh\_joint\_stiffness}, & \\
& \textit{shin\_joint\_stiffness}, \textit{foot\_joint\_stiffness}) & $(240,180,120)$ \\
& front \textit{joint\_damping} = (\textit{thigh\_joint\_damping}, & \\
& \textit{shin\_joint\_damping}, \textit{foot\_joint\_damping}) & $(4.5,3.0,1.5)$ \\
& back \textit{joint\_damping} = (\textit{thigh\_joint\_damping}, & \\
& \textit{shin\_joint\_damping}, \textit{foot\_joint\_damping}) & $(6.0,4.5,3.0)$ \\
\hline
\end{tabular}
\end{center}\caption{Details of model parameters for \textit{Hopper} and \textit{Half-Cheetah} environments.} \label{tab:model-parameters}
\end{table}
\vspace{-0.5cm}
For the \textit{Hopper}, we compare RFQI algorithm against the non-robust RL algorithms FQI and TD3 \citep{fujimoto2018addressing}. We trained FQI and RFQI algorithms on the mixed dataset $\mathcal{D}_{\textrm{m}}$ (a detailed description of dataset provided in Appendix \ref{appendix-sub:rfqi-prac-implementation}). We note that we do not compare with soft robust RL algorithms because of its poor performance on MuJoCo environments in the rest of our figures. We test the robustness of the algorithm by introducing action perturbations, and by changing the model parameters \textit{actuator\_ctrlrange}, \textit{foot\_joint\_stiffness}, and \textit{leg\_joint\_stiffness}.
Fig. \ref{fig:hopper_leg} and Fig. \ref{fig:hopper-mixed} shows RFQI algorithm is consistently robust compared to the non-robust algorithms. We note that we found $\rho=0.5$ is the best from grid-search for RFQI algorithm. The average episodic reward of RFQI remains almost the same initially, and later decays much less and gracefully when compared to FQI and TD3 algorithms. For example, in plot 3 in Fig. \ref{fig:hopper-mixed}, at the \textit{foot\_joint\_stiffness} parameter value $15$, the episodic reward of FQI is only around $1400$ whereas RFQI achieves an episodic reward of $3200$. Similar robust performance of RFQI can be seen in other plots as well. We also note that TD3 \citep{fujimoto2019off} is a powerful off-policy policy gradient algorithm that relies on large $10^6$ replay buffer of online data collection, unsurprisingly performs well initially with less perturbation near the nominal models.
In order to verify the effectiveness and consistency of our algorithm across different offline dataset, we repeat the above experiments, on additional OpenAI Gym MuJoCo \citep{brockman2016openai} environment \textit{Half-Cheetah}, using D4RL dataset $\mathcal{D}_{\mathrm{d}}$ (a detailed description of dataset provided in Appendix \ref{appendix-sub:rfqi-prac-implementation}) which are benchmark in offline RL literature \citep{fu2020d4rl,levine2020offline,liu2020provably} than our mixed dataset $\mathcal{D}_{\mathrm{m}}$. Since D4RL dataset is a benchmark dataset for offline RL algorithms, here we focus only on the comparison between the two offline RL algorithms we consider, our RFQI algorithm and its non-robust counterpart FQI algorithm. We now showcase the results on \textit{Hopper} and \textit{Half-Cheetah} for this setting.
\begin{figure*}[t]
\centering
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/hopper_gravity.pdf}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/hopper_joint_damping.pdf}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/hopper_joint_frictionloss.pdf}
\end{minipage}
\captionof{figure}{ \textit{Hopper evaluation simulation results on offline dataset $\mathcal{D}_{\textrm{d}}$.} Average cumulative reward in $20$ episodes versus different model parameter perturbations mentioned in the respective titles.}
\label{fig:hopper-d4rl}
\centering
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/halfcheetah_front_joint_stiffness.pdf}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/halfcheetah_back_joint_stiffness.pdf}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/halfcheetah_back_actuator_ctrlrange.pdf}
\end{minipage}
\captionof{figure}{ \textit{Half-Cheetah evaluation simulation results on offline dataset $\mathcal{D}_{\textrm{d}}$.} Average cumulative reward in $20$ episodes versus different model parameter perturbations mentioned in the respective titles.}
\label{fig:half-cheetah-d4rl}
\end{figure*}
For the \textit{Hopper}, we test the robustness by changing the model parameters \textit{gravity}, \textit{joint\_damping}, and \textit{joint\_frictionloss}.
Fig. \ref{fig:hopper-d4rl} shows
RFQI algorithm is consistently robust compared to the non-robust FQI algorithm. We note that we found $\rho=0.5$ is the best from grid-search for RFQI algorithm.
The average episodic reward of RFQI remains almost the same initially, and later decays much less and gracefully when compared to FQI algorithm. For example, in plot 2 in Fig. \ref{fig:hopper-d4rl}, for the $30\%$ change in \textit{joint\_damping} parameter, the episodic reward of FQI is only around $1400$ whereas RFQI achieves an episodic reward of $3000$ which is almost the same as for unperturbed model. Similar robust performance of RFQI can be seen in other plots as well.
For the \textit{Half-Cheetah}, we test the robustness by changing the model parameters \textit{joint\_stiffness} of front and back joints, and \textit{actuator\_ctrlrange} of back joint.
Fig. \ref{fig:half-cheetah-d4rl} shows
RFQI algorithm is consistently robust compared to the non-robust FQI algorithm. We note that we found $\rho=0.3$ is the best from grid-search for RFQI algorithm.
For example, in plot 1 in Fig. \ref{fig:half-cheetah-d4rl}, RFQI episodic reward stays at around $5500$ whereas FQI drops faster to $4300$ for more than $50\%$ change in the nominal value. Similar robust performance of RFQI can be seen in other plots as well.
\begin{figure*}[h]
\begin{minipage}{.47\textwidth}
\begin{tabular}{@{}c@{}}
\includegraphics[width=0.5\linewidth]{figures/hopper-d4rl-actuator-ctrlrange.pdf}
\includegraphics[width=0.5\linewidth]{figures/hopper-d4rl-thigh-joint-stiffness.pdf}
\end{tabular}
\captionof{figure}{Similar performance of RFQI and FQI in Hopper on dataset $\mathcal{D}_{\textrm{d}}$ w.r.t. parameters \textit{actuator\_ctrlrange} and \textit{thigh\_joint\_stiffness}.}
\label{fig:hopper-d4rl-same-robustness}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}{.47\textwidth}
\begin{tabular}{@{}c@{}}
\includegraphics[width=0.5\linewidth]{figures/halfcheetah_joint_damping.pdf}
\includegraphics[width=0.5\linewidth]{figures/halfcheetah_joint_frictionloss.pdf}
\end{tabular}
\captionof{figure}{Similar performance of RFQI and FQI in Half-Cheetah on dataset $\mathcal{D}_{\textrm{d}}$ w.r.t. parameters \textit{joint\_damping} and \textit{joint\_frictionloss}.}
\label{fig:half-cheetah-d4rl-same-robustness}
\end{minipage}
\end{figure*}
As part of discussing the limitations of our work, we also provide two instances where RFQI and FQI algorithm behave similarly. RFQI and FQI algorithms trained on the D4RL dataset $\mathcal{D}_{\textrm{d}}$ perform similarly under the perturbations of the \textit{Hopper} model parameters \textit{actuator\_ctrlrange} and \textit{thigh\_joint\_stiffness} as shown in Fig. \ref{fig:hopper-d4rl-same-robustness}. We also make similar observations under the perturbations of the \textit{Half-Cheetah} model parameters \textit{joint\_damping} (both front \textit{joint\_damping} and back \textit{joint\_damping}) and \textit{joint\_frictionloss} as shown in Fig. \ref{fig:half-cheetah-d4rl-same-robustness}. We observed that the robustness performance can depend on the offline data available, which was also observed for non-robust offline RL algorithms \citep{liu2020provably,fu2020d4rl,levine2020offline}. Also, perturbing some parameters may make the problem really hard especially if the data is not representative with respect to that parameter. We believe that this is the reason for the similar performance of RFQI and FQI w.r.t. some parameters. We believe that this opens up an exciting area of research on developing online policy gradient algorithms for robust RL, which may be able to overcome the restriction and challenges due to offline data. We plan to pursue this goal in our future work.
\section{Introduction}
\label{sec:introduction}
Reinforcement learning (RL) algorithms often require a large number of data samples to learn a control policy. As a result, training them directly on the real-world systems is expensive and potentially dangerous. This problem is typically overcome by training them on a simulator (online RL) or using a pre-collected offline dataset (offline RL). The offline dataset is usually collected either from a sophisticated simulator of the real-world system or from the historical measurements. The trained RL policy is then deployed assuming that the training environment, the simulator or the offline data, faithfully represents the model of the real-world system. This assumption is often incorrect due to multiple factors such as the approximation errors incurred while modeling, changes in the real-world parameters over time and possible adversarial disturbances in the real-world. For example, the standard simulator settings of the sensor noise, action delay, friction, and mass of a mobile robot can be different from that of the actual real-world robot, in addition to changes in the terrain, weather conditions, lighting, and obstacle densities of the testing environment.
Unfortunately, the current RL control policies can fail dramatically when faced with even mild changes in the training and testing environments \citep{sunderhauf2018limits, tobin2017domain, peng2018sim}.
The goal in robust RL is to learn a policy that is robust against the model parameter mismatches between the training and testing environments. The robust planning problem is formalized using the framework of Robust Markov Decision Process (RMDP) \citep{iyengar2005robust,nilim2005robust}. Unlike the standard MDP which considers a single model (transition probability function), the RMDP formulation considers a set of models which is called the \textit{uncertainty set}. The goal is to find an optimal robust policy that performs the best under the worst possible model in this uncertainty set.
The minimization over the uncertainty set makes the robust MDP and robust RL problems significantly more challenging than their non-robust counterparts.
In this work, we study the problem of developing a robust RL algorithm with provably optimal performance for an RMDP with arbitrarily large state spaces, using only offline data with function approximation. Before stating the contributions of our work, we provide a brief overview of the results in offline and robust RL that are directly related to ours. We leave a more thorough discussion on related works to Appendix~\ref{sec:app:related-works}.
\looseness=-1\textbf{Offline RL:} Offline RL considers the problem of learning the optimal policy only using a pre-collected (offline) dataset. Offline RL problem has been addressed extensively in the literature \citep{antos2008learning, bertsekas2011approximate, lange2012batch, chen2019information, xie2020q, levine2020offline, xie2021bellman}. Many recent works develop deep RL algorithms and heuristics for the offline RL problem, focusing on the algorithmic and empirical aspects \citep{fujimoto2019off, kumar2019stabilizing, kumar2020conservative, yu2020mopo, zhang2021towards}. A number of theoretical work focus on analyzing the variations of Fitted Q-Iteration (FQI) algorithm \citep{gordon1995stable, ernst2005tree}, by identifying the necessary and sufficient conditions for the learned policy to be approximately optimal and characterizing the performance in terms of sample complexity \citep{munos08a, farahmand2010error, lazaric2012finite, chen2019information, liu2020provably, xie2021bellman}. All these works assume that the offline data is generated according to a single model and the goal is to find the optimal policy for the MDP with the same model. In particular, none of these works consider the \textit{offline robust RL problem} where the offline data is generated according to a (training) model which can be different from the one in testing, and the goal is to learn a policy that is robust w.r.t.~an uncertainty set.
\textbf{Robust RL:} The RMDP framework was first introduced in~\citet{iyengar2005robust, nilim2005robust}. The RMDP problem has been analyzed extensively in the literature \citep{xu2010distributionally, wiesemann2013robust, yu2015distributionally, mannor2016robust, russel2019beyond} providing computationally efficient algorithms, but these works are limited to the {planning} problem. Robust RL algorithms with provable guarantees have also been proposed \citep{lim2013reinforcement,tamar2014scaling,roy2017reinforcement,panaganti2020robust, wang2021online}, but they are limited to tabular or linear function approximation settings and only provide asymptotic convergence guarantees. Robust RL problem has also been addressed using deep RL methods \citep{pinto2017robust, derman2018soft, derman2020bayesian, Mankowitz2020Robust, zhang2020robust}. However, these works do not provide any theoretical guarantees on the performance of the learned policies.
The works that are closest to ours are by \citet{zhou2021finite,yang2021towards,panaganti22a} that address the robust RL problem in a tabular setting under the generative model assumption. Due to the generative model assumption, the offline data has the same uniform number of samples corresponding to each and every state-action pair, and tabular setting allows the estimation of the uncertainty set followed by solving the planning problem. Our work is significantly different from these in the following way: $(i)$ we consider a robust RL problem with arbitrary large state space, instead of the small tabular setting, $(ii)$ we consider a true offline RL setting where the state-action pairs are sampled according to an arbitrary distribution, instead of using the generative model assumption, $(iii)$ we focus on a function approximation approach where the goal is to directly learn optimal robust value/policy using function approximation techniques, instead of solving the tabular planning problem with the estimated model. \textit{To the best of our knowledge, this is the first work that addresses the offline robust RL problem with arbitrary large state space using function approximation, with provable guarantees on the performance of the learned policy.}
\textbf{Offline Robust RL: Challenges and Our Contributions:} Offline robust RL is significantly more challenging than its non-robust counterpart mainly because of the following key difficulties. \\
$(i)$ Data generation: The optimal robust policy is computed by taking the infimum over all models in the uncertainty set $\mathcal{P}$. However, generating data according to all models in $\mathcal{P}$ is clearly infeasible. It may only be possible to get the data from a nominal (training) model $P^{o}$. \textit{How do we use the data from a nominal model to account for the behavior of all the models in the uncertainty set $\mathcal{P}$?} \\
$(ii)$ Optimization over the uncertainty set $\mathcal{P}$: The robust Bellman operator (defined in \eqref{eq:robust-bellman-primal}) involves a minimization over $\mathcal{P}$, which is a significant computational challenge. Moreover, the uncertainty set $\mathcal{P}$ itself is unknown in the RL setting. \textit{How do we solve the optimization over $\mathcal{P}$?} \\
$(iii)$ Function approximation: Approximation of the robust Bellman update requires a modified target function which also depends on the approximate solution of the optimization over the uncertainty set. \textit{How do we perform the offline RL update accounting for both approximations? }
\looseness=-1 As the \textit{key technical contributions} of this work, we first derive a dual reformulation of the robust Bellman operator which replaces the expectation w.r.t.~all models in the uncertainty set $\mathcal{P}$ with an expectation only w.r.t. the nominal (training) model $P^o$. This enables using the offline data generated by $P^o$ for learning, without relying on high variance importance sampling techniques to account for all models in $\mathcal{P}$. Following the same reformulation, we then show that the optimization problem over $\mathcal{P}$ can be further reformulated as functional optimization. We solve this functional optimization problem using empirical risk minimization and obtain performance guarantees using the Rademacher complexity based bounds. We then use the approximate solution obtained from the empirical risk minimization to generate modified target samples that are then used to approximate robust Bellman update through a generalized least squares approach with provably bounded errors. Performing these operations iteratively results in our proposed Robust Fitted Q-Iteration (RFQI) algorithm, for which we prove that its learned policy achieves non-asymptotic and approximately optimal performance guarantees.
\textbf{Notations:} For a set $\mathcal{X}$, we denote its cardinality as $|\mathcal{X}|$. The set of probability distribution over $\mathcal{X}$ is denoted as $\Delta(\mathcal{X})$, and its power set sigma algebra as $\Sigma(\mathcal{X})$. For any $x \in \mathbb{R}$, we denote $\max\{x, 0\}$ as $(x)_{+}$.
For any function $f:\cS\times\cA\rightarrow \mathbb{R}$, state-action distribution $\nu\in\Delta({\cS\times\cA})$, and real number $p\geq 1$, the $\nu$-weighted $p$-norm of $f$ is defined as $\|f\|_{p,\nu} = \mathbb{E}_{s,a\sim\nu}[|f(s,a)|^p]^{1/p}$.
\section{Preliminaries}
\label{sec:preliminaries}
A Markov Decision Process (MDP) is a tuple $(\cS, \mathcal{A}, r, P, \gamma, d_0)$, where $\cS$ is the state space, $\mathcal{A}$ is the action space, $r: \cS\times \mathcal{A}\rightarrow \mathbb{R}$ is the reward function, $\gamma \in (0, 1)$ is the discount factor, and $d_0\in \Delta({\cS})$ is the initial state distribution. The transition probability function $P_{s,a}(s')$ is the probability of transitioning to state $s'$ when action $a$ is taken at state $s$. In the literature, $P$ is also called the \textit{model} of the MDP. We consider a setting where $|\cS|$ and $|\mathcal{A}|$ are finite but can be arbitrarily large. We will also assume that $r(s,a) \in [0,1]$, for all $(s,a) \in \cS\times\cA$, without loss of generality. A policy $\pi : \cS \rightarrow \Delta({\mathcal{A}})$ is a conditional distribution over actions given a state. The value function $V_{\pi, P}$ and the state-action value function $Q_{\pi,P}$ of a policy $\pi$ for an MDP with model $P$ are defined as
\vspace{-0.1cm}
\begin{align*}
V_{\pi, P}(s) = \mathbb{E}_{\pi, P}[\sum^{\infty}_{t=0} \gamma^{t} r(s_{t}, a_{t}) ~|~ s_{0} = s], \quad Q_{\pi, P}(s, a) = \mathbb{E}_{\pi, P}[\sum^{\infty}_{t=0} \gamma^{t} r(s_{t}, a_{t}) ~|~ s_{0} = s, a_{0} = a],
\end{align*}
where the expectation is over the randomness induced by the policy $\pi$ and model $P$.
The optimal value function $V^{*}_{P}$ and the optimal policy $\pi^{*}_{P}$ of an MDP with the model $P$ are defined as $V^{*}_{P} = \max_{\pi} V_{\pi, P}$ and $\pi^{*}_{P} = \argmax_{\pi} V_{\pi, P}$.
The optimal state-action value function is given by $Q^{*}_{P} = \max_{\pi} Q_{\pi, P}$. The optimal policy can be obtained as $\pi^{*}_{P}(s) = \argmax_{a} Q_P^{*}(s, a)$. The discounted state-action occupancy of a policy $\pi$ for an MDP with model $P$, denoted as $d_{\pi, P} \in \Delta({\cS\times\cA})$, is defined as $d_{\pi,P}(s,a)=(1-\gamma) \mathbb{E}_{\pi,P}[\sum_{t=0}^\infty \gamma^{t} \mathbbm{1}(s_t=s,a_t=a)]$.
\textbf{Robust Markov Decision Process (RMDP):} Unlike the standard MDP which considers a single model (transition probability function), the RMDP formulation considers a set of models. We refer to this set as the \textit{uncertainty set} and denote it as $\mathcal{P}$. We consider $\mathcal{P}$ that satisfies the standard \textit{$(s,a)$-rectangularity condition} \citep{iyengar2005robust}. We note that a similar {uncertainty set} can be considered for the reward function at the expense of additional notations. However, since the analysis will be similar and the sample complexity guarantee will be identical up to a constant factor, without loss of generality, we assume that the reward function is known and deterministic.
We specify an RMDP as $M = (\cS, \mathcal{A}, r, \mathcal{P}, \gamma, d_0)$, where the uncertainty set $\mathcal{P}$ is typically defined as
\begin{align}
\label{eq:uncertainty-set}
\mathcal{P} = \otimes_{(s,a) \in \cS\times\cA }\, \mathcal{P}_{s,a}, \quad \text{where} \;\;
\mathcal{P}_{s,a} = \{ P_{s,a} \in \Delta(\cS)~:~ D(P_{s,a}, P^o_{s,a}) \leq \rho \},
\end{align}
$P^{o} = (P^o_{s,a}, (s, a) \in \cS\times\cA)$ is the \textit{nominal model}, $D(\cdot, \cdot)$ is a distance metric between two probability distributions, and $\rho>0$ is the radius of the uncertainty set that indicates the level of robustness. The nominal model $P^{o}$ can be thought as the model of the training environment. It is either the model of the simulator on which the (online) RL algorithm is trained, or in our setting, it is the model according to which the offline data is generated. The uncertainty set $\mathcal{P}$ \eqref{eq:uncertainty-set} is the set of all valid transition probability functions (valid testing models) in the neighborhood of the nominal model $P^{o}$, which by definition satisfies $(s,a)$-rectangularity condition \citep{iyengar2005robust}, where the neighborhood is defined using the distance metric $D(\cdot, \cdot)$ and radius $\rho$. In this work, we consider the \textit{Total Variation (TV) uncertainty set} defined using the TV distance, i.e., $D(P_{s,a}, P^{o}_{s,a}) = (1/2) \|P_{s,a} - P^{o}_{s,a} \|_{1}$.
The RMDP problem is to find the optimal robust policy which maximizes the value against the worst possible model in the uncertainty set $\mathcal P$. The \textit{robust value function} $V^{\pi}$ corresponding to a policy $\pi$ and the \textit{optimal robust value function} $V^{*}$ are defined as \citep{iyengar2005robust,nilim2005robust}
\begin{align}
\label{eq:robust-value-function}
V^{\pi} = \inf_{P \in \mathcal{P}} ~V_{\pi, P},\qquad V^{*} = \sup_{\pi} \inf_{P \in \mathcal{P}} ~V_{\pi, P} .
\end{align}
The \textit{optimal robust policy} $\pi^{*}$ is such that the robust value function corresponding to it matches the optimal robust value function, i.e., $V^{\pi^*} =V^{*} $. It is known that there exists a deterministic optimal policy \citep{iyengar2005robust} for the RMDP. The \textit{robust Bellman operator} is defined as \citep{iyengar2005robust}
\begin{align}
\label{eq:robust-bellman-primal}
(T Q)(s, a) = r(s, a) + \gamma \inf_{P_{s,a} \in \mathcal{P}_{s,a}} \mathbb{E}_{s' \sim P_{s,a}} [ \max_{b} Q(s', b)].
\end{align}
It is known that $T$ is a contraction mapping in the infinity norm and hence it has a unique fixed point $Q^{*}$ with $V^{*}(s) = \max_{a} Q^{*}(s,a)$ and $\pi^{*}(s) = \argmax_{a} Q^{*}(s,a)$ \citep{iyengar2005robust}. The \textit{Robust Q-Iteration (RQI)} can now be defined using the robust Bellman operator as $Q_{k+1} = T Q_{k}$. Since $T$ is a contraction, it follows that $Q_{k} \rightarrow Q^{*}$. So, RQI can be used to compute (solving the planning problem) $Q^{*}$ and $\pi^{*}$ in the tabular setting with a known $\mathcal{P}$. Due to the optimization over the uncertainty set $\mathcal{P}_{s,a}$ for each $(s,a)$ pair, solving the planning problem in RMDP using RQI is much more computationally intensive than solving it in MDP using Q-Iteration.
\textbf{Offline RL:} Offline RL considers the problem of learning the optimal policy of an MDP when the algorithm does not have direct access to the environment and cannot generate data samples in an online manner. For learning the optimal policy $\pi^{*}_{P}$ of an MDP with model $P$, the algorithm will only have access to an offline dataset $\mathcal{D}_{P}=\{(s_i,a_i,r_{i}, s'_{i})\}_{i=1}^N$, where $(s_{i}, a_{i}) \sim \mu$, $\mu \in \Delta (\cS\times\cA)$ is some distribution, and $s'_{i} \sim P_{s_{i},a_{i}}$. \textit{Fitted Q-Iteration (FQI)} is a popular offline RL approach which is amenable to theoretical analysis while achieving impressive empirical performance. In addition to the dataset $\mathcal{D}_{P}$, FQI uses a function class $\mathcal{F}=\{f:\cS\times\cA\to [0, 1/(1-\gamma)] \}$ to approximate $Q^{*}_{P}$. The typical FQI update is given by $ f_{k+1} = \argmin_{f \in \mathcal{F}} \sum^{N}_{i=1} (r(s_{i}, a_{i}) + \gamma \max_{b} f_{k}(s'_{i}, b) - f(s_{i}, a_{i}) )^{2}$, which aims to approximate the non-robust Bellman update using offline data with function approximation. Under suitable assumptions, it is possible to obtain provable performance guarantees for FQI \citep{szepesvari2005finite, chen2019information, liu2020provably}.
\section{Offline Robust Reinforcement Learning}
\label{sec:offline-robust-RL-basics}
The goal of an offline robust RL algorithm is to learn the optimal robust policy $\pi^{*}$ using a pre-collected offline dataset $\mathcal{D}$. The data is typically generated according to a nominal (training) model $P^{o}$, i.e., $\mathcal{D} =\{(s_i,a_i,r_{i}, s'_{i})\}_{i=1}^N$, where $(s_{i}, a_{i}) \sim \mu, \mu \in \Delta (\cS\times\cA)$ is some data generating distribution, and $s'_{i} \sim P^{o}_{s_{i},a_{i}}$. The uncertainty set $\mathcal{P}$ is defined around this nominal model $P^{o}$ as given in \eqref{eq:uncertainty-set} w.r.t. the total variation distance metric. We emphasize that the learning algorithm does not know the nominal model $P^{o}$ as it has only access to $\mathcal{D}$, and hence it also does not know $\mathcal{P}$. Moreover, the learning algorithm does not have data generated according to any other models in $\mathcal{P}$ and has to rely only on $\mathcal{D}$ to account for the behavior w.r.t. all models in $\mathcal{P}$.
Learning policies for RL problems with large state-action spaces is computationally intractable. RL algorithms typically overcome this issue by using function approximation. In this paper, we consider two function classes $\mathcal{F}=\{f:\cS\times\cA\to [0, 1/(1-\gamma)] \}$ and $\mathcal{G}=\{g :\cS\times\cA\to [0, 2/(\rho(1-\gamma))]\}$. We use $\mathcal{F}$ to approximate $Q^{*}$ and $\mathcal{G}$ to approximate the dual variable functions which we will introduce in the next section. For simplicity, we will first assume that these function classes are finite but exponentially large, and we will use the standard log-cardinality to characterize the sample complexity results, as given in Theorem \ref{thm:tv-guarantee}. We note that, at the cost of additional notations and analysis, infinite function classes can also be considered where the log-cardinalities are replaced by the appropriate notions of covering number.
Similar to the non-robust offline RL, we make the following standard assumptions about the data generating distribution $\mu$ and the representation power of $\mathcal{F}$.
\begin{assumption}[Concentratability]
\label{assum-concentra-condition}
There exists a finite constant $C>0$ such that for any $\nu\in \{ d_{\pi,P^o} ~|\text{ any policy}~ \pi\} \subseteq \Delta(\cS\times\cA)$, we have $\norm{\nu/\mu}_{\infty} \leq \sqrt{C}$.
\end{assumption}
Assumption \ref{assum-concentra-condition} states that the ratio of the distribution $\nu$ and the data generating distribution $\mu$, $\nu(s,a)/\mu(s,a)$, is uniformly bounded. This assumption is widely used in the offline RL literature \citep{munos2003error,agarwal2019reinforcement,chen2019information,wang2021what,xie2021bellman} in many different forms. We borrow this assumption from \citet{chen2019information}, where they used it for non-robust offline RL. In particular, we note that the distribution $\nu$ is in the collection of discounted state-action occupancies on model $P^o$ alone for the robust RL.
\begin{assumption}[Approximate completeness]
\label{assum-bellman-completion}
Let $\mu\in\Delta(\cS\times\cA)$ be the data distribution. Then, $\sup_{f\in\mathcal{F}} \inf_{f'\in\mathcal{F}} \| f' - T f \|_{2,\mu}^2 \leq \epsilon_{\mathrm{c}}.$
\end{assumption}
Assumption \ref{assum-bellman-completion} states that the function class $\mathcal{F}$ is approximately closed under the robust Bellman operator $T$. This assumption has also been widely used in the offline RL literature~\citep{agarwal2019reinforcement,chen2019information,wang2021what,xie2021bellman}.
One of the most important properties that the function class $\mathcal{F}$ should have is that there must exist a function $f' \in \mathcal{F}$ which well-approximates $Q^{*}$. This assumption is typically called \textit{approximate realizability} in the offline RL literature. This is typically formalized by assuming $\inf_{f \in \mathcal{F}} \norm{f - Tf}^{2}_{2, \mu} \leq \epsilon_{\mathrm{r}}$ \citep{chen2019information}. It is known that the approximate completeness assumption and the concentratability assumption imply the realizability assumption \citep{chen2019information,xie2021bellman}.
\section{Robust Fitted Q-Iteration: Algorithm and Main Results}
\label{sec:algorithm-results}
In this section, we give a step-by-step approach to overcome the challenges of the offline robust RL outlined in Section \ref{sec:introduction}. We then combine these intermediate steps to obtain our proposed RFQI algorithm. We then present our main result about the performance guarantee of the RFQI algorithm, followed by a brief description about the proof approach.
\subsection{Dual Reformulation of Robust Bellman Operator}
\label{sec:rob-Bellman-tv}
One key challenge in directly using the standard definition of the optimal robust value function given in \eqref{eq:robust-value-function} or of the robust Bellman operator given in \eqref{eq:robust-bellman-primal} for developing and analyzing robust RL algorithms is that both involve computing an expectation w.r.t. each model $P \in \mathcal{P}$. Given that the data is generated only according to the nominal model $P^{o}$, estimating these expectation values is really challenging. We show that we can overcome this difficulty through the dual reformulation of the robust Bellman operator, as given below.
\begin{proposition}
\label{prop:robust-bellman-dual}
Let $M$ be an RMDP with the uncertainty set $\mathcal{P}$ specified by~\eqref{eq:uncertainty-set} using the total variation distance $D(P_{s,a}, P^{o}_{s,a}) = (1/2) \|P_{s,a} - P^{o}_{s,a} \|_{1}$. Then, for any $Q: \cS\times\cA\to [0, 1/(1-\gamma)]$, the robust Bellman operator $T$ given in \eqref{eq:robust-bellman-primal} can be equivalently written as
\begin{align}
\label{eq:robust-bellman-dual}
(T Q)(s, a) = r(s, a) - \gamma ~ \inf_{\eta\in[0, \frac{2}{\rho(1-\gamma)}]} (\mathbb{E}_{s' \sim P^{o}_{s,a}}[(\eta -V(s'))_+] - \eta + \rho (\eta - \inf_{s''} V(s''))_+ ),
\end{align}
where $V(s)=\max_{a\in\mathcal{A}} Q(s,a)$. Moreover, the inner optimization problem in \eqref{eq:robust-bellman-dual} is convex in $\eta$.
\end{proposition}
This result mainly relies on \citet[Section 3.2]{shapiro2017distributionally} and~\citet[Proposition 1]{duchi2018learning}.
Note that in \eqref{eq:robust-bellman-dual}, the expectation is now only w.r.t.~the nominal model $P^{o}$, which opens up the possibility of using empirical estimates obtained from the data generated according to $P^{o}$. This avoids the need to use importance sampling based techniques to account for all models in $\mathcal{P}$, which often have high variance, and thus, are not desirable.
While \eqref{eq:robust-bellman-dual} provides a form that is amenable to estimation using offline data, it involves finding $ \inf_{s''} V(s'')$. Though this computation is straightforward in a tabular setting, it is infeasible in a function approximation setting. In order to overcome this issue, we make the following assumption.
\begin{assumption}[Fail-state]
\label{assum-fail-state}
The RMDP $M$ has a `fail-state' $s_{f}$, such that $r(s_{f}, a) = 0$ and $P_{s_{f}, a}(s_{f}) = 1,\;\forall a \in \mathcal{A},\;\forall P \in \mathcal{P}$.
\end{assumption}
We note that this is not a very restrictive assumption because such a `fail-state' is quite natural in most simulated or real-world systems. For example, a state where a robot collapses and is not able to get up, either in a simulation environment like MuJoCo or in real-world setting, is such a fail state.
Assumption~\ref{assum-fail-state} immediately implies that $V_{\pi, P}(s_{f}) = 0,\;\forall P \in \mathcal{P}$, and hence $V^{*}(s_{f}) = 0$ and $Q^{*}(s_{f}, a) = 0,\;\forall a \in \mathcal{A}$. It is also straightforward to see that $Q_{k+1}(s_{f}, a) = 0,\;\forall a \in \mathcal{A}$, where $Q_{k}$'s are the RQI iterates given by the robust Bellman update $Q_{k+1} = T Q_{k}$ with the initialization $Q_{0}=0$. By the contraction property of $T$, we have $Q_{k} \rightarrow Q^{*}$. So, under Assumption \ref{assum-fail-state}, without loss of generality, we can always keep $Q_{k}(s_{f}, a) = 0,\;\forall a \in \mathcal{A}$ and for all $k$ in RQI (and later in RFQI). So, in the light of the above description, for the rest of the paper we will use the robust Bellman operator $T$ by setting $ \inf_{s''} V(s'') = 0$. In particular, for any function $f : \cS\times\cA \rightarrow [0,1/(1-\gamma)]$ with $f(s_{f}, a) = 0$, the robust Bellman operator $T$ is now given by
\begin{align}
\label{eq:robust-bellman-dual-2}
(T f) (s, a) = r(s, a) - \gamma \inf_{\eta\in[0, \frac{2}{(\rho(1-\gamma))}]} ~ (\mathbb{E}_{s' \sim P^{o}_{s,a}}[(\eta -\max_{a'}f(s',a'))_+] - (1-\rho)\eta ).
\end{align}
\subsection{Approximately Solving the Dual Optimization using Empirical Risk Minimization}
Another key challenge in directly using the standard definition of the optimal robust value function given in \eqref{eq:robust-value-function} or of the robust Bellman operator given in \eqref{eq:robust-bellman-primal} for developing and analyzing robust RL algorithms is that both involve an optimization over $\mathcal{P}$. The dual reformulation given in \eqref{eq:robust-bellman-dual-2} partially overcomes this challenge also, as the optimization over $\mathcal{P}$ is now replaced by a {convex} optimization over a scalar $\eta\in[0, 2/(\rho (1-\gamma))]$. However, this still requires solving an optimization for each $(s, a)\in\cS\times\cA$, which is clearly infeasible even for moderately sized state-action spaces, not to mention the function approximation setting. Our key idea to overcome this difficulty is to reformulate this as a functional optimization problem instead of solving it as multiple scalar optimization problems. This functional optimization method will make it amenable to approximately solving the dual problem using an empirical risk minimization approach with offline data.
Consider the probability (measure) space $(\cS\times\cA,\Sigma(\cS\times\cA),\mu)$ and let $L^{1}(\cS\times\cA,\Sigma(\cS\times\cA),\mu)$ be the set of all absolutely integrable functions defined on this space.\footnote{In the following, we will simply denote $L^{1}(\cS\times\cA,\Sigma(\cS\times\cA),\mu)$ as $L^{1}$ for conciseness. } In other words, $L^{1}$ is the set of all functions $g : \cS\times\cA \rightarrow \mathcal{C} \subset \mathbb{R}$, such that $\norm{g}_{1,\mu}$ is finite. We set $\mathcal{C} = [0, 2/\rho (1 - \gamma)]$, anticipating the solution of the dual optimization problem~\eqref{eq:robust-bellman-dual-2}. We also note $\mu$ is the data generating distribution which is a $\sigma$-finite measure.
For any given function $f : \cS\times\cA \rightarrow [0,1/(1-\gamma)]$, we define the loss function $L_{\mathrm{dual}}(\cdot; f)$ as
\begin{align}
\label{L-f-eta-loss}
L_{\mathrm{dual}}(g; f) = \mathbb{E}_{s,a\sim\mu} [ \mathbb{E}_{s'\sim P^o_{s,a}} [(g(s,a) - \max_{a'} f(s',a'))_+] - (1-\rho) g(s,a)], \quad \forall g\in L^1.
\end{align}
In the following lemma, we show that the scalar optimization over $\eta$ for each $(s,a)$ pair in \eqref{eq:robust-bellman-dual-2} can be replaced by a single functional optimization w.r.t.~the loss function $L_{\mathrm{dual}}$.
\begin{lemma}
\label{lem:interchanging-lemma-1}
Let $L_{\mathrm{dual}}$ be the loss function defined in \eqref{L-f-eta-loss}. Then, for any function $f : \cS\times\cA \rightarrow [0,1/(1-\gamma)]$, we have
\begin{align}
\label{eq:interchanging-lemma-1}
\inf_{g \in L^{1}} L_{\mathrm{dual}}(g; f) = \mathbb{E}_{s,a\sim\mu} \Big[ \inf_{\eta\in[0, \frac{2}{(\rho(1-\gamma))}]} \Big(\mathbb{E}_{s'\sim P^o_{s,a}} \big[\big(\eta - \max_{a'} f(s',a')\big)_+\big] - (1-\rho)\eta\Big)\Big].
\end{align}
\end{lemma}
Note that the RHS of \eqref{eq:interchanging-lemma-1} has minimization over $\eta$ for each $(s, a)$ pair and minimization is inside the expectation $\mathbb{E}_{s,a\sim\mu}[\cdot]$. However, the LHS of \eqref{eq:interchanging-lemma-1} has a single functional minimization over $g \in L^{1}$ and this minimization is outside the expectation. For interchanging the expectation and minimization, and for moving from point-wise optimization to functional optimization, we use the result from \citet[Theorem 14.60]{rockafellar2009variational}, along with the fact that $L^{1}$ is a decomposable space. We also note that this result has been used in many recent works on distributionally robust optimization \citep{shapiro2017distributionally,duchi2018learning} (see Appendix \ref{appendix:useful-tech-results} for more details).
We can now define the empirical loss function $\widehat{L}_{\mathrm{dual}}$ corresponding to the true loss $L_{\mathrm{dual}}$ as
\begin{align} \label{eq:L-dual-empirical-loss}
\widehat{L}_{\mathrm{dual}}(g;f) = \frac{1}{N} \sum^{N}_{i=1} (g(s_{i},a_{i}) - \max_{a'} f(s'_{i},a'))_+ - (1-\rho) g(s_{i},a_{i}).
\end{align}
Now, for any given $f$, we can find an approximately optimal dual function through the \textit{empirical risk minimization} approach as $\inf_{g \in L^{1}} \widehat{L}_{\mathrm{dual}}(g;f)$.
As we mentioned in Section \ref{sec:offline-robust-RL-basics}, our offline robust RL algorithm is given an input function class $\mathcal{G}=\{g :\cS\times\cA\to [0, 2/(\rho(1-\gamma))]\}$ to approximate the dual variable functions. So, in the empirical risk minimization, instead of taking the infimum over all the functions in $L^{1}$, we can only take the infimum over all the functions in $\mathcal{G}$. For this to be meaningful, $\mathcal{G}$ should have sufficient representation power. In particular, the result in Lemma \ref{lem:interchanging-lemma-1} should hold approximately even if we replace the infimum over $L^{1}$ with infimum over $\mathcal{G}$.
One can see that this is similar to the realizability requirement for the function class $\mathcal{F}$ as described in Section \ref{sec:offline-robust-RL-basics}. We formalize the representation power of $\mathcal{G}$ in the following assumption.
\begin{assumption}[Approximate dual realizability] \label{assum-dual-realizability}
For all $f \in \mathcal{F}$, there exists a uniform constant $\epsilon_{\textrm{dual}}$ such that $ \inf_{g \in \mathcal{G}} L_{\mathrm{dual}}(g; f) - \inf_{g \in L^{1}} L_{\mathrm{dual}}(g; f) \leq \epsilon_{\textrm{dual}}$
\end{assumption}
Using the above assumption, for any given $f\in\mathcal F$, we can find an approximately optimal dual function $\widehat{g}_{f} \in \mathcal{G}$ through the \textit{empirical risk minimization} approach as $\widehat{g}_{f} = \argmin_{g \in \mathcal{G}} \widehat{L}_{\mathrm{dual}}(g;f)$.
In order to characterize the performance of this approach, consider the operator $T_g$ for any $g\in\mathcal{G}$ as
\begin{align}
\label{eq:Tg}
(T_g f) (s, a) = r(s, a) - \gamma (\mathbb{E}_{s' \sim P^{o}_{s,a}}[(g(s,a) -\max_{a'}f(s',a'))_+] - (1-\rho)g(s,a) ),
\end{align} for all $f\in\mathcal{F}$ and $(s,a)\in\cS\times\cA$. We will show in Lemma \ref{lem:erm-high-prob-bound} in Appendix \ref{appendix:thm:tv-guarantee} that the error $\sup_{f\in\mathcal{F}} \|T f - T_{\widehat{g}_f} f\|_{1,\mu}$ is $\mathcal{O}(\log(|\mathcal{F}|/\delta)/\sqrt{N})$ with probability at least $1-\delta$.
\subsection{Robust Fitted Q-iteration}
\label{sec:rofqi-algorithm}
The intuitive idea behind our robust fitted Q-iteration (RFQI) algorithm is to approximate the exact RQI update step $Q_{k+1} = T Q_{k}$ with function approximation using offline data. The exact RQI step requires updating each $(s,a)$-pair separately, which is not scalable to large state-action spaces. So, this is replaced by the function approximation as $Q_{k+1} = \argmin_{f \in \mathcal{F}} \norm{T Q_{k} - f}^{2}_{2,\mu}$. It is still infeasible to perform this update as it requires to exactly compute the expectation (w.r.t. $P^{o}$ and $\mu$) and to solve the dual problem accurately. We overcome these issues by replacing both these exact computations with empirical estimates using the offline data. We note that this intuitive idea is similar to that of the FQI algorithm in the non-robust case. However, RFQI has unique challenges due to the nature of the robust Bellman operator $T$ and the presence of the dual optimization problem within $T$.
Given a dataset $\mathcal{D}$, we also follow the standard non-robust offline RL choice of least-squares residual minimization \citep{chen2019information,xie2021bellman,wang2021what}. Define the empirical loss of $f$ given $f'$ (which represents the $Q$-function from the last iteration) and dual variable function $g$ as
\begin{align}
\label{eq:loss-rfqi-1}
\widehat{L}_{\mathrm{RFQI}}(f;f', g) = \frac{1}{N} \sum^{N}_{i=1} \bigg( \begin{array}{ll} r(s_i, a_i) + \gamma \big(-(g(s_{i},a_{i}) -\max_{a'}f'(s'_{i},a'))_+ \\ \hspace{2cm}+ \;(1-\rho) g(s_{i},a_{i})\big) - f(s_{i}, a_{i}) \end{array} \bigg)^{2}.
\end{align}
The correct dual variable function to be used in \eqref{eq:loss-rfqi-1} is the optimal dual variable $g^{*}_{f'} = \argmin_{g \in \mathcal{G}} L_{\mathrm{dual}}(g; f')$ corresponding to the last iterate $f'$, which we will approximate it by $\widehat{g}_{f'} = \argmin_{g \in \mathcal{G}} \widehat{L}_{\mathrm{dual}}(g; f')$. The RFQI update is then obtained as $\argmin_{f \in \mathcal{F}} \widehat{L}_{\mathrm{RFQI}}(f;f', \widehat{g}_{f'})$.
Summarizing the individual steps described above, we formally give our RFQI algorithm below.
\begin{algorithm}[H]
\caption{Robust Fitted Q-Iteration (RFQI) Algorithm}
\label{alg:RFQI-Algorithm}
\begin{algorithmic}[1]
\STATE \textbf{Input:} Offline dataset $\mathcal{D} =(s_i,a_i,r_{i}, s'_{i})_{i=1}^N$, function classes $\mathcal{F}$ and $\mathcal{G}$.
\STATE \textbf{Initialize:} $Q_{0}\equiv 0 \in \mathcal{F}$.
\FOR {$k=0,\cdots,K-1$ }
\STATE \textbf{Dual variable function optimization:} Compute the dual variable function corresponding to $Q_{k}$ through empirical risk minimization as $g_{k} = \widehat{g}_{Q_{k}} = \argmin_{g \in \mathcal{G}} \widehat{L}_{\mathrm{dual}}(g; Q_{k})\;\;$ (see \eqref{eq:L-dual-empirical-loss}).
\STATE \textbf{Robust Q-update:} Compute the next iterate $Q_{k+1}$ through least-squares regression as
$\quad Q_{k+1}=\argmin_{Q\in\mathcal{F}}\widehat{L}_{\mathrm{RFQI}}(Q;Q_{k}, g_{k})\;\;$ (see \eqref{eq:loss-rfqi-1}).
\ENDFOR
\STATE \textbf{Output:} $\pi_{K} = \argmax_a Q_{K}(s,a)$
\end{algorithmic}
\end{algorithm}
Now we state our main theoretical result on the performance of the RFQI algorithm.
\begin{theorem}
\label{thm:tv-guarantee}
Let Assumptions \ref{assum-concentra-condition}-\ref{assum-dual-realizability} hold. Let $\pi_{K}$ be the output of the RFQI algorithm after $K$ iterations. Denote $J^{\pi}=\mathbb{E}_{s\sim d_0}[V^{\pi}(s)]$ where $d_0$ is initial state distribution. Then, for any $\delta\in(0,1)$, with probability at least $1 - 2 \delta$, we have
\begin{align*}
J^{\pi^*} - J^{\pi_{K}} \leq \frac{\gamma^K}{(1-\gamma)^2} &+ \frac{\sqrt{C}(\sqrt{6\epsilon_{\textrm{c}}} + \gamma \epsilon_{\textrm{dual}}) }{(1-\gamma)^2} + \frac{16}{\rho(1-\gamma)^3} \sqrt{\frac{18 C \log(2|\mathcal{F}||\mathcal{G}|/\delta)}{N}}. \end{align*}
\end{theorem}
\begin{remark}
Theorem \ref{thm:tv-guarantee} states that the RFQI algorithm can achieve approximate optimality. To see this, note that with $K\geq\mathcal{O}(\frac{1}{ \log(1/\gamma)} \log(\frac{1}{\epsilon (1-\gamma)}))$, and neglecting the second term corresponding to (inevitable) approximation errors $\epsilon_{\textrm{c}}$ and $\epsilon_{\textrm{dual}}$, we get $J^{\pi^*} - J^{\pi_{K}} \leq {\epsilon}/{(1-\gamma)}$ with probability greater than $1-2\delta$ for any $\epsilon, \delta \in (0,1)$, as long as the number of samples $N\geq\mathcal{O}(\frac{1}{(\rho\epsilon)^{2} (1-\gamma)^{4}} \log \frac{|\mathcal{F}| |\mathcal{G}|}{\delta })$. So, the above theorem can also be interpreted as a \textbf{sample complexity} result.
\end{remark}
\begin{remark}
The known sample complexity of robust-RL in the tabular setting is $\widetilde{\mathcal{O}}(\frac{|\cS|^2 |\mathcal{A}|}{(\rho\epsilon)^{2} (1-\gamma)^{4}})$ \citep{yang2021towards,panaganti22a}. Considering $\widetilde{\mathcal{O}}(\log(|\mathcal{F}| | \mathcal{G}|))$ to be $\widetilde{\mathcal{O}}(|\cS| |\mathcal{A}|)$, we can recover the same bound as in the tabular setting (we save $|\cS|$ due to the use of Bernstein inequality).
\end{remark}
\begin{remark}
Under similar Bellman completeness and concentratability assumptions, RFQI sample complexity is comparable to that of a non-robust offline RL algorithm, i.e.,~$\mathcal{O}(\frac{1}{\epsilon^{2} (1-\gamma)^{4}} \log \frac{|\mathcal{F}|}{\delta })$~\citep{chen2019information}. As a consequence of robustness, we have $\rho^{-2}$ and $\log(|\mathcal{G}|)$ factors in our bound.
\end{remark}
\subsection{Proof Sketch} \label{sec:proof-sketch}
Here we briefly explain the key ideas used in the analysis of RFQI for obtaining the optimality gap bound in Theorem \ref{thm:tv-guarantee}. The complete proof is provided in Appendix \ref{appendix:thm:tv-guarantee}.
\textit{Step 1:} To bound $J^{\pi^*} - J^{\pi_K}$, we connect it to the error $\|Q^{\pi^*} - Q_{K}\|_{1,\nu }$ for any state-action distribution $\nu$. While the similar step follows almost immediately using the well-known performance lemma in the analysis of non-robust FQI, such a result is not known in the robust RL setting. So, we derive the basic inequalities to get a recursive form and to obtain the bound $J^{\pi^*} - J^{\pi_K} \leq 2\|Q^{\pi^*} - Q_{K}\|_{1,\nu }/(1-\gamma)$ (see~\eqref{eq:thm-bound-part-1} and the steps before in Appendix~\ref{appendix:thm:tv-guarantee}). \\
\textit{Step 2:} To bound $\|Q^{\pi^*} - Q_{K}\|_{1,\nu }$ for any state-action distribution $\nu$ such that $\norm{\nu/\mu}_{\infty}\leq\sqrt{C}$, we decompose it to get a recursion, with approximation terms based on the least-squares regression and empirical risk minimization. Recall that $\widehat{g}_f$ is the dual variable function from the algorithm for state-action value function $f\in\mathcal{F}$. Denote $\widehat{f}_{g}$ as the least squares solution from the algorithm for the state-action value function $f\in\mathcal{F}$ and dual variable function $g\in\mathcal{G}$, i.e., $\widehat{f}_{g}=\argmin_{Q\in\mathcal{F}}\widehat{L}_{\mathrm{RFQI}}(Q;f, g)$. By recursive use of the obtained inequality \eqref{eq:pf-sketech-stp4-1} (see Appendix \ref{appendix:thm:tv-guarantee}) and using uniform bound, we get
\begin{align*}
\|Q^{\pi^*} - Q_{K}\|_{1,\nu } \leq \frac{\gamma^K }{1-\gamma} + \frac{\sqrt{C} }{1-\gamma} \sup_{f\in\mathcal{F}} \|T f - T_{\widehat{g}_f} f\|_{1,\mu}
+ \frac{\sqrt{C} }{1-\gamma} \sup_{f\in\mathcal{F}} \sup_{g\in\mathcal{G}} \| T_{g} f - \widehat{f}_{g}\|_{2,\mu}.
\end{align*}
\textit{Step 3:} We recognize that $\sup_{f\in\mathcal{F}} \|T f - T_{\widehat{g}_f} f\|_{1,\mu}$ is an empirical risk minimization error term. Using Rademacher
complexity based bounds, we show in Lemma~\ref{lem:erm-high-prob-bound} that this error is $\mathcal{O}(\log(|\mathcal{F}|/\delta)/\sqrt{N})$ with high probability. \\
\textit{Step 4:} Similarly, we also recognize that $ \sup_{f\in\mathcal{F}} \sup_{g\in\mathcal{G}} \| T_{g} f - \widehat{f}_{g}\|_{2,\mu}$ is a least-squares regression error term. We also show that this error is $\mathcal{O}(\log(|\mathcal{F}||\mathcal{G}|/\delta)/\sqrt{N})$ with high probability. We adapt the generalized least squares regression result to accommodate the modified target functions resulting from the robust Bellman operator to obtain this bound (see Lemma~\ref{lem:least-squares-generalization-bound}). \\
The proof is complete after combining steps 1-4 above.
\section{Experiments}
\label{sec:main-experiments}
\begin{figure*}[h]
\centering
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/CartPolePerturbed-v0_force_mag.pdf}
\captionof{figure}{CartPole}
\label{fig:cart_force-in}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/CartPolePerturbed-v0_action.pdf}
\captionof{figure}{CartPole}
\label{fig:cart_action-in}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/Hopper_leg_joint.pdf}
\captionof{figure}{Hopper}
\label{fig:hopper_leg}
\end{minipage}
\end{figure*}
\looseness=-1 Here, we demonstrate the robust performance of our RFQI algorithm by evaluating it on \textit{Cartpole} and \textit{Hopper} environments in OpenAI Gym \citep{brockman2016openai}. In all the figures shown, the quantity in the vertical axis is averaged over $20$ different seeded runs depicted by the thick line and the band around it is the $\pm 0.5$ standard deviation. \textit{A more detailed description of the experiments, and results on additional experiments, are deferred to Appendix \ref{appendix:simulations}.} We provide our code in \textbf{github webpage} \url{https://github.com/zaiyan-x/RFQI} containing instructions to reproduce all results in this paper.
\looseness=-1 For the \textit{Cartpole}, we compare RFQI algorithm against the non-robust RL algorithms FQI and DQN, and the soft-robust RL algorithm proposed in \cite{derman2018soft}. We test the robustness of the algorithms by changing the parameter \textit{force\_mag} (to model external force disturbance), and also by introducing action perturbations (to model actuator noise). Fig. \ref{fig:cart_force-in} and Fig. \ref{fig:cart_action-in} shows superior robust performance of RFQI compared to the non-robust FQI and DQN. The RFQI performance is similar to that of soft-robust DQN. We note that soft-robust RL algorithm (here soft-robust DQN) is an online deep RL algorithm (and not an offline RL algorithm) and has no provable performance guarantee. Moreover, soft-robust RL algorithm requires generating online data according a number of models in the uncertainty set, whereas RFQI only requires offline data according to a single nominal training model.
For the \textit{Hopper}, we compare RFQI algorithm against the non-robust RL algorithms FQI and TD3 \citep{fujimoto2018addressing}, and the soft-robust RL (here soft-robust DDPG) algorithm proposed in \cite{derman2018soft}. We test the robustness of the algorithms by changing the parameter \textit{leg\_joint\_stiffness}. Fig. \ref{fig:hopper_leg} shows the superior performance of our RFQI algorithm against the non-robust algorithms and soft-robust DDPG algorithm. The average episodic reward of RFQI remains almost the same initially, and later decays much less and gracefully when compared to the non-robust FQI and TD3.
\section{Conclusion}
In this work, we presented a novel robust RL algorithm called Robust Fitted Q-Iteration algorithm with provably optimal performance for an RMDP with arbitrarily large state space, using only offline data with function approximation. We also demonstrated the superior performance of the proposed algorithm on standard benchmark problems.
\looseness=-1 One limitation of our present work is that, we considered only the uncertainty set defined with respect to the total variation distance. In future work, we will consider uncertainty sets defined with respect to other $f$-divergences such as KL-divergence and Chi-square divergence. Finding a lower bound for the sample complexity and relaxing the assumptions used are also important and challenging problems.
\section{Acknowledgements}
This work was supported in part by the National Science Foundation (NSF) grants NSF-CAREER- EPCN-2045783 and NSF ECCS 2038963. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsoring agencies.
\section*{Checklist}
\begin{enumerate}
\item For all authors...
\begin{enumerate}
\item Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
\answerYes{See \textbf{contributions} in the Introduction.}
\item Did you describe the limitations of your work?
\answerYes{The discussions on the assumptions describes the limitations.}
\item Did you discuss any potential negative societal impacts of your work?
\answerNA{}
\item Have you read the ethics review guidelines and ensured that your paper conforms to them?
\answerYes{}
\end{enumerate}
\item If you are including theoretical results...
\begin{enumerate}
\item Did you state the full set of assumptions of all theoretical results?
\answerYes{See Sections \ref{sec:offline-robust-RL-basics}-\ref{sec:rofqi-algorithm} }
\item Did you include complete proofs of all theoretical results?
\answerYes{We provide proof sketch \ref{sec:proof-sketch} in main paper and the complete proof in Appendix with self-contained material.}
\end{enumerate}
\item If you ran experiments...
\begin{enumerate}
\item Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)?
\answerYes{}
\item Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)?
\answerYes{Described in the Appendix.}
\item Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)?
\answerYes{Described in the main paper and the Appendix.}
\item Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)?
\answerYes{Mentioned in the Appendix.}
\end{enumerate}
\item If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
\begin{enumerate}
\item If your work uses existing assets, did you cite the creators?
\answerYes{}
\item Did you mention the license of the assets?
\answerYes{}
\item Did you include any new assets either in the supplemental material or as a URL?
\answerNA{}
\item Did you discuss whether and how consent was obtained from people whose data you're using/curating?
\answerNA{}
\item Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content?
\answerNA{}
\end{enumerate}
\item If you used crowdsourcing or conducted research with human subjects...
\begin{enumerate}
\item Did you include the full text of instructions given to participants and screenshots, if applicable?
\answerNA{}
\item Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable?
\answerNA{}
\item Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation?
\answerNA{}
\end{enumerate}
\end{enumerate}
\section*{Appendix}
\section{Useful Technical Results}
\label{appendix:useful-tech-results}
In this section, we state some existing results from concentration inequalities, generalization bounds, and optimization theory that we will use later in our analysis. We first state the Berstein's inequality that utilizes second-moment to get a tighter concentration inequality.
\begin{lemma}[Bernstein's inequality \hspace{-0.1cm} \text{\citep[Theorem 2.8.4]{vershynin2018high}}] \label{lem:bernstein-ineq}
Let $X_1,\cdots, X_T$ be independent random variables. Assume that $|X_t - \mathbb{E}[X_t]| \leq M$, for all $t$. Then, for any $\epsilon > 0$, we have
\[
\mathbb{P} \left( \Big|\frac{1}{T} \sum_{t=1}^T (X_t - \mathbb{E}[X_t])\Big| \geq \epsilon \right) \leq 2\exp\left( -\frac{T^2\epsilon^2}{2\sigma^2 + \frac{2MT\epsilon}{3}} \right),
\]
where $\sigma^2 = \sum_{t=1}^T \mathbb{E}[X_t^2]$. Furthermore, if $X_1,\cdots, X_T$ are independent and identically distributed random variables, then for any $\delta\in(0,1)$, we have
\[
\Big|\mathbb{E}[X_1] - \frac{1}{T} \sum_{t=1}^T X_t \Big| \leq \sqrt{\frac{2 \mathbb{E}[X_1^2] \log(2/\delta)}{T}} +\frac{M\log(2/\delta)}{3T},
\]
with probability at least $1-\delta$.
\end{lemma}
We now state a result
for the generalization bounds on empirical risk minimization (ERM) problems. This result is adapted from \citet[Theorem 26.5, Lemma 26.8, Lemma 26.9]{shalev2014understanding}.
\begin{lemma}[ERM generalization bound] \label{lem:erm-generalization-bound}
Let $P$ be the data generating distribution on the space $\mathcal{X}$ and let $\mathcal{H}$ be a given hypothesis class of functions. Assume that for all $x\in\mathcal{X}$ and $h \in \mathcal{H}$ we have that $|l(h,x)| \leq c_1$ for some positive constant $c_1>0$. Given a dataset $\mathcal{D}=\{X_i\}_{i=1}^N$, generated independently from $P$, denote $\hat{h}$ as the ERM solution, i.e. $\hat{h}=\argmin_{h\in\mathcal{H}} (1/N) \sum_{i=1}^N l(h, X_i)$. For any fixed $\delta\in(0,1)$ and $h^*\in\argmin_{h\in\mathcal{H}} \mathbb{E}_{X\sim P}[l(h,X)]$, we have
\begin{align} \label{eq:erm-generalization-bound-part-1} \mathbb{E}_{X\sim P}[l(\hat{h},X)] - \mathbb{E}_{X\sim P}[l(h^*,X)] \leq 2 R(l\circ \mathcal{H} \circ \mathcal{D}) + 5 c_1 \sqrt{\frac{2\log(8/\delta)}{N}},
\end{align}
with probability at least $1-\delta$, where $R(\cdot)$ is the Rademacher complexity of $l\circ \mathcal{H}$ given by
\[
R(l\circ \mathcal{H} \circ \mathcal{D}) = \frac{1}{N} \mathbb{E}_{\{\sigma_i\}_{i=1}^N} \left( \sup_{g\,\in\, l\circ \mathcal{H}}~ \sum_{i=1}^N \sigma_i g(X_i) \right),
\]
in which $\sigma_i$'s are independent from $X_i$'s and are independently and identically distributed according to the Rademacher random variable $\sigma$, i.e. $\mathbb{P}(\sigma=1) = 0.5 = \mathbb{P}(\sigma=-1)$.
Furthermore, if $\mathcal{H}$ is a finite hypothesis class, i.e. $|\mathcal{H}|<\infty$, with $|h\circ x| \leq c_2$ for all $h\in\mathcal{H}$ and $x\in\mathcal{X}$, and $l(h,x)$ is $c_3$-Lipschitz in $h$, then we have
\begin{align} \label{eq:erm-generalization-bound-part-2}
\mathbb{E}_{X\sim P}[l(\hat{h},X)] - \mathbb{E}_{X\sim P}[l(h^*,X)] \leq 2 c_2 c_3 \sqrt{\frac{2\log(|\mathcal{H}|)}{N}} + 5 c_1 \sqrt{\frac{2\log(8/\delta)}{N}},
\end{align}
with probability at least $1-\delta$.
\end{lemma}
We now mention two important concepts from variational analysis \citep{rockafellar2009variational} literature that is useful to relate minimization of integrals and the integrals of pointwise minimization under special class of functions.
\begin{definition}[Decomposable spaces and Normal integrands \text{\citep[Definition 14.59, Example 14.29]{rockafellar2009variational}}] \label{def:decomposable-normal-integrand}
A space $\mathcal{X}$ of measurable functions is a {decomposable space} relative to an underlying measure space $(\Omega, \mathcal{A}, \mu)$, if for every function $x_0 \in \mathcal{X}$, every set $A \in \mathcal{A}$ with $\mu(A) < \infty$, and any bounded measurable function $x_1\,:\,A\to\mathbb{R}$, the function $x(\omega) = x_0(\omega)\mathbbm{1}(\omega \notin A) + x_1(\omega)\mathbbm{1}(\omega \in A)$ belongs to $\mathcal{X}$. A function $f : \Omega \times \mathbb{R} \to \mathbb{R}$ (finite-valued) is a {normal integrand}, if and only if $f(\omega,x)$ is $\mathcal{A}$-measurable in $\omega$ for each $x$ and is continuous in $x$ for each $\omega$.
\end{definition}
\begin{remark} \label{remark:decomposable}
A few examples of decomposable spaces are $L^p(\cS\times\cA,\Sigma(\cS\times\cA),\mu)$ for any $p\geq 1$ and $\mathcal{M}(\cS\times\cA,\Sigma(\cS\times\cA))$, the space of all $\Sigma(\cS\times\cA)$-measurable functions. \end{remark}
\begin{lemma}[\text{\citealp[Theorem 14.60]{rockafellar2009variational}}] \label{lem:integral-min-exchange-rockafellar}
Let $\mathcal{X}$ be a space of measurable functions from $\Omega$ to $\mathbb{R}$ that is decomposable relative to a $\sigma$-finite measure $\mu$ on the $\sigma$-algebra $\mathcal{A}$. Let $f : \Omega \times \mathbb{R} \to \mathbb{R}$ (finite-valued) be a normal integrand. Then, we have
\[
\inf_{x\in\mathcal{X}} \int_{\omega\in\Omega} f(\omega,x(\omega)) \mu(\mathrm{d}\,\omega) = \int_{\omega\in\Omega} \left( \inf_{x\in\mathbb{R}} f(\omega,x) \right) \mu(\mathrm{d}\,\omega).
\]
Moreover, as long as the above infimum is not $-\infty$, we have that
\begin{align*}
x' \in \argmin_{x\in\mathcal{X}} \int_{\omega\in\Omega} f(\omega,x(\omega)) \mu(\mathrm{d}\,\omega),
\end{align*}
if and only if $\;x'(\omega) \in \argmin_{x\in\mathbb{R}} f(\omega,x)\cdot\mu\;$ almost surely.
\end{lemma}
We now give one result from distributioanlly robust optimization. The $f$-divergence between the distributions $P$ and $P^o$ is defined as
\begin{align}
\label{eq:f-divergence-defn}
D_{f}(P \| P^o) = \int f(\frac{\mathrm{d}\, P}{\mathrm{d}\, P^o})\mathrm{d}\, P^o,
\end{align}
where $f$ is a convex function \citep{csiszar1967information,moses2011further}. We obtain different divergences for different forms of the function $f$, including some well-known divergences. For example,
$f(t) = |t-1|/2$ gives Total Variation (TV), $f(t) = t\log t$ gives Kullback-Liebler (KL), $f(t) = (t-1)^2$ gives Chi-square, and $f(t) = (\sqrt{t}-1)^2$ gives squared Hellinger divergences.
Let $P^o$ be a distribution on the space $\mathcal{X}$ and let $l: \mathcal{X} \to \mathbb{R}$ be a loss function. We have the following result from the \textit{distributionally robust optimization} literature, see e.g.,~\citet[Section 3.2]{shapiro2017distributionally} and~\citet[Proposition 1]{duchi2018learning}.
\begin{proposition}
\label{prop:dro-inner-problem-solution}
Let $D_{f}$ be the $f$-divergence as defined in \eqref{eq:f-divergence-defn}. Then,
\begin{align}
\label{eq:dro-inner-problem-solution}
\sup_{D_f(P\|P^o) \leq \rho} \mathbb{E}_{P}[l(X)] = \inf_{\lambda>0, \eta\in\mathbb{R}} ~~ \mathbb{E}_{P^o} \left[ \lambda f^* \left(\frac{l(X)-\eta}{\lambda}\right)\right] + \lambda\rho + \eta,
\end{align}
where $f^*(s) = \sup_{t\geq 0} \{st - f(t)\}$ is the Fenchel conjugate.
\end{proposition}
Note that on the right hand side of \eqref{eq:dro-inner-problem-solution}, the expectation is taken only with respect to $P^{o}$. We will use the above result to derive the dual reformulation of the robust Bellman operator.
\section{Proof of the Proposition \ref{prop:robust-bellman-dual}}
\label{appendix:prop:robust-bellman-dual}
As the first step, we adapt the result given in Proposition~\ref{prop:dro-inner-problem-solution} in two ways: $(i)$ Since Proposition~\ref{prop:robust-bellman-dual} considers the TV uncertainty set, we will derive the specific form of this result for the TV uncertainty set, $(ii)$ Since Proposition~\ref{prop:robust-bellman-dual} considers the minimization problem instead of the maximization problem, unlike in Proposition~\ref{prop:dro-inner-problem-solution}, we will derive the specific form of this result for minimization.
\begin{lemma}
\label{lem:TV-inner-problem}
Let $D_{f}$ be as defined in \eqref{eq:f-divergence-defn} with $f(t)=|t-1|/2$ corresponding to the TV uncertainty set. Then,
\begin{align*}
\inf_{D_f(P\|P^o) \leq \rho} \mathbb{E}_{P}[l(X)] = -\inf_{\eta \in\mathbb{R}}\;\mathbb{E}_{P^o}[(\eta-l(X))_+] + (\eta - \inf_{x\in\mathcal{X}} l(x))_+ \times \rho - \eta,
\end{align*}
\end{lemma}
\begin{proof}
First, we will compute the Fenchel conjugate of $f(t)=|t-1|/2$. We have
\begin{align*}
f^*(s) &= \sup_{t\geq 0} ~ \{st - \frac{1}{2}|t-1|\}
= \max \big\{\!\!\sup_{t\in[0,1]} \!\{(s+\frac{1}{2})t - \frac{1}{2}\} \;,\; \sup_{t> 1} \!\{(s-\frac{1}{2}) t +\frac{1}{2}\} \big\}.
\end{align*}
It is easy to see that for $s>1/2$, we have $f^*(s)=+\infty$, and for $s\leq -1/2$, we have $f^*(s)=-1/2$. For $s \in [-1/2,1/2]$, we have
\begin{align*}
f^*(s) &= \max \big\{ \sup_{t\in[0,1]}\{(s+\frac{1}{2})t - \frac{1}{2}\} \;,\; \sup_{t> 1} (\{(s-\frac{1}{2}) t +\frac{1}{2}\} \big\} \\
&= \max \big\{ ((s+\frac{1}{2})\cdot 1 - \frac{1}{2}),~ ((s-\frac{1}{2})\cdot 1 +\frac{1}{2}) \big\} = s.
\end{align*}
Thus, we have
\begin{equation*}
f^*(s) =
\begin{cases}
-\frac{1}{2} & s\leq -\frac{1}{2},\\
s & s\in [-\frac{1}{2},\frac{1}{2}]\\
+\infty & s > \frac{1}{2}.
\end{cases} .
\end{equation*}
From Proposition \ref{prop:dro-inner-problem-solution}, we obtain
\begin{align*}
\sup_{D_f(P\|P^o) \leq \rho} \mathbb{E}_{P}[l(X)] &= \inf_{\lambda>0, \eta\in\mathbb{R}} ~ \mathbb{E}_{P^o} [\lambda f^*(\frac{l(X)-\eta}{\lambda})] + \lambda\rho + \eta \\
&= \inf_{\lambda,\eta: \lambda>0, \eta\in\mathbb{R}, \frac{\sup_{x\in\mathcal{X}}l(x)-\eta}{\lambda} \leq \frac{1}{2}} ~ \mathbb{E}_{P^o} [\lambda \max\{ \frac{l(X)-\eta}{\lambda}, -\frac{1}{2} \}] + \lambda\rho + \eta \\
&= \inf_{\lambda,\eta: \lambda>0, \eta\in\mathbb{R}, \frac{\sup_{x\in\mathcal{X}}l(x)-\eta}{\lambda} \leq \frac{1}{2}} ~ \mathbb{E}_{P^o} [\max\{ l(X)-\eta, -\lambda/2 \}] + \lambda\rho + \eta \\
&= \inf_{\lambda,\eta: \lambda>0, \eta\in\mathbb{R}, \frac{\sup_{x\in\mathcal{X}}l(x)-\eta}{\lambda} \leq \frac{1}{2}} ~ \mathbb{E}_{P^o} [ (l(X)-\eta +\lambda/2)_+ ] - \lambda/2 + \lambda\rho + \eta \\
&= \inf_{\lambda,\eta': \lambda>0, \eta'\in\mathbb{R}, \frac{\sup_{x\in\mathcal{X}}l(x)-\eta'}{\lambda} \leq 1} ~ \mathbb{E}_{P^o} [ (l(X)-\eta')_+] + \lambda\rho + \eta'.
\end{align*}
The second equality follows since $f^*(\frac{l(X)-\eta}{\lambda})=+\infty$ whenever $\frac{l(X)-\eta}{\lambda}>\frac{1}{2}$, which can be ignored as we are minimizing over $\lambda$ and $\eta$. The fourth equality follows form the fact that $\max\{x,y\} = (x-y)_+ + y$ for any $x,y\in\mathbb{R}$. Finally, the last equality follows by making the substitution $\eta'=\eta-\lambda/2$. Taking the optimal value of $\lambda$, i.e.,~$\lambda=(\sup_{x\in\mathcal{X}}l(x)-\eta')_+$, we get
\begin{align*}
\sup_{D_f(P\|P^o) \leq \rho} \mathbb{E}_{P}[l(X)] = \inf_{\eta\in\mathbb{R}} \mathbb{E}_{P^o}[(l(X)-\eta)_+] + (\sup_{x\in\mathcal{X}}l(x) - \eta)_+ \rho + \eta.
\end{align*}
Now,
\begin{align*}
\inf_{D_f(P\|P^o) \leq \rho} \mathbb{E}_{P}[l(X)] &= -\sup_{D_f(P\|P^o) \leq \rho} \mathbb{E}_{P}[-l(X)] \\
&= -\inf_{\eta\in\mathbb{R}} ~~ \mathbb{E}_{P^o}[(-l(X)-\eta)_+] + (\sup_{x\in\mathcal{X}} -l(x) - \eta)_+ \rho + \eta \\
&= -\inf_{\eta' \in\mathbb{R}} ~~ \mathbb{E}_{P^o}[(\eta'-l(X))_+] + (\eta' - \inf_{x\in\mathcal{X}} l(x))_+ \rho - \eta',
\end{align*}
which completes the proof.
\end{proof}
We are now ready to prove Proposition \ref{prop:robust-bellman-dual}.
\begin{proof}[\textbf{Proof of Proposition~\ref{prop:robust-bellman-dual}}]
For each $(s, a)$, the optimization problem in~\eqref{eq:robust-bellman-primal} is given by $\min_{P_{s,a} \in \mathcal{P}_{s,a}} \mathbb{E}_{s' \sim P_{s,a}} [ V(s')]$, and our focus is on the setting where $\mathcal{P}_{s,a}$ is given by the TV uncertainty set. So, $\mathcal{P}_{s,a}$ can be equivalently defined using the $f$-divergence with $f(t) = |t-1|/2$ as $\mathcal{P}_{s,a} = \{P_{s,a} : D_{f}(P_{s,a}||P^{o}_{s,a}) \leq \rho\}$. We can now use the result of Lemma~\ref{lem:TV-inner-problem} to get
\begin{align*}
\inf_{P_{s,a} \in \mathcal{P}_{s,a}} \mathbb{E}_{s' \sim P_{s,a}} [ V(s')] = -\inf_{\eta \in\mathbb{R}} ~~ \mathbb{E}_{s' \sim P^o_{s,a}}[(\eta-V(s'))_+] + (\eta - \inf_{s'' \in \cS} V(s''))_+ \rho - \eta.
\end{align*}
From Proposition~\ref{prop:dro-inner-problem-solution}, the function $h(\eta)=\mathbb{E}_{s' \sim P^{o}_{s,a}}[(\eta -V(s'))_+] + \rho (\eta- \inf_{s''} V(s''))_+ - \eta $ is convex in $\eta$. Since $V(s') \geq 0,$ $h(\eta)=-\eta \geq 0$ when $\eta \leq 0$. So, $\inf_{\eta \in (-\infty, 0]} h(\eta)$, achieved at $\eta = 0$. Also, since $V(s)\leq 1/(1-\gamma)$, we have
\begin{align*}
h(\frac{2}{\rho(1-\gamma)}) &= \mathbb{E}_{s' \sim P^{o}_{s,a}}[\frac{2}{\rho(1-\gamma)} -V(s')] + \rho (\frac{2}{\rho(1-\gamma)}- \inf_{s''} V(s'')) - \frac{2}{\rho(1-\gamma)} \\
&\geq -\frac{1}{(1-\gamma)} + \rho (\frac{2}{\rho(1-\gamma)}- \frac{1}{(1-\gamma)} ) = \frac{2}{(1-\gamma)} - \frac{(1+\rho)}{(1-\gamma)} \geq 0.
\end{align*}
So, it is sufficient to consider $\eta \in [0, \frac{2}{\rho (1-\gamma)}]$ for the above optimization problem.
Using these, we get
\begin{align*}
(T Q)&(s, a) = r(s, a) + \gamma \inf_{P_{s,a} \in \mathcal{P}_{s,a}} \mathbb{E}_{s' \sim P_{s,a}} [ V(s')] \\
&= r(s, a) + \gamma \cdot -1 \cdot \inf_{\eta \in \eta \in [0, \frac{2}{\rho (1-\gamma)}]} ~~ \mathbb{E}_{s' \sim P^o_{s,a}}[(\eta-V(s'))_+] + (\eta - \inf_{s'' \in \cS} V(s''))_+ \rho - \eta.
\end{align*}
This completes the proof of Proposition~\ref{prop:robust-bellman-dual}.
\end{proof}
\section{Proof of Theorem \ref{thm:tv-guarantee}}
\label{appendix:thm:tv-guarantee}
We start by proving Lemma \ref{lem:interchanging-lemma-1} which mainly follows from Lemma~\ref{lem:integral-min-exchange-rockafellar} in Appendix~\ref{appendix:useful-tech-results}.
\begin{proof}[Proof of Lemma \ref{lem:interchanging-lemma-1}]
Let $h((s,a),\eta)=\mathbb{E}_{s'\sim P^o_{s,a}} ((\eta - \max_{a'} f(s',a'))_+ - (1-\rho)\eta)$. We note that $h((s,a),\eta)$ is $\Sigma(\cS\times\cA)$-measurable in $(s,a)\in\cS\times\cA$ for each $\eta\in[0,{1}/{(\rho(1-\gamma))}]$ and is continuous in $\eta$ for each $(s,a)\in\cS\times\cA$. Now it follows that $h((s,a),\eta)$ is a normal integrand (see Definition~\ref{def:decomposable-normal-integrand} in Appendix~\ref{appendix:useful-tech-results}). We now note that $L^{1}(\cS\times\cA,\Sigma(\cS\times\cA),\mu)$ is a decomposable space (Remark \ref{remark:decomposable} in Appendix~\ref{appendix:useful-tech-results}). Thus, this lemma now directly follows from Lemma~\ref{lem:integral-min-exchange-rockafellar}.
\end{proof}
Now we state a result and provide its proof for the empirical risk minimization on the dual parameter.
\begin{lemma}[Dual Optimization Error Bound] \label{lem:erm-high-prob-bound}
Let $\widehat{g}_{f}$ be the dual optimization parameter from the algorithm (Step 4) for the state-action value function $f$ and let $T_{g}$ be as defined in \eqref{eq:Tg}. With probability at least $1-\delta$, we have
\[
\sup_{f\in\mathcal{F}} \|T f - T_{\widehat{g}_{f}} f\|_{1,\mu} \leq \frac{4\gamma(2-\rho)}{\rho(1-\gamma)}\sqrt{\frac{2\log(|\mathcal{G}|)}{N}} + \frac{25\gamma}{\rho(1-\gamma)} \sqrt{\frac{2\log(8|\mathcal{F}|/\delta)}{N}} + \gamma \epsilon_{\textrm{dual}}.
\]
\end{lemma}
\begin{proof}
Fix an $f\in\mathcal{F}$. We will also invoke union bound for the supremum here. We recall from~\eqref{eq:L-dual-empirical-loss} that $\widehat{g}_f = \argmin_{g \in \mathcal{G}} \widehat{L}_{\mathrm{dual}}(g; f)$. From the robust Bellman equation, we directly obtain
\begingroup
\allowdisplaybreaks
\begin{align*}
&\|T_{\widehat{g}_{f}} f - T f \|_{1,\mu} = \gamma (\mathbb{E}_{s,a\sim\mu} | \mathbb{E}_{s'\sim P^o_{s,a}} ((\widehat{g}_{f}(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)\widehat{g}_{f}(s,a)) \\&\hspace{2cm}- \inf_{\eta \in[0, {2}/{(\rho(1-\gamma))}]} \mathbb{E}_{s'\sim P^o_{s,a}} ((\eta - \max_{a'} f(s',a'))_+ - (1-\rho) \eta) | ) \\
&\stackrel{(a)}{=} \gamma (\mathbb{E}_{s,a\sim\mu}\mathbb{E}_{s'\sim P^o_{s,a}} ((\widehat{g}_{f}(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)\widehat{g}_{f}(s,a)) \\&\hspace{1cm}- \mathbb{E}_{s,a\sim\mu} [ \inf_{\eta \in[0, {2}/{(\rho(1-\gamma))}]} \mathbb{E}_{s'\sim P^o_{s,a}} ((\eta - \max_{a'} f(s',a'))_+ - (1-\rho) \eta) ] ) \\
&\stackrel{(b)}{=} \gamma (\mathbb{E}_{s,a\sim\mu,s'\sim P^o_{s,a}} ((\widehat{g}_{f}(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)\widehat{g}_{f}(s,a)) \\&\hspace{1cm}- \inf_{g \in L^{1}} \mathbb{E}_{s,a\sim\mu,s'\sim P^o_{s,a}} ((g(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)g(s,a)) )\\
&= \gamma (\mathbb{E}_{s,a\sim\mu,s'\sim P^o_{s,a}} ((\widehat{g}_{f}(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)\widehat{g}_{f}(s,a)) \\&\hspace{1cm}- \inf_{g\in \mathcal{G}} \mathbb{E}_{s,a\sim\mu,s'\sim P^o_{s,a}} ((g(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)g(s,a)) )
\\ &\hspace{1cm} + \gamma (\inf_{g\in \mathcal{G}} \mathbb{E}_{s,a\sim\mu,s'\sim P^o_{s,a}} ((g(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)g(s,a))
\\&\hspace{1.5cm}- \inf_{g \in L^{1}} \mathbb{E}_{s,a\sim\mu,s'\sim P^o_{s,a}} ((g(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)g(s,a)) )\\
&\stackrel{(c)}{\leq} \gamma (\mathbb{E}_{s,a\sim\mu,s'\sim P^o_{s,a}} ((\widehat{g}_{f}(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)\widehat{g}_{f}(s,a)) \\&\hspace{1cm}- \inf_{g\in \mathcal{G}} \mathbb{E}_{s,a\sim\mu,s'\sim P^o_{s,a}} ((g(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)g(s,a)) ) + \gamma \epsilon_{\textrm{dual}} \\
&\stackrel{(d)}{\leq} 2\gamma R(l\circ \mathcal{G} \circ \mathcal{D}) + \frac{25\gamma}{\rho(1-\gamma)} \sqrt{\frac{2\log(8/\delta)}{N}} + \gamma \epsilon_{\textrm{dual}}
\\&\stackrel{(e)}{\leq} \frac{4\gamma(2-\rho)}{\rho(1-\gamma)}\sqrt{\frac{2\log(|\mathcal{G}|)}{N}} + \frac{25\gamma}{\rho(1-\gamma)} \sqrt{\frac{2\log(8/\delta)}{N}} + \gamma \epsilon_{\textrm{dual}}.
\end{align*}
\endgroup
$(a)$ follows since $\inf_{g} h(g) \leq h(\widehat{g}_{f})$. $(b)$ follows from Lemma \ref{lem:interchanging-lemma-1}. $(c)$ follows from the approximate dual realizability assumption (Assumption \ref{assum-dual-realizability}).
For $(d)$, we consider the loss function $l(g,(s,a,s'))=(g(s,a) - \max_{a'} f(s',a'))_+ - (1-\rho)g(s,a)$ and dataset $\mathcal{D}=\{s_i,a_i,s_i'\}_{i=1}^N$. Note that $|l(g,(s,a,s'))|\leq 5/(\rho(1-\gamma))$ (since $f\in\mathcal{F}$ and $g\in\mathcal{G}$). Now, we can apply the empirical risk minimization result \eqref{eq:erm-generalization-bound-part-1} in Lemma \ref{lem:erm-generalization-bound} to get $(d)$, where $R(\cdot)$ is the Rademacher complexity.
Finally, $(e)$ follows from~\eqref{eq:erm-generalization-bound-part-2} in Lemma \ref{lem:erm-generalization-bound} when combined with the facts that $l(g,(s,a,s'))$ is $(2-\rho)$-Lipschitz in $g$ and $g(s,a) \leq 2/(\rho(1-\gamma))$, since $g\in\mathcal{G}$.
With union bound, with probability at least $1-\delta$, we finally get \begin{align*}
&\sup_{f\in\mathcal{F}}\|T f - T_{\widehat{g}_{f}} f\|_{1,\mu} {\leq} \frac{4\gamma(2-\rho)}{\rho(1-\gamma)}\sqrt{\frac{2\log(|\mathcal{G}|)}{N}} + \frac{25\gamma}{\rho(1-\gamma)} \sqrt{\frac{2\log(8|\mathcal{F}|/\delta)}{N}} + \gamma \epsilon_{\textrm{dual}},
\end{align*}
which concludes the proof.
\end{proof}
We next prove the least-squares generalization bound for the RFQI algorithm.
\begin{lemma}[Least squares generalization bound] \label{lem:least-squares-generalization-bound}
Let $\widehat{f}_{g}$ be the least-squares solution from the algorithm (Step 5) for the state-action value function $f$ and dual variable function $g$. Let $T_{g}$ be as defined in \eqref{eq:Tg}. Then, with probability at least $1-\delta$, we have
\begin{align*}
\sup_{f\in\mathcal{F}} \sup_{g\in\mathcal{G}} \| T_{g} f - \widehat{f}_{g} \|_{2,\mu}
&\leq \sqrt{6 \epsilon_{\textrm{c}}} + \frac{16}{\rho(1-\gamma)} \sqrt{\frac{18 \log(2|\mathcal{F}||\mathcal{G}|/\delta)}{N}}.
\end{align*}
\end{lemma}
\begin{proof}
We adapt the least-squares generalization bound given in~\citet[Lemma A.11]{agarwal2019reinforcement} to our setting. We recall from~\eqref{eq:loss-rfqi-1} that $\widehat{f}_{g}=\argmin_{Q\in\mathcal{F}}\widehat{L}_{\mathrm{RFQI}}(Q;f, g)$. We first fix functions $f\in\mathcal{F}$ and $g\in\mathcal{G}$. For any function $f'\in\mathcal{F}$, we define random variables $z_i^{f'}$ as
\begin{align*}
z_i^{f'} = \left( f'(s_i,a_i) - y_{i} \right)^2 - \left( (T_{g} f)(s_i,a_i) - y_{i} \right)^2,
\end{align*}
where $y_{i} = r_i - \gamma(g(s_i,a_i) - \max_{a'} f(s_i',a'))_+ + \gamma (1-\rho)g(s_i,a_i)$, and $(s_{i}, a_{i}, s'_{i}) \in \mathcal{D}$ with $(s_{i}, a_{i}) \sim \mu, s'_{i} \sim P^{o}_{s_{i}, a_{i}}$. It is straightforward to note that for a given $(s_{i}, a_{i})$, we have $ \mathbb{E}_{s'_i \sim P^o_{s_i,a_i}} [y_{i}] = (T_{g} f)(s_{i}, a_{i})$.
Also, since $g(s_i,a_i)\leq 2/(\rho(1-\gamma))$ (because $g\in\mathcal{G}$) and $f(s_i,a_i), f'(s_i,a_i)\leq 1/(1-\gamma)$ (because $f, f'\in\mathcal{F})$, we have $(T_{g}f)(s_i,a_i) \leq 5/(\rho(1-\gamma))$. This also gives us that $y_{i} \leq 5/(\rho(1-\gamma))$.
Using this, we obtain the first moment and an upper-bound for the second moment of $z_i^{f'}$ as follows:
\begin{align*}
\mathbb{E}_{s'_i \sim P^o_{s_i,a_i}} [z_i^{f'}] &= \mathbb{E}_{s'_i \sim P^o_{s_i,a_i}} [( f'(s_i,a_i) - (T_{g} f)(s_i,a_i)) \cdot (f'(s_i,a_i) + (T_{g} f)(s_i,a_i) -2 y_{i} ) ] \\
&= ( f'(s_i,a_i) - (T_{g} f)(s_i,a_i))^{2}, \\
\mathbb{E}_{s'_i \sim P^o_{s_i,a_i}} [(z_i^{f'})^2] &= \mathbb{E}_{s'_i \sim P^o_{s_i,a_i}} [( f'(s_i,a_i) - (T_{g} f)(s_i,a_i))^{2} \cdot (f'(s_i,a_i) + (T_{g} f)(s_i,a_i) -2 y_{i} )^{2} ] \\
&= ( f'(s_i,a_i) - (T_{g} f)(s_i,a_i))^{2} \cdot \mathbb{E}_{s'_i \sim P^o_{s_i,a_i}} [ (f'(s_i,a_i) + (T_{g} f)(s_i,a_i) -2 y_{i} )^{2} ] \\
&\leq C_{1} ( f'(s_i,a_i) - (T_{g} f)(s_i,a_i))^{2} ,
\end{align*}
where $C_{1}={16^2}/{(\rho^2(1-\gamma)^2)}$. This immediately implies that
\begin{align*}
\mathbb{E}_{s_i,a_i\sim \mu, s'_i \sim P^o_{s_i,a_i}} [z_i^{f'}] &= \norm{T_{g}f -f'}^{2}_{2,\mu}, \\
\mathbb{E}_{s_i,a_i\sim \mu, s'_i \sim P^o_{s_i,a_i}} [(z_i^{f'})^2] &\leq C_{1} \norm{T_{g}f -f'}^{2}_{2,\mu}.
\end{align*}
From these calculations, it is also straightforward to see that $|z_i^{f'}-\mathbb{E}_{s_i,a_i\sim \mu, s'_i \sim P^o_{s_i,a_i}} [z_i^{f'}]| \leq 2 C_{1}$ almost surely.
Now, using the Bernstein's inequality (Lemma \ref{lem:bernstein-ineq}), together with a union bound over all $f'\in \mathcal{F}$, with probability at least $1-\delta$, we have
\begin{align}
\label{eq:least-squares-bern-main}
|\| T_{g} f - f' \|_{2,\mu}^2 - \frac{1}{N} \sum_{i=1}^N z_i^{f'}| \leq \sqrt{\frac{2 C_{1} \| T_{g} f - f' \|_{2,\mu}^2 \log(2|\mathcal{F}|/\delta)}{N}} + \frac{2C_{1} \log(2|\mathcal{F}|/\delta)}{3N},
\end{align}
for all $f'\in \mathcal{F}$. Setting $f'=\widehat{f}_{g}$, with probability at least $1-\delta/2$, we have
\begin{align}
\label{eq:least-squares-bern-1}
\| T_{g} f - \widehat{f}_{g} \|_{2,\mu}^2 \leq \frac{1}{N} \sum_{i=1}^N z_i^{\widehat{f}_{g}}+ \sqrt{\frac{2 C_{1} \| T_{g} f - \widehat{f}_{g} \|_{2,\mu}^2 \log(4|\mathcal{F}|/\delta)}{N}} + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{3N}.
\end{align}
Now we upper-bound $(1/N) \sum_{i=1}^N z_i^{\widehat{f}_{g}}$ in the following. Consider a function $\widetilde{f} \in \argmin_{h\in\mathcal{F}} \| h - T_{g} f \|_{2,\mu}^2$. Note that $\widetilde{f}$ is independent of the dataset. We note that our earlier first and second moment calculations hold true for $\widetilde{f}$, replacing $f'$, as well. Now, from \eqref{eq:least-squares-bern-main} setting $f'=\widetilde{f}$, with probability at least $1-\delta/2$ we have
\begin{align} \label{eq:least-squares-z-main}
\frac{1}{N} \sum_{i=1}^N z_i^{\widetilde{f}} -\| T_{g} f - \widetilde{f} \|_{2,\mu}^2 \leq \sqrt{\frac{2 C_{1} \| T_{g} f - \widetilde{f} \|_{2,\mu}^2 \log(4|\mathcal{F}|/\delta)}{N}} + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{3N}.
\end{align}
Suppose $(1/N) \sum_{i=1}^N z_i^{\widetilde{f}} \geq 2C_{1}\log(4|\mathcal{F}|/\delta)/N$ holds, then from \eqref{eq:least-squares-z-main} we get
\begin{align} \label{eq:least-squares-z-main-2}
\frac{1}{N} \sum_{i=1}^N z_i^{\widetilde{f}} -\| T_{g} f - \widetilde{f} \|_{2,\mu}^2 \leq \sqrt{ \| T_{g} f - \widetilde{f} \|_{2,\mu}^2 \cdot \frac{1}{N} \sum_{i=1}^N z_i^{\widetilde{f}}} + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{N}.
\end{align}
We note the following algebra fact: Suppose $x^2 - ax + b \leq 0$ with $b>0$ and $a^2\geq 4b$, then we have $x\leq a$. Taking $x=(1/N) \sum_{i=1}^N z_i^{\widetilde{f}}$ in this fact, from~\eqref{eq:least-squares-z-main-2} we get
\begin{align} \label{eq:least-squares-z-part-1}
\frac{1}{N} \sum_{i=1}^N z_i^{\widetilde{f}} \leq 3\| T_{g} f - \widetilde{f} \|_{2,\mu}^2 + \frac{4C_{1} \log(4|\mathcal{F}|/\delta)}{3N} \leq 3\| T_{g} f - \widetilde{f} \|_{2,\mu}^2 + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{N}.
\end{align}
Now suppose $(1/N) \sum_{i=1}^N z_i^{\widetilde{f}} \leq 2 C_{1}\log(4|\mathcal{F}|/\delta)/N$, then \eqref{eq:least-squares-z-part-1} holds immediately. Thus,~\eqref{eq:least-squares-z-part-1} always holds with probability at least $1-\delta/2$. Furthermore, recall $\widetilde{f} \in \argmin_{h\in\mathcal{F}} \| h - T_{g} f \|_{2,\mu}^2$, we have
\begin{align}
\frac{1}{N} \sum_{i=1}^N z_i^{\widetilde{f}} &\leq 3\| T_{g} f - \widetilde{f} \|_{2,\mu}^2 + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{N} \nonumber \\
&= 3\min_{h\in\mathcal{F}} \| h - T_{g} f \|_{2,\mu}^2 + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{N} \leq 3 \epsilon_{\textrm{c}} + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{N}
\label{eq:least-squares-z-part-2},
\end{align}
where the last inequality follows from the approximate robust Bellman completion assumption (Assumption \ref{assum-bellman-completion}).
We note that since $\widehat{f}_{g}$ is the least-squares regression solution, we know that $(1/N) \sum_{i=1}^N z_i^{\widehat{f}_{g}} \leq (1/N) \sum_{i=1}^N z_i^{\widetilde{f}}$. With this note in \eqref{eq:least-squares-z-part-2}, from \eqref{eq:least-squares-bern-1}, with probability at least $1-\delta$, we have
\begin{align*}
\| T_{g} f - \widehat{f}_{g} \|_{2,\mu}^2 &\leq 3 \epsilon_{\textrm{c}} + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{N}
\\&+ \sqrt{\frac{2 C_{1} \| T_{g} f - \widehat{f}_{g} \|_{2,\mu}^2 \log(4|\mathcal{F}|/\delta)}{N}} + \frac{2C_{1} \log(4|\mathcal{F}|/\delta)}{3N} \\
&\leq 3 \epsilon_{\textrm{c}} + \frac{3C_{1} \log(4|\mathcal{F}|/\delta)}{N} + \sqrt{\frac{3 C_{1} \| T_{g} f - \widehat{f}_{g} \|_{2,\mu}^2 \log(4|\mathcal{F}|/\delta)}{N}}.
\end{align*}
From the earlier algebra fact, taking $x=\| T_{g} f - \widehat{f}_{g} \|_{2,\mu}^2$, with probability at least $1-\delta$, we have
\begin{align*}
\| T_{g} f - \widehat{f}_{g} \|_{2,\mu}^2
&\leq 6 \epsilon_{\textrm{c}} + \frac{9C_{1} \log(4|\mathcal{F}|/\delta)}{N}.
\end{align*}
From the fact $\sqrt{x+y}\leq \sqrt{x}+\sqrt{y}$, with probability at least $1-\delta$, we get
\begin{align*}
\| T_{g} f - \widehat{f}_{g} \|_{2,\mu}
&\leq \sqrt{6 \epsilon_{\textrm{c}}} + \sqrt{\frac{9C_{1} \log(4|\mathcal{F}|/\delta)}{N}}.
\end{align*}
Using union bound for $f\in\mathcal{F}$ and $g\in\mathcal{G}$, with probability at least $1-\delta$, we finally obtain
\begin{align*}
\sup_{f\in\mathcal{F}} \sup_{g\in\mathcal{G}} \| T_{g} f - \widehat{f}_{g} \|_{2,\mu}
&\leq \sqrt{6 \epsilon_{\textrm{c}}} + \sqrt{\frac{18 C_{1} \log(2|\mathcal{F}||\mathcal{G}|/\delta)}{N}},
\end{align*}
which completes the least-squares generalization bound analysis.
\end{proof}
We are now ready to prove the main theorem.
\begin{proof}[\textbf{Proof of Theorem \ref{thm:tv-guarantee}}]
We let $V_k(s) = Q_k(s,\pi_k(s))$ for every $s\in\cS$. Since $\pi_k$ is the greedy policy w.r.t $Q_k$, we also have $V_k(s) = Q_k(s,\pi_k(s)) = \max_a Q_k(s,a)$. We recall that $V^*=V^{\pi^*}$ and $Q^*=Q^{\pi^*}$. We also recall from Section~\ref{sec:preliminaries} that $Q^{\pi^*}$ is a fixed-point of the robust Bellman operator $T$ defined in~\eqref{eq:robust-bellman-primal}. We also note that the same holds true for any stationary deterministic policy $\pi$ from~\citet{iyengar2005robust} that $Q^{\pi}$ satisfies $Q^{\pi}(s, a) = r(s, a) + \gamma \min_{P_{s,a} \in \mathcal{P}_{s,a}} \mathbb{E}_{s' \sim P_{s,a}} [ V^{\pi}(s')].$ We can now further use the dual form \eqref{eq:robust-bellman-dual-2} under Assumption \ref{assum-fail-state}. We first characterize the performance decomposition between $V^{\pi^*}$ and ${V}^{\pi_K}$. For a given $s_0\in\cS$, we observe that
\begin{align*}
&V^{\pi^*}(s_0) - {V}^{\pi_K}(s_0) = (V^{\pi^*}(s_0) - V_K(s_0) ) - (V^{\pi_K}(s_0) - V_K(s_0) )\\
&= (Q^{\pi^*}(s_0,\pi^*(s_0)) - Q_K(s_0,\pi_K(s_0)) ) - (Q^{\pi_K}(s_0,\pi_K(s_0)) - Q_K(s_0,\pi_K(s_0))) \\
&\stackrel{(a)}{\leq} Q^{\pi^*}(s_0,\pi^*(s_0)) - Q_K(s_0,\pi^*(s_0)) + Q_K(s_0,\pi_K(s_0)) - Q^{\pi_K}(s_0,\pi_K(s_0)) \\
&= Q^{\pi^*}(s_0,\pi^*(s_0)) - Q_K(s_0,\pi^*(s_0)) + Q_K(s_0,\pi_K(s_0)) - Q^{\pi^*}(s_0,\pi_K(s_0))
\\&\hspace{5cm}+ Q^{\pi^*}(s_0,\pi_K(s_0)) - Q^{\pi_K}(s_0,\pi_K(s_0))
\\
&\stackrel{(b)}{\leq} Q^{\pi^*}(s_0,\pi^*(s_0)) - Q_K(s_0,\pi^*(s_0)) + Q_K(s_0,\pi_K(s_0)) - Q^{\pi^*}(s_0,\pi_K(s_0)) \\&\hspace{2cm}+ \gamma \sup_{\eta} (\mathbb{E}_{s_1\sim P^o_{s_0,\pi_{K}(s_0)}}( (\eta- V^{\pi_K}(s_1))_+ - (\eta-V^{\pi^*}(s_1))_+)) \\
&\stackrel{(c)}{\leq} |Q^{\pi^*}(s_0,\pi^*(s_0)) - Q_K(s_0,\pi^*(s_0))| + |Q^{\pi^*}(s_0,\pi_K(s_0)) - Q_K(s_0,\pi_K(s_0))|
\\&\hspace{4cm}+ \gamma \mathbb{E}_{s_1\sim P^o_{s_0,\pi_{K}(s_0)}}( |V^{\pi^*}(s_1)- V^{\pi_K}(s_1)| ).
\end{align*}
$(a)$ follows from the fact that $\pi_{K}$ is the greedy policy with respect to $Q_{K}$. $(b)$ follows from the Bellman optimality equations and the fact $|\sup_{x} f(x)-\sup_{x} g(x)| \leq \sup_{x} |f(x) - g(x)|$. Finally, $(c)$ follows from the facts $(x)_+ - (y)_+ \leq (x-y)_+$ and $(x)_+ \leq |x|$ for any $x,y\in\mathbb{R}$.
We now recall the initial state distribution $d_0$. Thus, we have
\begin{align*} &\mathbb{E}_{s_0\sim d_0}[{V}^{\pi^*}] - \mathbb{E}_{s_0\sim d_0}[V^{\pi_K}] \leq
\\&\hspace{0.5cm} \mathbb{E}_{s_0\sim d_0} \bigg[ |Q^{\pi^*}(s_0,\pi^*(s_0)) - Q_{K}(s_0,\pi^*(s_0))| + |Q^{\pi^*}(s_0,\pi_K(s_0)) - Q_{K}(s_0,\pi_K(s_0))|
\\&\hspace{3cm}+ \gamma \mathbb{E}_{s_1\sim P^{o}_{s_0,\pi_{K}(s_0)}} (|V^{\pi^*}(s_1) - {V}^{\pi_K}(s_1) |) \bigg].
\end{align*}
Since $V^{\pi^*}(s) \geq V^{\pi_K}(s)$ for any $s\in\cS$, by telescoping we get
\begin{align}
&\mathbb{E}_{s_0\sim d_0}[{V}^{\pi^*}] - \mathbb{E}_{s_0\sim d_0}[V^{\pi_K}] \leq \sum_{h=0}^\infty \gamma^h \times \nonumber \\
&\hspace{1cm} \bigg(\mathbb{E}_{s\sim d_{h,\pi_{K}}} [ |Q^{\pi^*}(s,\pi^*(s)) - Q_{K}(s,\pi^*(s))| + |Q^{\pi^*}(s,\pi_K(s)) - Q_{K}(s,\pi_K(s))| ] \bigg), \label{eq:perf-diff-1}
\end{align}
where $d_{h,\pi_K}\in \Delta(\cS)$ for all natural numbers $h\geq 0$ is defined as
\begin{equation*}
d_{h,\pi_{K}} =
\begin{cases}
d_0 & \text{if $h=0$},\\
P^{o}_{s',\pi_{K}(s')} & \text{otherwise, with } s'\sim d_{h-1,\pi_{K}}.
\end{cases}
\end{equation*}
We emphasize that the state distribution $d_{h,\pi_K}$'s are different from the discounted state-action occupancy distributions. We note that a similar state distribution proof idea is used in \citet{agarwal2019reinforcement}.
Recall $\|f\|_{p,\nu}^2 = (\mathbb{E}_{s,a\sim\nu} |f(s,a)|^p)^{1/p}$, where $\nu\in\Delta(\cS\times\cA)$. With this we have
\begin{align}
\label{eq:thm-bound-part-1}
&\mathbb{E}_{s_0\sim d_0}[{V}^{\pi^*}] - \mathbb{E}_{s_0\sim d_0}[V^{\pi_K}] \leq \sum_{h=0}^\infty \gamma^h \bigg( \|Q^{\pi^*} - Q_{K}\|_{1,d_{h,\pi_{K}}\circ \pi^* } + \|Q^{\pi^*} - Q_{K}\|_{1,d_{h,\pi_{K}}\circ \pi_K } \bigg),
\end{align}
where the state-action distributions $d_{h,\pi_{K}}\circ \pi^*(s,a)\propto d_{h,\pi_{K}}(s) \mathbbm{1}\{ a=\pi^*(s)\}$ and $d_{h,\pi_{K}}\circ \pi_K(s,a)\propto d_{h,\pi_{K}}(s) \mathbbm{1}\{ a=\pi_K(s)\}$ directly follows by comparing with \eqref{eq:perf-diff-1}.
We now bound one of the RHS terms above by bounding for any state-action distribution $\nu$ satisfying Assumption \ref{assum-concentra-condition} (in particular the following bound is true for $d_{h,\pi_{K}}\circ \pi^*$ or $d_{h,\pi_{K}}\circ \pi_K$ in \eqref{eq:perf-diff-1}):
\begin{align}
&\|Q^{\pi^*} - Q_{K}\|_{1,\nu} \leq \|Q^{\pi^*} - T Q_{K-1}\|_{1,\nu} + \|T Q_{K-1} - Q_{K}\|_{1,\nu} \nonumber \\
&\stackrel{(a)}{\leq} \|Q^{\pi^*} - T Q_{K-1}\|_{1,\nu} + \sqrt{C}\|T Q_{K-1} - Q_{K}\|_{1,\mu} \nonumber \\
&= (\mathbb{E}_{s,a\sim\nu} |Q^{\pi^*}(s,a) - T Q_{K-1}(s,a)|) + \sqrt{C}\|T Q_{K-1} - Q_{K}\|_{1,\mu} \nonumber \\
&\stackrel{(b)}{\leq} (\mathbb{E}_{s,a\sim\nu} \gamma \sup_{\eta}|
\mathbb{E}_{s'\sim P^o_{s,a}} ( (\eta-\max_{a'} Q_{K-1}(s',a'))_+ - (\eta-\max_{a'} Q^{\pi^*}(s',a'))_+)
|)
\nonumber \\&\hspace{6cm}+ \sqrt{C}\|T Q_{K-1} - Q_{K}\|_{1,\mu} \nonumber \\
&\stackrel{(c)}{\leq} (\mathbb{E}_{s,a\sim\nu} |
\mathbb{E}_{s'\sim P^o_{s,a}} (\max_{a'} Q^{\pi^*}(s',a')-\max_{a'} Q_{K-1}(s',a'))_+|) + \sqrt{C}\|T Q_{K-1} - Q_{K}\|_{1,\mu} \nonumber \\
&\stackrel{(d)}{\leq} \gamma (\mathbb{E}_{s,a\sim\nu} \mathbb{E}_{s'\sim {P}^o_{s,a}} \max_{a'} |Q^{\pi^*}(s',a') - Q_{K-1}(s',a')| ) + \sqrt{C}\|T Q_{K-1} - Q_{K}\|_{1,\mu} \nonumber\\
&\stackrel{(e)}{\leq} \gamma \|Q^{\pi^*} - Q_{K-1}\|_{1,\nu'} + \sqrt{C}\|T Q_{K-1} - Q_{K}\|_{1,\mu} \nonumber \\
\label{eq:pf-sketech-stp4-1}
&\stackrel{(f)}{\leq} \gamma \|Q^{\pi^*} - Q_{K-1}\|_{1,\nu'} + \sqrt{C}\|T_{g_{K-1}} Q_{K-1} - Q_{K}\|_{2,\mu}+ \sqrt{C}\|T Q_{K-1} - T_{g_{K-1}} Q_{K-1}\|_{1,\mu},
\end{align}
where $(a)$ follows by the concentratability assumption (Assumption \ref{assum-concentra-condition}), $(b)$ from Bellman equation, operator $T$, and the fact $|\sup_{x} p(x)-\sup_{x} q(x)| \leq \sup_{x} |p(x) - q(x)|$,
$(c)$ from the fact $|(x)_+ - (y)_+| \leq |(x-y)_+|$ for any $x,y\in\mathbb{R}$,
$(d)$ follows by Jensen's inequality and by the facts $|\sup_{x} p(x)-\sup_{x} q(x)| \leq \sup_{x} |p(x) - q(x)|$ and $(x)_+ \leq |x|$ for any $x,y\in\mathbb{R}$, and $(e)$ by defining the distribution $\nu'$ as $\nu'(s',a')=\sum_{s,a} \nu(s,a) {P}^o_{s,a}(s') \mathbbm{1} \{ a'= \argmax_b |Q^{\pi^*}(s',b)- Q_{K-1}(s',b)| \}$, and $(f)$ using the fact that $\|\cdot\|_{1,\mu}\leq \|\cdot\|_{2,\mu}$.
Now, by recursion until iteration 0, we get \begin{align} \|&Q^{\pi^*} - Q_{K}\|_{1,\nu} \leq \gamma^K \sup_{\bar{\nu}} \|Q^{\pi^*} - Q_{0}\|_{1,\bar{\nu}} + \sqrt{C} \sum_{t=0}^{K-1} \gamma^t \|T Q_{K-1-t} - T_{g_{K-1-t}} Q_{K-1-t}\|_{1,\mu} \nonumber
\\&\hspace{4cm}+ \sqrt{C} \sum_{t=0}^{K-1} \gamma^t \| T_{g_{K-1-t}} Q_{K-1-t} - Q_{K-t}\|_{2,\mu}
\nonumber\\
&\stackrel{(a)}{\leq} \frac{\gamma^K }{1-\gamma} + \sqrt{C} \sum_{t=0}^{K-1} \gamma^t \|T Q_{K-1-t} - T_{g_{K-1-t}} Q_{K-1-t}\|_{1,\mu}
\nonumber\\&\hspace{4cm}+ \sqrt{C} \sum_{t=0}^{K-1} \gamma^t \| T_{g_{K-1-t}} Q_{K-1-t} - Q_{K-t}\|_{2,\mu}
\nonumber\\&\stackrel{(b)}{\leq} \frac{\gamma^K }{1-\gamma} + \frac{\sqrt{C} }{1-\gamma} \sup_{f\in\mathcal{F}} \|T f - T_{\widehat{g}_{f}} f\|_{1,\mu}
+ \frac{\sqrt{C} }{1-\gamma} \sup_{f\in\mathcal{F}} \| T_{\widehat{g}_{f}} f - \widehat{f}_{\widehat{g}_{f}}\|_{2,\mu}
\nonumber\\&\leq \frac{\gamma^K }{1-\gamma} + \frac{\sqrt{C} }{1-\gamma} \sup_{f\in\mathcal{F}} \|T f - T_{\widehat{g}_{f}} f\|_{1,\mu}
+ \frac{\sqrt{C} }{1-\gamma} \sup_{f\in\mathcal{F}} \sup_{g\in\mathcal{G}} \| T_{g} f - \widehat{f}_{g}\|_{2,\mu}. \label{eq:thm-bound-part-2} \end{align} where $(a)$ follows since $|Q^{\pi^*}(s,a)|\leq 1/(1-\gamma), Q_{0}(s,a)= 0$, and $(b)$ follows since $\widehat{g}_{f}$ is the dual variable function from the algorithm for the state-action value function $f$ and $\widehat{f}_{g}$ as the least squares solution from the algorithm for the state-action value function $f$ and dual variable function $g$ pair.
The proof is now complete combining \eqref{eq:thm-bound-part-1} and \eqref{eq:thm-bound-part-2} with Lemma \ref{lem:erm-high-prob-bound} and Lemma \ref{lem:least-squares-generalization-bound}.
\end{proof}
\section{Related Works}
\label{sec:app:related-works}
Here we provide a more detailed description of the related work to complement what we listed in the introduction (Section~\ref{sec:introduction}).
\textbf{Offline RL:} The problem of learning the optimal policy only using an offline dataset is first addressed under the generative model assumption \citep{singh1994upper, AzarMK13, haskell2016empirical, sidford2018near, agarwal2020model,li2020breaking, kalathil2021empirical}. This assumption requires generating the same uniform number of next-state samples for each and every state-action pairs.
To account for large state spaces, there are number of works \citep{antos2008learning, bertsekas2011approximate, lange2012batch, chen2019information, xie2020q, levine2020offline, xie2021bellman} that utilize function approximation under similar assumption, concentratability assumption \citep{chen2019information} in which the data distribution $\mu$ sufficiently covers the discounted state-action occupancy. There is rich literature \citep{munos08a, farahmand2010error, lazaric2012finite, chen2019information, liu2020provably, xie2021bellman} in the conquest of identifying and improving these necessary and sufficient assumptions for offline RL that use variations of Fitted Q-Iteration (FQI) algorithm \citep{gordon1995stable, ernst2005tree}.
There is also rich literature \citep{fujimoto2019off, kumar2019stabilizing, kumar2020conservative, yu2020mopo, zhang2021towards} that develop offline deep RL algorithms focusing on the algorithmic and empirical aspects and propose multitude heuristic approaches to advance the field. All these results assume that the offline data is generated according to a single model and the goal is to find the optimal policy for the MDP with the same model. In particular, none of these works consider the \textit{offline robust RL problem} where the offline data is generated according to a (training) model which can be different from the one in testing, and the goal is to learn a policy that is robust w.r.t.~an uncertainty set.
\textbf{Robust RL:} To address the parameter uncertainty problem,~\citet{iyengar2005robust} and~\citet{nilim2005robust} introduced the RMDP framework.~\citet{iyengar2005robust} showed that the optimal robust value function and policy can be computed using the robust counterparts of the standard value iteration and policy iteration algorithms. To tackle the parameter uncertainty problem, other works considered distributionally robust setting \citep{xu2010distributionally}, modified policy iteration \citep{kaufman2013robust}, and {more general uncertainty set} \citep{wiesemann2013robust}. These initial works mainly focused on the planning problem (known transition probability dynamics) in the tabular setting.~\citet{tamar2014scaling} proposed linear function approximation method to solve large RMDPs. Though this work suggests a sampling based approach, a general model-free learning algorithm and analysis was not included.~\citet{roy2017reinforcement} proposed the robust versions of the classical model-free reinforcement learning algorithms, such as Q-learning, SARSA, and TD-learning in the tabular setting. They also proposed function approximation based algorithms for the policy evaluation. However, this work does not have a policy iteration algorithm with provable guarantees for learning the optimal robust policy.~\citet{derman2018soft} introduced soft-robust actor-critic algorithms using neural networks, but does not provide any global convergence guarantees for the learned policy.~\citet{tessler2019action} proposed a min-max game framework to address the robust learning problem focusing on the tabular setting.~\citet{lim2019kernel} proposed a kernel-based RL algorithm for finding the robust value function in a batch learning setting.~\citet{Mankowitz2020Robust} employed an entropy-regularized policy optimization algorithm for continuous control using neural network, but does not provide any provable guarantees for the learned policy.~\citet{panaganti2020robust} proposed least-squares policy iteration method to handle large state-action space in robust RL, but only provide asymptotic policy evaluation convergence guarantees whereas \citet{panaganti2020robust} provide finite time convergence for the policy iteration to optimal robust value.
\textbf{Other robust RL related works:} Robust control is a well-studied area in the classical control theory~\citep{zhou1996robust, dullerud2013course}. Recently, there are some interesting works that address the robust RL problem using this framework, especially focusing on the linear quadratic regulator setting \citep{zhang2020policy}. Risk sensitive RL algorithms \citep{borkar2002q,prashanth2016variance,fei2021exponential} and adversarial RL algorithms \citep{pinto2017robust,zhang2021provably,huang2022robust} also address the robustness problem implicitly under different frameworks which are independent from RMDPs. \citep{zhang2022corruption} addresses the problem of \textit{corruption-robust} offline RL problem, where an adversary is allowed to change a fraction of the samples of an offline RL dataset and the goal is to find the optimal policy for the nominal linear MDP model (according to which the offline data is generated).
Our framework and approach of robust MDP is significantly different from these line of works.
\textbf{This work:} The works that are closest to ours are by \citet{zhou2021finite,yang2021towards,panaganti22a} that address the robust RL problem in a tabular setting under the generative model assumption. Due to the generative model assumption, the offline data has the same uniform number of samples corresponding to each and every state-action pair, and tabular setting allows the estimation of the uncertainty set followed by solving the planning problem. Our work is significantly different from these in the following way: $(i)$ we consider a robust RL problem with arbitrary large state space, instead of the small tabular setting, $(ii)$ we consider a true offline RL setting where the state-action pairs are sampled according to an arbitrary distribution, instead of using the generative model assumption, $(iii)$ we focus on a function approximation approach where the goal is to directly learn optimal robust value/policy using function approximation techniques, instead of solving the tabular planning problem with the estimated model. \textit{To the best of our knowledge, this is the first work that addresses the offline robust RL problem with arbitrary large state space using function approximation, with provable guarantees on the performance of the learned policy.}
\section{Experiment Details}
\label{appendix:simulations}
We provide more detailed and practical version of our RFQI algorithm (Algorithm \ref{alg:RFQI-Algorithm}) in this section. We also provide more experimental results evaluated on \textit{Cartpole}, \textit{Hopper}, and \textit{Half-Cheetah} OpenAI Gym Mujoco \citep{brockman2016openai} environments.
We provide our code in \textbf{github webpage} \url{https://github.com/zaiyan-x/RFQI} containing instructions to reproduce all results in this paper.
We implemented our RFQI algorithm based on the architecture of Batch Constrained deep Q-learning (BCQ) algorithm \citep{fujimoto2019off} \footnote{Available at \url{https://github.com/sfujim/BCQ}} and Pessimistic Q-learning (PQL) algorithm \citep{liu2020provably} \footnote{Available at \url{https://github.com/yaoliucs/PQL}}. We note that PQL algorithm (with $b=0$ filtration thresholding \citep{liu2020provably}) and BCQ algorithm are the practical versions of FQI algorithm with neural network architecture.
\subsection{RFQI Practical Algorithm} \label{appendix-sub:rfqi-prac-implementation}
We provide the practical version of our RFQI algorithm in Algorithm \ref{alg:prac-RFQI-Algorithm} and highlight the difference with BCQ and PQL algorithms in \bluetext{blue} (steps 8 and 9).
\begin{algorithm}[t]
\caption{RFQI Practical Algorithm}
\label{alg:prac-RFQI-Algorithm}
\begin{algorithmic}[1]
\STATE \textbf{Input:} Offline dataset $\mathcal{D}$, radius of robustness $\rho$, maximum perturbation $\Phi$, target update rate $\tau$, mini-batch size $N$, maximum number of iterations $K$, number of actions $u$.
\STATE \textbf{Initialize:} Two state-action neural networks $Q_{\theta_1}$ and $Q_{\theta_2}$, \bluetext{one dual neural network $g_{\theta_3}$} \\policy (perturbation) model: $\xi_{\phi} \in [-\Phi, \Phi]),$ and action VAE $G^a_{\omega}$, with random parameters $\theta_1$, $\theta_2$, $\phi$, $\omega$, and target networks $Q_{\theta'_1}, Q_{\theta'_2}$, $\xi_{\phi'}$ with $\theta'_1 \leftarrow \theta_1, \theta'_2 \leftarrow \theta_2$, $\phi' \leftarrow \phi$
\FOR {$k=1,\cdots,K$ }
\STATE Sample a minibatch $B$ with $N$ samples from $\mathcal{D}$.
\STATE Train $\omega \leftarrow\argmin_{\omega} ELBO(B;G^a_{\omega}).$ Sample $u$ actions $a_i'$ from $G^a_{\omega} (s')$ for each $s'$.
\STATE Perturb $u$ actions $a_i'=a_i' + \xi_{\phi}(s',a_i')$.
\STATE Compute next-state value target for each $s'$ in $B$: $$V_t = \max_{a'_i} (0.75\cdot \min\{ Q_{\theta'_1}, Q_{\theta'_2} \} + 0.25\cdot \max\{ Q_{\theta'_1}, Q_{\theta'_2} \}).$$
\STATE \vspace{-0.3cm} \bluetext{$\theta_3 \leftarrow \argmin_{\theta} \sum [\max\{g_{\theta}(s,a) - V_t(s'), 0\} - (1-\rho) g_{\theta}(s,a)].$}
\STATE \bluetext{Compute next-state Q target for each $(s, a, r, s' )$ pair in $B$: $$ Q_t(s,a) = r - \gamma \cdot \max\{g_{\theta_3}(s,a) - V_t(s'), 0\} + \gamma (1-\rho) g_{\theta_3}(s,a). $$}
\STATE \vspace{-0.3cm} $\theta \leftarrow \argmin_{\theta} \sum (Q_t(s,a) - Q_{\theta}(s,a))^2.$
\STATE Sample $u$ actions $a_i$ from $G^a_{\omega} (s)$ for each $s$.
\STATE $\phi \leftarrow \argmax_{\phi} \sum \max_{a_i} Q_{\theta_1}(s,a_i+\xi_{\phi}(s,a_i))$.
\STATE Update target network: $\theta'=(1-\tau)\theta' + \tau\theta, \phi'=(1-\tau)\phi' + \tau\phi$.
\ENDFOR
\STATE \textbf{Output policy:} Given $s$, sample $u$ actions $a_i$ from $G^a_{\omega} (s)$. Select action $a = \argmax_{a_i} Q_{\theta_1}(s,a_i+\xi_{\phi}(s,a_i))$.
\end{algorithmic}
\end{algorithm}
\textbf{RFQI algorithm implementation details}: The Variational Auto-Encoder (VAE) $G^a_\omega$ \citep{kingma2013auto} is defined by two networks, an encoder $E_{\omega_1}(s,a)$ and decoder $D_{\omega_2}(s, z)$, where $\omega = \{\omega_1, \omega_2\}$. The encoder outputs mean and standard deviation, $(\mu, \sigma) = E_{\omega_1}(s,a)$, of a normal distribution.
A latent vector $z$ is sampled from the standard normal distribution and for a state $s$, the decoder maps them to an action $D_{\omega_2}:(s, z)\mapsto\tilde{a}$. Then the evidence lower bound ($ELBO$) of VAE is given by $ELBO(B;G^a_{\omega})=\sum_{B} (a - \tilde{a})^2 + D_{\text{KL}}(\mathcal{N}(\mu, \sigma),\mathcal{N}(0,1)),$ where $\mathcal{N}$ is the normal distribution with mean and standard deviation parameters. We refer to \citep{fujimoto2019off} for more details on VAE. We also use the default VAE architecture from BCQ algorithm \citep{fujimoto2019off} and PQL algorithm \citep{liu2020provably} in our RFQI algorithm.
We now focus on the additions described in blue (steps 8 and 9) in Algorithm \ref{alg:prac-RFQI-Algorithm}. For all the other networks we use default architecture from BCQ algorithm \citep{fujimoto2019off} and PQL algorithm \citep{liu2020provably} in our RFQI algorithm.\\
(1) In each iteration $k$, we solve the dual variable function $g_\theta$ optimization problem (step 4 in Algorithm \ref{alg:RFQI-Algorithm}, step 8 in Algorithm \ref{alg:prac-RFQI-Algorithm}) implemented by ADAM \citep{kingma2014adam} on the minibatch $B$ with the learning rate $l_1$ mentioned in Table \ref{tab:hp}.\\
(2) Our state-action value target function corresponds to the robust state-action value target function described in \eqref{eq:loss-rfqi-1}. This is reflected in step 9 of Algorithm \ref{alg:prac-RFQI-Algorithm}. The state-action value function $Q_\theta$ optimization problem (step 5 in Algorithm \ref{alg:RFQI-Algorithm}, step 9 in Algorithm \ref{alg:prac-RFQI-Algorithm}) is implemented by ADAM \citep{kingma2014adam} on the minibatch $B$ with the learning rate $l_2$ mentioned in Table \ref{tab:hp}.
\begin{table}[H]
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
Environment & Discount & Learning rates & Q Neural nets & Dual Neural nets \\
& $\gamma$ & $[l_1,l_2]$ & $\theta_1=\theta_2=[h_1,h_2]$ & $\theta_3=[h_1,h_2]$ \\
\hline & &&& \vspace{-0.3cm}\\
CartPole & $0.99$ & $[10^{-3},10^{-3}]$ & $[400, 300]$ & $[64, 64]$\\
Hopper & $0.99$ & $[10^{-3},8\times 10^{-4}]$ & $[400, 300]$ & $[64, 64]$\\
& & $[3\times 10^{-4},6\times10^{-4}]$ & & \\
Half-Cheetah & $0.99$ & $[10^{-3},8\times 10^{-4}]$ & $[400, 300]$ & $[64, 64]$\\
& & $[3\times 10^{-4},6\times10^{-4}]$ & & \\
\hline
\end{tabular}
\end{center}\caption{Details of hyper-parameters in FQI and RFQI algorithms experiments.} \label{tab:hp}
\end{table}
\textbf{Hyper-parameters details}: We now give the description of hyper-parameters used in our codebase in Table \ref{tab:hp}. We use same hyper-parameters across different algorithms. Across all learning algorithms we use $\tau=0.005$ for the target network update, $K=5\times10^5$ for the maximum iterations, $|\mathcal{D}|=10^6$ for the offline dataset, $|B|=1000$ for the minibatch size. We used grid-search for $\rho$ in $\{0.2,0.3,\cdots,0.6\}$. We also picked best of the two sets of learning rates mentioned in Table \ref{tab:hp}. For all the other hyper-parameters we use default values from BCQ algorithm \citep{fujimoto2019off} and PQL algorithm \citep{liu2020provably} in our RFQI algorithm that can be found in our code.
\textbf{Offline datasets}: Now we discuss the offline dataset used in the our training of FQI and RFQI algorithms. For the fair comparison in every plot, we train both FQI and RFQI algorithms on same offline datasets.
\textit{Cartpole dataset} $\mathcal{D}_{\textrm{c}}$: We first train proximal policy optimization (PPO) \citep{schulman2017proximal} algorithm, under default RL baseline zoo \citep{rl-zoo3} parameters. We then generate the Cartpole dataset $\mathcal{D}_{\textrm{c}}$ with $10^5$ samples using an $\epsilon$-greedy ($\epsilon=0.3$) version of this PPO trained policy. We note that this offline dataset contains non-expert behavior meeting the richness of the data-generating distribution assumption in practice.
\textit{Mixed dataset} $\mathcal{D}_{\textrm{m}}$: For the MuJoCo environments, \textit{Hopper} and \textit{Half-Cheetah}, we increase the richness of the dataset since these are high dimensional problems. We first train soft actor-critic (SAC) \citep{haarnoja2018soft} algorithm, under default RL baseline zoo \citep{rl-zoo3} parameters, with replay buffer updated by a fixed $\epsilon$-greedy ($\epsilon=0.1$) policy with the model parameter \textit{actuator\_ctrlrange} set to $[-0.85,0.85]$. We then generate the mixed dataset $\mathcal{D}_{\textrm{m}}$ with $10^6$ samples from this $\epsilon$-greedy ($\epsilon=0.3$) SAC trained policy. We note that such a dataset generation gives more diverse set of observations than the process of $\mathcal{D}_{\textrm{c}}$ generation for fair comparison between FQI and RFQI algorithms.
\textit{D4RL dataset} $\mathcal{D}_{\textrm{d}}$: We consider the \textit{hopper-medium} and \textit{halfcheetah-medium} offline datasets in \citep{fu2020d4rl} which are benchmark datasets in offline RL literature \citep{fu2020d4rl,levine2020offline,liu2020provably}. These `medium' datasets are generated by first training a policy online using Soft Actor-Critic \citep{haarnoja2018soft}, early-stopping the training, and collecting $10^6$ samples from this partially-trained policy. We refer to \citep{fu2020d4rl} for more details.
We end this section by mentioning the software and hardware configurations used. The training and evaluation is done using three computers with the following configuration. Operating system is Ubuntu 18.04 and Lambda Stack; main softwares are PyTorch, Caffe, CUDA, cuDNN, Numpy, Matplotlib; processor is AMD Threadripper 3960X (24 Cores, 3.80 GHz); GPUs are 2x RTX 2080 Ti; memory is 128GB RAM; Operating System Drive is 1 TB SSD (NVMe); and Data Drive is 4TB HDD.
\subsection{More Experimental Results}
\begin{figure*}[ht]
\vspace{-0.5cm}
\centering
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/CartPolePerturbed-v0_action.pdf}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/CartPolePerturbed-v0_force_mag.pdf}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/CartPolePerturbed-v0_length.pdf}
\end{minipage}
\captionof{figure}{ \textit{Cartpole simulation results on offline dataset $\mathcal{D}_{\textrm{c}}$.} Average cumulative reward in $20$ episodes versus different model parameter perturbations mentioned in the respective titles.}
\label{fig:cartpole}
\centering
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/hopper_action.pdf}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/hopper_actuator_ctrlrange.pdf}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/hopper_foot_joint_stiffness.pdf}
\end{minipage}
\captionof{figure}{ \textit{Hopper simulation results on offline dataset $\mathcal{D}_{\textrm{m}}$.} Average cumulative reward in $20$ episodes versus different model parameter perturbations mentioned in the respective titles.}
\label{fig:hopper-mixed}
\end{figure*}
Here we provide more experimental results and details in addition to Fig. \ref{fig:cart_force-in}-\ref{fig:hopper_leg} in Section \ref{sec:main-experiments}.
For the \textit{Cartpole}, we compare RFQI algorithm against the non-robust RL algorithms FQI and DQN, and the soft-robust RL algorithm proposed in \cite{derman2018soft}. We trained FQI and RFQI algorithms on the dataset $\mathcal{D}_{\textrm{c}}$ (a detailed description of data set is provided in Appendix \ref{appendix-sub:rfqi-prac-implementation}).
We test the robustness of the algorithms by changing the parameters \textit{force\_mag} (to model external force disturbance), \textit{length} (to model change in pole length), and also by introducing action perturbations (to model actuator noise). The nominal value of \textit{force\_mag} and \textit{length} parameters are $10$ and $0.5$ respectively. Fig. \ref{fig:cartpole} shows superior robust performance of RFQI compared to the non-robust FQI and DQN.
For example, consider the action perturbation performance plot in Fig. \ref{fig:cartpole} where RFQI algorithm improves by $75\%$ compared to FQI algorithm in average cumulative reward for a $40\%$ chance of action perturbation.
We note that we found $\rho=0.5$ is the best from grid-search for RFQI algorithm. The RFQI performance is similar to that of soft-robust DQN. We note that soft-robust DQN algorithm is an online deep RL algorithm (and not an offline RL algorithm) and has no provable performance guarantee. Moreover, soft-robust DQN algorithm requires generating online data according a number of models in the uncertainty set, whereas RFQI only requires offline data according to a single nominal training model.
Before we proceed to describe our results on the OpenAI Gym MuJoCo \citep{brockman2016openai} environments \textit{Hopper} and \textit{Half-Cheetah}, we first mention their model parameters and its corresponding nominal values in Table \ref{tab:model-parameters}. The model parameter names are self-explanatory, for example, stiffness control on the leg joint is the \textit{leg\_joint\_stiffness}, range of actuator values is the \textit{actuator\_ctrlrange}. The front and back parameters in Half-Cheetah are for the front and back legs. We refer to the perturbed environments provided in our code and the \textit{hopper.xml, halfcheetah.xml} files in the environment assets of OpenAI Gym MuJoCo \citep{brockman2016openai} for more information regarding these model parameters.
\begin{table}[H]
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
Environment & Model parameter & Nominal range/value \\
\hline & & \vspace{-0.3cm}\\
Hopper & \textit{actuator\_ctrlrange} & $[-1,1]$ \\
& \textit{foot\_joint\_stiffness} & $0$ \\
& \textit{leg\_joint\_stiffness} & $0$ \\
& \textit{thigh\_joint\_stiffness} & $0$ \\
& \textit{joint\_damping} & $1$ \\
& \textit{joint\_frictionloss} & $0$ \\
\hline & & \vspace{-0.3cm}\\
& \textit{joint\_frictionloss} & $0$ \\
Half-Cheetah & front \textit{actuator\_ctrlrange} & $[-1,1]$ \\
& back \textit{actuator\_ctrlrange} & $[-1,1]$ \\
& front \textit{joint\_stiffness} = (\textit{thigh\_joint\_stiffness}, & \\
& \textit{shin\_joint\_stiffness}, \textit{foot\_joint\_stiffness}) & $(180,120,60)$ \\
& back \textit{joint\_stiffness} = (\textit{thigh\_joint\_stiffness}, & \\
& \textit{shin\_joint\_stiffness}, \textit{foot\_joint\_stiffness}) & $(240,180,120)$ \\
& front \textit{joint\_damping} = (\textit{thigh\_joint\_damping}, & \\
& \textit{shin\_joint\_damping}, \textit{foot\_joint\_damping}) & $(4.5,3.0,1.5)$ \\
& back \textit{joint\_damping} = (\textit{thigh\_joint\_damping}, & \\
& \textit{shin\_joint\_damping}, \textit{foot\_joint\_damping}) & $(6.0,4.5,3.0)$ \\
\hline
\end{tabular}
\end{center}\caption{Details of model parameters for \textit{Hopper} and \textit{Half-Cheetah} environments.} \label{tab:model-parameters}
\end{table}
\vspace{-0.5cm}
For the \textit{Hopper}, we compare RFQI algorithm against the non-robust RL algorithms FQI and TD3 \citep{fujimoto2018addressing}. We trained FQI and RFQI algorithms on the mixed dataset $\mathcal{D}_{\textrm{m}}$ (a detailed description of dataset provided in Appendix \ref{appendix-sub:rfqi-prac-implementation}). We note that we do not compare with soft robust RL algorithms because of its poor performance on MuJoCo environments in the rest of our figures. We test the robustness of the algorithm by introducing action perturbations, and by changing the model parameters \textit{actuator\_ctrlrange}, \textit{foot\_joint\_stiffness}, and \textit{leg\_joint\_stiffness}.
Fig. \ref{fig:hopper_leg} and Fig. \ref{fig:hopper-mixed} shows RFQI algorithm is consistently robust compared to the non-robust algorithms. We note that we found $\rho=0.5$ is the best from grid-search for RFQI algorithm. The average episodic reward of RFQI remains almost the same initially, and later decays much less and gracefully when compared to FQI and TD3 algorithms. For example, in plot 3 in Fig. \ref{fig:hopper-mixed}, at the \textit{foot\_joint\_stiffness} parameter value $15$, the episodic reward of FQI is only around $1400$ whereas RFQI achieves an episodic reward of $3200$. Similar robust performance of RFQI can be seen in other plots as well. We also note that TD3 \citep{fujimoto2019off} is a powerful off-policy policy gradient algorithm that relies on large $10^6$ replay buffer of online data collection, unsurprisingly performs well initially with less perturbation near the nominal models.
In order to verify the effectiveness and consistency of our algorithm across different offline dataset, we repeat the above experiments, on additional OpenAI Gym MuJoCo \citep{brockman2016openai} environment \textit{Half-Cheetah}, using D4RL dataset $\mathcal{D}_{\mathrm{d}}$ (a detailed description of dataset provided in Appendix \ref{appendix-sub:rfqi-prac-implementation}) which are benchmark in offline RL literature \citep{fu2020d4rl,levine2020offline,liu2020provably} than our mixed dataset $\mathcal{D}_{\mathrm{m}}$. Since D4RL dataset is a benchmark dataset for offline RL algorithms, here we focus only on the comparison between the two offline RL algorithms we consider, our RFQI algorithm and its non-robust counterpart FQI algorithm. We now showcase the results on \textit{Hopper} and \textit{Half-Cheetah} for this setting.
\begin{figure*}[t]
\centering
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/hopper_gravity.pdf}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/hopper_joint_damping.pdf}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/hopper_joint_frictionloss.pdf}
\end{minipage}
\captionof{figure}{ \textit{Hopper evaluation simulation results on offline dataset $\mathcal{D}_{\textrm{d}}$.} Average cumulative reward in $20$ episodes versus different model parameter perturbations mentioned in the respective titles.}
\label{fig:hopper-d4rl}
\centering
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/halfcheetah_front_joint_stiffness.pdf}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/halfcheetah_back_joint_stiffness.pdf}
\end{minipage}
\begin{minipage}{.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/halfcheetah_back_actuator_ctrlrange.pdf}
\end{minipage}
\captionof{figure}{ \textit{Half-Cheetah evaluation simulation results on offline dataset $\mathcal{D}_{\textrm{d}}$.} Average cumulative reward in $20$ episodes versus different model parameter perturbations mentioned in the respective titles.}
\label{fig:half-cheetah-d4rl}
\end{figure*}
For the \textit{Hopper}, we test the robustness by changing the model parameters \textit{gravity}, \textit{joint\_damping}, and \textit{joint\_frictionloss}.
Fig. \ref{fig:hopper-d4rl} shows
RFQI algorithm is consistently robust compared to the non-robust FQI algorithm. We note that we found $\rho=0.5$ is the best from grid-search for RFQI algorithm.
The average episodic reward of RFQI remains almost the same initially, and later decays much less and gracefully when compared to FQI algorithm. For example, in plot 2 in Fig. \ref{fig:hopper-d4rl}, for the $30\%$ change in \textit{joint\_damping} parameter, the episodic reward of FQI is only around $1400$ whereas RFQI achieves an episodic reward of $3000$ which is almost the same as for unperturbed model. Similar robust performance of RFQI can be seen in other plots as well.
For the \textit{Half-Cheetah}, we test the robustness by changing the model parameters \textit{joint\_stiffness} of front and back joints, and \textit{actuator\_ctrlrange} of back joint.
Fig. \ref{fig:half-cheetah-d4rl} shows
RFQI algorithm is consistently robust compared to the non-robust FQI algorithm. We note that we found $\rho=0.3$ is the best from grid-search for RFQI algorithm.
For example, in plot 1 in Fig. \ref{fig:half-cheetah-d4rl}, RFQI episodic reward stays at around $5500$ whereas FQI drops faster to $4300$ for more than $50\%$ change in the nominal value. Similar robust performance of RFQI can be seen in other plots as well.
\begin{figure*}[h]
\begin{minipage}{.47\textwidth}
\begin{tabular}{@{}c@{}}
\includegraphics[width=0.5\linewidth]{figures/hopper-d4rl-actuator-ctrlrange.pdf}
\includegraphics[width=0.5\linewidth]{figures/hopper-d4rl-thigh-joint-stiffness.pdf}
\end{tabular}
\captionof{figure}{Similar performance of RFQI and FQI in Hopper on dataset $\mathcal{D}_{\textrm{d}}$ w.r.t. parameters \textit{actuator\_ctrlrange} and \textit{thigh\_joint\_stiffness}.}
\label{fig:hopper-d4rl-same-robustness}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}{.47\textwidth}
\begin{tabular}{@{}c@{}}
\includegraphics[width=0.5\linewidth]{figures/halfcheetah_joint_damping.pdf}
\includegraphics[width=0.5\linewidth]{figures/halfcheetah_joint_frictionloss.pdf}
\end{tabular}
\captionof{figure}{Similar performance of RFQI and FQI in Half-Cheetah on dataset $\mathcal{D}_{\textrm{d}}$ w.r.t. parameters \textit{joint\_damping} and \textit{joint\_frictionloss}.}
\label{fig:half-cheetah-d4rl-same-robustness}
\end{minipage}
\end{figure*}
As part of discussing the limitations of our work, we also provide two instances where RFQI and FQI algorithm behave similarly. RFQI and FQI algorithms trained on the D4RL dataset $\mathcal{D}_{\textrm{d}}$ perform similarly under the perturbations of the \textit{Hopper} model parameters \textit{actuator\_ctrlrange} and \textit{thigh\_joint\_stiffness} as shown in Fig. \ref{fig:hopper-d4rl-same-robustness}. We also make similar observations under the perturbations of the \textit{Half-Cheetah} model parameters \textit{joint\_damping} (both front \textit{joint\_damping} and back \textit{joint\_damping}) and \textit{joint\_frictionloss} as shown in Fig. \ref{fig:half-cheetah-d4rl-same-robustness}. We observed that the robustness performance can depend on the offline data available, which was also observed for non-robust offline RL algorithms \citep{liu2020provably,fu2020d4rl,levine2020offline}. Also, perturbing some parameters may make the problem really hard especially if the data is not representative with respect to that parameter. We believe that this is the reason for the similar performance of RFQI and FQI w.r.t. some parameters. We believe that this opens up an exciting area of research on developing online policy gradient algorithms for robust RL, which may be able to overcome the restriction and challenges due to offline data. We plan to pursue this goal in our future work.
|
\section{Introduction}
Ferrofluids are colloidal liquids consisting of non-conductive nanoscale ferromagnetic or ferrimagnetic particles suspended in carrier fluids, and have extensive applications in many technology and biomedicine fields ~\cite{Pankhurst2003, Zahn2001}.
There are two main ferrohydrodynamics (FHD) models which treat ferrofluids as homogeneous monophase fluids: the Rosensweig's model ~\cite{Rosensweig1985,Rosensweig1987} and the Shliomis' model~\cite{Shliomis1972Effective,Shliomis2002Ferrofluids}. The main difference between these two models lies in that the former one considers the internal rotation of the nanoparticles, while the latter deals with the rotation as a magnetic torque. We refer to \cite{Amirat2008,Amirat2008GlobalR,Amirat2009Strong,Amirat2010UniqueR,Nochetto2019} for some existence results of solutions to the two FHD models.
The FHD models are coupled nonlinear systems of the Maxwell equations and the incompressible Navier-Stokes equations.
There are limited works in the literature on the numerical analysis of the FHD models. In~\cite{Snyder2003Finite,Lavrova2006,Knobloch2010Numerical,Yoshikawa2010Numerical}, several numerical schemes were applied to solve reduced two dimensional FHD models where some nonlinear terms of the original models are dropped.
Nochetto et al. \cite{Nochetto2015The} showed the energy stability of the Rosensweig's model, proposed an energy-stable numerical scheme using finite element methods, and gave the existence and convergence of the numerical solutions. Recently, Wu and Xie \cite{YW2022} developed a class of energy-preserving mixed finite element methods for the Shliomis' FHD model, and derived optimal error estimates for both the the
semi- and fully discrete schemes.
We also note that in \cite{Zhang2021} an unconditionally energy-stable fully discrete finite element method was presented for a two-phase FHD model.
In this paper, we consider the Shliomis' FHD model with the assumption that the magnetization field is parallel to the magnetic field. Under this assumption, the magnetization equation in the Shliomis' model degenerates to the Langevin magnetization law~\cite{Lavrova2006,Rosensweig1985,Rosensweig1987}. We introduce some new variables to transform the model into two main decoupled subsystems, i.e., a nonlinear elliptic equation and the incompressible Navier-Stokes equations. We apply proper finite element spaces to discretize the nonlinear decoupled systems, prove the existence and, under some reasonable assumptions, uniqueness of the finite element solutions, and derive optimal error estimates. We also show that our scheme preserves the ferrofluids' nonconductive property ${\rm curl\,}\boldsymbol H = 0$ exactly.
The rest of this paper is organized as follows. In section 2, we introduce several Sobolev spaces, give the governing equation of the FHD model with magnetization paralleled to magnetic field, reform the FHD model equivalently, and construct the weak formulations. In section 3, we recall the finite element spaces, show the existence and uniqueness of solutions for the constructed finite element methods, and give the optimal order error estimates. In section 4, some numerical experiments will be given to verify our theoretical results.
\section{Preliminary}
In this section, we introduce several Sobolev spaces, give the governing equations of the FHD model with magnetization paralleled to magnetic field, derive the equivalent decoupled systems, and present the weak formulations.
\subsection{Sobolev spaces}
Let $\Omega \subset\mathbb R^d$ ($d = 2,3$) be a bounded and simply connected convex domain with Lipschitz boundary $\partial\Omega$, and $\boldsymbol n$ be the unit outward normal vector on $\partial\Omega$.
For any $p \geq 1$, we denote by $L^p(\Omega)$ the space of all power-$p$ integrable functions on $\Omega$ with norm $\|\cdot\|_{L^p}$. For any nonnegative integer $m$, denote by $H^m(\Omega)$ the usual $m$-th order Sobolev space with norm $\|\cdot\|_m$ and semi-norm $|\cdot|_m$. In particular, $H^0(\Omega) = L^2(\Omega)$ denotes the space of square integrable functions on $\Omega$, with the inner product $(\cdot,\cdot)$ and norm $\|\cdot\|$. For the vector spaces $\boldsymbol H^m(\Omega):=(H^m(\Omega))^d$ and $\boldsymbol L^2(\Omega) := (L^2(\Omega))^d$, we use the same notations of norm, semi-norm and inner product as those for the scalar cases.
We further introduce the Sobolev spaces
$$
\boldsymbol H({\rm curl\,}): = \left\{
\begin{array}{ll}
\{\boldsymbol v \in (L^2(\Omega))^2:~~{\rm curl\,}\boldsymbol v \in L^2(\Omega)\} & \text{if } d = 2,\\
\{\boldsymbol v \in (L^2(\Omega))^3:~~{\rm curl\,}\boldsymbol v \in (L^2(\Omega))^3\} & \text{if } d = 3
\end{array}
\right.
$$
and
$$
\boldsymbol H(\div) := \{\boldsymbol v \in (L^2(\Omega))^d:~~\div\boldsymbol v \in L^2(\Omega)\},
$$
and set
\begin{align*}
&H_0^1(\Omega) := \{ v \in H^1(\Omega):~v = 0\text{ on }\partial\Omega\},\\
&\boldsymbol H_0({\rm curl\,}) := \{\boldsymbol v \in \boldsymbol H({\rm curl\,}):~\boldsymbol v\times \boldsymbol n = 0\text{ on }\partial\Omega\},\\
&\boldsymbol H_0(\div) := \{\boldsymbol v \in \boldsymbol H(\div):~\boldsymbol v \cdot \boldsymbol n = 0\text{ on }\partial\Omega\}, \\
& L_0^2(\Omega) := \{ v\in L^2(\Omega):~\int_\Omega v\,{\rm d} \Omega = 0\},
\end{align*}
where
\begin{align*}
& {\rm curl\,} \boldsymbol v: = \left\{\begin{array}{ll}
\partial_x v_2 - \partial_y v_1 & \text{if }\boldsymbol v = ( v_1,~ v_2)^\intercal, \\
(\partial_y v_3 - \partial_z v_2,~\partial_z v_1 - \partial_x v_3,~\partial_x v_2 - \partial_y v_1)^\intercal & \text{if }\boldsymbol v = (v_1,~v_2,~v_3)^\intercal.
\end{array}
\right.
\\
& \div \boldsymbol v := \left\{
\begin{array}{ll}
\partial_x v_1 + \partial_y v_2 & \text{if }\boldsymbol v = (v_1,v_2)^\intercal, \\
\partial_x v_1 + \partial_y v_2 + \partial_z v_3 & \text{if } \boldsymbol v = (v_1,v_2,v_3)^\intercal,
\end{array}
\right.
\end{align*}
and $(x,y)$ and $(x,y,z)$ are the Cartesian coordinates in two and three dimensions, respectively.
For any Sobolev space $S$ with norm $\|\cdot\|_S$, we use $S^\prime$ to denote the dual space of $S$, and $\langle\cdot,\cdot\rangle$ to denote the dual product between $S^\prime$ and $S$. For any $f \in S^\prime$, the operator norm of $f$ is defined as $\|f\|_{S^\prime} = \sup\limits_{0\neq v \in S}\frac{\langle f,v\rangle}{\|v\|_S}$.
\subsection{Governing equations of the ferrofluid flow}
We consider the domain $\Omega$ filled with ferrofluid flow. On the macroscopic level, the mathematical model for describing the interactions between magnetic fields and ferrofluids consists of the Maxwell's equations and the Navier-Stokes equations~\cite{Rosensweig1987,Rosensweig1985,Shliomis1972Effective,Shliomis2002Ferrofluids}. Since the ferrofluid flow is nonconductive, the corresponding Maxwell's equations read as:
\begin{align}
\label{eq:maxwell1} & {\rm curl\,}\boldsymbol H = 0\qquad\qquad\qquad\,\text{in}\quad\Omega,\\
\label{eq:maxwell2} & \div\boldsymbol B = \div\boldsymbol H_e\qquad\qquad\text{in}\quad\Omega,
\end{align}
with the magnetic field $\boldsymbol H$ and the magnetic induction $\boldsymbol B$ satisfying the relation
\begin{equation}
\label{eq:rel-H-B}
\boldsymbol B = \mu_{0} (\boldsymbol H + \boldsymbol M)\qquad\qquad\text{in}\quad\Omega,
\end{equation}
where $\mu_{0}>0$ is the permeability constant, $\boldsymbol M$ is the magnetization, and $\boldsymbol H_e$ is the known external magnetic field that satisfies $\boldsymbol H_e\cdot\boldsymbol n = 0$ on $\partial\Omega$.
Under the assumption that the magnetization $\boldsymbol M$ of the ferrofluid flow is parallel to the magnetic field $\boldsymbol H$, it follows the nonlinear Langevin magnetization law~\cite{Rosensweig1987,Rosensweig1985,Lavrova2006}, i.e.,
\begin{equation}
\label{eq:langevin}
\boldsymbol M(\boldsymbol H) = M_{s}\left(\coth(\gamma H)- \frac{1}{\gamma H} \right)\frac{\boldsymbol H}{H},
\end{equation}
with the saturation magnetization $M_{s}>0$, the Lagevin parameter $\gamma= 3\chi_{0}/ M_{s}$, the initial susceptibility $\chi_{0}>0$, $H := |\boldsymbol H| = (\boldsymbol H \cdot\boldsymbol H)^{1/2} > 0$, and $\coth(x) = \frac{e^x+e^{-x}}{e^x-e^{-x}}$.
The hydrodynamic properties of the ferrofluid flow are described by the incompressible Navier-Stokes equations
\begin{align}
\label{eq:NS1} & \rho(\boldsymbol u\cdot \nabla)\boldsymbol u - \eta\Delta\boldsymbol u + \nabla p = \boldsymbol f + \mu_{0} (\boldsymbol M\cdot\nabla) \boldsymbol H\qquad\quad\text{in}\quad\Omega,\\
\label{eq:NS2} & \div\boldsymbol u = 0\,\,\,\,\,\quad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\text{in}\quad\Omega,
\end{align}
where $\rho$ denotes the fluid density, $\boldsymbol u$ the velocity field of the flow, $\eta$ the dynamic viscosity, and $\boldsymbol f$ the known volume force.
We consider the following homogenous boundary conditions for equation \eqref{eq:maxwell1}-\eqref{eq:NS2}:
\begin{equation}
\label{eq:bd} \boldsymbol u = 0\quad\text{ and }\quad \boldsymbol H \times \boldsymbol n = 0\qquad\text{on}\quad\partial\Omega.
\end{equation}
Using the fact that
$$
(\boldsymbol H \cdot\nabla)\boldsymbol H = \frac{1}{2}\nabla H^2 - \boldsymbol H \times ({\rm curl\,}\boldsymbol H) = \frac{1}{2} \nabla H^2
$$
and the Langevin magnetization law \eqref{eq:langevin}, we have
\begin{equation}
\label{eq:Kevinforce}
(\boldsymbol M\cdot\nabla)\boldsymbol H = \frac{M_s}{H}\left( \coth(\gamma H) - \frac{1}{\gamma H} \right) \frac{1}{2}\nabla H^2 = \frac{M_s}{\gamma} \nabla \ln \frac{\sinh(\gamma H)}{H},
\end{equation}
where $\sinh(x) = \frac{e^x-e^{-x}}{2}$. Denote
\begin{equation}
\label{eq:beta-def}
\beta(x): = \frac{M_s}{\gamma} \ln\frac{\sinh(\gamma x)}{x},
\end{equation}
and introduce two variables
\begin{equation}\label{eq:psi}
\psi :=\beta(H)
\end{equation}
and
\begin{equation}\label{eq:p}
\tilde p: = p - \mu_0\psi,
\end{equation}
then equation \eqref{eq:NS1} becomes
\begin{equation}
\label{eq:NS1-n}
\rho(\boldsymbol u\cdot\nabla)\boldsymbol u - \eta\Delta\boldsymbol u + \nabla \tilde p = \boldsymbol f\qquad\qquad\qquad\text{in }\Omega.
\end{equation}
Equation \eqref{eq:maxwell1} and the
boundary condition $\boldsymbol H \times \boldsymbol n|_{\partial\Omega} = 0$ in \eqref{eq:bd} imply that (cf. \cite{Arnold;Falk;Winther2006}) there exists $\phi \in H_{0}^{1}(\Omega)$ with
\begin{equation}
\label{eq:Hphi} \boldsymbol H = \nabla\phi,
\end{equation}
then combining \eqref{eq:maxwell2}, \eqref{eq:rel-H-B} and \eqref{eq:langevin} leads to
\begin{equation}
\label{eq:H} \nabla\cdot\left(\alpha(|\nabla\phi|)\nabla\phi\right) = \frac{1}{\mu_0}\div\boldsymbol H_e =: g\qquad\text{in}~~\Omega,
\end{equation}
with
\begin{equation}\label{eq:alpha-def}
\alpha(x) := 1+\frac{M_{s}}{x}\left(\coth(\gamma x) - \frac{1}{\gamma x}\right).
\end{equation}
The above equivalence transformation shows that the FHD model \eqref{eq:maxwell1} - \eqref{eq:NS2} with the boundary conditions \eqref{eq:bd} can be equivalently written as follows: Find $\phi \in H_0^1(\Omega)$, $\boldsymbol H \in\boldsymbol H_0({\rm curl\,})$, $\boldsymbol M \in \boldsymbol H_0({\rm curl\,})$, $\boldsymbol u \in \boldsymbol H_0^1(\Omega)$, $\tilde p \in L_0^2(\Omega)$, $\psi \in L_0^2(\Omega)$, and $p \in L_0^2(\Omega)$ such that
\begin{equation}
\label{eq:eq-sys}
\left\{
\begin{array}{ll}
\nabla\cdot(\alpha(|\nabla\phi|)\nabla\phi) = g & \text{in }\Omega,\\
\boldsymbol H - \nabla\phi = 0 & \text{in }\Omega, \\
\boldsymbol M = M_s\left( \coth(\gamma H) - \frac{1}{\gamma H} \right) \frac{\boldsymbol H}{H}=\big(\alpha(H)-1\big)\boldsymbol H & \text{in }\Omega, \\
\rho(\boldsymbol u\cdot\nabla)\boldsymbol u - \eta\Delta \boldsymbol u + \nabla\tilde p = \boldsymbol f & \text{in }\Omega,\\
\nabla\cdot\boldsymbol u = 0 & \text{in }\Omega,\\
\psi =\frac{M_s}{\gamma}\ln\frac{\sinh(\gamma H)}{H} = \beta(H) & \text{in }\Omega, \\
p = \tilde p + \mu_0\psi & \text{in }\Omega.
\end{array}
\right.
\end{equation}
\begin{remark}\label{rem2.1}
The system \eqref{eq:eq-sys} is a decoupled system. Firstly, we can solve the first nonlinear elliptic equation of \eqref{eq:eq-sys} to get $\phi$, and solve the Navier-Stokes equations, i.e. the fourth and fifth equations, to get $\boldsymbol u$ and $\tilde p$. Secondly, we can get $\boldsymbol H$ from the second equation of \eqref{eq:eq-sys}. Finally, we can obtain $\boldsymbol M$, $\psi$, and $p$ from the third, the sixth, and the seventh equations,
respectively.
\end{remark}
\subsection{ Preliminary estimates for nonlinear functions}
Notice that the decoupled system \eqref{eq:eq-sys} involves the nonlinear functions $\alpha(x)$ and $\beta(x)$ defined in \eqref{eq:alpha-def} and \eqref{eq:beta-def}, respectively.
The following basic estimates of these two functions will be used in later analysis.
\begin{lemma}\label{lem:alphaH}
\begin{itemize}
\item[(i)] There exist a positive constants $C_1$ and $C_\alpha$ such that for any $x > 0$, there hold
$$
1 <\alpha(x) \leq C_1, \qquad
|\alpha^\prime(x)| \leq C_\alpha \ ;
$$
\item[(ii)] For any $x > 0$, there holds
$$
|\beta^{\prime}(x)| \leq M_{s}.
$$
\end{itemize}
\end{lemma}
\begin{proof}
(i) We first show $
1 <\alpha(x) \leq C_1. $
On one hand, the L'Hopital law implies that
$$
\lim\limits_{y \rightarrow 0^+} y\coth(y) = 1,
$$
which, together with the fact that
$$
(y\coth(y))^\prime = \coth(y) - y{\rm csch}^2(y) = \frac{e^{2y} - e^{-2y} - 4y}{(e^y - e^{-y})^2}>0 \quad \forall~~ y > 0,
$$
shows
$$
y\coth(y) > 1\qquad \forall~~ y > 0.
$$
Therefore,
$$
\alpha(x) = 1 + \gamma M_s\frac{y\coth(y) -1 }{y^2} > 1\qquad\text{with }y = \gamma x,\text{ for }\forall~~x >0.
$$
On the other hand, we easily see that
$$
\lim\limits_{x\rightarrow 0^+} \alpha(x) = 1+ \frac{\gamma M_s}{3}\qquad\text{and}\qquad\lim\limits_{x\rightarrow + \infty} \alpha(x) = 1,
$$
which, together the continuity of $\alpha(x)$ for $x>0$, indicate that
there exists a positive constant $C_1$ such that
$$
\alpha(x) \leq C_1\qquad\forall~~x>0
$$
Second, let us show $
|\alpha^\prime(x)| \leq C_\alpha.$ It is easy to get
\begin{equation*}
\begin{split}
\alpha^\prime(x) & = \frac{2M_s}{\gamma x^3} - \frac{M_s}{x^2}\coth(\gamma x) - \frac{\gamma M_s}{x}{\rm csch}^2(\gamma x) \\
& = \frac{\gamma^2 M_s}{y^3}(2 - y^2{\rm csch}^2(y) - y\coth(y))\qquad\qquad\text{with } y = \gamma x.
\end{split}
\end{equation*}
The L'Hopital law implies that
$$
\lim\limits_{x\rightarrow 0^+}\alpha^\prime(x) = 0\qquad\text{and}\qquad \lim\limits_{x\rightarrow +\infty}\alpha^\prime(x) = 0.
$$
Since $\alpha^\prime(x)$ is continuous for $x>0$, we know that
there exists a positive constant $C_\alpha$ such that
$$
|\alpha^\prime(x)| \leq C_\alpha \qquad \forall x>0.
$$
As a result, the conclusions of (i) follow.
(ii) It is easy to find that
$$
\beta^{\prime}(x) = M_{s}\left(\coth(\gamma x) - \frac{1}{\gamma x}\right) = M_{s} \left(\coth(y) - \frac{1}{y}\right)
$$
with $y = \gamma x>0$. The L'Hopital law implies that
$$
\lim\limits_{x\rightarrow 0^{+}} \beta^{\prime}(x) = 0\quad\text{and}\quad\lim\limits_{x \rightarrow + \infty}\beta^{\prime}(x) = M_{s}.
$$
The fact that
$$
\left( \coth(y) - \frac{1}{y}\right)^{\prime} = \frac{1-y^{2}{\rm csch}^{2}(y)}{y^{2}} = \frac{1+y{\rm csch}(y)}{y^{2}}\frac{e^{y} - e^{-y} - 2y}{e^{y} - e^{-y}} > 0 \qquad\forall~~y>0
$$
implies that $\beta^{\prime}(x)$ is a monotonically increasing function on the interval $(0,+\infty)$. We conclude that
$$
0< \beta^{\prime}(x) \leq M_{s}.
$$
This finishes the proof.
\end{proof}
\subsection{Weak formulations}
Based on the FHD model \eqref{eq:eq-sys} and Remark \ref{rem2.1}, we consider the following
weak formulations: Find $\phi\in H_0^1(\Omega)$, $\boldsymbol H \in\boldsymbol H_0({\rm curl\,})$, $\boldsymbol u \in \boldsymbol H_{0}^{1}(\Omega)$ and $\tilde p \in L_0^2(\Omega)$
such that
\begin{align}
\label{eq:weak-maxwell-1}
& a(\phi;\phi,\tau) = -(g,\tau)\qquad\qquad\qquad\qquad\qquad\qquad\qquad\ \,\forall~~\tau \in H_{0}^1(\Omega),\\
\label{eq:weak-maxwell-2} & (\boldsymbol H,\boldsymbol C) - (\nabla\phi,\boldsymbol C) = 0\quad\qquad\qquad\qquad\qquad\qquad\qquad \forall~~\boldsymbol C \in \boldsymbol H_{0}({\rm curl\,}),\\
\label{eq:weak-NS-1}
& (\rho(\boldsymbol u\cdot\nabla)\boldsymbol u,\boldsymbol v) + \eta(\nabla\boldsymbol u,\nabla\boldsymbol v) -(\tilde p,\nabla\cdot\boldsymbol v)
= (\boldsymbol f,\boldsymbol v)
\qquad\forall~~\boldsymbol v \in \boldsymbol H_{0}^{1}(\Omega),\\
\label{eq:weak-NS-2} & (\nabla\cdot\boldsymbol u,q) = 0\,\quad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\forall~~q \in L_{0}^{2}(\Omega),
\end{align}
where $a(\cdot;\cdot,\cdot):~H^1(\Omega) \times H^1(\Omega) \times H^1(\Omega) \rightarrow \mathbb R$ is defined by
$$
a(w;\phi,\tau) := \int_{\Omega}\alpha(|\nabla w|) \nabla\phi \cdot\nabla \tau \,{\rm d} \Omega\qquad\forall~~w,~\phi,~\tau \in H^1(\Omega).
$$
\begin{remark} As shown in Remark \ref{rem2.1}, once the variables $\phi$, $\boldsymbol H$, $\boldsymbol u$ and $\tilde p$ are solved, the other variables, i.e. $\boldsymbol M$, $\psi$ and $p$, can immediately be obtained by the third, the sixth and the seventh equations of \eqref{eq:eq-sys}, respectively.
\end{remark}
In what follows, we discuss the existence and uniqueness of the solutions to the weak formulations \eqref{eq:weak-maxwell-1} - \eqref{eq:weak-NS-2}.
We first consider
the nonlinear equation \eqref{eq:weak-maxwell-1}.
It is easy to see that, for any given $w \in H^1(\Omega)$, $a(w;\cdot,\cdot)$ is a bilinear form on $H^1(\Omega)$. Recall that the Poincar\'e inequality
\begin{equation*}\label{eq:poin}
\|v\| \leq C_p \|\nabla v\| \qquad\forall~~v \in H_0^1(\Omega),
\end{equation*}
with $C_p > 0$ being a constant depending only on $\Omega$, means that the semi-norm $\|\nabla (\cdot)\|$ is also a norm on $H_0^1(\Omega)$. Then, from Lemma \ref{lem:alphaH} (i),
we easily obtain the following uniform coercivity and continuity results for $a(w;\cdot,\cdot)$:
\begin{lemma}\label{lem:cocervity}
For any $ w \in H_0^1(\Omega)$ and $\phi, \tau\in H_0^1(\Omega)$, we have
$$
a(w;\phi,\phi) \geq \|\nabla\phi\|^2
$$
and
$$
a(w;\phi,\tau) \leq C_1 \|\nabla\phi\|\|\nabla \tau\|.
$$
\end{lemma}
We have the following wellposedness results for equation \eqref{eq:weak-maxwell-1}.
\begin{theorem}\label{well-phi}
The nonlinear equation \eqref{eq:weak-maxwell-1} has at least one solution $\phi\in H_0^1(\Omega)$, and there holds
\begin{equation}\label{eq:phi-bd}
\|\nabla\phi\| \leq \frac{1}{\mu_0}\|\boldsymbol H_e\|.
\end{equation}
Moreover, \eqref{eq:weak-maxwell-1} admits at most one solution $\phi\in H_0^1(\Omega)$ satisfying
\begin{equation}\label{eq:ass-phi}
\|\nabla\phi\|_{L^\infty} < 1/C_\alpha.
\end{equation}
\end{theorem}
\begin{proof} The existence of a solution $\phi\in H_0^1(\Omega)$ follows from Lemma \ref{lem:cocervity} and \cite[Page 332, Theorem 2]{Boccardo1992} directly.
Recalling that $ g=\frac{1}{\mu_0}\div\boldsymbol H_e$,
from \eqref{eq:weak-maxwell-1} and the above lemma we have
$$
\|\nabla\phi\|^2\leq a(\phi;\phi,\phi)=-(g,\phi)=\frac{1}{\mu_0}(\boldsymbol H_e, \nabla \phi)\leq \frac{1}{\mu_0}\|\boldsymbol H_e\| \| \nabla \phi\|.
$$
This yields the estimate \eqref{eq:phi-bd}.
The thing left is to show the uniqueness of the solution under condition \eqref{eq:ass-phi}. In fact, let $\phi$ and $\tilde\phi$ be any two solutions of \eqref{eq:weak-maxwell-1} satisfying \eqref{eq:ass-phi}, we have
$$
a(\tilde\phi;\tilde\phi,\tau) = a(\phi;\phi,\tau)\qquad\forall~~\tau \in H_0^1(\Omega),
$$
which yields
\begin{align*}
a(\tilde\phi;\tilde\phi-\phi,\tau) = \int_\Omega \big(\alpha(|\nabla\phi|) - \alpha(|\nabla\tilde\phi|) \big)\nabla\phi\cdot\nabla\tau\,{\rm d}\Omega\qquad\forall~~\tau \in H_0^1(\Omega).
\end{align*}
Taking $\tau = \tilde\phi - \phi$ in the above equation and using Lemma \ref{lem:alphaH} give
\begin{align*}
\|\nabla(\tilde\phi - \phi)\|^2 & \leq a(\tilde \phi;\tilde \phi-\phi,\tilde\phi - \phi) \\
& = a(\phi;\phi,\tilde\phi - \phi) - a(\tilde\phi;\phi,\tilde\phi - \phi) \\
& \leq C_\alpha\|\nabla \phi\|_{L^\infty}\|\nabla(\tilde\phi - \phi)\|^2,
\end{align*}
which, together with \eqref{eq:ass-phi}, implies $\tilde \phi = \phi$. This finishes the proof.
\end{proof}
For equation \eqref{eq:weak-maxwell-2}, we have the following conclusion:
\begin{theorem}\label{H-well}
For any given $\phi \in H_0^1(\Omega)$, equation \eqref{eq:weak-maxwell-2} admits a unique solution $\boldsymbol H = \nabla\phi \in \boldsymbol H_0({\rm curl\,})$, which means $
{\rm curl\,} \boldsymbol H = 0.
$
\end{theorem}
\begin{proof}
The de Rham complex \cite{Arnold;Falk;Winther2006} on $2D$ and $3D$ domain implies that $\nabla\phi\in \boldsymbol H_0({\rm curl\,})$. Thus, taking $\boldsymbol C = \boldsymbol H - \nabla\phi \in \boldsymbol H_0({\rm curl\,})$ in \eqref{eq:weak-maxwell-2} yields
$
\boldsymbol H = \nabla\phi.
$
\end{proof}
Since \eqref{eq:weak-NS-1}-\eqref{eq:weak-NS-2} are the weak formulations of the Navier-Stokes equations, the following
existence and uniqueness results are standard (cf. \cite[Page 285-287, Theorems 2.1 and 2.2]{Girault1986}).
\begin{theorem}
Given $\boldsymbol f \in (H^{-1}(\Omega))^d$, there exists at least one pair $(\boldsymbol u,\tilde p) \in \boldsymbol H_0^1(\Omega) \times L_0^2(\Omega)$ satisfying \eqref{eq:weak-NS-1} - \eqref{eq:weak-NS-2}, and holds
$$
\|\nabla\boldsymbol u\| \leq \frac 1\eta \|\boldsymbol f\|_{-1},
$$
where $\|\boldsymbol f\|_{-1}: = \sup\limits_{\boldsymbol v \in \boldsymbol H_0^1(\Omega)}\frac{\int_\Omega \boldsymbol f \cdot \boldsymbol v\,{\rm d}\Omega}{\|\nabla \boldsymbol v\|}.$
In addition, if
$$
(\rho \mathcal N/\eta^2)\|\boldsymbol f\|_{-1} <1\quad\text{with}\quad
\mathcal N: = \sup\limits_{\boldsymbol u,\boldsymbol v,\boldsymbol w\in \boldsymbol H_0^1(\Omega)}\frac{\int_\Omega (\boldsymbol u\cdot\nabla)\boldsymbol v \cdot \boldsymbol w\,{\rm d}\Omega}{\|\nabla \boldsymbol u\|\|\nabla \boldsymbol v\|\|\nabla \boldsymbol w\|},
$$
then the solution pair $(\boldsymbol u,\tilde p)$ is unique.
\end{theorem}
\section{Finite element methods}
\subsection{Finite element spaces}
Let $\mathcal T_{h}$ be a quasi-uniform
simplicial decomposition of $\Omega$ with mesh size $h:=\max\limits_{K\in \mathcal T_h} h_K$, where $h_K$ denotes the diameter of $K$ for any $K \in \mathcal T_h$.
For an integer $l \geq 0$, let $\mathbb P_l(K)$ denote the set of polynomials, defined on $K\in \mathcal T_h$, of degree no more than $l$.
We introduce the following finite element spaces:
\begin{itemize}
\item $\boldsymbol V_{h} = (V_h)^d \ \big(\subset \text{or } \not\subset (H_{0}^{1}(\Omega))^d\big)$ is the Lagrange element space~\cite{Girault1986,Brezzi;Fortin1991,Girault;Raviart2012} for the velocity $\boldsymbol u$, with $\mathbb P_{l+1}(K) \subset V_h|_K$ for any $K \in \mathcal T_h$.
In particular, for the nonconforming case that $V_h\not\subset H_{0}^{1}(\Omega)$,
each $v_h \in V_h$ is required to satisfy the following conditions:
\begin{itemize}
\item[(i)] $v_h$ vanishes at all the nodes on $\partial \Omega$;
\item[(ii)]
$|v_h|_{1,h}:=(\sum\limits_{K\in \mathcal T_h} \|\nabla v_h\|_{K}^2)^{1/2}$ is a norm.
\end{itemize}
Note that the classical nonconforming Crouzeix-Raviart (CR) finite element ~\cite{Crouzeix1973CRele} is corresponding to the nonconforming case with $l=0$. In the nonconforming case, the gradient and divergence operators, $\nabla$ and $\nabla\cdot$, in the finite element scheme \eqref{eq:fem-3} - \eqref{eq:fem-4} will be understood as $\nabla_h$ and $\nabla_h\cdot$, respectively, which denote respectively the piecewise gradient and divergence operators acting on element by element in $\mathcal T_{h}$.
\item $W_{h} \subset L_{0}^{2}(\Omega)$ is the piecewise polynomial space
for the `pressure' variable
$\tilde p$, with $\mathbb P_l(K) \subset W_h|_K$ for any $K \in \mathcal T_h$.
\item $S_{h} \subset H_{0}^{1}(\Omega)$ is the Lagrange element space~\cite{Ciarlet1978}
for the new variable $\phi$, with $\mathbb P_{l+1}(K)\subset S_h|_K$ for any $K \in \mathcal T_h$.
\item $\boldsymbol U_{h} \subset \boldsymbol H_{0}({\rm curl\,})$ is the edge element space~\cite{Nedelec1980,Nedelec1986} for the magnetic field $\boldsymbol H$,
with $(\mathbb P_l(K))^d\subset \boldsymbol U_h|_K$ and $(\mathbb P_l(K))^{2d-3} \subset {\rm curl\,} \boldsymbol U_h|_K$
for any $K \in \mathcal T_h$.
\end{itemize}
In addition, we make the following assumptions for the above finite element spaces.
\begin{itemize}
\item[(A1)] There holds the inf-sup condition
\begin{equation}
\label{eq:inf-sup}
\inf\limits_{\boldsymbol v_h \in \boldsymbol V_{h}}\frac{(\nabla\cdot\boldsymbol v_h,q_h)}{\|\boldsymbol v_h\|_{1}} \geq \beta_{0} \|q_h\|\qquad\forall~~q_h \in W_{h},
\end{equation}
where $\beta_{0}>0 $ is a constant independent of $h$;
\item[ (A2)] The diagram
\begin{equation}\label{eq:exact-seq}
\begin{CD}
H_0^1@>{{\rm grad\,}}>> \boldsymbol H_0({\rm curl\,}) \\
@VV \pi_h^{s} V @ VV \pi ^{c}_h V \\\
S_{h} @>{{\rm grad\,}}>> \boldsymbol U_{h} \end{CD}
\end{equation}
is a commutative sequence. Here $\pi_h^s:~H_0^1(\Omega)\cap C^0(\overline\Omega)\rightarrow S_h$ and $\pi_h^c:~\boldsymbol H_0({\rm curl\,})\rightarrow \boldsymbol U_h$ are the classical interpolation operators, and ${\rm grad\,}$ denotes the gradient operator. Note that the diagram \eqref{eq:exact-seq} also indicates that
$$ {\rm grad\,} S_{h} \subset \boldsymbol U_{h} .$$
\end{itemize}
We recall the following inverse inequality:
\begin{equation}\label{inverse}
\|\nabla s_h\|_{\infty} \leq C_{inv} h^{-d/2}\|\nabla s_h\| \qquad \forall s_h \in S_h,
\end{equation}
where $ C_{inv} >0$ is a constant independent of $h$.
\begin{remark}
There are many finite element spaces that satisfy (A1) and (A2). For example, we can choose $\boldsymbol V_h$ and $W_h$ as the Taylor-Hood element pairs. And the spaces $S_h$ and $\boldsymbol U_h$ can be respectively chosen as the lowest order Lagrange finite element space and the lowest order N\'ed\'elec edge element space ~\cite{Arnold;Falk;Gopalakrishnan2012,Arnold;Falk;Winther2006,Arnold;Falk;Winther2010}.
\end{remark}
\begin{remark}
In this paper, we consider the conforming finite element spaces for the Navier-Stokes equation for simply of notations. In fact the nonconforming finite element spaces which satisfying the inf-sup condition (A1), with replace the global differential operators to element wise, are also work for the Navier-Stokes equations. For example the $CR$ - $\mathbb P_0$ finite element spaces are work for the Navier-Stokes equations.
\end{remark}
\subsection{Finite element scheme} In view of \eqref{eq:weak-maxwell-1}-\eqref{eq:weak-NS-2},
we consider the following finite element scheme for the FHD model: Find $\phi_h \in S_h$, $\boldsymbol H_h \in \boldsymbol U_h$,
$\boldsymbol u_h \in \boldsymbol V_h$ and $\tilde p_h\in W_h$,
such that
\begin{align}
\label{eq:fem-1}
& a(\phi_h;\phi_h,\tau_h) = -(g,\tau_h
&\forall~~\tau_h \in S_h,\\
\label{eq:fem-2}& (\boldsymbol H_h,\boldsymbol C_h) - (\nabla\phi_h,\boldsymbol C_h) = 0
&\forall~~\boldsymbol C_h \in \boldsymbol U_h,\\
\label{eq:fem-3} & \eta(\nabla\boldsymbol u_h,\nabla\boldsymbol v_h) +b(\boldsymbol u_h;\boldsymbol u_h,\boldsymbol v_h) - (\tilde p_h,\nabla\cdot\boldsymbol v_h) = (\boldsymbol f,\boldsymbol v_h)
&\forall~~\boldsymbol v_h \in \boldsymbol V_h, \\
\label{eq:fem-4} & (\nabla\cdot\boldsymbol u_h,q_h) = 0\
&\forall~~q_h \in W_h,
\end{align}
where the trilinear form $ b(\cdot;\cdot,\cdot) : \boldsymbol H^{1}(\Omega) \times\boldsymbol H^{1}(\Omega) \times \boldsymbol H^{1}(\Omega) \rightarrow \mathbb R$ is defined by
$$
b(\boldsymbol w;\boldsymbol u,\boldsymbol v) := \frac{\rho}{2}[((\boldsymbol w\cdot\nabla)\boldsymbol u,\boldsymbol v) - ((\boldsymbol w\cdot\nabla)\boldsymbol v,\boldsymbol u)].
$$
\begin{remark} \label{rem:sol}
Similar to \eqref{eq:weak-maxwell-1}-\eqref{eq:weak-NS-2}, the finite element scheme \eqref{eq:fem-1} - \eqref{eq:fem-4} is a decoupled system. We can first solve the nonlinear equation \eqref{eq:fem-1} to get $\phi_h$, and solve the Navier-Stokes system \eqref{eq:fem-3} - \eqref{eq:fem-4} to get $\boldsymbol u_h$ and $\tilde p_h$. Then we can get $\boldsymbol H_h$ from \eqref{eq:fem-2}. Finally, we can get $\boldsymbol M_h \in \boldsymbol U_h$, the approximation of the magnetization $\boldsymbol M$, from
\begin{equation}\label{eq:weak-maxwell-3-dis}
(\boldsymbol M_h,\boldsymbol F_h) = \big(
(\alpha(H_h)-1)\boldsymbol H_h,\boldsymbol F_h \big)
\,\,\,\,\,\qquad \quad \forall~~\boldsymbol F_h \in \boldsymbol U_h
\end{equation}
with $H_h:=|\boldsymbol H_h|$, get $\psi_h \in W_h$, the approximation of $\psi$, from
\begin{equation}\label{eq:fem-5}
(\psi_h,\chi_h) =\big(\beta(H_h),\chi_h\big)
\ \ \qquad\qquad\qquad\forall~~\chi_h \in W_h,
\end{equation}
and get $p_h \in W_h$, the approximation of the pressure $p$, from
\begin{equation}\label{eq:fem-6}
p_h = \tilde p_h + \mu_0 \psi_h.
\end{equation}
\end{remark}
\begin{remark}\label{rem:trilinear}
It is easy to see that the trilinear form $b(\cdot;\cdot,\cdot)$ is skew-symmetric with respect to the last two variables, i.e.,
\begin{equation}\label{eq:skew-s}
b(\boldsymbol w;\boldsymbol u ,\boldsymbol v ) = -b(\boldsymbol w ;\boldsymbol v ,\boldsymbol u )\qquad \forall ~\boldsymbol w ,~\boldsymbol u , ~\boldsymbol v \in \boldsymbol H^{1}(\Omega) ,
\end{equation}
and that
\begin{equation}\label{eq:breform}
b(\boldsymbol w;\boldsymbol u ,\boldsymbol v ) =\rho\big((\boldsymbol w\cdot\nabla)\boldsymbol u, \boldsymbol v\big) + \frac{\rho}{2}\big( (\nabla\cdot\boldsymbol w) \boldsymbol u,\boldsymbol v \big)\quad\forall~\boldsymbol w,~\boldsymbol u \in \boldsymbol H^{1}(\Omega) , \ \ \forall \ \boldsymbol v \in \boldsymbol H_0^{1}(\Omega).
\end{equation}
As a result, the following two relations hold
\begin{equation}
\label{eq:b-coer}
b(\boldsymbol w;\boldsymbol v,\boldsymbol v) =0 \qquad\forall~ \boldsymbol w, \boldsymbol v \in \boldsymbol H^{1}(\Omega),\\
\end{equation}
\begin{equation}
\label{eq:b-}
b(\boldsymbol w;\boldsymbol u ,\boldsymbol v ) =\rho\big((\boldsymbol w\cdot\nabla)\boldsymbol u, \boldsymbol v \big) \quad\forall~ \boldsymbol w, ~\boldsymbol u \in \boldsymbol H^{1}(\Omega) \text{ with } \div\boldsymbol w = 0,\ \forall \ \boldsymbol v \in \boldsymbol H_0^{1}(\Omega). \end{equation}
\end{remark}
Theorems \ref{lem:solu1-2}-\ref{lem:solu3-4} show the wellposedness of the finite element scheme \eqref{eq:fem-1} - \eqref{eq:fem-4}.
\begin{theorem}\label{lem:solu1-2}
The nonlinear discrete equation \eqref{eq:fem-1} has at least one solution $\phi_h \in S_h$, and there holds
\begin{equation}\label{eq:nabla-g}
\|\nabla\phi_h \| \leq \frac{1}{\mu_0}\|\boldsymbol H_e\|.
\end{equation}
Furthermore, if the external magnetic filed $\boldsymbol H_e$ satisfies
\begin{equation}\label{eq:ass-He}
\| \boldsymbol H_e\| < \mu_0 C_{inv}^{-1}C_\alpha^{-1} h^{d/2},
\end{equation}
then \eqref{eq:fem-1} admits a unique solution $\phi_h \in S_h$, and there holds
\begin{equation}\label{eq:phi-inf}
\|\nabla\phi_h\|_\infty < 1/ C_\alpha.
\end{equation}
\end{theorem}
\begin{proof}
Define an operator $A:~S_h \rightarrow S_h^\prime$ by
$$
\langle A(\phi_h),\tau_h\rangle = a(\phi_h;\phi_h,\tau_h)\qquad\forall~~\phi_h,~\tau_h\in S_h,
$$
and define $\Phi:~S_h \rightarrow S_h^\prime$ by
$$
\Phi(\phi_h) = A(\phi_h) + Q_h^\prime g\qquad\forall~~\phi_h \in S_h,
$$
where $Q_h^\prime:~(H_0^1(\Omega))^\prime \rightarrow S_h^\prime$ is given by
\begin{equation}\label{eq:Q_h}
\langle Q_h^\prime g,\tau_h \rangle = \langle g,\tau_h\rangle=-\frac{1}{\mu_0}(\boldsymbol H_e, \nabla \tau_h) \qquad \forall~~\tau_h \in S_h.
\end{equation}
Thus, \eqref{eq:fem-1} is equivalent to the equation
$$
\Phi(\phi_h) = 0.
$$
Lemma \ref{lem:cocervity} implies
$$
\langle A(\phi_h),\phi_h\rangle =a(\phi_h;\phi_h,\phi_h) \geq \|\nabla\phi_h\|^2\qquad\forall~~\phi_h \in S_h.
$$
The definition \eqref{eq:Q_h} and the Cauchy-Schwarz inequality
imply
$$
\langle Q_h^\prime g,\phi_h\rangle=\langle g ,\phi_h\rangle \leq \frac{1}{\mu_0}\|\boldsymbol H_e\|\|\nabla\phi_h\|\qquad\forall~~\phi_h \in S_h.
$$
Therefore, we have
$$
\langle \Phi(\phi_h),\phi_h\rangle =\langle A(\phi_h) + Q_h^\prime g ,\phi_h\rangle
\geq 0, \qquad\forall~~\phi_h \in S_h \text{ with }\|\nabla\phi_h\| = \frac{1}{\mu_0}\|\boldsymbol H_e\|.
$$
For any given $0 \neq \phi_{h0} \in S_h$
and $\epsilon >0$, denote
$\delta: = \min\left\{ \frac{\epsilon}{2C_1},~\frac{\epsilon}{2C_\alpha\|\nabla\phi_{h0}\|_\infty}\right\} $. Then for any $\phi_h \in S_h$ satisfying $\|\nabla (\phi_h - \phi_{h0})\| < \delta$, by Lemmas \ref{lem:alphaH} and \ref{lem:cocervity} we have
\begin{align*}
&\|\Phi(\phi_h) - \Phi(\phi_{h0})\|_{S_h^\prime} = \|A(\phi_h) - A(\phi_{h0})\|_{S_h^\prime} \\
= & \sup\limits_{\tau_h\in S_h} \frac{\langle A(\phi_h) - A(\phi_{h0}),\tau_h \rangle}{\|\nabla\tau_h\|} \\
= &\sup\limits_{\tau_h \in S_h} \frac{a(\phi_h;\phi_h,\tau_h) - a(\phi_{h0};\phi_{h0},\tau_h)}{\|\nabla\tau_h\|} \\
=& \sup\limits_{\tau_h \in S_h} \frac{a(\phi_h;\phi_h,\tau_h) - a(\phi_h;\phi_{h0},\tau_h) + a(\phi_h;\phi_{h0},\tau_h) - a(\phi_{h0};\phi_{h0},\tau_h)}{\|\nabla\tau_h\|} \\
\leq & C_1\|\nabla(\phi_h - \phi_{h0})\| + C_\alpha\|\nabla\phi_{h0}\|_\infty\|\nabla(\phi_h - \phi_{h0})\| \\
< & \epsilon.
\end{align*}
This means that $\Phi$ is continuous on the set $S_h\setminus\{0\}$. Notice that Lemma \ref{lem:cocervity} implies $\Phi$ is continuous at the point $0$. Hence, $\Phi$ is continuous on $S_h$.
On the other hand, the Riesz representation theorem implies that the spaces $S_h$ and $S_h^\prime$ are isometry. As a result, by \cite[Page 279, Corollary 1.1]{Girault1986} equation \eqref{eq:fem-1} admits at least one solution $\phi_h \in S_h$ satisfiying
\eqref{eq:nabla-g}
If $\boldsymbol H_e$ satisfies assumption \eqref{eq:ass-He},
then from \eqref{eq:nabla-g} and the inverse inequality \eqref{inverse} we ge
$$
\|\nabla\phi_h\|_\infty \leq C_{inv} h^{-d/2} \|\nabla\phi_h\| < 1/C_\alpha ,
$$
i.e. \eqref{eq:phi-inf} holds. Thus,
by following the same line as in the proof of the uniqueness of the weak solution $\phi$ in Theorem \ref{well-phi}, we can easily obtain the uniqueness of the discrete solution $\phi_h$.
\end{proof}
\begin{theorem}\label{Hh-well}
For any given $\phi_h \in S_h$, equation \eqref{eq:fem-2} admits a unique solution $\boldsymbol H_h = \nabla\phi_h \in \boldsymbol U_h$, which means $
{\rm curl\,} \boldsymbol H_h = 0.
$
\end{theorem}
\begin{proof}
For any given $\phi_h \in S_h$, assumption (A2)
on the finite element spaces implies that $\nabla\phi_h \in \boldsymbol U_h$. Thus, taking $\boldsymbol C_h = \boldsymbol H_h - \nabla\phi_h \in \boldsymbol U_h$ in \eqref{eq:fem-2} yields
$
\boldsymbol H_h = \nabla\phi_h.
$
\end{proof}
The following well-posedness results for the finite element scheme \eqref{eq:fem-3} - \eqref{eq:fem-4} of the Navier-Stokes equations are standard (cf. \cite{Girault1986}).
\begin{theorem}\label{lem:solu3-4}
The finite element scheme
\eqref{eq:fem-3} - \eqref{eq:fem-4} has at least one solution $(\boldsymbol u_h,\tilde p_h)\in \boldsymbol V_h \times W_h$, and there holds
$$
\|\nabla\boldsymbol u_h\| \leq \frac 1 \eta \|\boldsymbol f\|_{-1}.
$$ Further more, if
\begin{equation}\label{uni-NS}
(\tilde{\mathcal N}/\eta^2)\|\boldsymbol f\|_{-1} < 1\quad\text{with}\quad \tilde{\mathcal N} = \sup\limits_{\boldsymbol w_h,\boldsymbol u_h,\boldsymbol v_h \in \boldsymbol V_h}\frac{b(\boldsymbol w_h;\boldsymbol u_h,\boldsymbol v_h)}{\|\nabla \boldsymbol w_h\|\|\nabla \boldsymbol u_h\|\|\nabla \boldsymbol v_h\|},
\end{equation}
then the finite element solution $(\boldsymbol u_h,\tilde p_h)$ is unique.
\end{theorem}
\subsection{Error analysis}
In this subsection, we will give some error estimates for the finite element scheme \eqref{eq:fem-1} - \eqref{eq:fem-4}, under some rational assumptions.
Firstly, we have the following error estimate for the discrete solution $\phi_h$.
\begin{theorem}\label{lem:sigma-phi}
Let $\phi \in H_{0}^{1}(\Omega)\cap H^{l+2}(\Omega)$ be the solution of \eqref{eq:weak-maxwell-1}, and let $\phi_{h} \in S_{h}$ be the solution of \eqref{eq:fem-1}. Assume that $\phi$ satisfies \eqref{eq:ass-phi}, i.e. $
\|\nabla\phi\|_{L^\infty} < 1/C_\alpha,
$ then there exists a positive constant $C_s$, independent of $h$, such that
$$
\|\nabla(\phi - \phi_{h})\| \leq \frac{C_s(1+C_1) }{1 - C_\alpha\|\nabla\phi\|_{L^\infty}} h^{l+1} \|\phi\|_{l+2}.
$$
\end{theorem}
\begin{proof}
Let $\phi_h^* \in S_h$ be the classical interpolation of $\phi$, satisfying the estimate
\begin{equation}\label{phi-interpolation}
\|\nabla(\phi - \phi_h^*)\|\leq C_sh^{l+1}\|\phi\|_{l+2},
\end{equation}
where $C_s>0$ is a constant independent of $h$ and $\phi$.
In view of the inequality
$$
\|\nabla(\phi - \phi_h)\| \leq \|\nabla(\phi - \phi_h^*)\| + \|\nabla(\phi_h^* - \phi_h)\|,
$$
it suffices to estimate the term $ \|\nabla(\phi_h^* - \phi_h)\|$.
To this end, we apply
Lemma \ref{lem:cocervity} and the relation
$$
a(\phi;\phi,\tau_h) = a(\phi_h;\phi_h,\tau_h),\quad \forall \ \tau_h \in S_h,
$$
to get
\begin{eqnarray*}
\|\nabla(\phi_h^* - \phi_h)\|^2 & \leq& a(\phi_h;\phi_h^* - \phi_h,\phi_h^* - \phi_h) \\
&= &a(\phi_h;\phi_h^* - \phi,\phi_h^* - \phi_h) + a(\phi_h;\phi - \phi_h,\phi_h^* - \phi_h).
\end{eqnarray*}
Since
$$
|a(\phi_h;\phi_h^* - \phi,\phi_h^* - \phi_h)| \leq C_1\|\nabla(\phi - \phi_h^*)\|\|\nabla(\phi_h^* - \phi_h)\|
$$
and
\begin{align*}
a(\phi_h;\phi - \phi_h,\phi_h^* - \phi_h) =& a(\phi_h;\phi,\phi_h^* - \phi_h) - a(\phi;\phi,\phi_h^* - \phi_h) \\
=& \left( \big(\alpha(|\nabla \phi_h|) - \alpha(|\nabla\phi|)\big)\nabla\phi, \nabla(\phi_h^* - \phi_h)\right) \\
\leq &C_\alpha\|\nabla\phi\|_{L^\infty}\|\nabla(\phi - \phi_h)\|\|\nabla(\phi_h^* - \phi_h)\| \\
\leq & C_\alpha\|\nabla\phi\|_{L^\infty}\big(\|\nabla(\phi - \phi_h^*)\| + \|\nabla(\phi_h^* - \phi_h)\|\big)\|\nabla(\phi_h^* - \phi_h)\|,
\end{align*}
we have
\begin{eqnarray*}
\|\nabla(\phi_h^* - \phi_h)\| \leq C_1\|\nabla(\phi - \phi_h^*)\|+C_\alpha\|\nabla\phi\|_{L^\infty}\big(\|\nabla(\phi - \phi_h^*)\| + \|\nabla(\phi_h^* - \phi_h)\|\big),
\end{eqnarray*}
which, together with \eqref{eq:ass-phi}, implies
\begin{eqnarray*}
\|\nabla(\phi_h^* - \phi_h)\| \leq \frac{C_1 +C_\alpha\|\nabla\phi\|_{L^\infty}}{1 - C_\alpha\|\nabla\phi\|_{L^\infty}} \|\nabla(\phi - \phi_h^*)\|.
\end{eqnarray*}
This inequality, together with \eqref{phi-interpolation}, indicates the desired conclusion. \end{proof}
In light of Theorems \ref{H-well} and \ref{Hh-well}, we know that
$$
\boldsymbol H = \nabla\phi,\quad\boldsymbol H_{h} = \nabla\phi_{h} \quad \text{and} \quad {\rm curl\,} \boldsymbol H={\rm curl\,} \boldsymbol H_{h} =0.
$$
Thus, by Theorem \ref{lem:sigma-phi}, we immediately get the following error estimate for the discrete solution $\boldsymbol H_h$ of \eqref{eq:fem-2}.
\begin{theorem}\label{lem:H}
Let $\boldsymbol H \in \boldsymbol H_{0}({\rm curl\,})$ be the solution of \eqref{eq:weak-maxwell-2}, and let $\boldsymbol H_{h}\in \boldsymbol U_{h}$ be the solution of \eqref{eq:fem-2}. Then, under the same conditions as in Theorem \ref{lem:sigma-phi},
there holds
$$
\|\boldsymbol H - \boldsymbol H_{h}\|+\|{\rm curl\,}(\boldsymbol H - \boldsymbol H_{h})\| \leq \frac{C_s(1+C_1) }{1 - C_\alpha\|\nabla\phi\|_{L^\infty}} h^{l+1} \|\phi\|_{l+2}.
$$
\end{theorem}
The following error estimate for the Navier-Stokes system \eqref{eq:fem-3}-\eqref{eq:fem-4} is standard (cf. \cite{Girault1986}).
\begin{theorem}\label{lem:up}
Let $\boldsymbol u \in \boldsymbol H_{0}^{1}(\Omega) \cap\boldsymbol H^{l+2}(\Omega)$ and $\tilde p \in L_{0}^{2}(\Omega)\cap H^{l+1}(\Omega)$ be the solution of the Navier-Stokes system \eqref{eq:weak-NS-1}-\eqref{eq:weak-NS-2}, and let $\boldsymbol u_{h} \in \boldsymbol V_{h}$ and $\tilde p_{h} \in W_{h}$ be the solution of \eqref{eq:fem-3}-\eqref{eq:fem-4}. There exists a positive constant $C$, independent of $h$, such that
$$
\|\boldsymbol u - \boldsymbol u_{h}\|_{1} + \|\tilde p - \tilde p_{h}\| \leq C h^{l+1}(\|\boldsymbol u\|_{l+2} + \|\tilde p\|_{l+1}).
$$
\end{theorem}
As shown in Remark \ref{rem:sol}, for the the magnetization $\boldsymbol M=(\alpha(H)-1)\boldsymbol H$,
the new variable $\psi=\beta(H)$,
and the pressure $p=\tilde p + \mu_0 \psi$,
we can obtain their approximations $\boldsymbol M_h, \psi_h$ and $p_h$ from \eqref{eq:weak-maxwell-3-dis}, \eqref{eq:fem-5} and \eqref{eq:fem-6}, respectively. In view of Theorems \ref{lem:H} and \ref{lem:up}, we easily get the following error estimates for these three approximation solutions
\begin{corollary}\label{lem:M}
Assume that
$\boldsymbol M \in \boldsymbol H^{l+1}(\Omega)$ and
$\psi \in H^{l+1}(\Omega)$.
Under the same conditions as in Theorems \ref{lem:sigma-phi} and \ref{lem:up}, there exists positive constants $C_M$ and $C_\psi$, independent of $h$, such that
\begin{align}
&\|\boldsymbol M - \boldsymbol M_{h}\| \leq C_M h^{l+1}\left( \|\boldsymbol M\|_{l+1} + \frac{(1+C_1) (1+ C_1 + C_\alpha\|\nabla \phi\|_\infty ) }{1 - C_\alpha\|\nabla\phi\|_{L^\infty}} \|\phi\|_{l+2}\right),
\\
&\|\psi - \psi_{h}\| \leq \left( C_\psi \|\psi\|_{l+1} + \frac{M_sC_s(1+C_1) }{1 - C_\alpha\|\nabla\phi\|_{L^\infty}} \|\phi\|_{l+2} \right)h^{l+1}, \label{psi-h}
\\
&\|p - p_{h}\| \leq \left( C\|\boldsymbol u\|_{l+2} + C\|\tilde p\|_{l+1} + \mu_0C_\psi \|\psi\|_{l+1} + \frac{\mu_0M_sC_s(1+C_1) }{1 - C_\alpha\|\nabla\phi\|_{L^\infty}} \|\phi\|_{l+2} \right)h^{l+1}. \label{p-h}
\end{align}
\end{corollary}
\begin{proof}
Equation \eqref{eq:weak-maxwell-3-dis} implies that
$$
\boldsymbol M_h = Q_h^c((\alpha(H_h) - 1) \boldsymbol H_h),
$$
where $Q_h^c:~\boldsymbol L^2(\Omega) \rightarrow \boldsymbol U_h$ is the $L^2$ orthogonal projection operator. Using the Langevin law \eqref{eq:langevin}, we have
\begin{align*}
\boldsymbol M - \boldsymbol M_h = \boldsymbol M - Q_h^c\boldsymbol M + Q_h^c\big((\alpha(H) - 1)\boldsymbol H - (\alpha(H_h) - 1) \boldsymbol H_h\big).
\end{align*}
Thus, by Lemma \ref{lem:alphaH} (i), the boundedness of projection $Q_h^c$, and the relation $\boldsymbol H=\nabla \phi$, we get
\begin{align*}
\|\boldsymbol M - \boldsymbol M_h\| & \leq \|\boldsymbol M - Q_h^c\boldsymbol M\| + \|(\alpha(H) - \alpha(H_h))\boldsymbol H\| + \|(\alpha(H_h)-1)(\boldsymbol H - \boldsymbol H_h)\|\\
& \leq \|\boldsymbol M - Q_h^c\boldsymbol M\| + (C_\alpha\|\boldsymbol H\|_\infty + C_1 + 1)\|\boldsymbol H - \boldsymbol H_h\|\\
&\leq \|\boldsymbol M - Q_h^c\boldsymbol M\| + (C_\alpha\|\nabla \phi\|_\infty + C_1 + 1)\|\boldsymbol H - \boldsymbol H_h\|,
\end{align*}
which, together with Theorem \ref{lem:H} and the standard error estimation of the projection, gives the desired estimate for $\boldsymbol M_h$.
Similarly, equation \eqref{eq:fem-5} implies that
$$
\boldsymbol \psi_h = Q_h^w(\beta(H_h)),
$$
where $Q_h^w:~L^2(\Omega) \rightarrow W_h$ is the $L^2$ orthogonal projection. Then by Lemma \ref{lem:alphaH} (ii) we obtain
\begin{align*}
\|\psi - \psi_h\| &\leq \| \psi - Q_h^w\psi\| + \|Q_h^w\big(\beta(H) - \beta(H_H)\big)\|\\
&\leq \| \psi - Q_h^w\psi\| + \|\big(\beta(H) - \beta(H_h)\big)\|\\
&\leq \| \psi - Q_h^w\psi\| + M_s\|H - H_h \|\\
&\leq \| \psi - Q_h^w\psi\| + M_s\|\boldsymbol H - \boldsymbol H_h \|,
\end{align*}
which, together with Theorem \ref{lem:H} and the projection property, yields the desired estimate \eqref{psi-h}.
Finally, \eqref{eq:fem-6} means that
$$\|p-p_h\|\leq \| \tilde p-\tilde p_h\| + \mu_{0} \|\psi-\psi_h\|, $$
which, together with Theorem \ref{lem:up} and estimate \eqref{psi-h}, indicates the desired result \eqref{p-h}.
This completes the proof.
\end{proof}
\section{Numerical experiments}
This section is devoted to three numerical examples to verify the performance of the mixed finite element methods. In all the examples, we solve the nonlinear system \eqref{eq:fem-1} - \eqref{eq:fem-6} by using the {\em i}FEM package~\cite{Chen2009} and Algorithm \ref{alg}.
\begin{alg}\label{alg}
Given $\phi_h^0\in S_h$ and $\boldsymbol u_h^0 \in \boldsymbol V_h$, find $\phi_h \in S_h$, $\boldsymbol H_h \in \boldsymbol U_h$, $\boldsymbol M_h \in \boldsymbol U_h$, $\boldsymbol u_h \in \boldsymbol V_h$, $\tilde p_h \in W_h$, $\psi_h \in W_h$, and $p_h \in W_h$ through five steps:
\begin{enumerate}
\item[Step 1.] For $n = 1,2,\dots,L$ do
\begin{enumerate}
\item[(1)] Solving the nonlinear system \eqref{eq:fem-1} as: Find $\phi_h^n \in S_h$ such that
$$
a(\phi_h^{n-1};\phi_h^n,\tau_h) = -(g,\tau_h)\qquad\forall~\tau_h \in S_h.
$$
\item[(2)] Solving the nonlinear saddle point system \eqref{eq:fem-3}-\eqref{eq:fem-4} as: Find $\boldsymbol u_h^n \in \boldsymbol V_h$ and $\tilde p_h^n \in W_h$ such that
\begin{align*}
& \eta(\nabla\boldsymbol u_h^n,\nabla\boldsymbol v_h) + b(\boldsymbol u_h^{n-1};\boldsymbol u_h^n,\boldsymbol v_h) - (\tilde p_h^n,\nabla\cdot\boldsymbol v_h) = (\boldsymbol f,\boldsymbol v_h)
& \forall~\boldsymbol v_h \in \boldsymbol V_h,\\
& (\nabla\cdot\boldsymbol u_h^n,q_h) =
&\forall~q_h \in W_h.
\end{align*}
\end{enumerate}
\item[Step 2.] Let $\phi_h = \phi_h^L$, $\boldsymbol u_h = \boldsymbol u_h^L$ and $\tilde p_h = \tilde p_h^L$.
\item[Step 3.] Solving equation \eqref{eq:fem-2} to get $\boldsymbol H_h \in \boldsymbol U_h$.
\item[Step 4.] Solving equations \eqref{eq:weak-maxwell-3-dis} and \eqref{eq:fem-5} to get $\boldsymbol M_h \in \boldsymbol U_h$ and $\psi_h \in W_h$, respectively.
\item[Step 5.] Substituting $\tilde p_h $ and $\psi_h$ into \eqref{eq:fem-6} to get $p_h\in W_h$.
\end{enumerate}
\end{alg}
\begin{remark}
From the convergence theory of Newton-type methods~\cite{Gil2007,Suli2003}, we know that the iterative solution of Algorithm \ref{alg} will converge to the exact solution, provided that the iteration number $L$ is big enough and the initial guess is nearby the exact solution. In fact, in all the subsequent numerical examples we choose the initial guess $\phi_h^0$ as the corresponding finite element solution of the Poisson equation
$$
\Delta \tilde \phi = g \qquad\text{in }\ \Omega$$
with the same boundary condition as that of the exact solution $\phi$, and choose the initial guess $\boldsymbol u_h^0$ as the corresponding finite element solution of the Stokes equations
$$
\left\{
\begin{array}{ll}
-\eta\Delta\tilde{\boldsymbol u} + \nabla p = \boldsymbol f & \text{in }\Omega, \\
\nabla\cdot \tilde{\boldsymbol u} = 0 & \text{in }\Omega
\end{array}
\right.
$$
with the same boundary condition as that of the exact solution $\boldsymbol u$.
In so doing, the choice $L = 2$ works well in the algorithm.
\end{remark}
\begin{figure}[h!]
\centering
\includegraphics[width=0.70\textwidth]{2Dmesh.jpg}
\caption{The domain $\Omega = (0,1)^2$: $4\times 4$ (left) and $8 \times 8$ (right).}\label{fig:2Dmesh}
\end{figure}
\begin{example}
\label{ex:2D}
This is a 2D test. We take $\Omega = (0,1)^2$, and use $N \times N$ uniform triangular mesh (cf. Figure \ref{fig:2Dmesh}) with $N = 4,~8,~16,~32,~64,~128$.
The exact solution of the FHD model \eqref{eq:eq-sys} is given by
\begin{align*}
&\phi(x,y) = (x^2-x)(y^2-y),\\
&\boldsymbol H(x,y) = ((2x-1)(y^2-y),(x^2-x)(2y-1))^\intercal,\\
&\boldsymbol M(x,y) = M_s\left(\coth(\gamma H) - \frac{1}{\gamma H} \right)\frac{\boldsymbol H}{H}, \\
& \boldsymbol u(x,y) = (\sin(\pi y),\sin(\pi x))^\intercal,\\
&p(x,y) = 60x^2y - 20y^3 - 5,
\end{align*}
with the parameters $M_s$, $\rho$, $\eta$, $\mu_0$ and $\gamma$ all being chosen as $1$.
We use the conforming linear ($\mathbb P_1$) element to discretize the variable $\phi$, the lowest order edge element $NE_0$ to discretize the variables $\boldsymbol H$ and $\boldsymbol M$, the $CR$ (nonconforming-$\mathbb P_1$) element to discretize the variables $\boldsymbol u$, and the piecewise constant ($\mathbb P_0$) element to discretize the variables $\tilde p$, $\psi$ and $p$. Note that such a combination of finite element spaces corresponds to $l = 0$, then
we easily see from Theorems \ref{lem:sigma-phi} - \ref{lem:up} and Corollary \ref{lem:M} that the theoretical accuracy of the scheme is $\mathcal O(h)$.
Numerical results are listed in Table \ref{tab:2Dex-1}.
\begin{table}[h!]
\footnotesize
\begin{center}
\caption{Numerical results for Example \ref{ex:2D}.}
\label{tab:2Dex-1}
\begin{tabular}{c|c|c|c|c|c|c}
\hline
\hline $N$ & $\frac{ \|\nabla(\phi - \phi_{h})\|}{\|\nabla\phi\|} $ &$\frac{ \|\boldsymbol H - \boldsymbol H_{h}\|_c}{\|\boldsymbol H\|_{c}}$ & $\frac{\|\boldsymbol M - \boldsymbol M_{h}\|}{\|\boldsymbol M\|}$ & $\frac{\|\boldsymbol u - \boldsymbol u_{h}\|_{1,h}}{\|\boldsymbol u\|_{1}}$ & $\frac{\|p - p_h\|}{\|p\|}$ & $\|{\rm curl\,}\boldsymbol H_h\|_\infty$\\
\hline
$4$ & $0.3943$ & $0.3943$ & $0.3941$ & $0.7424$ & $0.3083$ & $0.0003e-12$\\
\hline
$8$ & $0.2023$ & $0.2023$ & $0.2023$ & $0.4385$ & $0.1524$ & $0.0013e-12$ \\
\hline
$16$ & $0.1018$ & $0.1018$ & $0.1018$ & $0.2352$ & $0.0717$ & $ 0.0083e-12$\\
\hline
$32$ & $0.0510$ & $0.0510$ & $0.0510$ & $0.1207$ & $0.0342$ & $ 0.0315e-12$\\
\hline
$64$ & $0.0255$ & $0.0255$ & $0.0255$ & $0.0609$ & $0.0167$ & $ 0.1485e-12$\\
\hline
$128$& $0.0128$ & $0.0128$ & $0.0128$ & $0.0305$ & $0.0083$ & $0.6481e-12 $\\
\hline
order & $0.9900$ & $0.9900$ & $0.9899$ & $0.9207$ & $1.0439$ & --\\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
\end{example}
\begin{figure}[!h]
\centering
\includegraphics[width=0.70\textwidth]{meshinfo.jpg}
\caption{The domain $\Omega = (0,1)^3$: $2\times 2\times 2$ (left) and $4\times 4\times 4$ (right).}\label{fig:mesh}
\end{figure}
The other two experiments, Examples \ref{ex-1} and \ref{ex-2}, are 3D tests, with $\Omega = (0,1)^3$ and use $N \times N \times N$ uniform tetrahedral meshes (cf. Figure \ref{fig:mesh}) with $N = 4,~8,~16$.
\begin{example} [The lowest order approximation]
\label{ex-1}
In this 3D test, the exact solution of the FHD model \eqref{eq:eq-sys} is given by
\begin{align*}
& \phi = \sin(\pi x)\sin(\pi y)\sin(\pi z), \\
&\boldsymbol H = \pi\begin{pmatrix}
\cos(\pi x)\sin(\pi y)\sin(\pi z)\\
\sin(\pi x)\cos(\pi y)\sin(\pi z) \\
\sin(\pi x)\sin(\pi y)\cos(\pi z)
\end{pmatrix}, \\
& \boldsymbol M = M_{s}\left(\coth(\gamma H)- \frac{1}{\gamma H} \right)\frac{\boldsymbol H}{H},\\
&\boldsymbol u = \begin{pmatrix}
\sin(\pi y), & \sin(\pi z) , & \sin(\pi x)
\end{pmatrix}^\intercal,\\
& p = 120x^2yz-40y^3z-40yz^3,
\end{align*}
with the parameters $M_s$, $\gamma$, $\rho$ and $\eta$ all being choosen as $1$ and $\mu_0 = 10$.
We use the the conforming $\mathbb P_1$-element to discretize the variable $\phi$, the lowest order N\'ed\'elec finite element space $NE_0$ \cite{Nedelec1980,Nedelec1986} to discretize the variables $\boldsymbol H$ and $\boldsymbol M$, the nonconforming $CR$ element \cite{Crouzeix1973CRele} to discretize the variable $\boldsymbol u$, and the piecewise constant $\mathbb P_0$ element to discretize the variables $\tilde p$, $\psi$ and $p$. With these settings, from Theorems \ref{lem:sigma-phi} - \ref{lem:up} and Corollary \ref{lem:M} we see that the theoretical accuracy of the scheme is $\mathcal O(h)$. Numerical results are given in Table \ref{tab:ex-1}.
\begin{table}[h!]
\footnotesize
\begin{center}
\caption{Numerical results for Example \ref{ex-1}.}
\label{tab:ex-1}
\begin{tabular}{c|c|c|c|c|c|c}
\hline
\hline $N$ & $\frac{ \|\nabla(\phi - \phi_{h})\|}{\|\nabla\phi\|} $ &$\frac{ \|\boldsymbol H - \boldsymbol H_{h}\|_c}{\|\boldsymbol H\|_{c}}$ & $\frac{\|\boldsymbol M - \boldsymbol M_{h}\|}{\|\boldsymbol M\|}$ & $\frac{\|\boldsymbol u - \boldsymbol u_{h}\|_{1,h}}{\|\boldsymbol u\|_{1}}$ & $\frac{\|p - p_h\|}{\|p\|}$ & $\|{\rm curl\,}\boldsymbol H_h\|_\infty$\\
\hline
$4$ & $0.4739$ & $0.4739$ & $0.4755$ & $0.2066$ & $0.3665$ & $0.0213e-12$\\
\hline
$8$ & $0.2491$ & $0.2491 $ & $0.2506$ & $0.1082$ & $0.1803$ & $0.1048e-12$\\
\hline
$16$ & $0.1262$ & $0.1262$ & $0.1271$ &
$0.0551$ & $0.0835$ & $0.6821e-12$\\
\hline
order & $0.9545$ & $0.9545$ & $0.9515$ & $0.9529$ & $1.0671$ & $--$ \\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
\end{example}
\begin{example}
[A higher order approximation]
\label{ex-2}
In this 3D test, the exact solution of the FHD model \eqref{eq:eq-sys} is given by
\begin{align*}
& \phi(x,y,z) = (x^2+y^2+z^2)(x^2-x)(y^2-y)(z^2-z), \\
& \boldsymbol H (x,y,z) = \begin{pmatrix}
2x(x^2-x)(y^2-y)(z^2-z)+(2x-1)(x^2+y^2+z^2)(y^2-y)(z^2-z) \\
2y(x^2-x)(y^2-y)(z^2-z)+(2y-1)(x^2+y^2+z^2)(x^2-x)(z^2-z) \\
2z(x^2-x)(y^2-y)(z^2-z)+(2z-1)(x^2+y^2+z^2)(x^2-x)(y^2-y)
\end{pmatrix}, \\
& \boldsymbol M(x,y,z) = M_{s}\left(\coth(\gamma H)- \frac{1}{\gamma H} \right)\frac{\boldsymbol H}{H},\\
& \boldsymbol u(x,y,z) = \begin{pmatrix}
\sin(\pi y), & \sin(\pi z) , & \sin(\pi x)
\end{pmatrix}^\intercal,\\
& p(x,y,z) = 120x^2yz-40y^3z-40yz^3,
\end{align*}
where the parameters $M_s$, $\gamma$, $\rho$, $\eta$ and $\mu_0$ are all choosen as $1$.
We use the conforming quadratic ($\mathbb P_2$) element to discretize the variable $\phi$, the first order N\'ed\'elec edge element $NE_1$ \cite{Nedelec1980,Nedelec1986} to discretize the variables $\boldsymbol H$ and $\boldsymbol M$, the Taylor-Hood $\mathbb P_2$-$\mathbb P_1$ element to discretize the variables $\boldsymbol u$ and $\tilde p$, and the continuous linear ($\mathbb P_1$) element to discretize the variables $\psi$ and $p$. With these settings, we see
that the theoretical accuracy of the scheme is $\mathcal O(h^2)$.
We list numerical results in Table \ref{tab:ex-2}.
\begin{table}[h!]
\footnotesize
\begin{center}
\caption{Numerical results for Example \ref{ex-2}.}
\label{tab:ex-2}
\begin{tabular}{c|c|c|c|c|c|c}
\hline
\hline $N$ & $ \frac{\|\nabla(\phi - \phi_{h})\|}{\|\nabla\phi\|} $ &$\frac{ \|\boldsymbol H - \boldsymbol H_{h}\|_c}{\|\boldsymbol H\|_{c}}$ & $\frac{\|\boldsymbol M - \boldsymbol M_{h}\|}{\|\boldsymbol M\|}$ & $\frac{\|\nabla(\boldsymbol u - \boldsymbol u_{h})\|}{\|\nabla\boldsymbol u\|}$ & $\frac{\|p - p_h\|}{\|p\|}$ & $\|{\rm curl\,}\boldsymbol H_h\|$\\
\hline
$4$ & $0.1369$ & $0.1369$ & $0.1369$ & $0.0450$ & $0.0535$ & $0.2970e-08$\\
\hline
$8$ & $0.0372$ & $0.0372$ & $0.0372$ & $0.0081$ & $0.0135$ & $0.7640e-08$\\
\hline
$16$ & $0.0095$ & $0.0095$ & $0.0095$ &
$0.0016$ & $ 0.0034$ & $1.6060e-08$\\
\hline
order & $1.9245$ & $1.9245$ & $1.9245$ & $2.4069$ & $1.9880$ & $--$\\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
\end{example}
From Examples \ref{ex:2D}, \ref{ex-1} and \ref{ex-2}, we have the following observations:
\begin{itemize}
\item From Tables \ref{tab:2Dex-1} and \ref{tab:ex-1}, we see that the errors in $H^1$ semi-norm for $\phi$, $\boldsymbol H({\rm curl\,})$ norm for $\boldsymbol H$, $L^2$ norm for $\boldsymbol M$ and $p$, and discrete $H^1$ norm for $\boldsymbol u$, all have the first (optimal) order rate.
\item From Table \ref{ex-2}, we see that the errors in $H^1$ semi-norm for $\phi$ and $\boldsymbol u$, $\boldsymbol H({\rm curl\,})$ norm for $\boldsymbol H$, $L^2$ norm for $\boldsymbol M$ and $p$, all have the second (optimal) order rate.
\item From Tables \ref{tab:2Dex-1} - \ref{tab:ex-2}, we can see that the numerical scheme preserves ${\rm curl\,}\boldsymbol H_h = 0$ exactly.
\end{itemize}
|
\section{Introduction}
All topological spaces are assumed to be Hausdorff unless a stronger separation axiom is specified. A space is feebly compact if every locally finite family of non-empty open sets is finite. In the class of Tychonoff spaces feeble compactness is equivalent to pseudocompactness and throughout this paper a pseudocompact space is always assumed to be Tychonoff. As usual, $cl_X(A)$ (or $\overline{A}$) denote the closure of a subset $A$ of a topological space $X$. We denote by $\omega$ the collection of finite ordinal numbers. All undefined terms can be found in \cite{AT} or \cite{E}.
\begin{defn}\label{comboun}Let $X$ be a topological space,
\begin{enumerate}
\item $X$ is \textit{compact-bounded} if whenever $\mathcal U$ is a countable family of non-empty open subsets of $X$, there is some compact subset $K\subset X$ which meets each element of $\mathcal U$.
\smallskip
\item $X$ is \textit{weakly compact-bounded} if whenever $\mathcal U$ is a countable family of non-empty open subsets of $X$, there is a compact subset $K\subset X$ which intersects infinitely many elements of $\mathcal{U}$.
\end{enumerate}
\end{defn}
The compact-boundedness and weak compact-boundedness properties were introduced by J. Angoa, Y. F. Ortiz-Castillo and A. Tamariz-Mascar\'ua in the class of Tychonoff spaces under the name \textit{pseudo-$\omega$-boundedness} and \textit{weak pseudo-$\omega$-boundedness} respectively in \cite{AOT}.
In this paper we study compact-bounded and weakly compact-bounded spaces in the class of topological and paratopological groups. In section 2, we study basic properties of these spaces in the class of topological spaces. In section 3, we focus on finding conditions under which a paratopological group is a topological group, for this, we find a relation of compact-bounded spaces with the 2-pseudocompact spaces.
Given a subgroup $H$ of a topological group $G$,
in section 4, we study how to extend the properties of the quotient space $G/H$ to $G$, this, in the cases when $H$ is either a compact, locally compact or a (weakly) compact-bounded subgroup. In section 5, we find a condition on a Tychonoff space $X$ for which the free topological group $F(X)$ and the free abelian topological group $A(X)$ admits an open continuous homomorphism onto the circle group $\mathbb{T}$.
Finally, in the section 6, we study the $r$-weakly compact-bounded subsets of a topological space $X$ and find its relationship with the strongly $r$-pseudocompact, $r$-pseudocompact and $C$-compact subsets of $X$.
\section{Basic properties}
In this section we present basic properties of (weakly) compact-bounded spaces.
The next fact shows that the family $\mathcal{U}$ of the Definition \ref{comboun}-(2) can be considered as a family of mutually disjoint non-empty open sets of the topological space $X$.
\begin{prop}[Proposition 2.4 in \cite{AOT}]\label{disj} For any topological space X, the following are equivalent.
\begin{enumerate}
\item X is weakly compact-bounded;
\item For any countable family of mutually disjoint non-empty open subsets of $X$, there is a compact subspace $K\subset X$ which intersects infinitely many elements of $\mathcal{U}$.
\end{enumerate}
\end{prop}
A topological space with the property (2) of the previous proposition is called \textit{weakly cellular-compact} (\cite{AMW}).
Recall that a space is \textit{weakly regular} if every non-empty open set contains a non-empty regular closed subset. Using the Proposition \ref{disj}, the proof of the next result is identical to that of Proposition 4.9 of \cite{PTV}, in which complete regularity is assumed, but weak regularity is all that is required.
\begin{prop}\label{fk} A topological space $X$ is a weakly regular $k$-space, then $X$ is weakly compact-bounded if and only if $X$ is feebly compact.
\end{prop}
\begin{cor}\label{ksp} If G is a topological group k-space, then $G$ is weakly compact-bounded if and only if $G$ is pseudocompact.
\end{cor}
\begin{ques} Is a paratopological group feebly compact k-space (Fr\'echet), weakly compact-bounded?
\end{ques}
The following proposition is immediate from the definition of compact-boundedness.
\begin{prop}\label{isolated} If $X$ is compact-bounded, then the closure of a countable set of isolated points of $X$ is compact; in particular, if a compact-bounded space has a countable dense set of isolated points, then it is compact.
\end{prop}
The following implications trivially hold.
\smallskip
\begin{center}
compact-bounded $\Longrightarrow$ weakly compact-bounded $\Longrightarrow$ feebly compact.
\end{center}
\smallskip
The following examples illustrate the differences between these spaces.
\begin{exam} There is a pseudocompact topological group not weakly compact-bounded.
\end{exam}
\begin{proof} It is easy to check that every weakly compact-bounded space $X$ has the property: For each family $\{U_n:n\in \omega \}$ of mutually disjoint non-empty open sets in $X$ there is a subset $\{x_n : n\in \omega \}$ of points in $X$ such that $cl_X (\{x_n:n\in \omega \})\setminus (\bigcup \{U_n:n\in \omega \})\neq \emptyset$ and $x_n\in U_n$ for each $n\in \omega$, that is, the set $cl_X (\{x_n:n\in \omega \})$ is not closed.
The topological group described in Example 2.4 in \cite{GT} is a pseudocompact topological group and it doesn't have the previous property, thus this is not weakly cellular-compact.
\end{proof}
\begin{exam}\label{Mr} There is an weakly compact-bounded space not compact-bounded.
\end{exam}
\begin{proof} If $X=\Psi(\mathcal{A})$ is the Isbell-Mr\'owka space (where $\mathcal{A}$ is a maximal almost disjoint family on $\omega$), then $X$ is locally compact and pseudocompact, thus it is weakly compact-bounded space (Proposition \ref{fk}). But $\omega$ is a countable dense subset of isolated points of $X$, thus $X$ is not compact-bounded (Proposition \ref{isolated}).
\end{proof}
\begin{ques} Is there a weakly compact-bounded paratopological group not compact-bounded?
\end{ques}
\begin{prop}\label{imgreg} Suppose that X is a (weakly) compact-bounded space. Then
\begin{enumerate}[(a)]
\item any continuous image of X is (weakly) compact-bounded.
\item any regular closed subset of X is (weakly) compact-bounded.
\end{enumerate}
\end{prop}
\begin{proof} The proof of (1) is trivial. We will to show (2). Let $A$ be a regular closed subspace of $X$ and let $\{U_n \cap A: n\in \omega \}$ be a disjoint family of non-empty open subsets of $A$, where $U_n$ is open in $X$ for each $n\in \omega$.
Since $\{ U_n \cap{\rm int}(A):n\in \omega\}$ is a disjoint family of non-empty open subsets of $X$, then there is a compact subspace $K\subset X$ which intersects infinitely many elements of $\{ U_n \cap {\rm int}(A):n\in \omega\}$. Clearly, $K\cap A$ is a compact subspace of $A$ and so $A$ is (weakly) compact-bounded.
\end{proof}
Both parts of the following proposition have simple proofs which we omit.
\begin{prop}\label{finden} Given any space X,
\begin{enumerate}[(a)]
\item if X is a finite union of (weakly) compact-bounded spaces, then X is (weakly) compact-bounded.
\item if X contains a dense (weakly) compact-bounded space, then X is (weakly) compact-bounded.
\end{enumerate}
\end{prop}
\begin{prop}\label{subCB} For any topological space X, the following are equivalent:
\begin{enumerate}
\item X is weakly compact-bounded;
\item For any family $\mathcal{U}$ of non-empty open subsets of X, there is a weakly compact-bounded subspace $A\subset X$ such that $A$ meets infinitely many elements of $\mathcal{U}$.
\end{enumerate}
\end{prop}
\begin{proof}(1)$\Rightarrow$(2) is clear since a compact space is weakly compact-bounded.
(2)$\Rightarrow$(1) Let $\mathcal{U}=\{U_n:n\in \omega \}$ be a family of non-empty open subsets of $X$ and let $A$ be a weakly compact-bounded subspace of $X$ such that $A$ meets infinitely-many elements of $\mathcal{U}$, that is, there is a infinite subset $I\subset \omega$ such that $A\cap U_m\neq \emptyset$ for each $m\in I$. Since the family $\{A\cap U_m:m\in I\}$ is a family of non-empty open subsets of $A$, there is a compact subset $K_A$ in $A$ (thus in $X$) such that $\emptyset \neq K_A\cap A\cap U_m \subset K_A\cap U_m $ for each $m\in I$, that is, $K_A$ meets infinitely many elements of $\mathcal{U}$.
\end{proof}
The next fact has a proof similar to the previous proposition.
\begin{prop}\label{subCB} For any topological space X, the following are equivalent:
\begin{enumerate}
\item X is compact-bounded;
\item For any family $\mathcal{U}$ of non-empty open subsets of X, there is a compact-bounded subspace $A\subset X$ such that $A$ meets every element of $\mathcal{U}$.
\end{enumerate}
\end{prop}
A space $X$ is said to have a \textit{regular $G_\delta$-diagonal} if the diagonal $\Delta_X=\{(x,x): x\in X \}$ can be represented as the intersection of the closures of a countable family of open neighbourhoods of $\Delta_X$ in $X\times X$.
We end this section with the following fact about (weakly) compact-bounded and first-countable subsets of topological groups.
\begin{prop} Let G be a topological group. Then, the following assertions are equivalent.
\begin{enumerate}
\item every (weakly) compact-bounded subset of G is first-countable;
\item every (weakly) compact-bounded subset of G is metrizable.
\end{enumerate}
\end{prop}
\begin{proof} Clearly, it will suffice to show the sufficiency. Let $A$ be a weakly compact-bounded (compact-bounded) subset of $G$, and let
$g: G\times G \to G$ be a map defined by $g(x,y)=x^{-1}y$ for every $x,y \in G$ (we note that $g$ is a continuous mapping). By Theorem 4.4 in \cite{AMW}, we have that $A\times A$ is weakly compact-bounded (compact-bounded), it follows that $g(A\times A)=A^{-1}A$ is weakly compact-bounded (compact-bounded), and thus $A^{-1}A$ is first countable.
Since the neutral element $e$ is in $ A^{-1}A$, there is a family $\{W_n : n\in \omega\}$ of non-empty open subsets of $e$ in $A^{-1}A$ such that $\{ e\}=\bigcap \{W_n : n\in \omega\}=\bigcap \{cl_{ A^{-1}A}(W_n) : n\in \omega\}$. If $f=g|(A\times A)$, then $\Delta_A =f^{-1}(\{ e\})$. Now, we have
$$
f^{-1}(\{ e\})=\bigcap \{ f^{-1}(W_n): n\in \omega\} \subseteq \bigcap \{cl_{ A\times A} (f^{-1}(W_n)): n\in \omega\} \subseteq
$$
$$
\subseteq \bigcap \{ f^{-1}(cl_{A^{-1}A}(W_n)): n\in \omega\}= f^{-1} \left( \bigcap \{ cl_{A^{-1}A}(W_n): n\in \omega\} \right).
$$
\medskip
On the other hand, $f^{-1}(\{ e\})=f^{-1}(\bigcap \{ cl_{A^{-1}A}(W_n): n\in \omega\})$. Thus, $\Delta_A = f^{-1}(\{ e\})= \bigcap \{cl_{ A\times A} (f^{-1}(W_n)): n\in \omega\}$, that is, $A$ has regular $G_\delta$-diagonal by Corollary 4 in \cite{AB} we conclude that $A$ is compact and metrizable.
\end{proof}
\section{When a (weakly) compact-bounded paratopological group is a topological group}
The following examples illustrate that not every weakly compact-bounded paratopological group is a topological group:
\begin{exam}There exists a weakly compact-bounded paratopological group which is not a topological group.
\end{exam}
\begin{proof} The paratopological group $H$ described in the Example 2.7.10 in \cite{lib} is feebly compact and second countable, thus it is weakly compact-bounded (Proposition \ref{fk}), however it is not a topological group.
\end{proof}
For a paratopological group $(G,\tau)$, the family $\{U^{-1}: U\in \tau \}$ defines a topology $\tau^{-1}$, named the \textit{conjugate} of $\tau$ and the pair $(G,\tau^{-1})$ is a paratopological group homeomorpic to $(G,\tau)$.
We will denote by $\Delta_G $ the diagonal of the direct product $(G,\tau^{-1})\times (G,\tau)$. By Lemma 2.2 in \cite{AS}, we know that if $G$ is a paratopological group, then the diagonal $(\Delta_G, \tau^{-1} \times \tau)$ of the direct product $G\times G$ is a topological group. Moreover, $\Delta_G$ is closed in $(G,\tau)\times (G,\tau^{-1})$ whenever that $(G,\tau)$ is a paratopological group $T_1$.
\begin{prop} Let G be a paratopological group. If the diagonal $(\Delta_G, \tau^{-1} \times \tau)$ is (weakly) compact-bounded, then $G$ is a (weakly) compact-bounded topological group.
\end{prop}
\begin{proof} Suppose that $(\Delta_G, \tau^{-1} \times \tau)$ is (weakly) compact-bounded. Since $G$ is a paratopological group, we have that $(\Delta_G, \tau^{-1} \times \tau)$
is a topological group (Lemma 2.2 in \cite{AS}), furthermore, since $(\Delta_G, \tau^{-1} \times \tau)$ is (in particular) feebly compact, then $(\Delta_G, \tau^{-1} \times \tau)$ is pseudocompact, by Lemma 2.3 in \cite{AS}, we conclude that $G$ is a topological group.
Now, since $G$ is a topological group, clearly $\tau = \tau^{-1}$, it follows that $(\Delta_G, \tau^{-1} \times \tau)=(\Delta_G, \tau \times \tau)$. Since, the projection $\pi: (\Delta_G, \tau \times \tau) \to G$ defined by $\pi(x,x)=x$ is a topological isomorphic, and $(\Delta_G, \tau^{-1} \times \tau)$ is (weakly) compact-bounded, we have that $G$ is (weakly) compact-bounded (Proposition \ref{imgreg}-(1)).
\end{proof}
A paratopological group $G$ is \textit{2-pseudocompact} if whenever $\{U_n :n\in \omega\}$ is a decreasing sequence of non-empty open subsets of $G$, we have $\bigcap \{ cl_G(U_n^{-1}):n\in \omega \} \neq \emptyset$.
\begin{prop} Let $G$ be a compact-bounded paratopological group. If for every compact subset $K\subset G$, the set $K^{-1}$ is countably compact in $G$, then $G$ is 2-pseudocompact.
\end{prop}
\begin{proof} Let $\{U_n :n\in \omega\}$ be a decreasing sequence of non-empty open subsets of $G$. Then, there is a compact set $F$ of $G$ such that $F\cap U_n \neq \emptyset $ for each $n\in \omega$. We note that since $F$ is compact in $G$, we have that $cl_G(F^{-1})=F^{-1}$ (Lemma 2.3.5 in \cite{AT}). It follows that, $\emptyset \neq F^{-1}\cap U_n^{-1} \subset cl_G({F^{-1}\cap U_n^{-1}}) \subset cl_G({F^{-1}})\cap cl_G({U_n^{-1}}) \subset F^{-1}\cap cl_G({U_n^{-1}})$
for each $n\in \omega$. Clearly, the family $\{ F^{-1}\cap cl_G(U_n^{-1}) : n\in \omega\}$ is a decreasing sequence of non-empty closet subsets of the countably compact subset $K^{-1}$. Thus, we can conclude that
$$
\emptyset \neq \bigcap\{ F^{-1}\cap cl_G(U_n^{-1}): n\in \omega \}\subset \bigcap \{ cl_G(U_n^{-1}): n\in \omega \}.
$$
Therefore, $G$ is 2-pseudocompact.
\end{proof}
Since every 2-pseudocompact is a Baire space, by the propositions 3.15 and 3.24 in \cite{BR}, we have some conditions under which a compact-bounded paratopological group is a topological group.
\begin{cor} Let G be a compact-bounded paratopological group such that the set $K^{-1}$ is countably compact in G for each compact subset $K\subset G$. If at least one of the following condition holds:
\begin{enumerate}
\item G is $\omega$-precompact;
\item G is topologically periodic;
\item G has countable pseudocharacter.
\end{enumerate}
Then $G$ is a topological group.
\end{cor}
\section{Quotients}
E. Hewitt and K. A. Ross showed in \cite{HR} that if $G$ is a topological group and $H$ is a compact subgroup of $G$, then the quotient mapping $\pi:G\to G/H$ is a perfect map. This fact helps to extend some compactness-like properties from $G/H$ to $G$, which are inverse invariants of perfect mappings, such as local compactness, countable compactness, paracompactness, the Lindelöf property, $\sigma$-compactness and \v{C}ech-completeness.
The Mrówka space $\Psi(\mathcal{D})$ (where $\mathcal{D}$ is a maximal almost disjoint family on $\omega$) is a weakly compact-bounded space. The subspace $\mathcal{D}$ of $\Psi(\mathcal{D})$ is closed, discrete and infinite (thus, not pseudocompact). That is, weak compact-boundedness is not hereditary with respect to closed subspaces.
It follows, from Lemma \ref{imgreg}-(2) and Theorem 3.7.29 in \cite{E}, that these property is not inverse invariant of perfect mappings. Thus, it is not immediate the fact that if the group $G/H$ is weakly compact-bounded, then $G$ is weakly compact-bounded. However, the following fact shows a way to prove this.
\begin{prop}\label{quotient} Let G be a topological group and H is a compact subgroup of G such that the quotient space $G/H$ has any of the following properties:
\begin{enumerate}
\item G/H is weakly compact-bounded,
\item G/H is compact-bounded.
\end{enumerate}
Then, G has the same property.
\end{prop}
\begin{proof} We will to show (2) (the proof of (1) is similar). Let $\{U_n :n\in \omega\}$ be a family of non-empty open sets in $ G$. Then, $\{\pi (U_n) :n\in \omega\}$ is a family of non-empty open sets in $G/H$, thus there is a compact $E$ in $G/H$ such that $\pi(U_n) \cap E \neq \emptyset $ for each $n\in \omega$. We note that the set $K=\pi^{-1}(E)$ is compact in $ G$, because $\pi:G\to G/H$ is a perfect map (Theorem 3.7.2 in \cite{E}).
Since the map $\pi$ is surjective, we have that $ \pi(U_n) \cap E =\pi(U_n) \cap \pi(\pi^{-1}(E))=\pi(U_n) \cap \pi(K) \neq \emptyset $ for each $n\in \omega$. It follows that, for each $n\in \omega$, there is $x_n\in U_n$ and $k_n \in K$, such that $\pi(x_n)=\pi(k_n)$, thus $x_n \in k_n H\subset KH$ for each $n\in \omega$. Since $KH$ is compact in $ G$ (Proposition 1.4.31 in \cite{AT}), we have that the set $F=cl(\{x_n:n\in \omega\})$ is also compact in $G$ and clearly, $F\cap U_n \neq \emptyset $ for each $n\in \omega$. Therefore, $G$ is compact-bounded topological group.
\end{proof}
By Theorem 2.4 in \cite{A} and Proposition \ref{ksp}, we have:
\begin{cor} Let G be a topological group and H is a compact subgroup of G such that $G/H$ is pseudocompact and k-space, then G is weakly compact-bounded.
\end{cor}
The next fact is an immediate consequence of Proposition 6.3 in \cite{CR}-(b) and Proposition \ref{ksp}:
\begin{cor} Let G be a topological group k-space and H a closed, normal subgroup of G. If H and $G/H$ are pseudocompact spaces, then G is weakly compact-bounded.
\end{cor}
If $G$ is a topological group, $H$ is a locally compact subgroup of $G$, and $\pi:G \to G/H$ is the natural quotient mapping, by Corollary 2.1 in \cite{A}, there is an open neighbourhood $U$ of the neutral element $e$ such that the restriction $\pi | \overline{U}: \overline{U} \to \pi(\overline{U})$ is a perfect mapping. Moreover, $\pi(\overline{U})=\overline{\pi(U)}$ and it is a regular closed set in $G/H$.
Assuming that the subgroup $H$ of a topological group $G$ is locally compact, some local properties of compactness type on $G/H$ can be extended to $G$ (see Corollary 3.2.6 in \cite{AT}) using the fact that these properties are invariant under perfect mappings. In the case of locally (weakly) compact-bounded spaces it is also possible as shown by the following fact:
\begin{prop}\label{loc} Let G be a topological group and H is a locally compact subgroup of G such that the quotient space $G/H$ has some of the following properties:
\begin{enumerate}
\item $G/H$ is weakly compact-bounded;
\item $G/H$ is compact-bounded.
\end{enumerate}
Then, (1) G is locally weakly compact-bounded, (2) G is locally compact-bounded.
\end{prop}
\begin{proof} We will to show (2) (the proof of (1) is similar). As mentioned in the previous paragraph, there is an open neighbourhood $U$ of the neutral element $e$ such that the restriction $\pi | \overline{U}: \overline{U} \to \pi(\overline{U})$ is a perfect mapping.
Let $\{U_n :n\in \omega\}$ be a family of non-empty open sets in $ \overline{U}$. Then, $\{\pi (U_n) :n\in \omega\}$ is a family of non-empty open sets in $\pi(\overline{U})$, thus there is a compact $E$ in $\pi(\overline{U})$ such that $\pi(U_n) \cap E \neq \emptyset $ for each $n\in \omega$. Using a process analogous to that used in the proof of Proposition \ref{quotient}, we can find a compact subset $F$ of $\overline{U}$ such that $F\cap U_n \neq \emptyset $ for each $n\in \omega$. Thus, $\overline{U}$ is a compact-bounded neighbourhood of $e$.
\end{proof}
We say that a paratopological group $G$ is \textit{precompact} if for every neighborhood $U$ of the identity $e$ in $G$, there exists a finite subset $F$ of $G$ such that $G=FU=UF$.
\begin{cor}\label{cor:pre}
Let $G$ be a precompact topological group, and let $H$ be a locally compact subgroup of $G$. If $G/H$ has any of the following properties:
\begin{enumerate}
\item $G/H$ is weakly compact-bounded.;
\item $G/H$ is compact-bounded.
\end{enumerate}
Then $G$ has this property.
\end{cor}
\begin{proof}
By Proposition \ref{loc}, there is an open neighborhood $U$ of the neutral element $e$ of $G$ such that $\overline{U}$ is (weakly) compact-bounded. Since $G$ is precompact, there is a finite subset $A$ of $G$, such that $AU = G$. Then, $A\overline U = G$. Thus $G$ is the union of a finite collection of (weakly) compact-bounded subspaces. We conclude that $G$ is (weakly) compact-bounded topological group by applying Proposition \ref{finden}-(a). \end{proof}
A topological property $\mathcal{P}$ is said to be a \textit{three space property} if for every topological group $G$ and a closed invariant subgroup $H$ of $G$, the fact that both groups $H$ and $G/H$ have $\mathcal{P}$ implies that $G$ also has $\mathcal{P}$.
In the study of weakly compact-bounded spaces and their properties related to quotients, the following question immediately arises:
\begin{ques} Is the (weak) compact-boundedness a three spaces property?
\end{ques}
In attempting to answer this question, we take the following approach:
\begin{prop} Let $G$ be a topological group and $H$ a closed invariant subgroup of $G$ such that $G/H$ is a $k$-space. If $H$ and $G/H$ have any of the following properties:
\begin{enumerate}
\item $H$ and $G/H$ are weakly compact-bounded;
\item $H$ and $G/H$ are compact-bounded.
\end{enumerate}
Then, $G$ has this property.
\end{prop}
\begin{proof} We will to show (2) (the proof of (1) is similar). Let $\{U_n :n\in \omega\}$ be a family of non-empty open sets in $G$ and let $\pi:G \to G/H$ be the natural quotient mapping. Then, the set $\{\pi (U_n) :n\in \omega\}$ is a family of non-empty open sets in $G/H$, thus there is a compact $E$ in $G/H$ such that $\pi(U_n) \cap E \neq \emptyset $ for each $n\in \omega$. Since $G/H$ is a $k$-space, the set $A=\pi^{-1}(E)$ is a compact subset of $G$ (Theorem 3.7.18 in \cite{E}). Since $ \pi(U_n)\cap \pi(A)\neq \emptyset$, there is $x_n\in U_n$ and $a_n\in A$ such that $\pi(x_n)=\pi(a_n)$ for each $n\in \omega$. Thus $x_n \in a_n H$, and consequently $a_n H \cap U_n \neq \emptyset$ for each $n\in \omega$.
Now, we note that the set $a_n H \cap U_n$ is homeomorphic to $ H \cap a^{-1}_n U_n$ for each $n\in \omega$. Since $\{H\cap a^{-1}_n U_n :n\in \omega\}$ is a family of non-empty open sets in $H$, there is a compact subset $K$ of $H$ such that $K\cap (H\cap a^{-1}_n U_n)= K \cap a^{-1}_n U_n \neq \emptyset$ for each $n\in \omega$. Thus, we have that $a_n K\cap U_n\neq \emptyset$ for each $n\in \omega$, that is $A K\cap U_n\neq \emptyset$ for each $n\in \omega$. Since the set $AK$ is a compact subset of $G$ (Proposition 1.4.31 in \cite{AT}), we conclude that $G$ is compact-bounded.
\end{proof}
A topological property $\mathcal{P}$ will be called an \textit{inverse fiber property} if whenever $f:X\to Y$ is a continuous onto mapping such that the space $Y$ and the fibers of $f$ have $\mathcal{P}$, then $X$ also has $\mathcal{P}$. By Proposition 2.2 in \cite{BT2}, we know that every inverse fiber property is a three space property.
\begin{prop} To have all compact-bounded (weakly) subsets finite is a three space property.
\end{prop}
\begin{proof} Let $\mathcal{P}_{cb}$ be the property that all compact-bounded subsets of $X$ are finite. It will be sufficient to prove that $\mathcal{P}_{cb}$ is an inverse fiber property.
Let $f:X \to Y$ be a continuous onto map. Suppose that $Y$ and the fibers of $f$ have $\mathcal{P}_{cb}$. Just as in the proof of Proposition 2.4(a) in \cite{BT2}, let $A\subset X$ be a compact-bounded subset of $X$. Then $f(A)\subset Y$ is compact-bounded subset of $Y$ and thus $f(A)$ is finite. Thus, the family $\{A \cap f^{-1}(y):y\in f(A) \}$ is a finite partition of $A$ into disjoint clopen subsets of $A$. Then, each set $A \cap f^{-1}(y)$ is compact-bounded by Theorem \ref{imgreg}(b); thus $A\cap f^{-1}(y)$ is finite. Therefore, $A$ is finite.
\end{proof}
As usual, for a topological space $X$, $A(X)$ is the Abelian free topological group generated by $X$, and for a locally compact topological space $X$, $\alpha X$ is the one point compactification of $X$. In contrast to the previous result, we have the following example:
\begin{exam}[Example 4.6 in \cite{BT2}] ``To have all (weakly) compact-bounded subsets compact" is not a three spaces properties.
\end{exam}
\begin{proof} Let $X=\Psi(\mathcal{D})$ be the space described in the Example \ref{Mr} (where $\mathcal{D}$ is a maximal almost disjoint family on $\omega$). Let $f:\alpha X \to \omega +1$ be the mapping defined by $f(n)=n$ for each $n\in \omega$ and $f(x)=\omega$ for each $x\in \alpha X\setminus \omega$. Since $f$ is closed, it extends to a continuous open homomorphism $\bar{f} :A(\alpha X) \to A(\omega +1)$. Let $G$ be the subgroup of $A(\alpha X)$ generated by $X$, $N=\mbox{ker} \bar{f}$ and $K=G \cap N$. By Example 4.6 in \cite{BT2}, we know that $K$ is a closed subgroup of $G$ such that all pseudocompact subspaces of the groups $K$ and $G/K$ are compact (thus $K$ and $G/K$ satisfy the hypothesis property) and that $X$ is a closed subset non-compact of $G$.
\end{proof}
\begin{ques} Consider the following properties:
\begin{enumerate}
\item All weakly compact-bounded (compact-bounded) subsets are closed;
\item All weakly compact-bounded (compact-bounded) subsets are first-countable;
\item All weakly compact-bounded (compact-bounded) subsets are metrizable.
\end{enumerate}
Is any of these properties a three spaces property?
\end{ques}
\section{Free topological groups}
Let $F(X)$ be the free topological group and $A(X)$ be the free abelian topological group of a Tychonoff space $X$. We denote the groups $F(X)$ or $A(X)$ by $G(X)$. The set $\beta X$ denotes the Stone-\v{C}ech compactification of $X$.
Recall that a topological space $X$ is \textit{$\omega$-bounded} if the closure of every countable subset of $X$ is compact. Clearly, every $\omega$-bounded space is a compact-bounded space.
By Proposition 4.6 in \cite{LT}, we know that $\omega$-boundedness is a property of a Tychonoff topological space $X$ under which the following conditions are equivalent:
(1) The Stone-\v{C}ech compactification $\beta X$ is scattered;
(2) $X$ is scattered.
Clearly, (1) always implies (2). The following fact shows that compact-boundedness also makes the above conditions equivalent.
\begin{prop} Let X be a scattered compact-bounded Tychonoff space, then $\beta X$ is scattered.
\end{prop}
\begin{proof} Suppose that $\beta X$ is not scattered. Then, there is a continuous surjection $f:\beta X \to I$ (Theorem 8.5.4 in \cite{S}). Let $\{V_n:n\in \omega\}$ be a countable base of $I$, since $\{f^{-1}(V_n)\cap X:n\in \omega\}$ is a family of non-empty open sets in $X$, there is a compact subset $K$ of $X$ such that $K\cap (f^{-1}(V_n)\cap X)=K\cap f^{-1}(V_n) \neq \emptyset$ for each $n\in \omega$. Thus, $
\emptyset\neq f(K\cap f^{-1}(V_n))\subset f(K)\cap f( f^{-1}(V_n))\subset f(K)\cap V_n, $
for each $n\in \omega$. This implies that the set $f(K)$ is dense in $I$. Since $K$ is compact and scattered, the image $f(K)=cl_I(f(K))=I$ is a countable set (Corollary 8.5.6 in \cite{S}), a contradiction.
\end{proof}
With the previous fact we can extend the Theorem 4.7 in \cite{LT} to compact-bounded spaces.
\begin{prop} Let X be an compact-bounded space. Then the following conditions are equivalent:
\begin{enumerate}
\item The group $G(X)$ admits an open continuous homomorphism onto the circle group $\mathbb{T}$;
\item The group $G(X)$ admits a non-trivial metrizable quotient;
\item The group $G(X)$ admits a non-trivial separable metrizable quotient;
\item X is not scattered.
\end{enumerate}
\end{prop}
\begin{proof} Since $X$ is pseudocompact, by Theorem 4.5 in \cite{LT}, we have the implications (1)$\Leftrightarrow$(2)$\Leftrightarrow$(3). We note that (d) of Theorem 4.5 in \cite{LT} is equivalent to item (4), thus we can conclude this fact.
\end{proof}
\section{R-weakly compact-bounded subsets}
Let $X$ be a topological space. A subset $A$ of $X$ is \textit{r-pseudocompact} in $X$ if every infinite countable family of open sets in $X$ meeting $A$ has an accumulation point in $A$. A subset $A$ of $X$ is said to be \textit{strongly r-pseudocompact} in $X$ if the product $A\times B$ is $r$-pseudocompact in $X\times Y$ whenever $B$ is a $r$-pseudocompact subset of a space $Y$. A detailed study of properties on this class of subsets can be found in \cite{HST}, \cite{ST} and \cite{XY}.
A subset $A$ of a topological space $X$ is \textit{r-weakly compact-bounded} in $X$ if every infinite countable family $\mathcal{U}$ of open sets in $X$ meeting $A$, there is some compact subset $K_A\subset A$ which intersects infinitely many elements of $\mathcal{U}$. Clearly, every $r$-weakly compact-bounded subset is a $r$-pseudocompact subset.
\begin{prop}\label{rpseudo} Let $A$ be a subset of a space $X$. Then, the following conditions are equivalent:
\begin{enumerate}
\item A is strongly r-pseudocompact in X;
\item Every infinite countable family $\mathcal{U}$ of mutually disjoint open sets in $X$ meeting $A$, there is some compact subset $K_A\subset A$ which intersects infinitely many elements of $\mathcal{U}$.
\item $A\times Y$ is r-pseudocompact in $X\times Y$ whenever $Y$ is a feebly compact space.
\end{enumerate}
\end{prop}
\begin{proof}(1)$\Leftrightarrow$(2) A similar argument of the proof of Proposition \ref{disj} proofs this fact.
(1)$\Leftrightarrow$(3)The sufficiency is clear. Proof of the necessity: Let $B$ be a $r$-pseudocompact subset of a space $Y$. Clearly, $B$ is a feebly compact subset of $Y$, thus $A\times B$ is $r$-pseudocompact in $X\times B$ (and hence in $X\times Y$). Therefore, $A$ is strongly $r$-pseudocompact in $X$.
\end{proof}
The following result is well known.
\begin{lem}\label{fpc} If $X$ is feebly compact and $Y$ is compact, then $X\times Y$ is feebly compact.
\end{lem}
\begin{prop}\label{wcbsrp} Every r-weakly compact-bounded subset of a space X is strongly r-pseudocompact of X.
\end{prop}
\begin{proof} Let $A$ be a weakly compact-bounded subset of $X$ and $Y$ be a feebly compact space. By Proposition \ref{rpseudo}, it will suffice to show that $A\times Y$ is $r$-pseudocompact in $X\times Y$.
Let $\{U_n \times V_n:n\in \omega\}$ be a family of non-empty open subsets of $X\times Y$ where $\{U_n :n\in \omega\}$ is a family of non-empty open subsets of $X$ and $\{V_n :n\in \omega\}$ is a family of non-empty open subsets of $Y$ such that $(U_n \times V_n) \cap (A\times Y) \neq \emptyset$ for each $n\in \omega$. Since $\{U_n \cap A :n\in \omega\}$ is a family of non-empty open subsets of $A$, there is a compact subset $K_A$ of $A$ and an infinite subset $I\subset \omega$ such that $U_m \cap K_A \neq \emptyset$ for each $m\in I$, clearly the family $\{(U_m \cap K_A)\times V_m :m\in I\}$ is a family of non-empty open subsets of $K_A\times Y$. Since $K_A\times Y$ is feebly compact (Lemma \ref{fpc}) we have that $\{(U_m \cap K_A)\times V_m :m\in I\}$ has a accumulation point in $K_A\times Y$, which is also an accumulation point of $\{U_n \times V_n:n\in \omega\}$ in $X\times Y$.
\end{proof}
The following example shows the difference between a subset weakly compact-bounded and a $r$-weakly compact-bounded subset in a Tychonoff space $X$.
\begin{exam} There is a r-weakly compact-bounded in a Tychonoff space X such that it is not pseudocompact.
\end{exam}
\begin{proof} Let $X=\Psi(\mathcal{D})$ be the space described in the Example \ref{Mr} (where $\mathcal{D}$ is a maximal almost disjoint family on $\omega$). Since every point in $\omega$ is isolated in $X$, $\mathcal{D}$ is $r$-pseudocompact in $X$, also $\mathcal{D}$ is closed in $X$, thus it is locally compact, that is, $\mathcal{D}$ is $r$-weakly compact-bounded in $X$ (Proposition \ref{rpse}). However, $\mathcal{D}$ is a closed discrete subspace of $X$, and hence $\mathcal{D}$ is not pseudocompact.
\end{proof}
\begin{exam} There is a $r$-pseudocompact subset in a Tychonoff space X such that it is not a r-weakly compact-bounded subset of $X$.
\end{exam}
\begin{proof} The dense pseudocompact subspace $P$ of $\beta \omega$ described in the Example 2.4-(4) in \cite{HST} is $r$-pseudocompact but it is not strongly $r$-pseudocompact in $\beta \omega$ (Example 2.4-(5) in \cite{HST}) thus, it is not $r$-weakly compact-bounded (Proposition\ref{wcbsrp}).
\end{proof}
\begin{ques} Is there a strongly $r$-pseudocompact in a topological space $X$ not $r$-weakly compact-bounded?
\end{ques}
The following fact shows sufficient conditions under which a subset $r$-pseudocompact of a topological space $X$ is a $r$-weakly compact-bounded subset.
\begin{prop}\label{rpse} Let X be a weakly regular space. Every r-pseudocompact k-space subset of X is a r-weakly compact-bounded subset of X.
\end{prop}
\begin{proof} Let $A$ be a $r$-pseudocompact $k$-space subset of $X$ and let $\{U_n:n\in \omega\}$ be a family of mutually disjoint non-empty open subsets of $X$ that meet $A$. By weak regularity of $X$, there is open subsets $V_n$ of $X$ such that $V_n\cap A \subset \overline{V_n}\cap A \subset U_n\cap A$ for each $n\in \omega$. Since $A$ is $r$-pseudocompact, the family $\{V_n:n\in \omega\}$ has a accumulation point in $A$. Using the same argument of the proof of Proposition 4.9 in \cite{PTV} and the fact that $A$ is a $k$-space, we can conclude that there is a compact $K_A$ of $A$ such that $K_A$ meets infinitely many elements of $\{\overline{V_n}: n \in \omega\}$ (and thus of $\{U_n:n\in \omega\}$). By Proposition \ref{disj}, we conclude that $A$ is a $r$-weakly compact-bounded subset of $X$.
\end{proof}
By Proposition \ref{wcbsrp} and the previous proposition we have the next fact.
\begin{cor}\label{eqwr} Let X be a weakly regular. Then the following conditions are equivalent for a k-space subset A of X:
\begin{enumerate}
\item A is $r$-weakly compact-bounded in X;
\item A is strongly r-pseudocompact;
\item A is r-pseudocompact.
\end{enumerate}
\end{cor}
We may drop the condition of weak regularity required in the previous fact. A argument similar to that used in the proof of Proposition 3.4 and Proposition 3.6 in \cite{AMW} gives:
\begin{prop} Let $A$ be a subset of a topological space $X$ which satisfies one of the following conditions:
\begin{enumerate}
\item $A$ is first countable and feebly compact;
\item $A$ is sequential and countably compact.
\end{enumerate}
Then, $A$ is $r$-weakly compact-bounded in X (thus, strongly r-pseudocompact in X).
\end{prop}
\begin{prop} Let $A_i$ be a $r$-weakly compact-bounded subset of a space $X_i$, where $i\in \omega$. Then, the product $\prod_{i\in \omega} A_i$ is $r$-weakly compact-bounded in $\prod_{i\in \omega} X_i$.
\end{prop}
\begin{proof} We denote by $A$ and $X$ the products $\prod_{i\in \omega} A_i$ and $\prod_{i\in \omega} X_i$, respectively. Let $\mathcal{U}= \{U_m:m\in \omega \}$ be a family of non-empty open subsets of $X$ that meets $A$. Then, $\pi_0[\mathcal{U}]=\{\pi_0[U_m]:m\in \omega \}$ is a countable family of non-empty open sets in $X_0$ that meets $A_0$. Thus, there is a compact space $K_0\subset A_0$ which meets infinitely many elements of $\pi_0[\mathcal{U}]$, say $\mathcal{U}_0=\{ U\in \mathcal{U}: K_0\cap \pi_0[U]\neq \emptyset\}$. We can assume that $\pi_1[\mathcal{U}]$ is a countable family of non-empty open sets in $X_1$ that meets $A_1$, thus there is a compact space $K_1\subset A_1$ which meets infinitely many elements of $\pi_1[\mathcal{U}]$, say $\mathcal{U}_1=\{ U\in \mathcal{U}: K_1\cap \pi_1[U]\neq \emptyset\}$, iterating this argument, we obtain a decreasing sequence of infinite subsets $\{J_n:n\in \omega \}$ of $\omega$, a decreasing sequence $\{\mathcal{U}_n:n\in \omega \}$ of infinite subsets of $\mathcal{U}$ where $\mathcal{U}_n=\{ U_m:m\in J_n\}$ and compact subspaces $K_n\subset A_n$ such that $K_n$ meets each element of the family of non-empty open sets $\pi_n[\mathcal{U}]$. Let $J\subset \omega$ be an infinite set which is almost contained in each set $J_n$. Let $\mathcal{V}=\{ U_j \in \mathcal{U}:j\in J\}$. For each $n\in \omega$ and each $m\in J\setminus J_n$ we can take a finite set $F_n\subset A_n$ which meets each $\mathcal{V}\setminus \mathcal{U}_n$ and then $E_n=K_n \cup F_n$ is a compact subspace which meets each element of the family $\pi_n[\mathcal{U}]$ for each $n\in J$ and so $\prod_{i\in \omega} E_i$ is the required compact subspace of $A$.
\end{proof}
\begin{cor} Let $A_i$ be a $r$-pseudocompact k-space subset of a weakly regular space $X_i$, where $i\in \omega$. Then, the product $\prod_{i\in \omega} A_i$ is $r$-weakly compact-bounded in $\prod_{i\in \omega} X_i$.
\end{cor}
A subset $A$ of a topological space $X$ is said to be \textit{C-compact} in $X$ if the image $f(A)$ is compact, for any continuous real-valued function $f$ on $X$.
Now, the following implications are clear:
\smallskip
\begin{center}
$r$-weakly compact-bounded $\Rightarrow$ strongly $r$-pseudocompact $\Rightarrow$ $r$-pseudocompact $\Rightarrow$ $C$-compact
\end{center}
\smallskip
The Example 2.4-(2) in \cite{HST} shows a example of a $C$-compact subset of a compact space not $r$-pseudocompact.
By Corollary 3.11 in \cite{HST}, we know that $C$-compactness and strong $r$-pseudocompactness coincide in the class of topological groups.
Thus, a consequence of Corollary \ref{eqwr} is the following:
\begin{prop}\label{Cwcc} Let G be a topological group and let $A$ be a k-space subset of G. Then, the following conditions are equivalent:
\begin{enumerate}
\item A is $r$-weakly compact-bounded in G;
\item A is strongly r-pseudocompact in G;
\item A is r-pseudocompact in G;
\item A is C-compact in G.
\end{enumerate}
\end{prop}
\begin{cor} Let $A_i$ be a $C$-compact k-space subset of a topological group $G_i$, where $i\in \omega$. Then, the product $\prod_{i\in \omega} A_i$ is $r$-weakly compact-bounded in $\prod_{i\in \omega} G_i$.
\end{cor}
If we want to know when the $r$-weakly compact-bounded subsets of a paratopological group coincide with the $C$-compact subsets, we have the class of weakly $\omega$-admissible paratopological groups (\cite{XY}) that answer this question. By Theorem 3.20 in \cite{XY} and Corollary \ref{eqwr} the next fact is immediate:
\begin{prop} Let G be a weakly regular paratopological group. If G is weakly $\omega$-admissible, then the following conditions are equivalent for a k-space subset A of G:
\begin{enumerate}
\item A is $r$-weakly compact-bounded in G;
\item A is strongly r-pseudocompact in G;
\item A is r-pseudocompact in G;
\item A is C-compact in G.
\end{enumerate}
\end{prop}
We call a paratopological group $G$ \textit{totally $\omega$-narrow} if the topological group $G^*$ associated to $G$ is $\omega$-narrow.
It is possible to find some properties of paratopological groups under which a $C$-compact subset becomes a subset $r$-weakly compact-bounded. By Corollary 4.3 in \cite{ST} and Proposition \ref{rpse}, we have the next fact:
\begin{prop} Let A be a C-compact k-space subset of a Tychonoff paratopological group G which satisfies one of the following conditions:
\begin{enumerate}
\item G is totally $\omega$-narrow;
\item G is commutative and has countable Hausdorff number;
\item G is precompact.
\end{enumerate}
Then A is $r$-weakly compact-bounded in G.
\end{prop}
|
\section{Introduction}
Whether acting as a teacher, an editor, or a creative partner, both artificially intelligent systems (AI) and human users can learn from each other to improve not only their own design and stylistic techniques, but also their ability to evaluate their own work for quality.
Interactive evolution can act as a bridge for the creative process between human design and automated, generative design by allowing a two-way communication between the user and the AI. Together they can collaboratively evolve and create new content - such as literature, art, music, and game design artifacts - all while gradually improving their design skills.
However, finding a common understanding between a user's preferences and what the AI is trained to generate towards is difficult to achieve. Much like talking to another person, communicating preference for how the content should look, feel, and perform while defining what content is ``high" and ``low" quality takes time, repeated effort, and a willingness from both parties to understand and learn the other's own creative process. Teaching and directing an AI system towards a particular design preference is also an exhausting task in itself, especially if the environment for teaching the system is unfamiliar or overwhelmingly complex. Many content generation systems only focus on generating content that fits an easily measured criteria - such as how well a prompt is matched for literature, the level of consonance or dissonance for music, and how playable a level is for procedurally generated game design levels. More work can be done to ensure how aesthetically pleasing the content generated is - whether on a personal scale or globally.
This paper introduces the Aesthetic Bot system, a novel procedural content generation system that focuses exclusively on improving the appeal of the visual design - henceforth referred to as visual aesthetics - of a game map by learning general user preferences of design.
To learn this preference, this system retrieves user feedback in a quick, minimal effort method of votes in an A vs. B comparison. We use the social media site Twitter as the host for this system, as it has a large (allegedly majority human \cite{muskTwitter}) user base that frequently interacts with the platform and is able to view and engage with content made by other accounts. With this platform, we are able to have Twitter users vote directly on the bot's generated maps or real, user-made maps in an interaction that takes only a few seconds of evaluation time within their regularly updated content timeline. After the polling, the system learns from the rating to generate new maps that are optimized to win future preference pollings and improve its overall ability to determine aesthetic quality in a map.
\section{Background}
\subsection{Aesthetic Design in Content Generation}
The Stanford Encyclopedia of Philosophy describes aesthetics as a subjective taste in a sense of beauty and ugliness. There is no hard truth to what aesthetic is and they do not necessarily need to be related to visual tastes \cite{zangwill2003aesthetic}. David and Globe note that the visual aesthetics of a system, such as one used for user interfaces and human computer interaction, can influence a user's perception of ease of use for the system overall, and also infer how much a user may engage with it. Well designed visual aesthetics of a system can also foster collaboration between users, as used in Google Docs and online learning tools. Thus, the study of aeshetic design is critical in order to create collaborative systems that not only allow for a more friendly first-impression of the system but also encourage frequent user engagement. \cite{david2010impact}.
Games often add extra visuals effects to make it more visually aesthetically pleasing and appealing to the player. This can be done through decorating the world itself or by adding effects to make the game experience more interactive and responsive called game feel \cite{swink2008game}.
Previous work has been done to automate this process, such as in the Juicer project \cite{johansen2021challenges}.
For this paper, we tactfully bow out of the complexities of pinning down visual aesthetics by operationally defining good visual aesthetics as ``visuals that a user prefers when given multiple options of content to choose from".
Aesthetic design is subjective to a human experience - determining what looks good is harder to describe or quantify than analyzing raw metrics of performance. However, we argue that judging aesthetic design could be emulated; much like how human decision making processes and evaluations have been emulated by machines already. While games and AI research have focused on narrative \cite{alvarez2022tropetwist, aarseth2012narrative} or mechanic \cite{hunicke2004mda} aesthetics of a system in the past, in the Aesthetic Bot system we focus exclusively on the visual output of the maps and ignore their playabilities, mechanic activations, or any other functionality constraints typically involved with game design.
\subsection{Procedural Content Generation}
Procedural content generation (PCG) for level design enables, among other things,
unique experiences for players. Most notably, PCG has been used across multiple genres of games - such as in Minecraft, Spelunky, Borderlands 2, and the classic Rogue - and in game AI research including top-down grid worlds \cite{earle2021learning,sandhu2021tileterror}, platformers \cite{bhaumik2021lode,sarkar2021procedural,volz2018evolving}, puzzle games \cite{charity2020baba,khalifa2020multi}, open-world 3d environments \cite{salge2018generative}, and more.
The generation process for PCG levels typically focuses on generating towards an objective or fitness that can be calculated based on some metric (i.e. playability, solution length, entropy.) While they may be successful in their endeavor of producing levels that guarantee completion of this objective, most PCG systems do so without taking into account the look or visual appeal of the level. As such, many of these levels look artificially made, and lack the meticulous, intentional design that human-made levels have. A system that is capable of maintaining the visual aesthetic of the level - either parallel to the main generation objective or as a post-processing step - could improve the visual quality of the level overall and make a much more engaging experience for the player.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{imgs/Mario_PCG.png}
\caption{Two levels from the Mario AI framework. The left is user-made and the right is generated. While both are playable, the user-made level arguably has more aesthetic.}
\label{fig:pcg_mario}
\end{figure}
\subsection{Interactive Evolution and Co-Creativity}
Contrary to pure PCG systems, interactive evolution (IE) is a content generation method where the evaluation of the system's content output is dependent on the user's selection \cite{dawkins1986blind}. Picbreeder is such an example of a system where users are able to select from a set of different images generated by an evolutionary algorithm. Images are evolved and removed based on a subset of images selected by a single user. This guides the system towards evolving the population of images to either a singular goal image or to fit the user's personal preference \cite{secretan2008picbreeder}. Bontrager et al.'s work used a similar system but instead generating the images with a deep learning model and evolved latent vectors to accomplish the same goal \cite{bontrager2018deep}. Liapis et al. also looked to use machine learning and evolutionary techniques with constrained, limited user selection to generate content based on visual and aesthetic appeal of the designer in various projects ranging from spaceship designs to map terrains \cite{Liapis_adaptive,liapis2012limitations,Liapis_optimizing,liapis2013sentient,Liapis_boosting}.
However, most of these systems mentioned are unable to retain a sense of general aesthetic preferences learned from users.
They require many iterations that start from scratch in order to reach a state where the content produced is satisfactory - and such content is only made for a single user. They cannot learn the majority preference of multiple users or transfer their knowledge across sessions, as they are isolated to the singular interaction with the current user. IE systems can also exhaust a user and lead to participant fatigue - especially when these types of systems require so much feedback in order to effectively learn \cite{kamalian2006reducing}.
The Aesthetic Bot attempts to addresses these issues by receiving polled feedback from multiple users all at once in short interactions while limiting the selection pool.
\subsection{Twitter Bots}
Twitter bots are artificial profiles made on the social media platform Twitter that post content either unprompted with automated posts or in response to other users on the platform \cite{veale2018twitterbots}. Many bots are made to produce content for mostly entertainment purposes - such as a bot that combines emojis together to form new ones (@EmojiMashupBot). Other bots exist to interact with users once prompted to do so - such as generating on-demand haikus (@haikookies) \cite{pichlmair2020procedural}. However, both types of bots only act as one-way posters - they do not learn from the a two-way communication between the user and the bot itself. Much in the same way Twitter's human profiles can gain information by posing questions and receiving replies or posting polls to receive votes on a subject, so too can a bot learn from these posts. With the Twitter API, Twitter can be used as a means of retrieving feedback in the same way as the previous interactive evolution co-creative systems but in a system that is already familiar and regularly used by participants.
\section{Aesthetic Bot System}
The Aesthetic Bot system is a pipeline comprising of a convolutional neural network called the Aesthetic Prediction Model (APM), an evolutionary algorithm, and the Twitter platform interface. Figure \ref{fig:pipeline} shows the entire learning pipeline generating and evaluating a new map that is sent to Twitter.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{imgs/Tilesets.png}
\caption{The experiment's usable tilesets from left to right: Zelda, Pokemon, Among Us, Pacman, Dungeon}
\label{fig:tilesets}
\end{figure}
For this experiment, we created 5 original pixel art tilesets (shown in Figure \ref{fig:tilesets}) that we designed to resemble the visual themes of the following games: Zelda, Pokemon, Among Us, Pac-Man, and a generic dungeon crawler (called Dungeon). Each tileset contains 16 8x8 pixel tiles - some tiles were purposefully designed for inter-connectivity and to be placed in such a way to form a larger pattern of tiles (i.e. the tree tiles in Zelda.) We chose to use very small tileset visual dimensions so that users would ideally have a lower amount of visual aesthetic bias towards a particular art style or game due to higher resolutions or more complex graphics. The smaller resolutions are also much easier to view at a glance when the pixels are scaled up.
\begin{figure}[b]
\centering
\includegraphics[width=1.0\linewidth]{imgs/Aesthetic_Bot_Pipeline.png}
\caption{Learning pipeline of the Aesthetic Bot. Dark grey arrows indicate the recipient will update based on the output, while light grey arrows indicate to only process the output}
\label{fig:pipeline}
\end{figure}
\subsection{Aesthetic Prediction Model}
The convolutional neural network known as the Aesthetic Prediction Model (APM) acts as the ``judge" for the system and tries to estimate the output of a Twitter poll based on an input map. It trains and updates based on the actual results of the Twitter polls. For this experiment, the model consists of an extremely small neural network with 3 convolutional layers, each followed by a batch normalization layer, a ReLU activation function and finally a MaxPooling layer (except at the last ReLU which has a flattened, dense layer.) Instead of passing the entire map to the network, the map is divided into one-hot encoded 4x4 tile windows with 16 channels (for each tile in the tileset). Due to the small input size the model only has 15,185 parameters and output shapes of (2,2,64), (1,1,32), (1,1,16), (16), and (1).
The smaller sliding window of 4x4 helps to accommodate for the various sizes of the maps and helps the model learn and judge on patterns of tiles instead of entire maps which have more distinction and variance between each other. If a neural network with a fixed size were used instead (such as the smallest map size of 6 or the largest of 12), this could limit the learning capabilities of the APM, as some maps may not be able to be processed for being too small while others may lose noticeable aesthetic features from being cutoff by the smaller input window. We later compare the prediction capabilties of a fixed input network against the APM's sliding window input network in the 'Post Experiments' section of this paper.
The model is trained by passing the encoded 4x4 sliding subsection windows of a map as its input and using the percentage of votes the input map received in the Twitter pairing as its output y value. Figure \ref{fig:cnn_pipeline} shows this process of taking two maps from a polling and processing them for training the APM. We train the model on the percentage of votes instead of the votes directly because we want to encourage the model to learn how to estimate which maps will win in any given pairing rather than on how many votes it would receive.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{imgs/CNN_Diagram.png}
\caption{Diagram of the APM learning process. Each window from the map receives the same y value and the network learns to associate a window pattern with a vote percentage.}
\label{fig:cnn_pipeline}
\end{figure}
The votes themselves are variable with the Twitter polls and harder to distinguish relatively in a map's terms of quality, except when paired with another map. The output value of a map is assigned as the y value for every subsection pattern of the map. As such, it is possible for the APM to learn 2 different contradictory y values for the same pattern, however the context and values of the other patterns will help distinguish the quality of the map overall.
After training, the weights are frozen and the APM is used as the evolutionary model's fitness function to evaluate the population of maps. When the APM evaluates a map, the map is separated into the same 4x4 sliding windows and passed through the APM. The average of the set of window patterns' outputs is used as the final fitness value for the entire map. Therefore, the more higher rated patterns a map contains, the more likely it will perform better in a pairing and conversely for lower rated patterns.
\subsection{Evolutionary Algorithm}
To generate new maps, the pipeline implements a $\mu+\lambda$ evolutionary algorithm - where $\mu$ maps with the highest fitness are retained in the next population and $\lambda$ maps are randomly selected from the population and mutated for
evaluation in the next iteration. This experiment uses 10\% for mu and 90\% for lambda. The algorithm starts from a uniformly randomized population of maps of varying dimensions.
The sizes of the maps are limited to a square range of 6-12 tiles in dimension. As mentioned in the previous subsection, the output of the APM acts as the fitness function for the evolutionary system. As a mutation function, each 2x2 subsection of tiles in a map have a random chance to be mutated and replaced with a 2x2 subsection pattern from the real user maps saved to the system. We found that this method encourages the occurrence and retention of more ``user"-made patterns with the direct replacement rather than randomly mutating singular tiles. The dimensions of the map also have a small (0.1\%) chance to change.
After a set number of iterations, the map with the highest fitness value (as determined by the APM) is selected from the evolutionary algorithm's population. This experiment runs the evolutionary process for multiple trials and saves the best maps from each trial's population in a separate list to be evaluated in a final selection phase. This final selection evaluates the maps based on some preset function (entropy, fake Twitter function value, both, etc) in case the evolutionary algorithm plateaued in its search for a high quality map. Figure \ref{fig:final_selection} shows a set of the final maps generated. This final map is saved to the database as the tileset's generated map and associated with the trained APM that helped with the evaluation to produce it. This is to keep track of which maps came as the results of a particular APM and monitor its improvements (or regression.)
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{imgs/final_selection3.png}
\caption{A set of generated maps from the pipeline. The map submitted to the database will be decided based on a secondary function (labeled at the top.)}
\label{fig:final_selection}
\end{figure}
\subsection{Twitter Interaction}
In a separate process performed parallel to the map generation (in order to monitor generated levels before pairing),
a Python script selects a submitted user map and a generated map that uses the same tileset from their respective databases (typically the oldest unevaluated map from both) and pairs them together. This pairing renders the two maps together, side-by-side, in an image with the label ``Which do you like better?" underneath and saves the pairing information to another database table. The image is then posted to the Aesthetic Bot Twitter account with a poll as a threaded reply post using the Twitter API. The poll is held for 3 hours to allow as many users as possible to vote on the pairing.
At the completion of the poll, the votes are recorded to the pairing's database row. The tileset's associated APM is trained and updated with the results of the post's pairing. This is how the APM updates past its initial training stage and improves dynamically with each new pairing. This updated APM is used again as the evaluator for the tileset's next generated map and cycles through the pipeline. Finally, as a way of giving closure to curious users, the system reveals in a final post in the poll thread which map shown in the pairing was the user-made map as well as the map author if one was provided. A link to the Aesthetic Bot map editor - described in more detail in the 'Live Training' section - is included in the post to encourage participants to upload their own map to the database to be used in a future pairing.
A later modification of the system also included a small chance to make homogeneous author-type pairings as well to prevent biases from occurring in the APM and the participating users. Through this method, the network still learns which maps outperformed others, not just in a user-gen pairing situation. This would also cause voting participants to think more critically about the map pairings being shown and why they actually preferred a map over another - and not just to try to discern which of the two maps was user made - as it could not be one of them, both of them, or neither.
\subsection{Pre-training Stage}
Before pairing posts were made for the Aesthetic Bot Twitter account, the APMs for each tileset went through a pre-training stage to prevent completely random maps from being shown in the pairings and provide a ``decent" looking map as comparison for users. A pseudo-Twitter evaluation function was used to emulate Twitter votes on a map and train the neural network. This function used a modification of Lucas and Volz's Tile Pattern Kullback-Leibler Divergence (TPKL-Div) algorithm to evaluate maps \cite{lucas2019tile} by taking the map's Kullback-Leibler divergence score - or relative entropy - to compare it to a training set's tile patterns and determine the similarity in pattern compositions. This was to encourage more ``user-made" looking maps while also encouraging novel design and structure. In addition to the TPKL-Div score, raw tile entropy probabilities were calculated (this was to discourage maps with little variation in tiles or mostly empty maps) and multiplied together with the TPKL-Div score. The final value from this function was meant to simulate the number of votes a map received. Like the real Twitter polls, each map in a pairing received their votes (artificially calculated with the entropy and TPKL-Div) and the difference in votes was used as the output y value to train the neural network. By using an artificial function to simulate a base for user aesthetic, the neural network was able to train in a much shorter amount of time than the weeks or months it would take to train a randomly initialized network using live Twitter data.
Initially, the networks were pre-trained with a set of pairings of user-made and randomly generated maps so that the network would have a baseline output and learn to distinguish pure randomness before evaluating populations. Afterwards, the new maps were evolved in the evolutionary algorithm as normal with the pseudo-Twitter function acting as the Twitter evaluator to update the APM. On completion of these iterations, the pre-trained APM was implemented into the actual pipeline and continued to update and learn using the real data polling from Twitter.
\subsection{Live Training}
The Aesthetic Bot was officially announced and revealed publicly on Twitter on April 28th, 2022. The bot would post a new poll every 3 hours. The database of submitted and generated maps along with the results of the pair polls were backed up to a separate server so any new models made could be trained and updated to the current state of the pollings.
On the Aesthetic Bot bio information status for the account, as well as at the end of every poll, a link to the Aesthetic Bot map editor was made available to Twitter users. The site\footnote{http://aesthetic-bot.xyz/} allows participants to submit their own map designs to the server database to be used in a future pairing. Users could select from the five tilesets available and draw tiles onto the JavaScript canvas. The map editor interface - as shown in figure \ref{fig:editor} - was designed to be mobile friendly as well so that users on the Twitter app could seamlessly submit a map. Upon submission of a new map, users were also given the option to provide their Twitter handle so that they could have authorship over the map. The user would be tagged/credited in the closing thread post of the poll that included their map.
On the developer end, we built a map deleter to monitor the maps made by users and make sure any offensive or graphic looking maps or offensive fake usernames were not included in the pairings.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{imgs/Level_Editor.png}
\caption{The map editor site where users can design and submit their own maps that Aesthetic Bot will learn from.}
\label{fig:editor}
\end{figure}
\section{Results}
As of May 30th, 2022, there were a total of 399 user maps submitted with an average map size of 9.92 tiles squared. The distribution of maps included 101 user maps made for the Zelda tileset, 108 Pokemon maps, 31 Among Us maps, 69 Pacman maps, and 90 Dungeon maps.
The Among Us tileset's unpopularity may have been because the actual game is the least familiar and more recently made of the 5 games the other tilesets were based on. The authors with the most submitted maps (excluding anonymously authored maps that made up 1/3 of the user maps) were @Nifflas, @MasterMilkX, dginev, @3phen, @AndresZarta, and @charphinB. Out of the hybrid pairings, the average number of total votes per poll were 27.5 votes. Of the 220 total polls, the user-made maps won 74\% of matchups, while the generated maps won 24\%. 2\% resulted in ties between the 2 maps.
However, the percentage of votes the user maps have received slowly decreased over time - as shown in Figure \ref{fig:win_perc_chart}. The trend line (in red) shows the user maps slowly decreasing in vote majority. This implies that the generated maps are starting to receive more user votes over time and rival user-made maps in aesthetic quality.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{imgs/win_perc_chart_long2.png}
\caption{Percentage of votes received for the user maps. }
\label{fig:win_perc_chart}
\end{figure}
A validation test was done on the APMs to test how well it could distinguish between ``good" and ``bad" maps over time. 3 different models at various stages of training: the pre-trained model before release to Twitter, the model 2 weeks after release, and the model 6 weeks after release. These were evaluated on 2 sets of maps: randomly generated (unevolved) maps that would be unlikely to win a polling and unseen user maps hand chosen and predicted to have a high aesthetic. Every tileset model - including those trained on Twitter were able to successfully predict the random maps at a lower winning rate (less than 30\% of votes) - except the 2 week Pacman model. The models were also able to predict high vote rate (more than 60\% of votes) for the unseen user maps - except the Among Us tileset, most likely because of the considerable lack of trained pairings. Thus, we can infer that most of the APMs are still able retain a high accuracy for predicting how likely a particular map would win a polling even after receiving new data from real polls. Table \ref{label:valid_exp} shows the results of the averaged predictions from each model type on each of the tilesets.
\begin{table}[ht!]
\resizebox{\columnwidth}{!}{%
\begin{tabular}{|l|l|l|l|l|l|}
\hline
Tileset & Map type & U & P & 2W & 6W \\ \hline
& random & 0.50 & 0.27 & 0.23 & 0.23 \\ \cline{2-6}
\multirow{-2}{*}{Zelda} & user & 0.50 & 0.52 & 0.67 & 0.66 \\ \hline
& random & 0.50 & 0.22 & 0.15 & 0.22 \\ \cline{2-6}
\multirow{-2}{*}{Pokemon} & user & 0.50 & 0.56 & 0.68 & 0.62 \\ \hline
& random & 0.48 & 0.30 & 0.33 & 0.31 \\ \cline{2-6}
\multirow{-2}{*}{Among Us} & user & 0.48 & 0.41 & 0.38 & 0.48 \\ \hline
& random & 0.50 & 0.13 & 0.49 & 0.35 \\ \cline{2-6}
\multirow{-2}{*}{Pacman} & user & 0.49 & 0.52 & 0.67 & 0.74 \\ \hline
& random & 0.49 & 0.33 & 0.24 & 0.28 \\ \cline{2-6}
\multirow{-2}{*}{Dungeon} & user & 0.49 & 0.50 & 0.71 & 0.72 \\ \hline
\end{tabular}%
}
\caption{Average prediction values for each model trained on the pairing data ([U]ntrained, [P]retrained, [2 W]eeks, [6 W]eeks) for each tileset. The values indicate the percentage of votes the maps are expected to receive.}
\label{label:valid_exp}
\end{table}
\section{Post Experiments}
By August 4th, over 3 months since the release of the Aesthetic Bot, further data was collected from the polling results and more experiments were performed. The total number of maps nearly doubled from 399 to 733. Understandably, the average number of votes per poll decreased over time from 27.5 to 18.4. The user win percentage has increased to 78\% as well, most likely due to the increase in quality from the user-made maps. The author @Nifflas has uploaded 135 maps out of the 733 submitted maps that have a poll win rate of 94\% against generated maps and other user maps.
Initially, we designed the APM architecture with the sliding window inputs to account for the varying sizes of the maps. However, because we have substantially more training data since the initial experiments and release of the system, we decided to validate the original APM architecture against a convolutional neural network with a fixed input size. The internal architecture of the network remained the same as the original APM, only instead of taking 4x4 windows of the input map and averaging the output value of each window for a final evaluation measure of the map, the fixed input convolutional neural network (referred henceforth as the POST network) takes in the entire map as input and outputs a single value as the evaluation score. For this experiment we calculate the percent error from the predicted rating of the original APM network architecture and the POST network architecture against the real pollings from Twitter. The poll data was randomly split into a train/test set (90\% / 10\%) to train and evaluate both networks as if they were deployed to Twitter. We trained the networks on 2 map different sizes: 6 (the smallest map size) and 10 (the most common map size.)
Maps that were not exactly 6x6 or 10x10 tiles large were not evaluated by either network. Each network was also pretrained with real-user levels with a vote of 100\% and randomly generated levels with a vote of 0\% to establish a baseline before being fed the real poll data.
Table \ref{label:post_exp} shows the results of this experiment averaged over 5 trials. The APM has a slightly lower relative error for map prediction than the POST network for all tilesets for both map sizes. Not only is the original APM model more flexible to accommodate for varying map sizes, but it also has a better prediction rate than a fixed-size CNN model.
\begin{table}[ht!]
\resizebox{\columnwidth}{!}{%
\begin{tabular}{|cl|ll|ll|}
\hline
\multicolumn{2}{|c|}{Map Size} & \multicolumn{2}{c|}{6} & \multicolumn{2}{c|}{10} \\ \hline
\multicolumn{2}{|l|}{Network} & \multicolumn{1}{l|}{APM} & POST & \multicolumn{1}{l|}{APM} & POST \\ \hline
\multicolumn{1}{|c|}{} & \multicolumn{1}{c|}{Zelda} & \multicolumn{1}{c|}{0.1363} & 0.1564 & \multicolumn{1}{l|}{0.1061} & 0.1492 \\ \cline{2-6}
\multicolumn{1}{|c|}{} & Pokemon & \multicolumn{1}{l|}{0.1596} & 0.2011 & \multicolumn{1}{l|}{0.1157} & 0.1319 \\ \cline{2-6}
\multicolumn{1}{|c|}{} & Among Us & \multicolumn{1}{l|}{0.2136} & 0.2272 & \multicolumn{1}{l|}{0.1150} & 0.1159 \\ \cline{2-6}
\multicolumn{1}{|c|}{} & Pacman & \multicolumn{1}{l|}{0.1309} & 0.1488 & \multicolumn{1}{l|}{0.1419} & 0.1616 \\ \cline{2-6}
\multicolumn{1}{|c|}{\multirow{-5}{*}{Tileset}} & Dungeon & \multicolumn{1}{l|}{0.1327} & 0.1761 & \multicolumn{1}{l|}{0.1545} & 0.1657 \\ \hline
\end{tabular}%
}
\caption{Average percent error on the test dataset for the original APM network architecture and the POST network architecture. The APM slightly outperforms the POST network for poll prediction accuracy.}
\label{label:post_exp}
\end{table}
\section{Discussion}
Outside of the general statistics from this experiment, we noticed many interesting behaviors from both the userbase and the Aesthetic Bot that either emerged over time or were initially overlooked as a possibility.
\subsection{Types of Aesthetic}
As defined before, aesthetics can take on many different forms. What is pleasing to the eye can be defined based on both a globally subjective quality or personal preference. Some examples of aesthetic in terms of level design can include minimalism, symmetry, the variance of tiles shown in a single screen, sprite arrangement, and more.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{imgs/aesthetic_poll.png}
\caption{An example of the bot's map design (B) with more tile variation that lost to the user's more ``minimalistic" (A)}
\label{fig:poll}
\end{figure}
Both the users and the Aesthetic Bot are capable of producing different styles of aesthetic. Figure \ref{fig:poll} shows a user-gen pairing that could arguably have different aesthetic designs. Most of the polls held by Aesthetic Bot also had a small number of users who would vote for the minority map
- causing these polls to not have a unanimous vote for a single map. We carefully chose the wording of the question to be ``Which do you like better" and not ``Which map is user-made?" or ``Which looks the most like a real level?," so that users evaluated the maps on their own personal aesthetic preference. The votes of the poll are also hidden until a user votes themselves, so they are unable to simply choose the majority answer without consideration to the map design.
In the end, the bot learns to weigh the map with the majority of votes as the aesthetic preference and tries to generate maps that can achieve a similar aesthetic, but this may not necessarily mean the minority map had ``bad" aesthetic - the bot instead associates with a lower vote rate.
\subsection{``Hacked" and ``Non-Ludic" Map Designs}
With the map designer tool, we expected users to make their own maps that could look like stand-alone levels or partial subsections of a level for the tilemap's game - referred here as ``level"-like maps. From this, the Aesthetic Bot would learn general level design patterns (i.e. constructing fully made houses or trees or having paths that lead to other objects) as well as potential ``decorating" techniques that could add overall aesthetic to the map (i.e. replacing empty spaces with flowers, grass, or dirt tiles with moderation.)
Some users used the tilesets in a more creative manner to make more complex structures and designs with the maps - outside of the intention of the author's designs. This included overlapping certain tiles to make patterns or clusters of objects or using tiles to look like other structures such as bridges. We call these more complex designs ``hacked" maps; some examples include multi-story houses using the Pokemon tilemap, a ``forest" design in the Zelda tilemap, and a interwoven knot design with the Pacman tilemap.
Thirdly, there were a few maps created by users that instead took advantage of the patterns and art style of the tilemaps themselves to create maps that look more like art than game levels. We call these maps ``non-ludic" maps. Nonetheless, these maps could also be considered aesthetically pleasing, but in a different way than intended. Some maps simply looked like images of other objects (such as a duck or a face) while other maps contain words or messages (like 'Hi' or 'Yo'.)
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{imgs/All_Level_Types.png}
\caption{Different groups of user made maps that were submitted to the database. From top to bottom are ``level"-like maps, ``hacked" maps, and ``non-ludic" maps.}
\label{fig:user_level_types}
\end{figure}
These types of maps - demonstrated in Figure \ref{fig:user_level_types} - received high votes in their paired pollings. As a result, the Aesthetic Bot also tried to emulate these designs in its own generated maps. However, the ``hacked" and ``non-ludic" designs have occurred infrequently (so far), thus the bot tries to combine these techniques with the more common ``level-like" designs. Figure \ref{fig:hack_gen} shows some examples of generated maps that most likely have been influenced by previously evaluated user maps; the Pokemon model tries to recreate the protruding structure of the house, the Among Us model tries to arrange groups of characters by color, and the Dungeon model tries to make maze-like paths through the map.
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{imgs/Hacked_Generations.png}
\caption{Some examples of user maps influencing future map generations. The bot tries to recreate the more complex styles and designs in the user maps.}
\label{fig:hack_gen}
\end{figure}
\section{Future Work and Conclusion}
We use Twitter to quickly retrieve votes and ratings from users on static image game maps made by Aesthetic Bot and the users themselves. The A-vs-B voting system posted along with the initial image of the map pairing was intended to try to retain users attention spans as they scroll through the feed of content while still remaining in low effort for evaluation on the users' ends. In the future, we would like to use Twitter or another social media site with a large user base for more PCG evaluation - such as for
GIFs of agents playing game levels or links to small generated micro-games.
We hope that more Twitter bots emerge that have a two-way feedback with their user-base where the bots can adapt to user preference instead of only posting content without learning from their audience.
The Aesthetic Bot will continue to live on Twitter and improve its ``understanding" of aesthetic
Once the bot has attained a level of suitable generation quality, we would like to apply Aesthetic Bot as a post-process generator/decorator to real game levels. Ideally, Aesthetic Bot would be able to work around essential game elements and place tiles and elements that are not necessary for gameplay and instead act as decorations for the level generation.
The base level of the game could be designed either by a human or procedurally designed by another level generator - like many already existing PCG systems - and work as a collaborative tool.
We introduce a novel interactive evolutionary system called that focuses on evolving the visual aesthetic design of game maps based on the results of user ratings retrieved from polls posted on the social media platform Twitter. We show that the system is capable of improving the design of the maps over a relatively short amount of time and that user engagement through voting and submission of new maps is retained throughout. We hope that this system will be the first of many interactive evolution systems that offer a new form of two-way feedback and communication between the AI generative system and their user-base and foster a better creative process for collaborative design projects.
\section{Introduction}
Whether acting as a teacher, an editor, or a creative partner, both artificially intelligent systems (AI) and human users can learn from each other to improve not only their own design and stylistic techniques, but also their ability to evaluate their own work for quality.
Interactive evolution can act as a bridge for the creative process between human design and automated, generative design by allowing a two-way communication between the user and the AI. Together they can collaboratively evolve and create new content - such as literature, art, music, and game design artifacts - all while gradually improving their design skills.
However, finding a common understanding between a user's preferences and what the AI is trained to generate towards is difficult to achieve. Much like talking to another person, communicating preference for how the content should look, feel, and perform while defining what content is ``high" and ``low" quality takes time, repeated effort, and a willingness from both parties to understand and learn the other's own creative process. Teaching and directing an AI system towards a particular design preference is also an exhausting task in itself, especially if the environment for teaching the system is unfamiliar or overwhelmingly complex. Many content generation systems only focus on generating content that fits an easily measured criteria - such as how well a prompt is matched for literature, the level of consonance or dissonance for music, and how playable a level is for procedurally generated game design levels. More work can be done to ensure how aesthetically pleasing the content generated is - whether on a personal scale or globally.
This paper introduces the Aesthetic Bot system, a novel procedural content generation system that focuses exclusively on improving the appeal of the visual design - henceforth referred to as visual aesthetics - of a game map by learning general user preferences of design.
To learn this preference, this system retrieves user feedback in a quick, minimal effort method of votes in an A vs. B comparison. We use the social media site Twitter as the host for this system, as it has a large (allegedly majority human \cite{muskTwitter}) user base that frequently interacts with the platform and is able to view and engage with content made by other accounts. With this platform, we are able to have Twitter users vote directly on the bot's generated maps or real, user-made maps in an interaction that takes only a few seconds of evaluation time within their regularly updated content timeline. After the polling, the system learns from the rating to generate new maps that are optimized to win future preference pollings and improve its overall ability to determine aesthetic quality in a map.
\section{Background}
\subsection{Aesthetic Design in Content Generation}
The Stanford Encyclopedia of Philosophy describes aesthetics as a subjective taste in a sense of beauty and ugliness. There is no hard truth to what aesthetic is and they do not necessarily need to be related to visual tastes \cite{zangwill2003aesthetic}. David and Globe note that the visual aesthetics of a system, such as one used for user interfaces and human computer interaction, can influence a user's perception of ease of use for the system overall, and also infer how much a user may engage with it. Well designed visual aesthetics of a system can also foster collaboration between users, as used in Google Docs and online learning tools. Thus, the study of aeshetic design is critical in order to create collaborative systems that not only allow for a more friendly first-impression of the system but also encourage frequent user engagement. \cite{david2010impact}.
Games often add extra visuals effects to make it more visually aesthetically pleasing and appealing to the player. This can be done through decorating the world itself or by adding effects to make the game experience more interactive and responsive called game feel \cite{swink2008game}.
Previous work has been done to automate this process, such as in the Juicer project \cite{johansen2021challenges}.
For this paper, we tactfully bow out of the complexities of pinning down visual aesthetics by operationally defining good visual aesthetics as ``visuals that a user prefers when given multiple options of content to choose from".
Aesthetic design is subjective to a human experience - determining what looks good is harder to describe or quantify than analyzing raw metrics of performance. However, we argue that judging aesthetic design could be emulated; much like how human decision making processes and evaluations have been emulated by machines already. While games and AI research have focused on narrative \cite{alvarez2022tropetwist, aarseth2012narrative} or mechanic \cite{hunicke2004mda} aesthetics of a system in the past, in the Aesthetic Bot system we focus exclusively on the visual output of the maps and ignore their playabilities, mechanic activations, or any other functionality constraints typically involved with game design.
\subsection{Procedural Content Generation}
Procedural content generation (PCG) for level design enables, among other things,
unique experiences for players. Most notably, PCG has been used across multiple genres of games - such as in Minecraft, Spelunky, Borderlands 2, and the classic Rogue - and in game AI research including top-down grid worlds \cite{earle2021learning,sandhu2021tileterror}, platformers \cite{bhaumik2021lode,sarkar2021procedural,volz2018evolving}, puzzle games \cite{charity2020baba,khalifa2020multi}, open-world 3d environments \cite{salge2018generative}, and more.
The generation process for PCG levels typically focuses on generating towards an objective or fitness that can be calculated based on some metric (i.e. playability, solution length, entropy.) While they may be successful in their endeavor of producing levels that guarantee completion of this objective, most PCG systems do so without taking into account the look or visual appeal of the level. As such, many of these levels look artificially made, and lack the meticulous, intentional design that human-made levels have. A system that is capable of maintaining the visual aesthetic of the level - either parallel to the main generation objective or as a post-processing step - could improve the visual quality of the level overall and make a much more engaging experience for the player.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{imgs/Mario_PCG.png}
\caption{Two levels from the Mario AI framework. The left is user-made and the right is generated. While both are playable, the user-made level arguably has more aesthetic.}
\label{fig:pcg_mario}
\end{figure}
\subsection{Interactive Evolution and Co-Creativity}
Contrary to pure PCG systems, interactive evolution (IE) is a content generation method where the evaluation of the system's content output is dependent on the user's selection \cite{dawkins1986blind}. Picbreeder is such an example of a system where users are able to select from a set of different images generated by an evolutionary algorithm. Images are evolved and removed based on a subset of images selected by a single user. This guides the system towards evolving the population of images to either a singular goal image or to fit the user's personal preference \cite{secretan2008picbreeder}. Bontrager et al.'s work used a similar system but instead generating the images with a deep learning model and evolved latent vectors to accomplish the same goal \cite{bontrager2018deep}. Liapis et al. also looked to use machine learning and evolutionary techniques with constrained, limited user selection to generate content based on visual and aesthetic appeal of the designer in various projects ranging from spaceship designs to map terrains \cite{Liapis_adaptive,liapis2012limitations,Liapis_optimizing,liapis2013sentient,Liapis_boosting}.
However, most of these systems mentioned are unable to retain a sense of general aesthetic preferences learned from users.
They require many iterations that start from scratch in order to reach a state where the content produced is satisfactory - and such content is only made for a single user. They cannot learn the majority preference of multiple users or transfer their knowledge across sessions, as they are isolated to the singular interaction with the current user. IE systems can also exhaust a user and lead to participant fatigue - especially when these types of systems require so much feedback in order to effectively learn \cite{kamalian2006reducing}.
The Aesthetic Bot attempts to addresses these issues by receiving polled feedback from multiple users all at once in short interactions while limiting the selection pool.
\subsection{Twitter Bots}
Twitter bots are artificial profiles made on the social media platform Twitter that post content either unprompted with automated posts or in response to other users on the platform \cite{veale2018twitterbots}. Many bots are made to produce content for mostly entertainment purposes - such as a bot that combines emojis together to form new ones (@EmojiMashupBot). Other bots exist to interact with users once prompted to do so - such as generating on-demand haikus (@haikookies) \cite{pichlmair2020procedural}. However, both types of bots only act as one-way posters - they do not learn from the a two-way communication between the user and the bot itself. Much in the same way Twitter's human profiles can gain information by posing questions and receiving replies or posting polls to receive votes on a subject, so too can a bot learn from these posts. With the Twitter API, Twitter can be used as a means of retrieving feedback in the same way as the previous interactive evolution co-creative systems but in a system that is already familiar and regularly used by participants.
\section{Aesthetic Bot System}
The Aesthetic Bot system is a pipeline comprising of a convolutional neural network called the Aesthetic Prediction Model (APM), an evolutionary algorithm, and the Twitter platform interface. Figure \ref{fig:pipeline} shows the entire learning pipeline generating and evaluating a new map that is sent to Twitter.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{imgs/Tilesets.png}
\caption{The experiment's usable tilesets from left to right: Zelda, Pokemon, Among Us, Pacman, Dungeon}
\label{fig:tilesets}
\end{figure}
For this experiment, we created 5 original pixel art tilesets (shown in Figure \ref{fig:tilesets}) that we designed to resemble the visual themes of the following games: Zelda, Pokemon, Among Us, Pac-Man, and a generic dungeon crawler (called Dungeon). Each tileset contains 16 8x8 pixel tiles - some tiles were purposefully designed for inter-connectivity and to be placed in such a way to form a larger pattern of tiles (i.e. the tree tiles in Zelda.) We chose to use very small tileset visual dimensions so that users would ideally have a lower amount of visual aesthetic bias towards a particular art style or game due to higher resolutions or more complex graphics. The smaller resolutions are also much easier to view at a glance when the pixels are scaled up.
\begin{figure}[b]
\centering
\includegraphics[width=1.0\linewidth]{imgs/Aesthetic_Bot_Pipeline.png}
\caption{Learning pipeline of the Aesthetic Bot. Dark grey arrows indicate the recipient will update based on the output, while light grey arrows indicate to only process the output}
\label{fig:pipeline}
\end{figure}
\subsection{Aesthetic Prediction Model}
The convolutional neural network known as the Aesthetic Prediction Model (APM) acts as the ``judge" for the system and tries to estimate the output of a Twitter poll based on an input map. It trains and updates based on the actual results of the Twitter polls. For this experiment, the model consists of an extremely small neural network with 3 convolutional layers, each followed by a batch normalization layer, a ReLU activation function and finally a MaxPooling layer (except at the last ReLU which has a flattened, dense layer.) Instead of passing the entire map to the network, the map is divided into one-hot encoded 4x4 tile windows with 16 channels (for each tile in the tileset). Due to the small input size the model only has 15,185 parameters and output shapes of (2,2,64), (1,1,32), (1,1,16), (16), and (1).
The smaller sliding window of 4x4 helps to accommodate for the various sizes of the maps and helps the model learn and judge on patterns of tiles instead of entire maps which have more distinction and variance between each other. If a neural network with a fixed size were used instead (such as the smallest map size of 6 or the largest of 12), this could limit the learning capabilities of the APM, as some maps may not be able to be processed for being too small while others may lose noticeable aesthetic features from being cutoff by the smaller input window. We later compare the prediction capabilties of a fixed input network against the APM's sliding window input network in the 'Post Experiments' section of this paper.
The model is trained by passing the encoded 4x4 sliding subsection windows of a map as its input and using the percentage of votes the input map received in the Twitter pairing as its output y value. Figure \ref{fig:cnn_pipeline} shows this process of taking two maps from a polling and processing them for training the APM. We train the model on the percentage of votes instead of the votes directly because we want to encourage the model to learn how to estimate which maps will win in any given pairing rather than on how many votes it would receive.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{imgs/CNN_Diagram.png}
\caption{Diagram of the APM learning process. Each window from the map receives the same y value and the network learns to associate a window pattern with a vote percentage.}
\label{fig:cnn_pipeline}
\end{figure}
The votes themselves are variable with the Twitter polls and harder to distinguish relatively in a map's terms of quality, except when paired with another map. The output value of a map is assigned as the y value for every subsection pattern of the map. As such, it is possible for the APM to learn 2 different contradictory y values for the same pattern, however the context and values of the other patterns will help distinguish the quality of the map overall.
After training, the weights are frozen and the APM is used as the evolutionary model's fitness function to evaluate the population of maps. When the APM evaluates a map, the map is separated into the same 4x4 sliding windows and passed through the APM. The average of the set of window patterns' outputs is used as the final fitness value for the entire map. Therefore, the more higher rated patterns a map contains, the more likely it will perform better in a pairing and conversely for lower rated patterns.
\subsection{Evolutionary Algorithm}
To generate new maps, the pipeline implements a $\mu+\lambda$ evolutionary algorithm - where $\mu$ maps with the highest fitness are retained in the next population and $\lambda$ maps are randomly selected from the population and mutated for
evaluation in the next iteration. This experiment uses 10\% for mu and 90\% for lambda. The algorithm starts from a uniformly randomized population of maps of varying dimensions.
The sizes of the maps are limited to a square range of 6-12 tiles in dimension. As mentioned in the previous subsection, the output of the APM acts as the fitness function for the evolutionary system. As a mutation function, each 2x2 subsection of tiles in a map have a random chance to be mutated and replaced with a 2x2 subsection pattern from the real user maps saved to the system. We found that this method encourages the occurrence and retention of more ``user"-made patterns with the direct replacement rather than randomly mutating singular tiles. The dimensions of the map also have a small (0.1\%) chance to change.
After a set number of iterations, the map with the highest fitness value (as determined by the APM) is selected from the evolutionary algorithm's population. This experiment runs the evolutionary process for multiple trials and saves the best maps from each trial's population in a separate list to be evaluated in a final selection phase. This final selection evaluates the maps based on some preset function (entropy, fake Twitter function value, both, etc) in case the evolutionary algorithm plateaued in its search for a high quality map. Figure \ref{fig:final_selection} shows a set of the final maps generated. This final map is saved to the database as the tileset's generated map and associated with the trained APM that helped with the evaluation to produce it. This is to keep track of which maps came as the results of a particular APM and monitor its improvements (or regression.)
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{imgs/final_selection3.png}
\caption{A set of generated maps from the pipeline. The map submitted to the database will be decided based on a secondary function (labeled at the top.)}
\label{fig:final_selection}
\end{figure}
\subsection{Twitter Interaction}
In a separate process performed parallel to the map generation (in order to monitor generated levels before pairing),
a Python script selects a submitted user map and a generated map that uses the same tileset from their respective databases (typically the oldest unevaluated map from both) and pairs them together. This pairing renders the two maps together, side-by-side, in an image with the label ``Which do you like better?" underneath and saves the pairing information to another database table. The image is then posted to the Aesthetic Bot Twitter account with a poll as a threaded reply post using the Twitter API. The poll is held for 3 hours to allow as many users as possible to vote on the pairing.
At the completion of the poll, the votes are recorded to the pairing's database row. The tileset's associated APM is trained and updated with the results of the post's pairing. This is how the APM updates past its initial training stage and improves dynamically with each new pairing. This updated APM is used again as the evaluator for the tileset's next generated map and cycles through the pipeline. Finally, as a way of giving closure to curious users, the system reveals in a final post in the poll thread which map shown in the pairing was the user-made map as well as the map author if one was provided. A link to the Aesthetic Bot map editor - described in more detail in the 'Live Training' section - is included in the post to encourage participants to upload their own map to the database to be used in a future pairing.
A later modification of the system also included a small chance to make homogeneous author-type pairings as well to prevent biases from occurring in the APM and the participating users. Through this method, the network still learns which maps outperformed others, not just in a user-gen pairing situation. This would also cause voting participants to think more critically about the map pairings being shown and why they actually preferred a map over another - and not just to try to discern which of the two maps was user made - as it could not be one of them, both of them, or neither.
\subsection{Pre-training Stage}
Before pairing posts were made for the Aesthetic Bot Twitter account, the APMs for each tileset went through a pre-training stage to prevent completely random maps from being shown in the pairings and provide a ``decent" looking map as comparison for users. A pseudo-Twitter evaluation function was used to emulate Twitter votes on a map and train the neural network. This function used a modification of Lucas and Volz's Tile Pattern Kullback-Leibler Divergence (TPKL-Div) algorithm to evaluate maps \cite{lucas2019tile} by taking the map's Kullback-Leibler divergence score - or relative entropy - to compare it to a training set's tile patterns and determine the similarity in pattern compositions. This was to encourage more ``user-made" looking maps while also encouraging novel design and structure. In addition to the TPKL-Div score, raw tile entropy probabilities were calculated (this was to discourage maps with little variation in tiles or mostly empty maps) and multiplied together with the TPKL-Div score. The final value from this function was meant to simulate the number of votes a map received. Like the real Twitter polls, each map in a pairing received their votes (artificially calculated with the entropy and TPKL-Div) and the difference in votes was used as the output y value to train the neural network. By using an artificial function to simulate a base for user aesthetic, the neural network was able to train in a much shorter amount of time than the weeks or months it would take to train a randomly initialized network using live Twitter data.
Initially, the networks were pre-trained with a set of pairings of user-made and randomly generated maps so that the network would have a baseline output and learn to distinguish pure randomness before evaluating populations. Afterwards, the new maps were evolved in the evolutionary algorithm as normal with the pseudo-Twitter function acting as the Twitter evaluator to update the APM. On completion of these iterations, the pre-trained APM was implemented into the actual pipeline and continued to update and learn using the real data polling from Twitter.
\subsection{Live Training}
The Aesthetic Bot was officially announced and revealed publicly on Twitter on April 28th, 2022. The bot would post a new poll every 3 hours. The database of submitted and generated maps along with the results of the pair polls were backed up to a separate server so any new models made could be trained and updated to the current state of the pollings.
On the Aesthetic Bot bio information status for the account, as well as at the end of every poll, a link to the Aesthetic Bot map editor was made available to Twitter users. The site\footnote{http://aesthetic-bot.xyz/} allows participants to submit their own map designs to the server database to be used in a future pairing. Users could select from the five tilesets available and draw tiles onto the JavaScript canvas. The map editor interface - as shown in figure \ref{fig:editor} - was designed to be mobile friendly as well so that users on the Twitter app could seamlessly submit a map. Upon submission of a new map, users were also given the option to provide their Twitter handle so that they could have authorship over the map. The user would be tagged/credited in the closing thread post of the poll that included their map.
On the developer end, we built a map deleter to monitor the maps made by users and make sure any offensive or graphic looking maps or offensive fake usernames were not included in the pairings.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{imgs/Level_Editor.png}
\caption{The map editor site where users can design and submit their own maps that Aesthetic Bot will learn from.}
\label{fig:editor}
\end{figure}
\section{Results}
As of May 30th, 2022, there were a total of 399 user maps submitted with an average map size of 9.92 tiles squared. The distribution of maps included 101 user maps made for the Zelda tileset, 108 Pokemon maps, 31 Among Us maps, 69 Pacman maps, and 90 Dungeon maps.
The Among Us tileset's unpopularity may have been because the actual game is the least familiar and more recently made of the 5 games the other tilesets were based on. The authors with the most submitted maps (excluding anonymously authored maps that made up 1/3 of the user maps) were @Nifflas, @MasterMilkX, dginev, @3phen, @AndresZarta, and @charphinB. Out of the hybrid pairings, the average number of total votes per poll were 27.5 votes. Of the 220 total polls, the user-made maps won 74\% of matchups, while the generated maps won 24\%. 2\% resulted in ties between the 2 maps.
However, the percentage of votes the user maps have received slowly decreased over time - as shown in Figure \ref{fig:win_perc_chart}. The trend line (in red) shows the user maps slowly decreasing in vote majority. This implies that the generated maps are starting to receive more user votes over time and rival user-made maps in aesthetic quality.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{imgs/win_perc_chart_long2.png}
\caption{Percentage of votes received for the user maps. }
\label{fig:win_perc_chart}
\end{figure}
A validation test was done on the APMs to test how well it could distinguish between ``good" and ``bad" maps over time. 3 different models at various stages of training: the pre-trained model before release to Twitter, the model 2 weeks after release, and the model 6 weeks after release. These were evaluated on 2 sets of maps: randomly generated (unevolved) maps that would be unlikely to win a polling and unseen user maps hand chosen and predicted to have a high aesthetic. Every tileset model - including those trained on Twitter were able to successfully predict the random maps at a lower winning rate (less than 30\% of votes) - except the 2 week Pacman model. The models were also able to predict high vote rate (more than 60\% of votes) for the unseen user maps - except the Among Us tileset, most likely because of the considerable lack of trained pairings. Thus, we can infer that most of the APMs are still able retain a high accuracy for predicting how likely a particular map would win a polling even after receiving new data from real polls. Table \ref{label:valid_exp} shows the results of the averaged predictions from each model type on each of the tilesets.
\begin{table}[ht!]
\resizebox{\columnwidth}{!}{%
\begin{tabular}{|l|l|l|l|l|l|}
\hline
Tileset & Map type & U & P & 2W & 6W \\ \hline
& random & 0.50 & 0.27 & 0.23 & 0.23 \\ \cline{2-6}
\multirow{-2}{*}{Zelda} & user & 0.50 & 0.52 & 0.67 & 0.66 \\ \hline
& random & 0.50 & 0.22 & 0.15 & 0.22 \\ \cline{2-6}
\multirow{-2}{*}{Pokemon} & user & 0.50 & 0.56 & 0.68 & 0.62 \\ \hline
& random & 0.48 & 0.30 & 0.33 & 0.31 \\ \cline{2-6}
\multirow{-2}{*}{Among Us} & user & 0.48 & 0.41 & 0.38 & 0.48 \\ \hline
& random & 0.50 & 0.13 & 0.49 & 0.35 \\ \cline{2-6}
\multirow{-2}{*}{Pacman} & user & 0.49 & 0.52 & 0.67 & 0.74 \\ \hline
& random & 0.49 & 0.33 & 0.24 & 0.28 \\ \cline{2-6}
\multirow{-2}{*}{Dungeon} & user & 0.49 & 0.50 & 0.71 & 0.72 \\ \hline
\end{tabular}%
}
\caption{Average prediction values for each model trained on the pairing data ([U]ntrained, [P]retrained, [2 W]eeks, [6 W]eeks) for each tileset. The values indicate the percentage of votes the maps are expected to receive.}
\label{label:valid_exp}
\end{table}
\section{Post Experiments}
By August 4th, over 3 months since the release of the Aesthetic Bot, further data was collected from the polling results and more experiments were performed. The total number of maps nearly doubled from 399 to 733. Understandably, the average number of votes per poll decreased over time from 27.5 to 18.4. The user win percentage has increased to 78\% as well, most likely due to the increase in quality from the user-made maps. The author @Nifflas has uploaded 135 maps out of the 733 submitted maps that have a poll win rate of 94\% against generated maps and other user maps.
Initially, we designed the APM architecture with the sliding window inputs to account for the varying sizes of the maps. However, because we have substantially more training data since the initial experiments and release of the system, we decided to validate the original APM architecture against a convolutional neural network with a fixed input size. The internal architecture of the network remained the same as the original APM, only instead of taking 4x4 windows of the input map and averaging the output value of each window for a final evaluation measure of the map, the fixed input convolutional neural network (referred henceforth as the POST network) takes in the entire map as input and outputs a single value as the evaluation score. For this experiment we calculate the percent error from the predicted rating of the original APM network architecture and the POST network architecture against the real pollings from Twitter. The poll data was randomly split into a train/test set (90\% / 10\%) to train and evaluate both networks as if they were deployed to Twitter. We trained the networks on 2 map different sizes: 6 (the smallest map size) and 10 (the most common map size.)
Maps that were not exactly 6x6 or 10x10 tiles large were not evaluated by either network. Each network was also pretrained with real-user levels with a vote of 100\% and randomly generated levels with a vote of 0\% to establish a baseline before being fed the real poll data.
Table \ref{label:post_exp} shows the results of this experiment averaged over 5 trials. The APM has a slightly lower relative error for map prediction than the POST network for all tilesets for both map sizes. Not only is the original APM model more flexible to accommodate for varying map sizes, but it also has a better prediction rate than a fixed-size CNN model.
\begin{table}[ht!]
\resizebox{\columnwidth}{!}{%
\begin{tabular}{|cl|ll|ll|}
\hline
\multicolumn{2}{|c|}{Map Size} & \multicolumn{2}{c|}{6} & \multicolumn{2}{c|}{10} \\ \hline
\multicolumn{2}{|l|}{Network} & \multicolumn{1}{l|}{APM} & POST & \multicolumn{1}{l|}{APM} & POST \\ \hline
\multicolumn{1}{|c|}{} & \multicolumn{1}{c|}{Zelda} & \multicolumn{1}{c|}{0.1363} & 0.1564 & \multicolumn{1}{l|}{0.1061} & 0.1492 \\ \cline{2-6}
\multicolumn{1}{|c|}{} & Pokemon & \multicolumn{1}{l|}{0.1596} & 0.2011 & \multicolumn{1}{l|}{0.1157} & 0.1319 \\ \cline{2-6}
\multicolumn{1}{|c|}{} & Among Us & \multicolumn{1}{l|}{0.2136} & 0.2272 & \multicolumn{1}{l|}{0.1150} & 0.1159 \\ \cline{2-6}
\multicolumn{1}{|c|}{} & Pacman & \multicolumn{1}{l|}{0.1309} & 0.1488 & \multicolumn{1}{l|}{0.1419} & 0.1616 \\ \cline{2-6}
\multicolumn{1}{|c|}{\multirow{-5}{*}{Tileset}} & Dungeon & \multicolumn{1}{l|}{0.1327} & 0.1761 & \multicolumn{1}{l|}{0.1545} & 0.1657 \\ \hline
\end{tabular}%
}
\caption{Average percent error on the test dataset for the original APM network architecture and the POST network architecture. The APM slightly outperforms the POST network for poll prediction accuracy.}
\label{label:post_exp}
\end{table}
\section{Discussion}
Outside of the general statistics from this experiment, we noticed many interesting behaviors from both the userbase and the Aesthetic Bot that either emerged over time or were initially overlooked as a possibility.
\subsection{Types of Aesthetic}
As defined before, aesthetics can take on many different forms. What is pleasing to the eye can be defined based on both a globally subjective quality or personal preference. Some examples of aesthetic in terms of level design can include minimalism, symmetry, the variance of tiles shown in a single screen, sprite arrangement, and more.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{imgs/aesthetic_poll.png}
\caption{An example of the bot's map design (B) with more tile variation that lost to the user's more ``minimalistic" (A)}
\label{fig:poll}
\end{figure}
Both the users and the Aesthetic Bot are capable of producing different styles of aesthetic. Figure \ref{fig:poll} shows a user-gen pairing that could arguably have different aesthetic designs. Most of the polls held by Aesthetic Bot also had a small number of users who would vote for the minority map
- causing these polls to not have a unanimous vote for a single map. We carefully chose the wording of the question to be ``Which do you like better" and not ``Which map is user-made?" or ``Which looks the most like a real level?," so that users evaluated the maps on their own personal aesthetic preference. The votes of the poll are also hidden until a user votes themselves, so they are unable to simply choose the majority answer without consideration to the map design.
In the end, the bot learns to weigh the map with the majority of votes as the aesthetic preference and tries to generate maps that can achieve a similar aesthetic, but this may not necessarily mean the minority map had ``bad" aesthetic - the bot instead associates with a lower vote rate.
\subsection{``Hacked" and ``Non-Ludic" Map Designs}
With the map designer tool, we expected users to make their own maps that could look like stand-alone levels or partial subsections of a level for the tilemap's game - referred here as ``level"-like maps. From this, the Aesthetic Bot would learn general level design patterns (i.e. constructing fully made houses or trees or having paths that lead to other objects) as well as potential ``decorating" techniques that could add overall aesthetic to the map (i.e. replacing empty spaces with flowers, grass, or dirt tiles with moderation.)
Some users used the tilesets in a more creative manner to make more complex structures and designs with the maps - outside of the intention of the author's designs. This included overlapping certain tiles to make patterns or clusters of objects or using tiles to look like other structures such as bridges. We call these more complex designs ``hacked" maps; some examples include multi-story houses using the Pokemon tilemap, a ``forest" design in the Zelda tilemap, and a interwoven knot design with the Pacman tilemap.
Thirdly, there were a few maps created by users that instead took advantage of the patterns and art style of the tilemaps themselves to create maps that look more like art than game levels. We call these maps ``non-ludic" maps. Nonetheless, these maps could also be considered aesthetically pleasing, but in a different way than intended. Some maps simply looked like images of other objects (such as a duck or a face) while other maps contain words or messages (like 'Hi' or 'Yo'.)
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{imgs/All_Level_Types.png}
\caption{Different groups of user made maps that were submitted to the database. From top to bottom are ``level"-like maps, ``hacked" maps, and ``non-ludic" maps.}
\label{fig:user_level_types}
\end{figure}
These types of maps - demonstrated in Figure \ref{fig:user_level_types} - received high votes in their paired pollings. As a result, the Aesthetic Bot also tried to emulate these designs in its own generated maps. However, the ``hacked" and ``non-ludic" designs have occurred infrequently (so far), thus the bot tries to combine these techniques with the more common ``level-like" designs. Figure \ref{fig:hack_gen} shows some examples of generated maps that most likely have been influenced by previously evaluated user maps; the Pokemon model tries to recreate the protruding structure of the house, the Among Us model tries to arrange groups of characters by color, and the Dungeon model tries to make maze-like paths through the map.
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{imgs/Hacked_Generations.png}
\caption{Some examples of user maps influencing future map generations. The bot tries to recreate the more complex styles and designs in the user maps.}
\label{fig:hack_gen}
\end{figure}
\section{Future Work and Conclusion}
We use Twitter to quickly retrieve votes and ratings from users on static image game maps made by Aesthetic Bot and the users themselves. The A-vs-B voting system posted along with the initial image of the map pairing was intended to try to retain users attention spans as they scroll through the feed of content while still remaining in low effort for evaluation on the users' ends. In the future, we would like to use Twitter or another social media site with a large user base for more PCG evaluation - such as for
GIFs of agents playing game levels or links to small generated micro-games.
We hope that more Twitter bots emerge that have a two-way feedback with their user-base where the bots can adapt to user preference instead of only posting content without learning from their audience.
The Aesthetic Bot will continue to live on Twitter and improve its ``understanding" of aesthetic
Once the bot has attained a level of suitable generation quality, we would like to apply Aesthetic Bot as a post-process generator/decorator to real game levels. Ideally, Aesthetic Bot would be able to work around essential game elements and place tiles and elements that are not necessary for gameplay and instead act as decorations for the level generation.
The base level of the game could be designed either by a human or procedurally designed by another level generator - like many already existing PCG systems - and work as a collaborative tool.
We introduce a novel interactive evolutionary system called that focuses on evolving the visual aesthetic design of game maps based on the results of user ratings retrieved from polls posted on the social media platform Twitter. We show that the system is capable of improving the design of the maps over a relatively short amount of time and that user engagement through voting and submission of new maps is retained throughout. We hope that this system will be the first of many interactive evolution systems that offer a new form of two-way feedback and communication between the AI generative system and their user-base and foster a better creative process for collaborative design projects.
|
\section{Results}
\subsection{Quantitative Results}
\subsubsection{Overall Impact of Reminder Messages on Homework Performance}
\begin{figure}[htb]
\centering
\includegraphics[scale=0.45]{sections/figures/AttemptRateEffectAcrossWeeksCI.pdf}
\includegraphics[scale=0.45]{sections/figures/PerformanceTrendCI.pdf}
\caption{The top figure shows the proportion of students who started the online homework each week from weeks 5 - 8 for students who received a reminder message (circle) and those who did not (diamond). The bottom figure shows the average homework performance each week from weeks 5 - 8 for students who received a reminder message (circle) and those who did not (diamond). The dotted interval at each point indicates the standard error of the mean.}
\label{fig:attempt_across_weeks}
\end{figure}
\ignore{
\begin{figure}[htb]
\centering
\includegraphics[scale=0.5]{sections/figures/PerformanceTrendCI.pdf}
\caption{Connected line plot that shows the average homework performance each week from weeks 5 - 8 for students who received a reminder message (circle) and those who did not (diamond). The dotted interval at each point indicates the standard error of the mean.}
\label{fig:performance_across_weeks}
\end{figure}
}
The messaging intervention had an impact in starting homework. Students who received the reminder messages consistently attempted homework at a higher rate than students who did not as shown by Figure \ref{fig:attempt_across_weeks}. The attempt rate indicates how many students tried the homework, regardless of the number of questions completed or the success. To aggregate the effects of reminder messages across 4 weeks of experimentation we employ a panel data regression model. Table \ref{tab:impact_rate} shows that the reminder messages increased the number of students who started the homework by around 4 percentage points (coefficient = 0.0319; \textit{z} = 2.86; \textit{p} < .01; panel regression with time and participant effects).
The increase of attempt rate is reflected in the homework performance rates shown in Figure \ref{fig:attempt_across_weeks}. Students who received the reminder messages consistently performed better than students in the control group by attempting or completing more homework problems. We employ a panel data regression model to aggregate the weekly effects. In Table \ref{tab:impact_rate}, we observe that the reminder messages increased the homework average by 4 percentage points (coefficient = 0.0388; \textit{z} = 2.80; \textit{p} < .005; panel regression with time effects and participant effects).
Although the reminder messages impacted students' propensity to start the homework, their impact on students' start time was more limited. As shown in Table \ref{tab:impact_rate}, receiving email reminders do not result in a statistically significant impact on start time (coefficient = 1.105; \textit{z} = 0.86; \textit{p} > .10; panel regression with time and participant effects) or on end time (coefficient = -0.118; \textit{z} = -0.11; \textit{p} > .10; panel regression with time and participant effects).
\subsubsection{Impact of the Reminder Design on Homework Performance}
\begin{figure*}[htb]
\centering
\includegraphics[scale=0.39]{sections/figures/box-starttime.PNG}
\caption{Box plots of student homework start time for conditions in each variation. \textit{p}-values are derived from the Komogorov-Smirnov equality of distributions test, in each variation, compares the distribution of homework start time to the control group. Students who did not attempt the homework were removed from this analysis.}
\label{fig:box-start}
\end{figure*}
We want to address the effect of each design variation on student action rate. As noted in Figure \ref{fig:box-start}, the distribution of homework start times shows that students tend to start closer to the deadline, thereby creating a skewed start time distribution. As a result, the mean start time and statistical tests based on the mean are not wholly representative of the effect that the emails conditions have. We chose to use box plots to visualize the median and quartiles of start times between conditions and the control group. Furthermore, start time distributions of control-condition within each variation were compared using the Komogorov-Smirnov test. Significant \textit{p}-values for a control-condition comparison means students' start times between the conditions and control are different; in other words, the reminder has an effect, even if not immediately apparent.
Our results show a significant (\textit{p}<.05) difference in start time distribution for sender (control vs. instructor), recommendation (control vs. no prompt), and advice (control vs. no prompt) variations. Marginal significance (.05$\leq{p}$<.1) was observed for information (control vs. showing the prompt), and advice (control vs. showing the prompt) variations. We did not observe statistical significance for start time distributions for any of the subject line conditions.
The box plots show an approximate 10-hour increase in median start time in the sender (control vs. instructor), information (control vs. yes - showing the prompt), and recommendation (control vs. no prompt) variations. Smaller increases in median start time are seen the sender (control vs. teaching assistant), information (control vs. no prompt), subject line (control vs. summary), recommendation (control vs. yes - showing the prompt), and both advice conditions over control. We note the median start time for the control group was lower or at least as low as conditions in all cases.
\begin{table*}[]
\label{tab:my-table}
\begin{tabular}{@{}lllllllllll@{}}
\toprule
Comparison & Estimate & $P$-Value & Estimate & $P$-Value & Estimate & \multicolumn{1}{l|}{$P$-Value} & Estimate & \multicolumn{1}{l|}{$P$-Value} & Estimate & $P$-Value \\ \midrule
& \multicolumn{2}{l}{Sender-instructor} & \multicolumn{2}{l}{Information} & \multicolumn{2}{l}{Subject line-prompt} & \multicolumn{2}{l}{Recommendation} & \multicolumn{2}{l}{Advice} \\
No - Control & 0.003 & 0.988 & 0.0207 & 0.419 & 0.041 & .211 & $0.081^{**}$ & 0.003 & 0.029 & .338 \\
Yes - Control & 0.015 & 0.538 & -0.004 & 0.853 & 0.047 & .148 & $0.093^{**}$ & 0.001 & $0.081^{**}$ & .008 \\ \bottomrule
\end{tabular}
\caption{The table presents differences in propensity to start the homework for each reminder message variation relative to the control group who did not receive the message. The corresponding p-values are obtained from two independent samples proportions z-test.}
\label{table: attempt_rate}
\end{table*}
Due to the definition of a student attempting a homework problem as a Bernoulli variable, the mean and subsequent \textit{z}-tests in Table \ref{table: attempt_rate} represent the impact of emails on student attempt rate. We note statistically significant (\textit{p}<.05) attempt rate increases for all conditions of recommendation over control (est. 8.1\% no prompt/9.3\% showing the prompt) and advice (control vs. showing the prompt, est. 8.1\%).
\section{Introduction}
The ability to hold and sustain student interest and engagement in course learning activities is a persistent challenge for educators. In particular, studies of online courses have evidence of issues related to high withdrawal rates and low levels of course evaluation and satisfaction when compared to more traditional on-campus-based offerings \cite{bawa2016retention}. As well, student participation in learning activities has been shown to be a significant factor influencing academic performance in online courses \cite{finn2012student, you2016identifying}. Thus, educators are faced with the problem of developing economical yet scalable practices that can assist students in sustaining their engagement in learning activities, especially in online learning modes. Many instructors send emails and announcements to contact students. How can we use emails as a channel for helping students stay on track with their homework, either online or in-person? In other contexts, reflecting and making plans about the next steps improves the chances that people will follow through with their goals \cite{rodriguez2018students, VANDENBOOM2004551}. In this paper, we explore different ways of using emailed homework reminders to prompt students to take steps that might increase students' engagement with the homework.
Past work has identified situations in which particular interventions can be promising-- like particular implementations of reflection prompts \cite{VANDENBOOM2004551} and reminders \cite{edwards2015examining} -- although in other contexts similar interventions do not have an effect \cite{baker2016}, underscoring the importance of empirically testing potential alternative implementations in a context. We explored building alternative versions of emails to students by combining various components: The subject line prompts to spend a few minutes planning, reading homework, or thinking of advice to others to start early, links to external sources or learning resources. Given the broad and complex space of emails that instructors and researchers can write, we design and deploy randomized A/B comparisons (also called randomized controlled trials/experiments) that compare alternative emails, providing examples of how to test the impact and perception of alternative ideas for how to help students keep engaged.
We quantified the potential impact on students by looking at students’ academic performance such as homework grades, attempt rate, and start time. We collected qualitative data from students about the advantages and disadvantages of various design choices, how these reminders fit into their lives and how these reminders are received \cite{kizilcec2017diverse, lim2021impact}. We also collected qualitative data from instructors to contrast their predictions of the impact of various design choices with the actual impact observed. We found that many intuitive interventions predicted to have an impact (e.g. emails from instructors vs. teaching assistants) did not change student behaviour. However, we did find that on average, repeated emails did have an impact over an extended period of time in getting students to attempt homework more often (generalizing past work where we saw only an immediate benefit \cite{zavaleta2021investigating}). Our broader goal is for the details of the experimental design process in these case studies to provide insight into how instructors and researchers can design, deploy, and quantitatively and qualitatively evaluate a wide range of alternative approaches to messaging students.
\section{Related Work}
\textbf{Embedded experimentation:} Motz et al. highlighted the importance of scalable embedded experimentation in providing an accurate characterization of education in real-world scenarios, while also being able to make causal inferences. They also highlighted the necessity of working with instructors in the process of designing experimentation to further the applicability of results \cite{motz2018causal}. Our study addresses this framework through the deployment of A/B comparisons to isolate the effect of deliberate design factors. Furthermore, our case study demonstrates how to instantiate this form of embedded experimentation into practice. Though the involvement of instructors is important, we also seek to address a quantitative approach to evaluating instructor intuitions by comparing predictions to data collected.
\textbf{Student engagement and study habits:} Procrastination is one of the biggest barriers to academic performance \cite{yilmaz2017relation}. Procrastination can be reduced by student engagement in class and a positive learning environment \cite{katz2014ll}. Through student and instructor surveys, Martin and Bollieger found that instructors want to increase students' engagement, inside and outside class, because it motivates them to learn and improves their performance \cite{martin2018engagement}. More recently, Goodsett conducted a literature review about the best teaching practices in online learning and found that as online learning can be more self-paced compared to in-class delivery, instructors should consistently motivate students and engage them in active learning \cite{goodsett2020best}. Previous studies also show the importance of improving self-regulated learning strategies in flipped classrooms, and how it can be achieved with goal-setting skills and reflection prompts \cite{falkner2014identifying, schwarzenberg2020supporting, VANDENBOOM2004551}. Van den Boom et al. investigated the effects of reflection prompts in distance education and web-based learning environments and found that they encouraged students to take action on a particular task
\cite{VANDENBOOM2004551}. Similarly, Rodriguez et al. used self-reporting and self-testing practices (i.e. self-evaluation through practice problems) to improve study habits, which they found to be correlated with course grades \cite{rodriguez2018students}.
\textbf{Mitigating procrastination:} In online courses, reflection prompts that provide self-regulated learning scaffolds, like motivation and time management, have a human factor that encourages closer student-instructor interaction \cite{effects2020}. Song and Kim concluded that the human factor is helpful in self-regulated learning and course performance, encouraging students to participate in the course \cite{effects2020}. Azfal and Jami recommended that university instructors explore different possibilities to promote time management and self-regulation skills to increase student motivation for homework completion \cite{afzal2018prevalence}. Klassen et al. recommended that building students’ confidence is also an efficient way to address procrastination after assessing students' responses from the Motivated Strategies for Learning Questionnaire \cite{klassen2008academic}.
The benefits of addressing procrastination and suggesting different study strategies in online courses are also reflected in students' performance \cite{michinov2011procrastination, hartwig2012study}. Michinov et al. evaluated student time management and performance in online learning and provided empirical evidence of the negative relationship between procrastination and performance as students who are high procrastinators tend to engage less with the course material in general \cite{michinov2011procrastination}. Moreover, in a study conducted by Hartwig and Dunlosky in a large undergraduate course that evaluated students' study strategies, those who reported scheduling their work in advance had higher grade point average (GPA) in their overall university academic performance than students who had trouble managing deadlines \cite{hartwig2012study}.
\textbf{Reminder messages:} Researchers have deployed similar interventions to the one described in this paper, with the goal of reducing procrastination by sending email and text messages reminders. Edwards et al. sent students an email with their previous homework performance, which caused significantly reduced rates of late submissions and increased rates of early submissions \cite{edwards2015examining}. Humphry et al. focused on the format of the reminders that students received and used randomized A/B comparisons to analyze the effect of receiving an email reminder compared to a text message \cite{humphrey2019exploring}. The results of this study showed that text reminders had a positive impact on on-time assignment submissions and overall performance in the class \cite{humphrey2019exploring}. In a Massive Open Online Course (MOOC), Baker et al. prompted students to schedule a reminder nudge about their course videos to measure their engagement with the course material \cite{baker2016}. Although they found no impact on student attitudes in their settings, the authors suggested that these type of interventions would have success in a context where students are highly motivated by course credits \cite{baker2016}.
The research by Nikolayeva et al. is particularly relevant for the current study \cite{nikolayeva2020does}. They also conducted randomized control trials to test out four different versions of email reminders. Their study found that simple, non-personalized email reminders to complete optional online quiz exercises in a blended Physics course had a greater impact than longer, more personalized emails that specifically prompted students’ time management goals. While their study contributed to research on effective reminder emails for self-regulation with respect to time management, it stopped short of providing important information about the reasons for this effect. As noted by the researchers, they did not solicit students’ reactions to the email content - this would have provided greater insights into the quantitative results, especially, how the more sophisticated versions of reminders were perceived by students. Moreover, the email reminders targeted optional quizzes rather than required course tasks; as noted by the authors, it would be interesting to investigate how different email reminders would affect students’ completion of course tasks that were assessed, which is the goal of our current study. Our study also tests a broader set of A/B comparisons over multiple weeks to maximize the number of comparisons we make and to evaluate effectiveness at specific time points. This helps narrow down conditions which have the potential for further exploration. In addition, we make use of homework start time as a more direct measurement of the effect conditions have in reducing student procrastination.
Even though work has been done to investigate students’ opinions about their behaviour towards online courses including homework performance and procrastination \cite{zheng2016ask}, and evidence about their course performance after receiving reminders, to our knowledge, there is not much work that focuses on students' perceptions of those reminders, or aims evaluate the difference between instructor intuition and quantitative results.
Our past work \cite{zavaleta2021investigating}, which analyzed students' attempt and completion rate for homework during a single week of the four weeks of intervention described in this paper, provided initial evidence that reminder messages may have motivated some students who would otherwise not have attempted the homework to at least start. The current study aims to extend these findings over a longer time frame. Our study was guided by the following research questions: 1) What is the overall impact of reminder messages on homework performance? 2) What is the impact of reminder design on homework performance? 3) What are students’ perceptions of reminder messages? 4) How does instructor intuition of the reminder messages differ from observed results?
\section{Conclusion}
This paper leverages time-stamped homework submissions, qualitative data from nearly 1000 students and two course instructors, and randomized A/B comparisons to assess the effectiveness of reminder messages in getting students to start their online homework early. We find evidence, over four weeks, to suggest that reminder messages have a modest effect in promoting more students to start their homework, but are not as observably effective in altering the time at which they start the homework. We also found suggestive evidence of effective email design conditions which could be further explored. Combining instructor predictions about email design variations, we found although instructor intuition is important, they might not always be able to confidently quantify this intuition, and using experimental A/B comparisons can help them test their ideas. Furthermore, from surveying and interviewing students we discover some negative aspects of the reminders such that students may find the email motivating but are unsure on how to make progress, some students can become further stressed being told about upcoming deadlines, and some students may become annoyed if they have already made significant progress on the homework. These trends in student perspective align with our experimental findings that reminder messages are helpful to a subset of students that are encouraged to attempt and start the homework due to the reminder, but they also provide insight into how to improve our reminders.
\section{Acknowledgements}
This work was partially supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) (No. RGPIN-2019-06968), as well as by the Office of Naval Research (ONR) (No. N00014-21-1-2576)
\section{Design of a Reminder Message to Encourage Online Homework Participation}
\section{Methods}
To test the reminder messages, we use A/B comparisons which have been used in similar experiments \cite{humphrey2019exploring, necamp2019beyond, yeomans2017planning, o2018can}, to find the most effective option from different factors \cite{williams2018enhancing}. A/B comparisons, also known as randomized control trials (RCT), are a common practice used in different areas to find the most effective option from different variations \cite{williams2018enhancing}. In this paper, we show how A/B comparisons for emails reminders are a low-cost and effective strategy to test multiple messages alternatives and find the one that works better for students, without the use of external tools or resources.
\subsection{Course Setting}
We deployed the reminder messages interventions during the second semester of the academic year in an introduction to Computer Science course (CS1) in a research-focused North American university. Our goal was to help students start earlier on the weekly graded homework. These multiple-choice and open-ended programming questions about Python were completed before the lecture (worth 0.5\% of their total marks and due on Mondays at 10 am) and after the lecture (worth 2\% of their total marks and due on Fridays at 5 pm), as part of a flipped/inverted classroom course. Each set of questions was released three to four days before the deadline and was due on the same day and time every week. The homework problems were submitted on the course's online homework platform, which provides immediate feedback to students. Students are allowed unlimited attempts for each question, with the best attempt being scored.
The reminder messages were deployed using the university’s Learning Management System, Canvas, which allows students to receive these reminders through its inbox and, by default, to receive an email notification. There were 1018 students enrolled in the course at the time of the interventions, 946 of whom provided explicit consent via a Qualtrics form and were included in the analysis.
\begin{figure*}[htb]
\centering
\includegraphics[scale=0.45]{sections/figures/email-template.png}
\caption{Example of reminder email sent during Week 5. Two variations were sent that week: sender and information content.}
\vspace*{-1.5em}
\label{fig:email-template}
\end{figure*}
\subsection{Data Collection}
Every week we randomized participants to the treatment group (receiving a reminder message through the Learning Management System with an email notification) or to the control group (a no-reminder message condition). This was double-blind in that neither instructors nor students knew which condition they were assigned to. We report on a sequence of interventions from weeks 5, 6, 7, and 8 in a 12-week course. The interventions started in week 5 as it was a point in the middle of the semester when students begin to fall behind as the course content becomes more challenging.
We collected time-stamped homework submissions data from the online homework platform. We then compared metrics constructed from this data across the control and treatment groups. To illustrate our findings in students’ homework participation, we discuss the four weeks of interventions to show students’ responses to the emails throughout the term. The weeks 5 and 8 deployments targeted the pre-lecture homework assignment, while the weeks 6 and 7 deployments were reminders for the post-lecture homework assignment. The average send time of the reminder messages was 84, 51, 40, and 47 hours before the deadline respectively for weeks 5, 6, 7, and 8. The week 5 and week 8 homework was posted 96 hours before the deadline, week 6 and 7 homework was 149 hours before the deadline.
Our goal was to measure the impact of reminder messages on student behaviour. Accordingly, we measured behavior changes across the following outcomes: 1) the number of hours before a deadline that students started homework (start time), and 2) homework attempt rate. The reminder messages intended to encourage students to start homework earlier by getting them to set a plan or encouraging them to begin. However, one could also imagine these reminder messages would not change students' behaviors as students would not find them useful or would simply ignore the messages.
We investigate this question by looking at the impact on students' starting time relative to the deadline, as well as exploring whether the emails might increase the tendency of students to even start the homework. The data for these analyses were collected directly from the online homework system and included the time at which the first problem was attempted (start time), the proportion who attempted at least one problem (attempt rate), and homework scores.
To understand students' perceptions and experience of receiving these messages, and what students found helpful or not about the reminders, we also conducted a post-intervention survey where we asked students four 7-point Likert scale questions to rate the helpfulness of the reminder messages. We also asked students to answer three open-ended questions about their feelings and the actions they took when receiving messages such as what they liked or did not liked about the reminders. The open-ended responses were manually coded by three researchers using thematic analysis to capture the insights about the intervention. While 935 students provided consent and responded to the Likert-scale questions, only 474 students answered the open-ended questions. We complemented this large collection of written responses with semi-structured interviews of a smaller sample of students (N=9) that probed similar questions in more depth. This sample of students was recruited by announcing the call for interviews after the course was completed, to only students who expressed interest in being contacted for a follow-up during the survey.
To evaluate educator perspectives on the A/B designs, we surveyed two computer science instructors who have been regularly teaching and coordinating introductory courses for the past 10 years to predict the effectiveness of the reminder emails. Instructors were not shown any data or analysis. We separately presented them with the two conditions for each variation as described in Section 3.3 below and asked which would be better to send in terms of changing student behaviour in terms of (a) getting students to start earlier, and (b) increasing the rate of students doing the homework after receiving the email. We also tasked instructors with giving a quantitative prediction of attention rate (percentage of students paying attention to the emails) and action rate (percentage of students who as a result of the message start their homework early) changes between conditions after students receive the email. Instructors were finally asked to rate their confidence in their predictions.
\begin{table*}
\begin{tabular}{l|l|cccc}
\hline
\multicolumn{2}{c}{Condition} & \multicolumn{1}{l}{W5} & \multicolumn{1}{l}{W6} & \multicolumn{1}{l}{W7} & \multicolumn{1}{l}{W8} \\ \hline
\multirow{2}{*}{Sender} & Instructor & \multirow{2}{*}{X} & \multirow{2}{*}{} & \multirow{2}{*}{} & \multirow{2}{*}{} \\ \cline{2-2}
& Teaching Assistant & & & & \\ \hline
\multirow{3}{*}{Information} & No prompt & \multirow{3}{*}{X} & \multirow{3}{*}{} & \multirow{3}{*}{} & \multirow{3}{*}{} \\ \cline{2-2}
& \multirow{2}{*}{\begin{tabular}[c]{@{}l@{}}"We have found that students who start earlier tend to learn more,\\ since if they get stuck, they have time to figure it out."\end{tabular}} & & & & \\
& & & & & \\ \hline
\multirow{2}{*}{Subject Line} & "When did you plan to do PCRS Perform? Can you start earlier?" & \multirow{2}{*}{} & \multirow{2}{*}{X} & \multirow{2}{*}{} & \multirow{2}{*}{} \\ \cline{2-2}
& "Remember to start early on PCRS Perform and finish before Fri at 5 pm" & & & & \\ \hline
\multirow{2}{*}{Recommendation} & No prompt & \multirow{2}{*}{} & \multirow{2}{*}{} & \multirow{2}{*}{X} & \multirow{2}{*}{} \\ \cline{2-2}
& \begin{tabular}[c]{@{}l@{}}"One tip for starting is to just spend 5 minutes on it. This can help you\\ overcome the procrastination barrier, and make it less likely you forget. "\end{tabular} & & & & \\ \hline
\multirow{4}{*}{Advice} & No prompt & \multirow{4}{*}{} & \multirow{4}{*}{} & \multirow{4}{*}{} & \multirow{4}{*}{X} \\ \cline{2-2}
& \multirow{3}{*}{\begin{tabular}[c]{@{}l@{}}"Take a moment to think about the following prompt: What advice would\\ you give to other students to try to start earlier? Take 30 seconds to just\\ say out loud to yourself what you would tell them"\end{tabular}} & & & & \\
& & & & & \\
\\
\end{tabular}%
\caption{Details of reminder email conditions for each variation. X indicates the week we tested the variation specified in each row.}
\label{tab:design_factors}
\end{table*}
\subsection{Reminder Message Design}
We designed the reminder messages using a broad set of models from work on the implementation intentions theory \cite{gollwitzer1999implementation} and with an emphasis on self-efficacy \cite{klassen2008academic}. Implementation intentions involve explicitly stating a goal that supports the transformation of good intentions into concrete behaviour change \cite{gollwitzer1999implementation}. Implementation intentions have been effective in self-regulatory strategies to bridge the gap between intentions and actions \cite{gollwitzer2006implementation, falkner2014identifying}, in improving homework performance by predicting grades \cite{sommer2012influences}, and habitual behaviour \cite{arbuthnott2009education}. The reminder messages used for these interventions encourage students to set a goal for when they will work on their homework and including students' past experiences on starting early and what they found useful. The reminders also contained a link to the homework problems. We created different versions of reminder messages which are similar in design with variations such as sender, subject line, and content (information, recommendation, and advice) presented to the students [Table \ref{tab:design_factors}]. We performed two manipulations in the first round of deployments, as we originally intended to explore the design space and potential impact of the emails. We tested one variation within the email content (information) and one external variation that involved the delivery method (sender). For the remaining deployments, we analyzed one new variation per email to focus on the prompts and nudges. An example of a full reminder email can be found in Figure \ref{fig:email-template}.
\subsubsection{Sender}
For Week 5, we randomized the sender of the reminder message. Students either received the reminder message from one of the course instructors or one of the teaching assistants [Table \ref{tab:design_factors}]. Moreover, the course instructor's learning management system's account had a profile picture associated with it for easy recognition, whereas the teaching assistant did not have a profile picture. We decided to randomize the sender to better understand the impact of authority on student behaviour and motivation. Receiving a message directly from an instructor can make students feel that they care about their learning and feel encouraged to engage with the course material \cite{goetz2021getting}. On the other hand, receiving a message from a teaching assistant to who students might be able to relate more, as they are students as well and could take them as peer role models, might make students feel the advice is more genuine \cite{clarke2018near}.
\subsubsection{Information}
For week 5, we also randomized if students would receive information about the impact of starting homework early [Table \ref{tab:design_factors}]. The rush and urgency of doing homework at the last minute can cause poor performance and harm long-term retention \cite{you2015examining}. Presenting students a piece of information stating why starting homework early can help them may encourage them to do so. However, it is also possible that the precise reason given may not resonate with the student, leading the information to reduce the attention that they give to the reminder.
\subsubsection{Subject Line}
For week 6, we randomized the type of subject line of the reminder message [Table \ref{tab:design_factors}]. We are aware that not all students will open and read the reminder message, and therefore, we want to get students to take action even if they only pay attention to the subject line. We decided to randomize between two types of subject lines: a reminder deadline summary and a prompt to plan their homework. Showing students the deadline their homework and reminding them to start early can give them a sense of urgency that the due date is coming up and they should get to do the work. However, giving students a prompt to plan when they will do their homework and asking them if they can start earlier may allow students to think about the actions they need to take to achieve this goal \cite{rogers2015beyond}. Prompting people to make concrete plans increases the chances they will act on them and complete their tasks \cite{rogers2015beyond}.
\subsubsection{Recommendation}
For week 7, we randomized if students would receive a recommendation on how to start the homework early [Table \ref{tab:design_factors}]. Even though it might sound trivial how to start the homework, one of the hardest steps to complete a difficult task is starting, or even finding the motivation to start. With this tip, we recommend students to spend a few minutes looking at the homework to help them visualize what they need to complete. Moreover, following the implementation intentions theory, we present to students a concrete action that they can take in order to achieve the goal of starting early \cite{gollwitzer1999implementation}.
\subsubsection{Advice}
For week 8, we randomized whether students would receive a prompt asking them to reflect on the advice they would give to another student if they are trying to start their homework early [Table \ref{tab:design_factors}]. It may be easier for students to give recommendations to peers, and it can be difficult to recognize who is in the same spot as you \cite{kross2012boosting}. With this prompt, we want students to think about their classmates and what can help them, and also to realize that other students might be facing similar struggles. This prompt encourages collaboration and the social aspect of learning \cite{laal2012benefits}.
\subsection{Student Assignment}
We randomized whether or not students received reminder messages, to investigate the impact on their homework start times, as well as whether or not they started the homework. We also randomized which version of reminder email students in the treatment group received, as we tested the different variations of our factorial design.
To evaluate the impact of the reminder messages, we conducted randomized A/B comparisons where only a portion of the class received the messages in any given week. For the first week, 2/3 of the students were assigned to the treatment group and received a reminder message; for the following three weeks the treatment group was re-randomized and extended to 3/4 of students so that more students could benefit from the reminder message after analyzing the first week's results. However, due to technical issues, 1/4 of the treatment group in week 7 did not receive a reminder message. The randomization of the control and treatment groups was done via Python's random shuffle function.
\subsection{Qualitative Results}
\subsubsection{Students' Perceptions of Reminder Messages}
Of the 935 students who responded to the Likert-scale survey questions, most students found it motivating to read the evidence about the benefits of starting the homework early (63\%) and instructor recommendations on how to start early (63\%). Moreover, the majority of students seemed to appreciate that the message included a link to the homework (56\%). We captured the reasons behind the students’ attitudes towards the reminder messages from an open-ended survey question asking about their initial reaction and if it affected their behaviours. When asked the question, \textit{``After receiving the message, did it motivate you to set aside time to work on [homework], or did it change your behaviour at all?''}, 474 students provided an answer, with most of them indicating a positive behaviour change after receiving the reminder (57\%). About half of students who indicated that the messages were motivating (27\% of the total) did not provide a reason for finding the messages motivational. Of those who responded, the majority (20\% of the total) shared that the messages helped them reduce procrastination by starting early and making a plan: \textit{``The message made me consider my time ahead, and plan when to do the [homework] questions.''}
Of the students who indicated that the reminder messages did not influence a behaviour change, most (28\% of the total) did not explain why. The majority of those who provided a reason noted that they already had a plan (7.5\%) or had already completed the work (2.9\%). For those students, the messages could be frustrating: \textit{``I was a little annoyed because I have already started doing [homework] for that week''} and \textit{``it becomes annoying ... I know I will do it before the deadline.''}
For other students, the message was appreciated and could potentially lead to helping them in some contexts (13.6\% of the total), but they noted that other factors got in the way of inducing a change: \textit{``Yes, but when I saw it I was busy and later on forgot about it.''}
\textbf{Students' Emotional or Attitudinal Responses to Emails}.
Whether or not the emails were effective, they often produced an emotional response, which suggests that emails need to be targeted carefully. Students who had not yet started sometimes felt that they were being cared for: \textit{``I felt like the [course code] team actually cares about me and my learning that it surprised me.''} (1.4\% of the total), but if they were unable to start or were struggling in the course, the extra attention could be stressful: \textit{``if anything it kind of stressed me out knowing I have something due and I haven't done it yet''} (6\% of the total). Negative reactions were more common from students who had already completed the work (4\% of the total). The messages caused them to question whether their work had been submitted correctly: \textit{``It made me panic not thinking I'd turned in [the] assignments.''}
We observed similar patterns in the interviews. Some students (6/9) reported in the interviews that they liked receiving the reminder messages as they were helpful for spurring homework completion. One student shared: \textit{``the instructor would send us emails ... and those were so helpful for me because I procrastinate a lot, so getting reminders by one of my favorite professors was actually so much fun for me.''}. The remaining 3/9 students had neutral opinions about the reminders. A few students (2/9) mentioned that having a notification on their phones was a positive aspect of the reminders, especially if the due date was in the header of the message.
Students (5/9) mentioned in the interviews that they appreciated the messages that included a recommendation about how to start early as they agreed that it motivated them to make a study plan or to consider starting early, which is the goal of the implementation intentions principle. One student highlighted that they were able to apply the recommended actions to other courses: \textit{``That piece of advice in that email actually helped me to like go to other classes, and like, read my other assignments before starting them. So that was really helpful for me and I learned from that because now I've actually like apply that for my other classes''}.
Students (2/9) said that the reminder messages gave them the feeling of being valued and cared for by the course team: \textit{``I planned everything beforehand. So it was really nice to receive this because most courses don't even tell you at all. There's no prompt, there's nothing. ... I think there's a lot more intention and care.''}. One student shared their reaction to the first email they received from the course instructor: \textit{``It seems like [the instructor] genuinely wanted to like do his best to help you learn and be there for you make sure that you know that there are resources that you can use and that you can reach out to him. He was a very caring professor and it seemed that this was his way of helping us. When I received the first email, it made me happy because you don't usually get those from professors''}.
\textbf{Student perceptions of the negative effects of reminder messages}.
Some comments from students' surveys and interviews about the reminder emails highlighted design aspects that should be considered for future work. For instance, students mentioned that the reminder message looked like a spam email.
Including the students' names brings up a sense of personalization that is appreciated by some students, but it might cause confusion and the feeling of getting targeted: \textit{``When I saw my name ... I panicked thinking I did something wrong. Then I realized it was sent to everyone.''} Several students commented that they replied to the reminder message providing their plan to start early, but they did not get a response which caused discomfort: \textit{``It stressed me out beyond belief! Also, nobody replied to them?''}. Students were prompted to reply to the email to motivate them to answer.
During the interviews, some students (3/9) mentioned that they consistently received the reminder messages after completing their homework, and felt patronized by the reminder. One interviewee said: \textit{``I always got [the email] after I finished everything...it was kind of like annoying. Cause I’m already done, why are you still nagging me about this?''} However, all students in the interviews who reported receiving the reminders after completing their homework mentioned that they categorize the reminders as beneficial for students who struggle with time management when asked about their overall opinions of the emails.
\begin{table*}[htbp]\centering
\begin{tabular}{l|c c c c c}
&\multicolumn{1}{c}{(1)}&\multicolumn{1}{c}{(2)}&\multicolumn{1}{c}{(3)}&\multicolumn{1}{c}{(4)}&\multicolumn{1}{c}{(5)}\\
& Email & Better condition & Email Attention & Student Action & Instructor Confidence\\
& Variation & Prediction & Rate Difference & Rate Difference & (-5 to +5)\\
\hline
Instructor 1& Control vs. Email & Email & - & 10\% & +2 \\
Instructor 2& (Week 5-8) & Email & - & 10\% & +4 \\
\hline
Instructor 1& Sender & Instructor & 5\% & 10\% & +2 \\
Instructor 2& (Week 5) & Instructor & unsure & 5\% & +1 \\
\hline
Instructor 1& Information & Include information & 0\% & 5\% & 0 \\
Instructor 2& (Week 5) & Include information & 10\% & 5\% & +2 \\
\hline
Instructor 1& Subject Line & Deadline summary & 0\% & 20\% & +2 \\
Instructor 2& (Week 6) & Deadline summary & 10\% & 5\% & +2 \\
\hline
Instructor 1& Recommendation & No recommendation & 0\% & 5\% & -2 \\
Instructor 2& (Week 7) & No recommendation & 10\% & 5\% & 0 \\
\hline
Instructor 1& Advice & Include advice & 0\% & 5\% & -1 \\
Instructor 2& (Week 8) & Include advice & 10\% & 5\% & +2 \\
\end{tabular}
\caption{Summary of instructor survey responses. Email Variation represents the email design factor the instructors considered. Better condition Prediction is the condition that each instructor predicted was best for getting students to start earlier and take action. Email Attention Rate Difference represents each instructor's prediction of student engagement increase given better condition. Student Action Rate Difference is each instructor's prediction of student action increase (doing their weekly homework) for the better condition. Instructor Confidence is how confident they were in their predictions, where -5 is not confident at all and +5 is extremely confident. Refer to Section 3.3 for the design of each condition.}
\vspace*{-2em}
\label{tab:instructor}
\end{table*}
\subsubsection{Instructor Predictions on Reminder Messages}
We summarized the data collected from the two instructor surveys in Table \ref{tab:instructor}. These surveys help visualize the differences between the effect observed and the predictions. For all variations, both instructors made the same prediction for which condition would be better. Instructors predicted a $\leq$10\% increase for all predictions, with Instructor 2 being unsure of how sender would affect engagement. Instructors also generally predicted a $\leq$10\% increase in students taking action to do weekly homework given the better condition, the only exception being Instructor 1, who emphasized that including the deadline in a subject line would have a large (20\%) impact. Instructor 1 was moderately confident when predicting variations outside the email body: Control vs. Treatment (+2), Sender(+2), Subject Line (+2), while less confident when predicting conditions in the email body: Information (0), Recommendation (-2), Advice (-1). Instructor 2 was moderately confident in their predictions (+2 or +4) for all variations, except for Sender (+1/expressed uncertainty) and Recommendation (0).
\subsubsection{Instructor Intuition Relative to Quantitative Results}
Based on the quantitative results, instructors can generally correctly predict the better condition for students to start early (corresponding to start time Figure \ref{fig:box-start}) and take action (corresponding to attempt rate Table \ref{table: attempt_rate}) and aggregated overall in Table \ref{tab:impact_rate}. However, instructors overestimated the effectiveness of some conditions. For example, the overall effectiveness of emails in Table \ref{tab:impact_rate} was <5\% for attempt rate (statistically significant) and <2 hour increase for start time (non-statistically significant), but both instructors predicted a 10\% action ate difference. We again note instructors are also not wholly confident in all their predictions, even if they did consistently pick the better condition.
\section{Limitations \& Future Work}
Students who found the reminder messages not helpful or motivating still received the messages, which can cause a negative effect as they can feel annoyed and stressed. To help students who find the reminder messages helpful, but at the same time, consider the students who do not find value in receiving these, we would like to use more adaptive and personalized messages in future deployments. If there were a closer integration between the homework system and the learning management system, we could provide more customization and not send out a message if the student has already started the homework. Moreover, obtaining more information about students’ prior knowledge and motivations may provide possible targets for personalization, which may influence their response to the interventions. For the students who plan their week ahead but would like to receive reminders, we could provide an option for changing the reminder messages’ time and frequency. Students who found that the messages did not fit their schedule could schedule one or multiple reminders based on when they wish to receive them, as these messages contain more than just reminders but also prompts that can motivate them to take action.
Students were aware of the format of the reminder messages after a certain number of occurrences, which could have led them to stop opening and reading these emails. If we chose to add more personalized information to the emails, such as current marks and performance rates \cite{edwards2015examining}, we could make the messages more novel and varied, to reduce the perception that the messages are formulaic and to attract readers. The question becomes if the reminder function is sufficient for students, and having a message attached to it might encourage students to pay attention to it.
We could not directly measure student engagement with the emails at the time of sending such as if and when they opened the emails, though we could get parts of this data through later interviews. Student engagement is important in making sure they get the full benefit of the reminder emails, while also ensuring the effect between design decision and student action are accurately inferred.
Future evaluation that could build on our work includes: (1) Exploring how these findings generalize to contexts beyond CS1; (2) Exploring in more depth the characteristics of students who appreciated the reminders and those who had negative opinions as the qualitative responses were mixed; (3) Investigating improvements on these reminder messages, such as alternative ways of getting students to start earlier, and emphasizing the importance of students planning how to complete their homework after starting; (4) Introducing additional measures of student engagement with the emails; (5) Furthering evidence through A/B comparisons on effective conditions independent of the week they were sent. Although sending emails to students is a ubiquitous everyday activity, it is important for instructors and researchers to understand how to use widely used communication channels to motivate and reach students, especially during a pandemic, and which of these observable measures of behaviour are effective.
\section{Discussion}
We explored alternative designs for using emails to prompt students to plan and stay on track with their homework. We show how we can use randomized A/B comparisons to test which ideas might impact behaviour. We present the particular dependent variables and metrics anyone can use to conduct such studies - start time and attempt rate. We also explore the design space of factors to test, which we hope can inspire a wide range of follow-up studies using a similar paradigm. Our results illustrate some of the tradeoffs instructors might consider, as well as empirical insights into what may be more or less effective, to guide the necessary future studies needed for expanding and generalizing this work to other contexts. The randomized A/B comparisons investigated whether and how different reflection/reminder messages might motivate students to complete their homework and start earlier. Procrastination and homework performance are key issues that instructors in online learning contexts need to address, and we provide a feasible intervention.
Some instructors were skeptical that students would change their behavior from a simple message – or even pay attention to them. However, identifying reminder messages that \textit{can} shift behavior is extremely valuable because it is a low-cost and widely applicable intervention, and one of the major ways instructors are reaching out to students outside of class time. Interestingly, we found that the emails prompting students to reflect on how they could start earlier did not increase how many hours before the due date students began. However, it did increase the number of students who actually started the homework (as many students in a week never started the homework) and, as a result, increased the average performance on the homework.
We used psychologically-grounded principles to create the emails in a way that they are helpful and encouraging to students. One of the psychological theories that we used to design these reminder messages and make them more motivating is the implementation intentions strategy \cite{gollwitzer1999implementation, klassen2008academic}, where students were asked to reflect on a plan with information on when and how to improve to transform good intentions into behavioural change. Furthermore, we also used practices motivated by self-efficacy and sharing mastery experiences with students.
We tried to understand students' experiences and perspectives on receiving the emails. We collected a total of 935 survey responses, and combined this with interviews with 9 students, to get converging perspectives of large-scale data and in-depth discussions. The overall positive impact of the reminder emails on students’ attempts of homework aligns with other research on email interventions \cite{lim2021impact, nikolayeva2020does, yeomans2017planning}. However, it is interesting that the intervention did not significantly lead students to start early. A possible explanation for this can be gleaned from the qualitative response of students - at least one student expressed that they did not receive feedback on their plan to start the homework early, with a suggestion that it affected their motivation. Possibly, a more dialogic or coaching approach might further encourage students to work out their plan, but this needs to be investigated.
For students who were less impacted by the intervention, i.e., did not attempt the homework even after reminders, it is possible that they believed that they were already self-regulating effectively, or did not feel confident about doing the homework. It may also be the case that these students are facing some issues that might need to be addressed in a different way, other than through emails. More information is needed about students who are not engaging or responding to the interventions. Future research can investigate students’ baseline self-regulated learning habits or self-efficacy, to obtain further insights into how email interventions could be effectively tailored to them.
The qualitative data collected from students sheds light on when and why such reminder messages can be seen as valuable by students versus annoying or excessive. The majority of students said they found the message helpful, and some felt that the intervention showed that the instructional team cared about their learning. Despite this, our quantitative metrics suggested only a fraction of those actually changed their behaviour, suggesting how future work can go deeper into understanding how exactly to craft the prompts to tackle the mechanisms of behaviour change. The qualitative finding around students’ appreciation of instructor care as demonstrated through the emails resonates with other research such as \cite{lim2021impact, arthars2019empowering}, showing the affordances of technology-mediated communication can foster a feeling of greater support in students. On the other hand, some students were stressed by a reminder about doing homework, and a few expressed irritation because they already had a plan for completing the homework or were already done.
In terms of how useful the results were in testing predictions from instructors, we found that though instructors were generally able to predict the best condition to send in an email (though not always confidently), they would often overestimate the effectiveness of the emails when translated into student attempt rate and start times. This reinforces the notion that involving instructor intuition in the design process of reminder emails would allow for better conditions to be picked \cite{motz2018causal}. At the same time, this demonstrates the need for data from real-life A/B comparison experiments to help instructors quantify their intuitions and better understand the effectiveness of their communications.
We found suggestive evidence some design choices motivated students to start earlier compared to other conditions during certain weeks such as emails from instructors, including information about starting early, and not including recommendations or advice prompts. Similarly, we also found evidence which suggests some design choices can help significantly increase student attempt rate such as including an advice prompt. In addition, students receiving certain conditions would significantly alter their start time distributions, indicating our design choices had an impact even if not directly summarized by the mean or median. Finally, on an aggregate level over the period of 4 weeks, it was more beneficial for students to receive a reminder email than not in terms of doing their homework and starting it early.
These results can guide instructors in understanding students' context and decision making, In the time of COVID-19, with channels like emails, instructors should learn how to better reach students, and about how to balance the benefit to some students with the inconvenience to others.
|
\section{Introduction}\label{sec:intro}
The Domain Name System (DNS) is an Internet protocol that maps human-readable
domain names to Internet Protocol (IP) addresses. Conventionally, DNS queries
have been unencrypted, leaving both queries and responses vulnerable to
passive eavesdropping and active manipulation. In recent years, major browser
and operating system vendors have begun to deploy protocols such as
DNS-over-TLS (DoT) and DNS-over-HTTPS (DoH), which encrypt queries and DNS.
With the emergence of this functionality, browsers now have different default
settings for encrypted DNS, as well as various configuration options to allow
users to customize how their browser performs DNS lookups. As is the case
with browser vendors, the ``Private DNS'' mode on Android uses DNS-over-TLS
(DoT) to encrypt DNS queries and responses in between the client and the DNS
resolver. Encrypted DNS settings are implemented in subtly different ways,
and the settings have a variety of defaults.
These differences lead to the setting behaving in ways with implications they may not understand. The obscurity of these settings may be further
exacerbated by the lack of information provided to users when they modify
these settings. Nevertheless, the implications of these settings are
significant: depending on how users configure encrypted DNS, for example, all
of a user's DNS queries may be sent to a single DNS provider, such as Google
or Cloudflare. Occasionally, such configurations are even changed without a
user's awareness, with browser vendors pushing changes to default settings
with browser version and operating system upgrades. Although users can
typically change these settings, doing so requires an awareness that
configuring encrypted DNS settings is possible, knowledge about different
configuration choices, and the ability to change these settings.
It is important to give careful consideration to default settings and interfaces in light of the fact that they often go unmodified by users~\cite{weinmann2016digital}.
Furthermore, previous research has observed that encrypted DNS is what Clark {\em et al.} have described as an Internet ``tussle space''~\cite{clark2002tussle}, where various Internet stakeholders may vie for control through technical protocol design.
Indeed, Hounsel et al. have identified encrypted DNS as a tussle
space~\cite{hounsel2021designing}, and highlighted the importance of designing the encrypted DNS infrastructure and user interfaces in ways that {\em preserve
user choice}.
Clark explains the profound importance of maintaining user choice in Internet
protocols as follows, ``It is important that protocols be designed in such a
way that all the parties to an interaction have the ability to express preference about which
other parties they interact with...it matters if the consequence of choice is
visible.''~\cite{clark2002tussle}.
The design choices in the encrypted DNS ecosystem have implications for market consolidation and power, corporate visibility into and control over user data, as well as user privacy on the Internet.
Given the significant implications of these settings, and users' ability to change them, it is critical to understand how users understand these settings and interact with them, which is inherently dependent on how they are presented by vendors in terms of options to match their preferences.
In this paper, we study encrypted DNS settings in Brave, Chrome, Edge,
Firefox, Opera, and the Android mobile operating system.
Previous small-scale (approximately ten users) research~\cite{nisenoff-21-PrivateDNS} included a preliminary study of these questions involving Android's Private DNS setting. In this study, we expand on these results to give
a larger scale look at how users interact with these settings in the context
of browsers.
We study the following questions:
\begin{itemize}
\itemsep=0pt
\item Are users aware of encrypted DNS settings in browsers and mobile
devices? (Section~\ref{sec:awareness})
\item What encrypted DNS settings do users have enabled in their browsers
and phones? (Section~\ref{sec:settings})
\item When shown encrypted DNS settings for different browsers, which settings do
users choose, and why? (Section~\ref{sec:selection})
\item When the technical aspects of these systems are explained to users,
do their preferences for settings change?
(Section~\ref{sec:technical})
\end{itemize}
\noindent
Many users have heard of DNS but do not know what it does. Even users who
believed they knew about DNS's function were often incorrect in describing it.
Users typically had the default settings enabled in their own browsers. When
users deviated from the default, they tended to select either Cloudflare or
Google as their recursive resolver or simply disabled the setting. In many
cases, the default setting opportunistically encrypts their DNS queries. In
mobile devices, a surprising number of users had the ``Private DNS'' setting
disabled, despite that not being the default setting.
More than 70\% of participants continued to use the default encrypted DNS
settings, and this tendency varied across different interfaces. Moreover, the
way the settings were described played a role in users' decision making. 37\%
of participants chose to change their encrypted DNS settings after receiving
an explanation of DNS and encrypted DNS. Although the defaults remained
popular, users did select a variety of settings after receiving more information.
The setting to enter a custom resolver was not often selected, but when
participants did attempt to specify a custom resolver, none of the participants
entered text that would have functioned properly. Participants seemed to
understand the settings better after an explanation, yet they still had
problems understanding the differences in functionality and privacy guarantees
among the different resolvers.
These results lead to a number of practical recommendations for designers of encrypted DNS
interfaces, standards bodies, and other policymakers, which we detail in
Section~\ref{sec:disc}. In particular, users need a basic primer on DNS,
its associated privacy risks, the guarantees that encrypted DNS can (and
cannot) provide, and an intuitive way to understand the implications of the
choices for different recursive resolvers. Users also need easier ways
to customize the choices for trusted recursive resolvers, once the
implications of these choices are clear. Future work can and should expand on
our initial findings to explore interfaces for configuring encrypted DNS that
provide users choices for configuration, as well as a clear
understanding of the implications of these choices.
\section{Background and Related Work}\label{sec:back}
This section surveys background and related work on encrypted DNS,
including the basic operation and privacy risks of DNS queries, the
functionality of encrypted DNS, and the history of the introduction of
encrypted DNS into modern browsers and operating systems. We also survey past
related work concerning users' awareness concerning privacy settings in
browsers, including encrypted DNS.
\subsection{Background: DNS and Encrypted DNS}
Most Internet connections are preceded by a Domain Name System (DNS) lookup,
which maps a human-readable name to an Internet protocol (IP) address that a
client can use to ultimately connect to a remote destination or service. DNS
queries and responses have historically been unencrypted, and they may contain
sensitive information including information about the site that a user is
visiting. Previous research has shown that observing a user's DNS queries can
allow users to be tracked across multiple
websites~\cite{bortzmeyer2015dns,Moura2020Centralized,herrmann2010analyzing}.
Beyond tracking users online, DNS traffic can also be used to infer what
``smart'' Internet of things (IoT) devices are present in a home, and may even
expose information on how people use
them~\cite{apthorpe2017EncIoTTraffic,apthorpe2017smart,le2019policy}. Some
Internet service providers have logged DNS queries to their resolvers and
shared them with third parties~\cite{DNSprivProb}. Because these queries are
typically made in plaintext, anyone who can observe a user's network traffic
could see the contents of these queries. Outside of the privacy implications
of the DNS, there are also potential concerns with integrity, where the
responses to DNS queries may be modified resulting in users receiving the
incorrect IP for a website; a censor may be able to implement this type of
manipulation to block access to a legitimate website or redirect users to
another website entirely~\cite{pearce2017global,bortzmeyer2015dns}.
To mitigate some of the security and privacy issues with DNS, Internet
operators and vendors have introduced DoT and DoH, both of which send DNS
queries using encrypted protocols, with subtle differences in their
implementations~\cite{hu2016DoTSpec, hoffman2018DoHSpec}. Encrypting DNS
queries and responses prevents passive eavesdroppers from observing the
content of users' DNS queries. DoT sends queries over a Transport Layer
Security (TLS) connection using port 853, while DoH uses HTTPS rather than TLS
for the transport protocol, typically on port 443. Because DoT uses a
dedicated port, it is easier to monitor and detect (and potentially block); on
the other hand, DoH uses port 443 for transport and thus DoH traffic tends to
be more difficult to identify, particularly when combined with other HTTPS
traffic~\cite{HvT}. Although both of these protocols encrypt DNS queries and
responses, they are still susceptible to various
attacks, ranging from downgrade attacks to traffic analysis-based inference
attacks~\cite{unpaddedDoH,Siby2020EncDNSTraffic,Huang2020DoHDowngrade,Houser2019DoTLeakage,vekshin2020dohML}.
Using DoH and DoT can provide users with many benefits, but there are
downsides to using these protocols as well. Many existing systems such as
parental filtering, safe search, or malware detection tools rely on access to
the content of DNS queries. However, encrypting those queries while using DoH
or DoT can sometimes inhibit the functioning of these
systems~\cite{DNSprivProb,ff1,lemos2013got, Hounsel2021Tussle,
Lyu2022SurveyEncDNS}. Similarly, because ISPs can be required by government
to block access to certain illegal content, implementation of encrypted DNS
may in some cases prevent ISPs from complying with these laws~\cite{hvt2}. On
the privacy front, because DoH and DoT need to communicate with resolvers that
support these protocols, encrypted DNS may result in users' queries being sent
to fewer resolvers, allowing them to see more of a user's online behavior, given that fewer entities can observe more of a user's DNS queries~\cite{borgolte2019ecosystem,
Hounsel2021Tussle}. This consolidation is evident in the growing consolidation of
hosting authoritative DNS resolution in
general~\cite{Wang2021Consolidation}.
Although DoH and DoT provide security for the queries while they are in transit,
these protocols do not prevent the operators of these DNS resolvers from learning about users'
queries or ensuring that the DNS resolver returns the correct response. Other
proposed improvements to DNS designed to address these issues include:
Oblivious DNS~\cite{ODNS}, Oblivious DNS over HTTPS~\cite{ODoH},
DNSSEC~\cite{eastlake1999DNSSEC,arends2005protocol}, QNAME
minimization~\cite{bortzmeyer2016QNAMEMin}, and dividing queries across
multiple resolvers~\cite{kresolver2020, Hounsel2021DistributingQueries}.
\subsection{User Awareness of Encryption and Privacy Settings}
This study focuses on user awareness concerning encrypted DNS and its
configuration through common interfaces, but many other studies have
investigated how to communicate security and privacy concepts to users. While
usually less buried in setting menus, private browsing modes are included in
browsers and do not store browsing history, cookies, or temporary files across
sessions. Research has shown that users have many misconceptions about what
these setting actually do. Much like encrypted DNS settings, each browser
provides a unique description of each setting, which both play a role in what
protections the users think the setting provides and have been shown to be
insufficient in correcting common misconceptions about what the settings
do~\cite{wu18privatebrowsing,gao14privatebrowsing}. Other research into the
communication of security risks to users have covered topics such as SSL
warnings~\cite{devdatta13warningland,felt15improveSSLwarnings,
sunshine09crywolf}, visual
icons~\cite{Felt16ConnectionIndicators,Habib21CCPAIcon}, privacy
policies~\cite{Wilson16AnnotatePrivacyPolicies,tsai11purchasingbehavior},
privacy notices~\cite{Schaub15EffectivePrivacyNotices}, and social media
privacy settings~\cite{Liu2011FbPrivSettings}. In the realm of encrypted DNS
settings, one small-scale exploratory study found that users do not understand
the impact of different setting options in the PrivateDNS setting on Android
and that most users would initially choose the default options, but when given more information on the
setting, some users did choose to modify their
choice~\cite{nisenoff-21-PrivateDNS}. In this paper, we expand on those ideas
and explore encrypted DNS settings in the browsers on a larger scale.
Several studies have shown that users make incorrect assumptions about the
security guarantees of encryption, have doubts about the protection from
adversaries, misunderstand phrases like end-to-end encryption, or have
incorrect mental models relative to protection provided in different
contexts~\cite{Wu18EncMentalModels,
Abu-Salma18e2eMentalModels,Stransky2021visEnc, Gerber18JohnnyCanEncrypt, Dechand19EncDontTrust,
Abu-Salma17ObstaclestoAdoption}. Technical jargon, paired with inconsistent
terminology, can also make tools that use encryption even more difficult for the
average user to use correctly~\cite{abu2017securityblanket}. Beyond how users
react to and interact with settings, it is also helpful to understand users'
perception of encryption in other contexts. If individuals use
encryption tools incorrectly, they can have a false sense of security or get themselves
into situations where they can no longer perform the tasks they were
originally attempting to do.
\section{Interfaces for Configuring Encrypted DNS}
\label{sec:interface-analysis}
Vendors have increasingly added support for encrypted DNS, including Web browsers and mobile devices.
In this section, we survey the current state of the interfaces for configuring encrypted DNS in common Web browsers.
We focus in particular on the interfaces for configuring encrypted DNS in five
different popular browsers---Chrome, Brave, Firefox, Microsoft Edge, and
Opera---and one mobile operating system, Android. We focus on how these
settings are presented to users in the United States; these interfaces may
differ in other regions or countries.
\begin{table*}[t]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{ l|ccc|c}
& \multicolumn{3}{c|}{\textbf{Browsers}} & \textbf{Mobile} \\ \hline
\textbf{\begin{tabular}[c]{c}Platform\end{tabular}} & \textbf{Chromium} & \textbf{Firefox} & \textbf{Opera} & \textbf{Android} \\
\hline
\textbf{\begin{tabular}[c]{c}Version\end{tabular}} & {Brave 1.7, Chrome 83, Edge 86} & {Firefox 73} & {Opera 65 Beta} & {Android 9 Pie} \\
\hline
\textbf{\begin{tabular}[c]{c}Setting Name\end{tabular}} & \shortstack{Secure DNS} & \shortstack{DNS over HTTPS} & \shortstack{DNS-over-HTTPS} & \shortstack{Private DNS} \\
\hline
\textbf{\begin{tabular}[c]{c}Protocol\end{tabular}} & DoH & DoH & DoH & DoT\\
\hline
\textbf{\begin{tabular}[c]{c}Default\end{tabular}} & Opportunistic & Cloudflare & Disabled & Opportunistic \\
\hline
\textbf{\shortstack{Support for Opportunistic\\use of Encrypted DNS}} & \newmoon & \fullmoon & \fullmoon & \newmoon\\
\hline
\textbf{\shortstack{Warning for Malformed\\Custom DNS Resolver URL}} & \newmoon & \newmoon & \fullmoon & \fullmoon\\
\hline
\textbf{\shortstack{Links to Privacy Polices\\for Resolvers Shown to Users}} & \newmoon & \fullmoon & \fullmoon & No resolvers shown \\
\bottomrule
\end{tabular}}
\caption{Summary of encrypted DNS Interfaces.}\label{tab:interfaces}
\end{table*}
\begin{table*}[t]
\centering
\begin{tabular}{l|cccccc}
\hline
Browser & Cloudflare & CleanBrowsing & Google & NextDNS & OpenDNS & Quad9 \\
\hline
Chrome & \newmoon & \newmoon & \newmoon & \newmoon & \newmoon & \fullmoon \\
Firefox & \newmoon & \fullmoon & \fullmoon& \newmoon & \fullmoon& \fullmoon \\
Edge\footnotemark[1] & \newmoon & \newmoon & \newmoon & \newmoon & \newmoon & \fullmoon \\
Opera\footnotemark[2] & \newmoon & \fullmoon & \newmoon & \fullmoon& \fullmoon& \fullmoon \\
Brave & \newmoon & \newmoon & \newmoon & \newmoon & \newmoon & \newmoon \\
\bottomrule
\end{tabular}
\caption{Resolvers listed in encrypted DNS settings interfaces by different browser vendors.}
\label{tab:SupportedResolvers}
\end{table*}
Encrypted DNS interfaces fall into two categories: (1)~Chromium-based
interfaces (Brave, Chrome, and Edge) and (2)~Firefox/Opera. Chromium-based
browsers refer to encrypted DNS as ``secure DNS'' and default to
opportunistically using the user's default DNS resolver, while falling back to
sending unencrypted DNS queries if the encrypted DNS resolver is unavailable.
Edge provides many of the same options for settings to users, but does so in a
slightly different way: When a user selects a resolver from the drop down
menu, Edge automatically fills the text box for selecting a custom resolver
with the resolver's URL allowing it to be edited by the user.
In contrast to Chromium-based browsers, Firefox and Opera refer to the
encrypted DNS setting as ``DNS-over-HTTPS'' and do not provide support for an
opportunistic mode, where the browser will encrypt queries if the DNS resolver they were already using supports DoH, as Chromium browsers do. Firefox's default behavior is to
use encrypted DNS with Cloudflare as the default resolver; in contrast, Opera
disables encrypted DNS by default. When encrypted DNS is enabled, Opera uses
Cloudflare as the default resolver. Only Opera explicitly mentions that DoH
uses third-party services; rather than falling back to unencrypted queries, it
alerts users that a page was inaccessible, mentioning that the DNS-over-HTTPS
setting may be to blame. As far as selection options of resolvers shown to users, Mozilla operates a Trusted Recursive Resolver (TRR) program, which ensures that DoH
providers recommended by Firefox best protect privacy by not over-collecting
and sharing data~\cite{ff-trr}.
Some of the resolvers shown in these descriptions offer additional features such as blocking malware or adult content.
\footnotetext[1]{After the survey was distributed Quad9 was removed from the list of resolvers in Edge.}
\footnotetext[2]{Opera offers multiple versions of the Cloudflare resolver, including versions that block adult content and malware.}
\addtocounter{footnote}{2}
``Private DNS'' in Android is the only mobile operating system that we include
in our analysis. Unlike the browser-based settings, ``Private DNS'' supports
DoT rather than DoH. By default, DoT encrypts DNS queries to whichever
resolver the user has selected; Android's private DNS will fall back to
unencrypted queries if DoT fails. In contrast to the settings in browsers,
Private DNS does not give the user any suggestions of resolvers, forcing the
user to input their own URL for a resolver. This mode will not fall back to
unencrypted queries and can cause web pages or resources to become unavailable
if, for example, the user inputs an invalid URL.
\floatstyle{boxed}
\restylefloat{figure}
\begin{figure}[t!]
\centering
\begin{subfigure}[t]{0.45\linewidth}
\includegraphics[width=\linewidth]{brave_interface.png}
\caption{Brave.}
\label{fig:brave}
\end{subfigure}
\begin{subfigure}[t]{0.45\linewidth}
\includegraphics[width=\linewidth]{chrome_interface.png}
\caption{Chrome.}
\label{fig:chrome}
\end{subfigure}
\begin{subfigure}[t]{0.45\linewidth}
\includegraphics[width=\linewidth]{edge_interface.png}
\caption{Edge.}
\label{fig:edge}
\end{subfigure}
\begin{subfigure}[t]{0.45\linewidth}
\includegraphics[width=\linewidth]{firefox_interface.png}
\caption{Firefox.}
\label{fig:firefox}
\end{subfigure}
\begin{subfigure}[t]{0.35\linewidth}
\includegraphics[width=\linewidth]{opera_interface.png}
\caption{Opera.}
\label{fig:opera}
\end{subfigure}
\caption{Encrypted DNS settings for different browsers.}
\label{fig:enc-dns}
\end{figure}
\floatstyle{plain}
\restylefloat{figure}
\section{Method}\label{sec:method}
To learn more about participants' understanding of encrypted DNS settings in
browsers, we designed a two-part survey. In the first part of the survey, we
asked participants about their usage of different browsers. Based on the
answers from the first part of the survey, the second part of the survey then
asked users to interact with a high-fidelity interface for encrypted DNS
settings designed to resemble their browser. We also asked users about encrypted DNS
settings in their own browser. In this section, we first describe the survey design and
recruitment methods; we then discuss the limitations of our survey design, as
well as the ethical considerations associated with our survey design. This
study was approved by our university's Institutional Review Board (IRB).
\subsection{Survey Design and Recruitment}\label{sec:method:survey_design}
\begin{figure}[t]
\centering \label{fig:method-diagram}
\resizebox{0.75\columnwidth}{!}{ \input{diagram} }
\caption{Overview of Survey Structure.} \end{figure}%
Before starting either of the surveys, respondents were asked to read a
consent form and agree to participate in the study. If a participant was not
eligible to participate, the participant was immediately redirected to a
survey termination page. In the initial survey, we asked 800 participants
about their use of different popular web browsers as well as filler questions
about their mobile and Internet service providers. Based on their responses
to these questions, we assigned participants to subgroups determined by the
browsers they reported using at least once a week. We included users who
reported using multiple browsers at least once a week in the grouping for the
least commonly used browser they mentioned. We excluded participants that
displayed a lack of effort, as evidenced through responses unrelated to the
questions. Assigning users to groups in this manner ensured that we recruited
participants who used a variety of major browsers that offer encrypted DNS.
After assigning participants to sub-groups according to browser, we invited up
to 50 participants from each group to participate in a second, longer survey.
For some of the less commonly used browsers (Brave and Opera), we were only
able to recruit a lower number of participants. In the second survey,
following the consent form, we asked participants if they had heard of the DNS
prior to this study, and if they knew what DNS did. To discourage dishonest
responses, we informed participants that they did not need to have prior
knowledge of the DNS to continue with the survey.
After this step, we randomly assigned participants to an anonymized version
of actual DNS settings that can be found in browsers. Rather than simply
showing them a screenshot of their assigned interface or just giving them a
multiple choice question with the different settings options, we embedded
interactive versions of the browser directly in the survey so that
participants could interact with them in the same way they would in their own
browser.
We recorded information about how they interacted with the interface
they were shown. We logged not only the final options that they chose, but
also how they interacted with the interface before making a final selection.
Before being shown the interface, each user was asked to select the setting
that they would choose if they encountered it in their own browser. These
interfaces were based on Chrome, Brave, Edge, Firefox, and Opera. We removed
any reference to the browser itself from the interface to avoid biasing
participants, but we opted to leave the actual company names for any
recommended resolvers as this best simulates a real browser interface.
We merged the Chrome and Brave interfaces because the interfaces used the same terminology and layout; we included all of the resolvers from both interfaces.
After interacting with the settings screen, we asked participants questions about
their rationale for their choosing different options, as well as their
understanding and perceptions of the different options.
We also showed participants an interactive version of the Private
DNS settings on Android and asked about the settings that they would choose.
After collecting participants' responses, we then described the DNS and
checked the participants' understanding through a multiple-choice question
about the basic functionality of DNS. To ensure that our analysis of this
part of the survey was based on responses from users who had a basic
understanding of DNS, we excluded participants who did not correctly answer
the multiple-choice question from the analysis, but we still allowed these
participants to complete the survey. We then explained, in more detail, how
the specific settings work, and given this additional information, we asked
their thoughts on the setting.
Next, we showed participants the interactive encrypted DNS settings screen
that they saw earlier in the survey and asked them once again to decide what
options they would choose and why. This allowed us to understand how
respondents changed their settings after gaining a high-level understanding
of what encrypted DNS does and the benefits it provides. We subsequently asked
participants if they had heard of encrypted DNS before the survey. (We
decided to ask this question after participants chose their setting in
the anonymized interfaces because we did not want to bias them while they
looked at the simulated settings interfaces that referred to encrypted DNS
through different names.) We also briefly asked participants about their
preferences between different ways of distributing DNS queries and other
security trade-offs to better understand how participants viewed tradeoffs
between usability and privacy associated with encrypted DNS.
The next short section of the survey asked participants about their knowledge
and trust in different DNS resolvers and other parties that are relevant to
DNS, such as Internet Service Providers (ISPs) and browsers. Subsequently, we
asked participants to report on the current DoH settings in a browser that
they regularly use. We also asked about their experience with these settings.
The browser that we asked the participants to look at depended on their
responses to the screening survey: If, for some reason, a participant did not
have access to the browser subgroup that we assigned them to, we reassigned
the participant to another browser that they did have access to. We then
asked participants about the operating system of their primary mobile phone.
If a participant reported using a phone with an operating system that supported
DoT, we asked the participant to check their DNS settings and answer questions
about their use of those settings.
We concluded the survey with basic demographics and an optional question for
participants to leave feedback.
We recruited participants via Prolific, where we required that they were at
least 18 years old, live in the US, and have completed at least 100 surveys before the
first survey with at least a 95\% approval rating. The recruitment text of
the survey was phrased as a set of surveys where individuals would be asked
about network settings and did not mention encrypted DNS or security to avoid
self-selection bias. Participants were also asked to complete the survey on a
desktop or laptop computer to avoid formatting issues. The first part of the
survey was designed to take approximately four minutes while the second part
was designed to take approximately 20 minutes. Participants were paid \$0.50
and \$3.30 for completing the first and second parts, respectively. Payments
were made within 72 hours of the participants completing the survey through
the Prolific platform. To avoid compensating Prolific participants who did
not honestly intend to complete the study, we only compensated participants
who completed the full survey. Table~\ref{tab:demographics} summarizes the
demographics of our participants.
\input{demographics}
\subsection{Limitations}\label{sec:Limitations}\label{sec:method:limitations}
This study design has several limitations. First, the encrypted DNS interfaces
were shown to participants in the context of a survey, rather than in the
context of other settings in their browser's settings menu. Showing users
these menu options in the specific context of a study and survey on encrypted
DNS might in some cases affect the nature of responses, given that users were
only presented the part of the configuration menus specifically associated
with encrypted DNS. Informing participants that their choice of encrypted DNS
setting would not affect the settings of their actual browser may also have
influenced the options that they chose in the survey, since users knew that
their choices in the survey would ultimately not have any practical effect on
their own user experience, or their privacy. The survey-based setup also
prevented users from experimenting with different settings. In an actual
browser, a user might experiment with several settings to see their impact,
which was not possible in this survey.
The participant sample also has some limitations, as the sample demographic
may not directly correspond to a population sample for the relevant target
population. Specifically, the sample of participants in this survey is more
representative of typical Prolific respondents and thus tends slightly
younger and more educated than the general population of the United States.
Our sample of participants was skewed towards male participants. This sample
bias may mean that, to some degree, the results of this survey may not
necessarily generalize to a broader population. Nevertheless, these survey
results are still useful because our goal in this work is not to make general
statistical claims about the broader population, but rather to gain insight
into how users make choices with regard to encrypted DNS settings. In this
case, a sample of the general population is likely to shed light on similar
issues and insights as they pertain to encrypted DNS. It may not be advisable
to cite percentages of respondents in this paper as they might pertain to the
general population, but we do expect that the trends that appear in this
study would also be {\em present} in the general population and thus, the
issues that come to light in this study are some that designers of encrypted
DNS interfaces should consider.
Another limitation of using Prolific (and survey recruitment platforms in
general) is the risk of participants simply clicking through the survey and
not giving thoughtful responses. We mitigate this potential concern by
removing participants that did not understand the description of DNS or gave
answers to free-response questions that were consistently unrelated to the
questions, as a form of attention check.
\subsection{Ethics}
\label{sec:method:ethics}
This study was approved by our university's Institutional Review Board (IRB)
and was designed in accordance with ethical principles outlined in the Belmont
Report, namely: respect for persons, beneficence, and justice. Respect for
humans was considered as part of the consent process and survey design: Due to
the short and concise nature of the screening and main surveys, participant
fatigue was not expected or reported. Before taking either the screening or
main surveys, participants provided their consent to participate via a form,
which informed them about the structure of the survey and their rights as a
participant. When participants were shown the simulated encrypted DNS
settings interface, they were informed that their choice of setting would not
affect the setting in their own browser. Participants potentially garnered
some benefit from taking this survey by having the opportunity to learn more
about the potential risks and benefits of encrypted DNS settings which might
enable them to make more informed DNS settings choices in the future. From
the perspective of justice, the benefits of this survey are likely to benefit
the same population as that of the respondents: Internet users who depend on
web browsers and mobile operating systems (and the underlying DNS) to access
websites and Internet services. Over the course of the survey, we did not
collect any personally identifiable information (PII) beyond general
demographic information.
\section{Results}\label{sec:results}
In this section, we present the results from our survey and highlight
prevailing themes that emerged during our analysis.
When discussing participant answers to qualitative, free-response questions,
we developed a codebook to identify themes in the responses. This process
involved a primary coder creating a codebook for all of the responses. A
secondary coder used the codebook developed by the primary coder, and all
disagreements were resolved through discussion. As an attention check, we
excluded participants that incorrectly answered the question about the
function of DNS after our explanation as discussed in Section~\ref{sec:method}
or answered free response questions with responses that were unrelated to the
prompts.
\subsection{Are users aware of encrypted DNS settings in browsers and mobile
devices?}\label{sec:awareness}
\begin{tcolorbox}[width=\linewidth,center]
Many participants had heard of DNS prior to the survey, yet fewer had heard of encrypted DNS.
Of the participants that had heard of DNS, most did not know what it did,
and many who thought they knew what it did were unable to correctly describe its primary function.
Participants were largely ambivalent about their trust in companies that
are not well-known in contexts other than DNS, while they had stronger
opinions on their trust in other well-known companies, as well as other
stakeholders in the DNS ecosystem with whom they had preexisting relationships (e.g., their primary browser or ISP).
\end{tcolorbox}
\noindent
Participants frequently reported having heard of DNS, but often did not know
what it did or had incorrect assumptions about its function. 73\% of all
participants reported having heard of DNS prior to the survey, with 36\% of
the participants believing that they knew the functionality that DNS provides.
Although participants could overstate their knowledge of DNS, we attempted to
mitigate this possibility by asking the question at the beginning of the
survey and clearly stating that individuals would be able to participate in
the survey regardless of whether they had any knowledge of the DNS.
When participants who claimed to know the purpose of DNS were asked to
describe it in their own words, fewer than half mentioned the DNS as being
responsible for translating domain names to IP addresses. The other
participants had varying misconceptions of its purpose, including identifying
computers on a network and connecting a device to the Internet. Thus, fewer
than 18\% of participants could describe what DNS does prior to the survey.
Of the participants who reported having heard of DNS, only 59.9\% had heard of
encrypted DNS.
Individuals who had heard of DNS without necessarily fully understanding it
may have encountered the term tangentially in other contexts not specific to
encrypted DNS. Examples of when users have encountered DNS in the popular
press may include: (1)~when problems with DNS cause Internet outages and make
popular websites and services inaccessible~\cite{nytimes-outage}; or (2)~when widespread
vulnerabilities are disclosed or a major security event
occurs~\cite{nytimes-mirai}. In these
cases, people may be exposed to the terminology and a brief description of the
DNS, yet we suspect they are likely focused on how
DNS affects them, rather than internalizing technical details.
As Figure~\ref{fig:heard_of_provider} demonstrates, most participants had
heard of Google and Comcast but were less familiar with Cloudflare, NextDNS, and Quad 9.
Further, Figure~\ref{fig:trust_provider} shows that participants
were ambivalent about their trust in NextDNS and Quad 9. Both of these factors may
be due to participants' lack of knowledge about these providers.
Participants were more divided on if they trusted the more well-known
companies, Google and Comcast, that also operate recursive resolvers. More participants reported that they trusted their primary browser, mobile service provider, and Internet service provider, which may
be related to their familiarity and previous interactions with them.
Because so few participants knew what the DNS was or what DNS queries do, it
is unreasonable to expect users to understand the implications that their
choices of setting could have on their security and general browsing
experience. Thus, even a small description or a link to a more in-depth description
of DNS and encrypted DNS could be useful when presenting these settings to
users.
\begin{figure}[t!]
\includegraphics[width=0.65\linewidth]{heard_of_provider.pdf}
\caption{Percentage of participants who had heard of different stake holders
in the encrypted DNS ecosystem.}
\label{fig:heard_of_provider}
\end{figure}
\begin{figure}[t!]
\includegraphics[width=0.80\linewidth]{trust_provider.pdf}
\caption{Percentage of participants who trust their provider.}
\label{fig:trust_provider}
\end{figure}
\subsection{\mbox{What encrypted DNS settings do users have enabled in their browsers and phones?}}\label{sec:settings}
\begin{tcolorbox}[width=\linewidth,center]
Most participants had the default settings enabled in their browsers,
which typically meant that participants' browsers were opportunistically
encrypting their DNS queries. The only deviations from the defaults were
to disable the setting, or to use the Cloudflare or Google resolvers. In
the mobile devices, to our surprise, there was a more even split between
users that had the default ``Automatic'' (opportunistic) setting and the
setting entirely disabled.
\end{tcolorbox}
\begin{table}[]
\resizebox{\linewidth}{!}{
\begin{tabular}{l|rl|ccccc}
Browser & \multicolumn{2}{c|}{\shortstack{Successfully\\reached setting\\(\% of participants)}} & \shortstack{With Current\\Service Provider} & Google & Cloudflare & Disabled & Unsure\\
\hline
Brave & 23 & (88.5\%) & \cellcolor[gray]{0.8}\textbf{21}& 0 & 1 & 1 & 0\\%26
Chrome & 58 & (81.7\%) & \cellcolor[gray]{0.8}\textbf{46} & 2 & 0 & 10 & 0\\%71
Edge & 24 & (70.6\%) & \cellcolor[gray]{0.8}\textbf{18} & 0 & 0 & 5 & 1 \\%34
Firefox & 43 & (91.5\%) & N/A & N/A & N/A & N/A & N/A\\%47
Opera & 3& (75.0\%) & \textbf{-} & 2 & 1 & \cellcolor[gray]{0.8}\textbf{0} & 0\\%4
\hline
\multicolumn{2}{l}{{\small\cellcolor[gray]{0.8} \textbf{default setting}}}\\
\end{tabular}}
\end{table}
Although most participants were able to access the encrypted DNS settings
menu, most still reported that they had the default options enabled.
108 participants reported that they could access one of the Brave, Chrome, Edge,
and Opera interfaces.\footnote{Due to an error in the survey, we do not have
the actual settings for individuals that looked at their settings in the
Firefox browser.}
One participant was unsure if the setting was enabled.
Although the survey provided written instructions and a video describing how
to access the encrypted DNS settings in each browser, it is possible that the
participants who could not access the encrypted DNS settings menu were using
older versions of the browser that did not provide support for these settings.
It is also possible that participants answered dishonestly since this was
towards the end of the survey. Additionally, participants may have simply
been unable to find the setting due to the complexity of browser settings
menus. Of the remaining 107 participants, 85 (79.4\%) had the default
settings for their browser selected.
Only six participants had encrypted DNS configured in such a way that all of
their queries went to a single resolver. All of those individuals were using
either Cloudflare's resolver or Google Public DNS. No participants reported
inputting a custom DNS resolver in their browser. This observation
underscores not only the importance of good defaults, but also the careful
selection of the suggested resolvers that are shown to users because those are
more likely to be selected than a custom DNS resolver. Many of the participants that had a setting other than the default reported not remembering changing the setting.
This observation could result in part from circumstances, such as not having the most updated browser or operating system,
or perhaps having had settings modified by another user on a shared machine.
\begin{figure}[t!]
\includegraphics[width=0.55\linewidth]{private_dns_actual.pdf}
\caption{Private DNS settings of participants.}
\label{fig:actual_mobile_choice}
\end{figure}
In the case of mobile devices, 109 participants reported having a phone that ran an android operating system and 83 reported being able to navigate to the ``Private DNS'' settings page.
Of these participants, 48 (57.8\%) had the ``Automatic'' option selected, 34 (41.0\%) had ``Off'' selected, and only one participant had ``Private DNS Provider hostname'' selected with the correct URL to send their queries to Cloudflare.
``Automatic'' is the current default, yet none of the participants that had the ``Off'' setting selected reported remembering changing this setting in the past.
This suggests that perhaps the default of ``Automatic'' was somehow not applied to them, so while their phone supported the opportunistic mode, they were not taking advantage of the potential benefits of the setting.
Participants who were unable to reach the Private DNS settings page may have
had older operating systems installed on their phone, and thus might not be able to access the page.
\subsection{When shown encrypted DNS settings for different browsers, which settings do users select, and why?}
\label{sec:selection}
\begin{tcolorbox}[width=\linewidth,center]
The majority of participants chose the default settings in the interface
shown to them. There were variations in the settings that users chose
based on which interface they were shown. For example, no participants
that saw a Chromium-based interface disabled the setting. On the other
hand, participants that saw the Opera interface deviated the most from the
defaults, perhaps because it was the only interface that disabled
encrypted DNS by default. No participants correctly entered a custom
trusted recursive DNS resolver. Finally, participants had conflicting
opinions of the effect of various settings, and their attitudes were
influenced by the way the settings were described.
\end{tcolorbox}
\begin{table*}[t]
\centering
\large
\resizebox{\linewidth}{!}{
\begin{tabular}{l|r|r|r|rrrrrr|r}
{\normalsize\textbf{Browser}}& {\normalsize\textbf{\#}} & {\small\textbf{Off}} & {\small\textbf{\shortstack{With Current\\Service Provider}}} & {\small\textbf{Cloudflare}} & {\small\textbf{\shortstack{Google\\(Public DNS)}}} & {\small\textbf{Quad9}} & {\small\textbf{NextDNS}} & {\small\textbf{\shortstack{CleanBrowsing\\(Family Filter)}}} & {\small\textbf{OpenDNS}} & {\small\textbf{Custom}}\\
\hline
{\normalsize\textbf{Chrome/Brave}} & 51 & 0 &\cellcolor[gray]{0.8}\textbf{37} & 6 & 2 & 0 & 0 & 0 & 1 & 3 \\
{\normalsize\textbf{Edge}} & 48 & 0 & \cellcolor[gray]{0.8}\textbf{41} & 3 & 3 & 0 & 0 & 1 & 1 & 0 \\
{\normalsize\textbf{Firefox}} & 45 & 6 & - & \cellcolor[gray]{0.8}\textbf{34} & - & - & 3 & - & - & 2 \\
{\normalsize\textbf{Opera}} & 40 & \cellcolor[gray]{0.8}\textbf{20} & - & 17$^*$ & 3 & - & - & - & - & 0 \\
\hline
{\normalsize\textbf{Android}} & 184 & 12 & \cellcolor[gray]{0.8}\textbf{158} & - & - & - & - & - & - & 18\\
\hline
{\small\cellcolor[gray]{0.8} \textbf{default setting}} &
\multicolumn{6}{l}{{\small - indicates that the option is not available for that browser}}\\
\multicolumn{11}{l}{{\small* The Opera interface offers three version of the Cloudflare resolver: the default (11), No Malware (2), and No Malware or Adult Content (4)}} \\
\end{tabular}}
\caption{Users' choice of encrypted DNS setting in the anonymized browser interfaces with no addition information about DNS or encrypted DNS.}
\label{tab:InitialChoice}
\end{table*}
\noindent
When shown the anonymized browser interfaces for encrypted DNS, 71.7\% of participants continued to use the default settings shown to them.
This percentage varied by browser, with Firefox having the highest percentage
of users continuing with the default setting (75.6\%) and Opera having the
lowest, with only 50\% of participants selecting the default.
Opera is the only browser that has encrypted DNS disabled by default, which could have been a factor for many people modifying the setting.
Firefox, on the other hand, has the fewest alternatives.
The predominant reasons participants gave for selecting the settings that they
did included: (1)~a perceived increase in security, simply because the setting
was the default; or (2)~because they just didn't know what the setting did. As one participant put it (P144), ``It was the default setting and also helps prevent low level attacks from hackers.''
Participants' choices for the ``Private DNS'' setting show a similar pattern, with 83.7\% of participants keeping the default ``Automatic'' setting.
This finding illustrates the importance of how browsers and mobile service providers configure defaults.
Looking at the number of participants that had the setting enabled overall also showed some interesting trends.
Notably, no participants disabled the encrypted DNS setting when shown any of the interfaces for the Chromium based browsers (Brave, Chrome, Edge).
Firefox users followed, with 86.7\% of users enabling the setting and Opera had only 50\% of participants that enabled the setting.
Chromium-based browsers have a small toggle that collapses the settings
window, it is possible that discouraged people from disabling the setting.
Furthermore, the option to use their current service provider might have been
seen as a lower risk alternative since they already have a relationship with
that company: ``I chose my current service provider because I trust that it is
a great choice since I am currently using it.'' was how one user expressed the
sentiment (P31).
When participants enabled encrypted DNS, but did not go with the default setting, they were much more likely to choose a resolver that was listed in the setting rather than specifying a custom resolver. Cloudflare was by far the most popular. Table~\ref{tab:InitialChoice} shows the initial choices of the users.
While very few participants chose to specify a custom resolver in any of the browser (2.7\%) or mobile (9.8\%) interfaces, the ones who did, entered text that would not function the way they might expect it to.
\textbf{In fact, no participant entered a custom DNS resolver that would have resulted in their queries being encrypted.}
For example ``McAfee,'' ``www.google.com,'' and ``1.1.1.1'' were entered by participants into browser and mobile interfaces.
In the case of ``1.1.1.1'', while it may seem correct, to actually use the Cloudflare DNS resolver through the custom resolver input field, the user would need to have entered ``1dot1dot1dot1.cloudflare-dns.com'' on android or ``https://cloudflare-dns.com/dns-query'' for any of the browsers.
Some interfaces do check whether the text entered will function, but it is often possible to click away from the setting before the warnings are shown. It would be beneficial to users if all of the interfaces, rather than only some, would give users actionable feedback on the validity of their inputs before they exit the page.
When asked about the advantages and disadvantages of enabling encrypted DNS,
many participants thought enabling the setting would result in general
improvements to security, although they were often unable to go into any detail as to what those security benefits might be.
Participants who saw interfaces that labeled DoH as ``secure DNS'' were more likely to mention security as a potential benefit of enabling DoH than people that saw the setting called ``DNS over HTTPS'' (p$<$0.05)
This connection between the name of the setting and the perceived impact
emphasizes the importance of this subtle difference in wording.
There was confusion among participants about how their choice of encrypted DNS
resolver might affect their browsing performance: 19.6\% of participants
mentioning that the setting could slow down Internet browsing, while 13.0\%
(not a significantly different number of participants) thought it might
improve performance. Such confusion is actually consistent with empirical
studies, which have shown that the relative performance improvement (or
degradation) of encrypted DNS actually depends quite a lot on the choice of
encrypted DNS resolver and client~\cite{hounsel2021can}.
Aside from the potential effects on speed, participants also mentioned concerns
that enabling encrypted DNS might reduce their access to websites with one
respondent stating, ``If it is security-related, it may restrict access to
certain domains.'' (P6). The desire to be able to access the Internet as they
normally do was also mentioned by several participants as the reason they
choose their respective settings. Participants overwhelmingly wanted an
explanation of the settings to assist with making the decision. More specific
requests included wanting definitions of different relevant terms shown on the
settings page, the pros and cons of different settings, and information on the
security benefits of each setting.
Interfaces that labeled the setting with the technical name ``DNS over HTTPS'' caused additional confusion in some participants.
\textbf{Instead of interpreting the name as meaning DNS \textit{using} the HTTPS protocol they interpreted DoH as meaning use DNS \textit{instead} of HTTPS.}
Of the participants who saw the Firefox or Opera interfaces, that label the setting in this way, 10.6\% mentioned an incorrect interpretation of the setting name as part of their reason for choosing the setting option that they did.
There may have been more participants who misinterpreted the setting name, but they did not mention it in their reasoning for choosing their preferred setting option.
One participant (P3) who saw the Firefox interface and chose to disable DoH stated, ``I have no earthly idea what DNS is, while I at least have a passing familiarity with HTTPS.'' Another participant (P30) said ``From the little I know I believe that HTTPS is more secure than DNS'' and chose to disable the setting in the Opera interface.
While most of the participants who misinterpreted the setting name in this way ended up opting to disable the setting, that decision was not universal.
This observation highlights the importance of avoiding technical jargon that could be easily misinterpreted by average users.
\begin{table*}[t]
\centering
\large
\resizebox{\linewidth}{!}{
\begin{tabular}{l|r|r|r|rrrrrr|r}
\hline
{\normalsize\textbf{Browser}}& {\normalsize\textbf{\#}} & {\small\textbf{Off}} & {\small\textbf{\shortstack{With Current\\Service Provider}}} & {\small\textbf{Cloudflare}} & {\small\textbf{\shortstack{Google\\(Public DNS)}}} & {\small\textbf{Quad9}} & {\small\textbf{NextDNS}} & {\small\textbf{\shortstack{CleanBrowsing\\(Family Filter)}}} & {\small\textbf{OpenDNS}} & {\small\textbf{Custom}}\\
\hline
{\normalsize\textbf{Chrome/Brave}} & 51& 1 & \cellcolor[gray]{0.8}\textbf{33} & 8 & 3 & 0 & 1 & 0 & 2 & 3\\
{\normalsize\textbf{Edge}} & 48 & 2 & \cellcolor[gray]{0.8}\textbf{29} & 6 & 4 & 0 & 1 & 1 & 3 & 2\\
{\normalsize\textbf{Firefox}} & 45 & 6 & - & \cellcolor[gray]{0.8}\textbf{28} & - & - & 7 & - & - & 4\\
{\normalsize\textbf{Opera}} & 40 & \cellcolor[gray]{0.8}\textbf{17} & - & 19$^*$ & 4 & - & - & - & - & -\\
\hline
{\small\cellcolor[gray]{0.8} \textbf{default setting}} &
\multicolumn{6}{l}{{\small - indicates that the option is not available for that browser}}\\
\multicolumn{11}{l}{{\small* The Opera interface offers three version of the Cloudflare resolver: the default (11), No Malware (5), and No Malware or Adult Content (3)}} \\
\end{tabular}}
\caption{Users' choice of encrypted DNS setting in the anonymized browser interfaces after DNS and encrypted DNS has been explained.}
\label{tab:SecondChoice}
\end{table*}
\begin{figure}[t!]
\includegraphics[width=.9\linewidth]{sankey.pdf}
\caption{How users' settings choices changed from their initial choice to their decision after DNS and encrypted DNS has been explained.}
\label{fig:sankey}
\end{figure}
\subsection{When the technical aspects of these systems are explained to users,
do their choices of settings change?}\label{sec:technical}
\begin{tcolorbox}[width=\linewidth,center]
Almost 40\% of participants modified their settings in some way after
being shown an explanation of DNS and encrypted DNS. The default settings
remained popular, but much less so than when participants had made
decisions
without extra information about encrypted DNS. Although participants
seemed to understand the settings better, they still had problems
understanding the differences in functionality or privacy guarantees between the different resolvers.
\end{tcolorbox}
\noindent
When asked to look at the same anonymous interface they were shown earlier in
the survey, after having been provided a description of what the DNS is used
for and what encrypted DNS does, 37.0\% of participants chose to modify their
settings in some way. 30 (16.3\%) participants reversed their choice: with
an equal number choosing to disable the setting that they had previously
enabled and enabling the setting that they had previously disabled.
Figure~\ref{fig:sankey} shows the relationship between the two settings that users chose.
The default settings were still the most selected option across all of the
interfaces (59.2\%), but the number of participants who selected the defaults decreased across every browser which is significantly different than the (71.7\%) of participants who chose the default before they were provided with additional information (p$<$0.5)
The largest change was seen among users of the Opera browser, with 47.5\% of participants choosing to modify their setting in some way.
Participants were still more likely to choose a resolver listed in their
respective interfaces. Although more people opted to enter their own custom
resolver hostnames, as with their initial choice, none of the text entered
would have functioned properly. Because the descriptions we provided to
participants did not provide instructions for how to select a resolver, but
mainly talked about the settings on a high level, such errors could be
expected and were consistent with how participants described their experience; for example, ``I understand more about DNS now, but still don't know who else I'd use if not my current service provider'' (P46).
Table~\ref{tab:SecondChoice} shows all of the setting options chosen by
participants after receiving the additional information about each setting.
Perceived benefits to security remained a predominant factor in many
participants' decision-making process. Unlike with their initial choice of
setting, privacy and encryption factored into participants' decision-making
processes at the end of the survey. Because our descriptions of DNS and
encrypted DNS discussed those topics, this result was expected to some degree.
Finally, prior knowledge of and trust in a company was a factor users
mentioned when explaining why they chose the setting options that they did.
When participants reported having heard of encrypted DNS, they were significantly more likely to enable encrypted DNS as their initial choice when they were not provided with any additional information about the setting (p$<$0.05).
On the other hand, after everyone had been provided with information about DNS and a brief description of encrypted DNS the differences were no longer significant.
\section{Discussion}\label{sec:disc}
The results from previous sections highlight several general takeaways and
point to a variety of recommendations, both for designers and policymakers
(e.g., standards bodies, or regulators who may wish to standardize how various
protocol options are presented to users).
\subsection{Takeaways from Results}
Our results highlighted that, although many users are generally aware of
private DNS and that various configuration options are possible, most users do
not change their browser or mobile OS settings from the defaults, and many
users also do not understand either how DNS (or encrypted DNS) works, or the
guarantees that encrypted DNS can provide. Users generally want more
information about the privacy benefits that encrypted DNS can provide, as well
as information about their options in configuring it. These observations are
consistent with Clark's discussions of ``tussle spaces'', which have noted
that the choices that users have (and are aware of) can have significant
consequences. As such, encrypted DNS (and interface) implementations should recognize these tussles and make choices available to users in ways that allow them to make appropriate tradeoffs between privacy and performance, according to their own preferences.
Additionally, users were concerned that their attempts to customize encrypted
DNS configurations could cause basic functionality to break, resulting in
their inability to use the Internet. As it turns out, such concerns are not
unfounded. For example, specifying an custom recursive resolve with incorrect
syntax in the mobile OS configuration does result in a silent connectivity
failure, with no error message to the user concerning the nature of the
misconfiguration.
Based on these observations, we provide a set of recommendations for designers
of encrypted DNS interfaces on user-facing devices (e.g., browsers, mobile
OSes) that could allow users to make more informed choices concerning the
configuration of encrypted DNS.
\subsection{Design Recommendations}
\paragraph{Provide a basic primer on DNS function (and privacy risks).} Many
users are not aware of the functions of DNS, as well as the privacy risks
associated with the ability of a third party to observe DNS traffic.
We thus recommend that application designers find a way to provide
information to users about DNS function, privacy risks, and the tradeoffs
associated with each setting.
Some of this information could be provided in a way that is independent of
the interface provided by the platform. Nonetheless, it may be useful to
augment the {\em interface} itself in a way that indicates to the user which
entities and organizations can see DNS traffic for different settings, in
simple terms (e.g., ``your ISP'', ``the coffee shop provider'', ``the web site
host''). The exact design of such an interface
could be a ripe topic for future work, as we discuss in more detail below.
\paragraph{Provide privacy policies for the resolvers.} In principle, users
have many choices for trusted recursive resolvers, from major providers (i.e.,
Google, Cloudflare), to medium-sized operators (i.e., Quad9), to smaller
independently operated recursive resolvers. A user's choice of trusted
recursive resolver has significant implications for privacy, since the
organization operating the trusted recursive resolver sees potentially all of
the user's DNS traffic (and hence may be able to infer much about the user,
from browsing patterns to other behavior). Our results indicated that while
participants understood the setting after it was described to them, they still struggled to understand the differences between the recursive resolvers.
Because users have significantly different levels of trust for the respective operators of recursive resolvers, the privacy practices and policies of operators should be more transparent to
users to ensure that users can make more informed choices.
\paragraph{Be thoughtful about the use of technical protocol terminology,
which may not map to users' mental models.} Some technical terminology to
describe encrypted DNS protocols can be confusing to users: In particular, we
found that many users misinterpreted the phrase ``DNS-over-HTTPS'', to mean
that DNS would be used instead of HTTPS, {\em not} that HTTPS was the
transport protocol over which DNS queries and responses were transmitted. In
such cases, understanding the assumptions that users make about functionality
based on language choice can help designers choose terminology and phrasing
that better reflects the properties that a protocol provides. User studies and
focus groups may be appropriate when deploying such protocols and variants,
both now and in the future.
\paragraph{Provide users with resolver options.} The large collection of data
by one or a few mainstream resolvers raises privacy concerns. Per the
suggestions of participants, browsers could also add more information about
the advantages and drawbacks of different choices of encrypted DNS resolvers,
which would allow them to make an educated decision about their browser
settings.
\paragraph{Provide users with the necessary format to select a custom
resolver, and check that the user specification is correct and functional.}
Participants expressed concerns about experimenting with the settings,
fearing that they would break elements of their browsers. Browsers could add
more apparent instructions, warnings, or guidelines to their interfaces to
provide more clarity for users. Many survey participants also maintained the
default setting, rather than changing it to best suit their needs.
\subsection{Future Research}\label{sec:disc:future_work}
Future work could replicate this study on a larger scale, and across a wider
range of demographics. Because encrypted DNS
is not limited to the United States, a larger study that captures a broader
cross-section of users could deepen our understanding of user perceptions by
including participants who live in other countries. Involving
more participants, could also provide data that may highlight broader themes,
including how various attitudes and awareness might vary according to user
demographics such as age, level of education, and geography. Furthermore, as
previously noted, asking participants to configure private DNS settings in
their own browser, rather than in the context of the survey might also result
in users being more cautious. Future studies could further explore the extent of user
behavior in the context of the browsers and mobile operating systems
themselves, rather than in the context of a survey. Finally, future studies
might evaluate the extent to which different phrasing and explanations of
settings and options (including the privacy implications associated with
different choices of trusted recursive resolver) might ultimately affect
users' attitudes and behaviors concerning encrypted DNS settings.
\section{Conclusion}
\label{sec:conclusion}
The increasing deployment of encrypted DNS in browsers and mobile operating
systems has significant consequences for privacy, performance, and
reliability---particularly as vendors change default settings (often without direct notification to users). Previous research has observed that encrypted DNS is a tussle space among users, Internet service providers, and content providers, because the parties who control DNS have more ability to optimize content and services and have access to potentially sensitive information about users' browsing behaviors and activities.
Given the significant stakes of encrypted DNS deployment, users should be able to
make informed choices about how it is configured. Interfaces should make
it easy for users to be aware of how encrypted DNS is configured, as well as
how to change default settings in accordance with their preferences. Our
findings in this research confirmed that work is needed in several areas,
including: to improve
user awareness about the privacy implications of DNS, to provide users with
information to better understand the implications of how encrypted DNS is
configured, and to design setting interfaces that make these options intuitive for users to customize. Although this paper does not offer the last word
on user attitudes and awareness about encrypted DNS, our hope is that it lays
the groundwork for more research in this area, to positively affect interfaces, standardization,
and policymaking.
\section{Appendix}\label{sec:appendix}
\subsection{Survey}
\subsubsection{Consent Form}
\emph {Consent form shown here}\\
I am over 18 years old.\\
\\\hspace*{0.35in}\fullmoon~~Yes
\\\hspace*{0.35in}\fullmoon~~No\\
I currently live in the United States of America. \\
\\\hspace*{0.35in}\fullmoon~~Yes
\\\hspace*{0.35in}\fullmoon~~No\\
I have read and understood this consent form.\\
\\\hspace*{0.35in}\fullmoon~~Yes
\\\hspace*{0.35in}\fullmoon~~No\\
I agree to participate in this online research study.\\
\\\hspace*{0.35in}\fullmoon~~Yes
\\\hspace*{0.35in}\fullmoon~~No\\
Please enter your Prolific ID here\\
\subsubsection{Screening Survey}
How often, if at all, do you use the Chrome browser on a laptop/desktop?\\
\\\hspace*{0.35in}\fullmoon~~At least once a day
\\\hspace*{0.35in}\fullmoon~~At least once a week
\\\hspace*{0.35in}\fullmoon~~At least once a month
\\\hspace*{0.35in}\fullmoon~~Less than once a month
\\\hspace*{0.35in}\fullmoon~~I do not use this browser\\
What online activities do you regularly do in your Chrome browser? Please select all that apply.\\
\\\hspace*{0.35in}\fullmoon~~Checking the news
\\\hspace*{0.35in}\fullmoon~~Participating in online video calls or conferences
\\\hspace*{0.35in}\fullmoon~~Streaming or downloading music, radio, podcasts, etc.
\\\hspace*{0.35in}\fullmoon~~Online shopping, making reservations, or using other online consumer services
\\\hspace*{0.35in}\fullmoon~~Working remotely
\\\hspace*{0.35in}\fullmoon~~Checking the weather
\\\hspace*{0.35in}\fullmoon~~Using social media
\\\hspace*{0.35in}\fullmoon~~Text messaging or instant messaging Emailing
\\\hspace*{0.35in}\fullmoon~~Watching video online
\\\hspace*{0.35in}\fullmoon~~Using online financial services (banking, investing, paying bills, etc.)
\\\hspace*{0.35in}\fullmoon~~Other
\\\hspace*{0.35in}\fullmoon~~None of the above
\\\hspace*{0.35in}\fullmoon~~Prefer not to answer\\
What, if any, browsers other than Chrome, Firefox, Microsoft Edge, Brave, Opera, or Safari do you use?\\
What would you say is the browser you use the most?\\
\\\hspace*{0.35in}\fullmoon~~Brave
\\\hspace*{0.35in}\fullmoon~~Safari
\\\hspace*{0.35in}\fullmoon~~Opera
\\\hspace*{0.35in}\fullmoon~~Firefox
\\\hspace*{0.35in}\fullmoon~~Microsoft Edge
\\\hspace*{0.35in}\fullmoon~~Chrome
\\\hspace*{0.35in}\fullmoon~~None
\\\hspace*{0.35in}\fullmoon~~Other\\
What would you say is the browser you use the most?\\
Who is your mobile service provider (the company that you pay for phone or cellular data service on your mobile devices)?\\
\\\hspace*{0.35in}\fullmoon~~Mint Mobile
\\\hspace*{0.35in}\fullmoon~~T-Mobile
\\\hspace*{0.35in}\fullmoon~~AT\&T
\\\hspace*{0.35in}\fullmoon~~Verizon
\\\hspace*{0.35in}\fullmoon~~Visible
\\\hspace*{0.35in}\fullmoon~~Other
\\\hspace*{0.35in}\fullmoon~~I do not have a mobile service provider
\\\hspace*{0.35in}\fullmoon~~Don't know
\\\hspace*{0.35in}\fullmoon~~Prefer not to answer\\
What is your Internet Service Provider (the company that you pay for Internet access in your home)?\\
\\\hspace*{0.35in}\fullmoon~~TDS Telecom
\\\hspace*{0.35in}\fullmoon~~Consolidated Communications Cox Communications
\\\hspace*{0.35in}\fullmoon~~Verizon
\\\hspace*{0.35in}\fullmoon~~Sparklight
\\\hspace*{0.35in}\fullmoon~~CenturyLink
\\\hspace*{0.35in}\fullmoon~~Frontier Communications Comcast (Xfinity)
\\\hspace*{0.35in}\fullmoon~~Charter Communications Windstream
\\\hspace*{0.35in}\fullmoon~~Mediacom
\\\hspace*{0.35in}\fullmoon~~AT\&T
\\\hspace*{0.35in}\fullmoon~~Other
\\\hspace*{0.35in}\fullmoon~~I do not have an internet service provider
\\\hspace*{0.35in}\fullmoon~~Don't know
\\\hspace*{0.35in}\fullmoon~~Prefer not to answer\\
During a typical week how many different places do you go where you connect to a wifi network (e.g. home, work, a public library, etc.)?\\
\\\hspace*{0.35in}\fullmoon~~0-1
\\\hspace*{0.35in}\fullmoon~~2-3
\\\hspace*{0.35in}\fullmoon~~4-5
\\\hspace*{0.35in}\fullmoon~~5+
\\\hspace*{0.35in}\fullmoon~~Don't know
\\\hspace*{0.35in}\fullmoon~~Prefer not to answer\\
\subsubsection{Participant Knowledge of DNS}
Before this survey have you ever heard of DNS?\\
You do not need to have any prior knowledge of DNS to take part in this survey.\\
\\\hspace*{0.35in}\fullmoon~~Yes, I have heard of DNS and I know what it does
\\\hspace*{0.35in}\fullmoon~~Yes, I have heard of DNS, but I do not know what it does
\\\hspace*{0.35in}\fullmoon~~No, I have not heard of DNS\\
\emph{If yes,} What do you think DNS does?\\
\subsubsection{Participants Interact with Browser Interface}
On the next page, you will shown a network setting that you could see in a browser on your laptop or desktop computer.\\
Please select the setting option that you would choose if you encountered this setting in your own browser.\\
The next button will appear after a few seconds to give you time to choose a setting option.\\
Please note that this page will not modify the settings in your browser.\\
\begin{figure}[t!]
\includegraphics[width=0.6\linewidth]{chrome_interface.png}
\caption{Chrome Browser Interface}
\label{fig:Chrome_Interface}
\end{figure}
\emph{Here, we only show the questions asked for the Chrome interface, but participants were shown either a Chrome, Edge, Firefox, Opera, or Brave interface.}
\emph{If secure DNS disabled,} Why did you choose to disable secure DNS?\\
\emph{If secure DNS enabled,} Why did you choose to enable secure DNS?\\
\emph{If current service provider chosen as resolver,} Why did you choose "your current service provider" as your resolver?\\
\emph{If resolver from dropdown chosen,} Why did you choose \{resolver name\} as your resolver?\\
\emph{If custom resolver chosen,} Why did you choose to use a custom resolver?\\
\emph{If custom resolver chosen,} Why did you choose \{custom resolver name\} as your custom resolver?\\
\subsubsection{General Questions about Setting Choices}
How certain are you that the setting you chose will allow you to access the Internet as you usually do?\\
\\\hspace*{0.35in}\fullmoon~~Very certain
\\\hspace*{0.35in}\fullmoon~~Somewhat certain
\\\hspace*{0.35in}\fullmoon~~Neither certain nor uncertain
\\\hspace*{0.35in}\fullmoon~~Somewhat uncertain
\\\hspace*{0.35in}\fullmoon~~Very uncertain\\
Why did you not choose any of the other setting options?\\
Regardless of your knowledge about this setting, what advantages do you think it might have?\\
Regardless of your knowledge about this setting, what disadvantages do you think it might have?\\
While you were choosing a setting, what additional information, if any, would you have wanted?\\
If you wanted to make a more informed choice about this setting where would you go to learn more? Please select all that apply\\
\\\hspace*{0.35in}\fullmoon~~Social media, e.g., Twitter or Facebook
\\\hspace*{0.35in}\fullmoon~~A security blog
\\\hspace*{0.35in}\fullmoon~~Ask a friend, family member, or coworker Look up the setting online
\\\hspace*{0.35in}\fullmoon~~A news outlet (e.g., TV, online)
\\\hspace*{0.35in}\fullmoon~~Reddit or other online forums
\\\hspace*{0.35in}\fullmoon~~Company website for your browser
\\\hspace*{0.35in}\fullmoon~~Other\\
\begin{figure}[!ht]
\includegraphics[width=0.6\linewidth]{DNS-Diagram-with-regions}
\caption{Diagram of DNS}
\label{fig:DNS_Diagram_Regions}
\end{figure}
Referring to the diagram above, at which places do you think your Internet Service Provider (the company that you pay for Internet access in your home) could know what websites you are visiting if the setting you saw on the last page was enabled and queries were sent to a single DNS resolver (regardless of what you actually chose)?\\
Please select all that apply.\\
If you are uncertain please just make your best guess.\\
\\\hspace*{0.35in}\fullmoon~~A
\\\hspace*{0.35in}\fullmoon~~B
\\\hspace*{0.35in}\fullmoon~~C
\\\hspace*{0.35in}\fullmoon~~D
\\\hspace*{0.35in}\fullmoon~~None of the above
\\\hspace*{0.35in}\fullmoon~~Don't know\\
Referring to the diagram above, at which places do you think a third party (someone not directly involved in you reaching a website) could know what websites you are visiting if the setting you saw on the last page\\
was enabled and queries were sent to a single DNS resolver (regardless of what you actually chose)?\\
Please select all that apply.\\
If you are uncertain please just make your best guess.\\
\\\hspace*{0.35in}\fullmoon~~A
\\\hspace*{0.35in}\fullmoon~~B
\\\hspace*{0.35in}\fullmoon~~C
\\\hspace*{0.35in}\fullmoon~~D
\\\hspace*{0.35in}\fullmoon~~None of the above
\\\hspace*{0.35in}\fullmoon~~Don't know\\
\subsubsection{Participants Interact with Mobile Interface}
On the next page, you will be shown a network setting that you could see on a mobile phone.\\
Please select the setting option that you would choose if you encountered this setting in your own browser.\\
The next button will appear after a few seconds to give you time to choose a setting option.\\
Please note that this page will not modify the settings on your phone.\\
\begin{figure}[!ht]
\includegraphics[width=0.6\linewidth]{mobile_custom.png}
\caption{Mobile Interface}
\label{fig:mobile_custom}
\end{figure}
\emph{If resolver from dropdown chosen,} Why did you select the \{resolver name\} setting for Private DNS?\\
\emph{If custom resolver chosen,} Why did you enter \{custom resolver name\} as your DNS Provider hostname?\\
\subsubsection{DNS Explained}
\begin{figure}[t!]
\includegraphics[width=0.6\linewidth]{DNS-Diagram-without-regions}
\caption{Diagram of DNS}
\label{fig:DNS_Diagram_No_Regions}
\end{figure}
DNS stands for Domain Name System.\\
It is the system that takes a human-readable name like \censor{noise.cs.uchicago.edu} and returns an Internet Protocol (IP) address, like \censor{128.135.24.19}.\\
Whenever your browser tries to contact a website (when you click on a link, when your browser loads an image, etc.), your browser sends a request over the Internet to a resolver to look up the IP address for that website allowing your computer find that site or resource.\\
A DNS request is like looking up the street address (e.g., \censor{5801 S Ellis Ave. Chicago, IL 60637}) of a particular location (e.g., \censor{The University of Chicago}).\\
Either can be used to refer to the place, but only the street address can be used to send a package.\\
Similarly, data on the Internet is sent between IP addresses (not between names).\\
Typically, DNS requests are not encrypted.\\
This means that your Internet Service Provider or any third-party on the path between your computer, laptop, phone, etc. and the other side of the communication may be able to see which websites you are about to access and potentially change what is returned.\\
What does DNS do?\\
\\\hspace*{0.35in}\fullmoon~~Encrypts all network traffic
\\\hspace*{0.35in}\fullmoon~~Optimizes the order of your search results
\\\hspace*{0.35in}\fullmoon~~Maps a website to an IP address\\
Normally these DNS requests are made in plain text, making it easy for your Internet Service Provider or third- parties to see what website you’re about to access.\\
When encrypted DNS is used, the request is encrypted, meaning that third parties and your Internet Service Provider can't see what website you are trying to access at this point.\\
However, encrypted DNS can be incompatible with parental controls and malware detectors, can make browsing the internet slower, and the place that translates the URL to an IP address (resolver) knows what site you are trying to reach.\\
\begin{figure}[t!]
\includegraphics[width=0.6\linewidth]{DNS-Diagram-with-regions}
\caption{Diagram of DNS}
\label{fig:DNS_Diagram_Regions1}
\end{figure}
Referring to the diagram above, at which places do you think your Internet Service Provider (the company that you pay for Internet access in your home) could know what websites you are visiting if the setting you saw on the last page was enabled and queries were sent to a single DNS resolver (regardless of what you actually chose)?\\
Please select all that apply.\\
If you are uncertain please just make your best guess.\\
\\\hspace*{0.35in}\fullmoon~~A
\\\hspace*{0.35in}\fullmoon~~B
\\\hspace*{0.35in}\fullmoon~~C
\\\hspace*{0.35in}\fullmoon~~D
\\\hspace*{0.35in}\fullmoon~~None of the above
\\\hspace*{0.35in}\fullmoon~~Don't know\\
Referring to the diagram above, at which places do you think a third party (someone not directly involved in you reaching a website) could know what websites you are visiting if the setting you saw on the last page was enabled and queries were sent to a single DNS resolver (regardless of what you actually chose)?\\
Please select all that apply.\\
If you are uncertain please just make your best guess.\\
\\\hspace*{0.35in}\fullmoon~~A
\\\hspace*{0.35in}\fullmoon~~B
\\\hspace*{0.35in}\fullmoon~~C
\\\hspace*{0.35in}\fullmoon~~D
\\\hspace*{0.35in}\fullmoon~~None of the above
\\\hspace*{0.35in}\fullmoon~~Don't know\\
Now that you have read about encrypted DNS, when it is enabled, how safe do you think while you are browsing the Internet?\\
\\\hspace*{0.35in}\fullmoon~~Very safe
\\\hspace*{0.35in}\fullmoon~~Somewhat safe
\\\hspace*{0.35in}\fullmoon~~Neither safe or unsafe
\\\hspace*{0.35in}\fullmoon~~Somewhat unsafe
\\\hspace*{0.35in}\fullmoon~~Very unsafe\\
Why?\\
\subsubsection{Participants Interact with Browser Interface Again}
On the next page, you will be shown the same network setting that you saw earlier in the survey.\\
Please select the setting option that you would choose if you encountered this setting in your own browser on your laptop or desktop computer given what you have read here today.\\
The next button will appear after a few seconds to give you time to choose a setting option.\\
Please note that this page will not modify the settings in your browser.\\
\begin{figure}[t!]
\includegraphics[width=0.6\linewidth]{chrome_interface}
\caption{Chrome Browser Interface}
\label{fig:Chrome_Interface2}
\end{figure}
\emph{If secure DNS disabled,} Why did you choose to disable secure DNS?\\
\emph{If secure DNS enabled,} Why did you choose to enable secure DNS?\\
\emph{If current service provider chosen as resolver,} Why did you choose "your current service provider" as your resolver?\\
\emph{If resolver from dropdown is chosen,} Why did you choose \{resolver name\} as your resolver?\\
\emph{If custom resolver is chosen,} Why did you choose to use a custom resolver?\\
Why did you choose \{custom resolver name\} as your custom resolver?\\
\subsubsection{Questions about Setting Choices}
Why did you not choose any of the other setting options?\\
How certain are you that the setting you chose will allow you to access the Internet as you usually do?\\
\\\hspace*{0.35in}\fullmoon~~Very certain
\\\hspace*{0.35in}\fullmoon~~Somewhat certain
\\\hspace*{0.35in}\fullmoon~~Neither certain nor uncertain
\\\hspace*{0.35in}\fullmoon~~Somewhat uncertain
\\\hspace*{0.35in}\fullmoon~~Very uncertain\\
\subsubsection{Encrypted DNS Preferences}
Prior to this survey, had you heard of encrypted DNS? You may have also heard it referred to as referred to as DNS over HTTPS (DoH), DNS over TLS (DoT), or Secure DNS.\\
\\\hspace*{0.35in}\fullmoon~~Yes
\\\hspace*{0.35in}\fullmoon~~No/Don't know\\
Where did you hear about encrypted DNS?\\
\\\hspace*{0.35in}\fullmoon~~A friend, family member, or coworker
\\\hspace*{0.35in}\fullmoon~~A security blog
\\\hspace*{0.35in}\fullmoon~~Social media, e.g., Twitter or Facebook
\\\hspace*{0.35in}\fullmoon~~Company website for your browser
\\\hspace*{0.35in}\fullmoon~~Reddit or other online forums
\\\hspace*{0.35in}\fullmoon~~A news outlet (e.g. TV, online)
\\\hspace*{0.35in}\fullmoon~~Other\\
\begin{figure}[t!]
\includegraphics[width=0.6\linewidth]{DNS-Diagram-with-regions}
\caption{Diagram of DNS}
\label{fig:DNS_Diagram_Regions2}
\end{figure}
Sometimes companies that host websites also run resolvers (e.g. B and D in the diagram above are run by the same company).\\
Please select the option that most closely reflects your preference.\\
\\\hspace*{0.35in}\fullmoon~~I would prefer that B and D were run by the same company.
\\\hspace*{0.35in}\fullmoon~~I would prefer that B and D were run by the different companies.
\\\hspace*{0.35in}\fullmoon~~Other
\\\hspace*{0.35in}\fullmoon~~No preference\\
There are many different companies that operate resolvers.\\
Please select the option that most closely reflects your preference when the companies that host websites you are trying to reach does not run a resolver.\\
\\\hspace*{0.35in}\fullmoon~~I would prefer that my computer always contact the same resolver (B) who would see all of my DNS requests.
\\\hspace*{0.35in}\fullmoon~~I would prefer that my computer distributes my requests over multiple resolvers (B) who would each see some of my DNS requests.
\\\hspace*{0.35in}\fullmoon~~Other
\\\hspace*{0.35in}\fullmoon~~Don't know\\
In the case that your chosen resolver does not support encrypted DNS or your encrypted DNS requests fail, please select the behavior that you would prefer.\\
\\\hspace*{0.35in}\fullmoon~~I would prefer to have my computer make an unencrypted query so the page can load.
\\\hspace*{0.35in}\fullmoon~~I would prefer to have my computer fail to reach the site/load the resource.
\\\hspace*{0.35in}\fullmoon~~Other
\\\hspace*{0.35in}\fullmoon~~No preference\\
Do you think this a setting you should be able to modify on your phone or computer browser?\\
\\\hspace*{0.35in}\fullmoon~~Yes
\\\hspace*{0.35in}\fullmoon~~No
\\\hspace*{0.35in}\fullmoon~~Don't know\\
Why?\\
\subsubsection{Trust in Specific DNS Providers}
\emph{Here, we only show the questions asked for Quad9, but participants were asked about either Quad9, Google, Cloudflare, NextDNS, or Comcast.}
Had you heard of Quad9 prior to this survey?\\
\\\hspace*{0.35in}\fullmoon~~Yes
\\\hspace*{0.35in}\fullmoon~~Maybe
\\\hspace*{0.35in}\fullmoon~~No\\
Please select the answer option that most accurately describes your agreement with the following statement.\\
I trust Quad9?\\
\\\hspace*{0.35in}\fullmoon~~Strongly agree
\\\hspace*{0.35in}\fullmoon~~Somewhat agree
\\\hspace*{0.35in}\fullmoon~~Neither agree nor disagree
\\\hspace*{0.35in}\fullmoon~~Somewhat disagree
\\\hspace*{0.35in}\fullmoon~~ Strongly disagree\\
\subsubsection{Trust in Providers and Browsers}
Please select the answer option that most accurately describes your agreement with the following statement. \\
I trust my Internet Service Provider (e.g. Xfinity, Time Warner, Spectrum)?\\
\\\hspace*{0.35in}\fullmoon~~Strongly agree
\\\hspace*{0.35in}\fullmoon~~Somewhat agree
\\\hspace*{0.35in}\fullmoon~~Neither agree nor disagree
\\\hspace*{0.35in}\fullmoon~~Somewhat disagree
\\\hspace*{0.35in}\fullmoon~~Strongly disagree
\\\hspace*{0.35in}\fullmoon~~I don't know who my Internet Service Provider is\\
Please select the answer option that most accurately describes your agreement with the following statement. \\
I trust my mobile service provider (e.g. AT\&T, Verizon, etc.)?\\
\\\hspace*{0.35in}\fullmoon~~Strongly agree
\\\hspace*{0.35in}\fullmoon~~Somewhat agree
\\\hspace*{0.35in}\fullmoon~~Neither agree nor disagree
\\\hspace*{0.35in}\fullmoon~~Somewhat disagree
\\\hspace*{0.35in}\fullmoon~~Strongly disagree
\\\hspace*{0.35in}\fullmoon~~I don't know who my mobile service provider is\\
Please select the answer option that most accurately describes your agreement with the following statement.\\
I trust my primary web browser (e.g. Chrome, Firefox, Safari)?\\
\\\hspace*{0.35in}\fullmoon~~Strongly agree
\\\hspace*{0.35in}\fullmoon~~Somewhat agree
\\\hspace*{0.35in}\fullmoon~~Neither agree nor disagree
\\\hspace*{0.35in}\fullmoon~~Somewhat disagree
\\\hspace*{0.35in}\fullmoon~~Strongly disagree\\
\subsubsection{Access Interfaces through Actual Browsers}
Which of the following browsers are you currently able to access on the device you are using to take this survey?\\
\\\hspace*{0.35in}\fullmoon~~Opera
\\\hspace*{0.35in}\fullmoon~~Chrome
\\\hspace*{0.35in}\fullmoon~~Microsoft Edge Brave
\\\hspace*{0.35in}\fullmoon~~Firefox
\\\hspace*{0.35in}\fullmoon~~None of the above\\
\emph{Here, we only show the questions asked for the Chrome interface, but participants were asked about either Chrome, Edge, Firefox, Opera, or Brave.}
Please open a new Chrome browser window and navigate to your browser's settings page by following the written instructions or video below:\\
At the top right of the window, click More\\
Click Settings\\
Under Privacy and security select Security\\
\emph{Please note that these instructions may not be compatible with old versions of Chrome.}
\emph{\{Video with instructions shown below\} }\\
Were you able to reach this page?
\\\hspace*{0.35in}\fullmoon~~Yes
\\\hspace*{0.35in}\fullmoon~~No\\
Can you find "Use secure DNS" on this page?
\emph{If you are using an old version of Chrome the secure DNS setting might not be available.}
\\\hspace*{0.35in}\fullmoon~~Yes
\\\hspace*{0.35in}\fullmoon~~No
Near the bottom of this page you should see an option called ''Use secure DNS''. Is this setting enabled?
\\\hspace*{0.35in}\fullmoon~~Yes
\\\hspace*{0.35in}\fullmoon~~No
\\\hspace*{0.35in}\fullmoon~~Don't know\\
\emph{If yes,} Which setting is selected under ''Use secure DNS?''
\\\hspace*{0.35in}\fullmoon~~With your current service provider
\\\hspace*{0.35in}\fullmoon~~With ---\\
\emph{If "With ---" is selected,} What DNS provider is selected?
\\\hspace*{0.35in}\fullmoon~~Custom
\\\hspace*{0.35in}\fullmoon~~CleanBrowsing (Family Filter)
\\\hspace*{0.35in}\fullmoon~~OpenDNS
\\\hspace*{0.35in}\fullmoon~~Quad9 (9.9.9.9)
\\\hspace*{0.35in}\fullmoon~~Cloudflare (1.1.1.1)
\\\hspace*{0.35in}\fullmoon~~NextDNS
\\\hspace*{0.35in}\fullmoon~~Google (Public DNS)\\
Do you remember seeing the "secure DNS" setting in Chrome prior to this survey?
\\\hspace*{0.35in}\fullmoon~~Yes
\\\hspace*{0.35in}\fullmoon~~No
\\\hspace*{0.35in}\fullmoon~~Don't Know\\
\emph{If yes,} Do you remember changing the "secure DNS" setting in Chrome?
\\\hspace*{0.35in}\fullmoon~~Yes
\\\hspace*{0.35in}\fullmoon~~No
\\\hspace*{0.35in}\fullmoon~~Don't Know\\
\emph{If yes,} Why did you decide to change the "secure DNS" setting in Chrome?\\
\emph{If no,} Why did you decide to not change, the "secure DNS" setting in Chrome?
\subsubsection{Access Interfaces through Mobile Device}
What is the operating system of your primary smartphone?\\
\\\hspace*{0.35in}\fullmoon~~Android
\\\hspace*{0.35in}\fullmoon~~iOS (Apple device)
\\\hspace*{0.35in}\fullmoon~~Other
\\\hspace*{0.35in}\fullmoon~~I do not use a smartphone\\
What version of iOS does your primary smartphone run?\\
The version of iOS can be found by going to Settings > General > About on your phone.\\
What version of Android does your primary smartphone run?\\
The version of Android can be found by going to Settings > System > Advanced > System update on your phone.\\
\emph{Here, we only show the questions asked for Android devices, but participants were asked about either iOS or Android.}
Open your phone's Settings app.\\
Tap Network \& internet > Advanced > Private DNS.\\
Were you able to make it to this page?\\
\\\hspace*{0.35in}\fullmoon~~Yes
\\\hspace*{0.35in}\fullmoon~~No\\
What is you current setting on this page?\\
\\\hspace*{0.35in}\fullmoon~~Off
\\\hspace*{0.35in}\fullmoon~~Automatic
\\\hspace*{0.35in}\fullmoon~~Private DNS Provider hostname\\
Do you remember visiting this page before now?\\
\\\hspace*{0.35in}\fullmoon~~Yes
\\\hspace*{0.35in}\fullmoon~~No/Don't know\\
Do you remember changing this setting?\\
\\\hspace*{0.35in}\fullmoon~~Yes
\\\hspace*{0.35in}\fullmoon~~No/Don't know\\
If you changed the setting or not, why did you make that decision?\\
\subsubsection{Demographics}
What is your age?\\
\\\hspace*{0.35in}\fullmoon~~18 - 24
\\\hspace*{0.35in}\fullmoon~~25 - 34
\\\hspace*{0.35in}\fullmoon~~35 - 44
\\\hspace*{0.35in}\fullmoon~~45 - 54
\\\hspace*{0.35in}\fullmoon~~55 - 64
\\\hspace*{0.35in}\fullmoon~~65 - 74
\\\hspace*{0.35in}\fullmoon~~75 - 84
\\\hspace*{0.35in}\fullmoon~~85 or older
\\\hspace*{0.35in}\fullmoon~~Prefer not to answer\\
What is your gender?\\
\\\hspace*{0.35in}\fullmoon~~Woman
\\\hspace*{0.35in}\fullmoon~~Man
\\\hspace*{0.35in}\fullmoon~~Non-Binary
\\\hspace*{0.35in}\fullmoon~~Prefer not to disclose
\\\hspace*{0.35in}\fullmoon~~Prefer to self-describe\\
What is the highest level of school you have completed or the highest degree you have received?\\
\\\hspace*{0.35in}\fullmoon~~Less than high school
\\\hspace*{0.35in}\fullmoon~~High school graduate
\\\hspace*{0.35in}\fullmoon~~Some college
\\\hspace*{0.35in}\fullmoon~~2 year degree
\\\hspace*{0.35in}\fullmoon~~4 year degree
\\\hspace*{0.35in}\fullmoon~~Professional degree
\\\hspace*{0.35in}\fullmoon~~Doctorate
\\\hspace*{0.35in}\fullmoon~~Prefer not to answer\\
Which of the following best describes your educational background or job field?\\
\\\hspace*{0.35in}\fullmoon~~I have an education in, or work in, the field of computer science, computer engineering or IT
\\\hspace*{0.35in}\fullmoon~~I DO NOT have an education in, nor do I work in, the field of computer science, computer engineering or IT4
\\\hspace*{0.35in}\fullmoon~~Prefer not to answer\\
\subsubsection{Feedback}
If you have any further feedback, questions, comments, concerns, or anything else you want to tell us, please leave a comment below.
|
\section{Introduction}
Blockchain, featured with its decentralized tamper-resistance, has been an active research topic to address the security issue of centralization, and has been widely adopted in many scenarios such as Internet-of-Things (IoT)~\cite{iot-1,iot-2,iot-3,iot-4,iot-5,iot-6}. A public Blockchain, often referred to as \textit{Token Chain}, can accomplish a decentralized tokenomics. Since \textit{Token Chains} are usually public, they require decentralized consensus algorithms, such as Proof of Work (PoW)~\cite{nakamoto2008bitcoin}. However, a decentralized consensus algorithm entails the comparatively low performance and efficiency, implying that those delay-sensitive applications or applications with massive deployment and data, such as IoT applications, cannot operate on \textit{Token Chains} properly. For such applications, consortium chains are proposed with partially decentralized and more efficient consensus algorithms, such as Byzantine Faulty Tolerance (BFT)~\cite{castro1999practical}.
However, \textit{Side Chains} have no economic value, which prevents the trading of IoT data.
We propose a cross-chain architecture consisting of a public \textit{Token Chain} and multiple associated consortium \textit{Side Chains}. In the cross-chain system, the \textit{Token Chain} issues the main tokens and maintains the official token ledger for the cryptocurrency ecosystem, while a number of \textit{Side Chains} import the tokens from the \textit{Token Chain} for the management of the massive records and transactions for applications, such as IoT. The key challenge is the design of cross-chain interactions, including token exchange between the \textit{Token Chain} and its associated \textit{Side Chains}. A set of nodes are defined as \textit{Witnesses} in the genesis block of each \textit{Side Chain}. These \textit{Witnesses} are responsible for the exchange of tokens between the \textit{Token Chain} and the \textit{Side Chains}. \\
The cross-chain system provides the flexibility to meet various capacity demands of heterogeneous applications, including but not limited to IoT applications. It achieves modularity in that the \textit{Side Chains} are decoupled from the \textit{Token Chain} through the design of \textit{Witnesses}. This system is not limited to Ethereum platform, and has no limitations on the number of \textit{IoT Chains}.\footnote{The theoretical maximum number of IoT chains is $16^{40}\approx 1.46\times 10^{48}$ based on the 20-byte address. It is a quite large number to exceed in practice.}
\section{Related Works}
There exist several cross-chain technologies, mainly classified into Relay, Hash-Lock and Notaries.
\begin{itemize}
\item Relay: Chain $A$ can prove the facts and behaviors on Chain $B$. Technologies using Relay are including but not limited to, the following solutions:
\begin{itemize}
\item BTC-Relay~\cite{hash-lock}. It is a technology that can inter-operate between Bitcoin and Ethereum. A light client that introduces Simple Payment Verfication (SPV) is implemented by a smart contract in Ethereum so that facts on Bitcoin can be proved in Ethereum. However, it also implies that Ethereum needs to store all headers of Bitcoin, which is quite expensive.
\item Non-Interactive Proof of PoW~\cite{nipopow}. Combined with Monte Carlo method and Bisection method, this proof tries to calculate the probability that a chosen puzzle with a high difficulty falls into a specific interval. As such the PoW on Chain $B$ can be proved by including this proof of PoW on Chain $A$. However, currently this method is not general and strictly dependent from PoW consensus.
\end{itemize}
\item Hash-Lock~\cite{hash-lock}. It acts as a personal trading platform. Fair and secure assets exchange can be accomplished with the aid of a time-out mechanism. However, it does not support assets portability and it is unfriendly to the interoperation between a public chain and a consortium chain because of the possible different token-scheme on this consortium chain.
\item Notaries~\cite{notary}. It is also known as \textit{Witnesses}. \textbf{Our invention belongs to this type.} In words, there exist some nodes which run clients on both Chain $A$ and $B$. Messages (Assets) exchanged is conducted through these nodes with a necessary consensus process. Notaries is the most frequently used technology so far. There are a number of Blockchain projects announcing their schemes, including but not limited to,
\begin{itemize}
\item Hcash. It is a typical design of Notaries-liked cross-chain technology. It collectively features most of the advantages of Notaries from other similar projects. We believe that comparing our invention with Hcash will be sufficiently significant to differential our invention with other similar designs. In our opinion, notaries act as a set of nodes running on both chains, as if multiple personal trading platforms exist with a necessary consensus process. The On/Off-chain multi-signature wallet takes critical responsibilities when assets exchanging. However, this design does require users to have assets prior to the process of exchanging, which implies that similar to Hash-Lock, it does not support consortium chain very well either. Hcash does not provide the solutions to address the implications of the native gas-oriented account-balance on the consortium chain, and to generate exchangeable token for a consortium chain with the native gas-oriented account-balance.
\item Walton Chain. Notaries are a set of nodes only running on the local chain. The request of generating a new side chain is executed by a smart contract by sending a transaction. In other words, notaries are the bridge to exchange the messages of the most external layer to the internal layer with a necessary consensus process. It implies that a side chain is only an abstractive structure in Walton Chain. Link structure is maintained on smart contracts in which the data storage is simplified as the state-oriented data to be stored in the local database. Such a design is not only strongly restricted by the throughput of the most external chain but also extremely increases the loading of the entire network.
\end{itemize}
\end{itemize}
\emph{There appears an existing national patent (CN 106447309 A) in China that is similar to our invention. Their patent implements the third-party chain that pegs to both the main chain and the side chain with a certain address on each chain as if an accessible API connects the main chain and the side chain. It implies that these APIs have an offline information relaying between the certain address on either the main chain or the side chain, and its third-party chain, which they did not precisely point out. In addition, the existence of the third-party chain exposes the issue of data privacy. It also makes no sense that the third-party is eligible to maintain multiple side chains due to the data privacy, centralization and scalability.} Therefore, we can conclude that our invention is more general and outplays their patent.
Our invention is targeting at such a case that multiple \textit{IoT Chains} with different consensus algorithms, a high throughput and a high volume are anchored with a \textit{Token Chain}. Therein each \textit{IoT Chain} has its own business groups charged with the audition of cross-chain process. One \textit{IoT Chain} can be totally a black-box for the others. Likewise, any behaviors on an \textit{IoT Chain} can be also a black-box for the \textit{Token Chain} apart from the amount of token exchanged. Either an \textit{IoT Chain} without the native gas or an \textit{IoT Chain} withe the native gas can be registered to the \textit{Token Chain} whenever a user requires token involved and afterwards provides the sufficient assets and proofs on the \textit{Token Chain}. In words, we believe that our invention can be a better solution to the separation between the financial-layer and data-layer in order for a more data-friendly and healthy Blockchain system.
\section{Brief Description of the Drawings}
The description of each step and module are shown in the following drawings.
\begin{enumerate}
\item Figure \ref {cross} shows the flow diagram for the entire architecture of our invention. It combines the rest of the Figures into a whole to elaborate the entire flow of the procedure.
\item Figure \ref {cross_1} shows the flow diagram for the pre-defined parameters required to be setup prior to the registration to \textit{Token Chain} for an \textit{IoT Chain} without the native gas.
\item Figure \ref {cross_2} shows the flow diagram for the first time registration to \textit{Token Chain} for an \textit{IoT Chain} without the native gas from the perspective of \textit{Token Chain}.
\item Figure \ref {cross_3} shows the flow diagram for the first time registration to \textit{Token Chain} for an \textit{IoT Chain} without the native gas from the perspective of \textit{IoT Chain}.
\item Figure \ref {cross_4} shows the flow diagram for transferring consumable token from \textit{IoT Chain} back to \textit{Token Chain} from the perspective of \textit{IoT Chain}.
\item Figure \ref {cross_5} shows the flow diagram for transferring consumable token from \textit{IoT Chain} back to \textit{Token Chain} from the perspective of \textit{Token Chain}.
\item Figure \ref {cross_6} shows the flow diagram for subsequently transferring from \textit{Token Chain} to \textit{IoT Chain} that has been registered from the perspective of \textit{Token Chain}.
\item Figure \ref {cross_7} shows the flow diagram for subsequently transferring from \textit{Token Chain} to \textit{IoT Chain} that has been registered from the perspective of \textit{IoT Chain}.
\item Figure \ref {cross_8} shows the flow diagram for transferring between \textit{Token Chain} and an \textit{IoT Chain} with the native gas.
\end{enumerate}
\section{Brief Description of the Smart Contracts and Witnesses}
\begin{itemize}
\item \textit{SC\_A}: A dedicated cross-chain smart contract for an \textit{IoT Chain \{A\}}, deployed on \textit{Token Chain}. It locks/unlocks the assets transferred. It introduces the consensus for the Validation of registration from witnesses, as well as the consensus for the assets transferred back to \textit{Token Chain}.
\item \textit{SC\_ID}: It is deployed on \textit{Token Chain}. A \textit{Token Chain} can only have one \textit{SC\_ID} and it offers 2-of-2 Mulsig (Note that it can be a n-of-m Mulsig that has been claimed in the last section. For the rest of place using Mulsig, 2-of-2 is implemented in order for the simple elaboration). It records the $Chain\_ID$ of \textit{IoT Chains} that have been successfully registered. It $Updates$ the list of $Chain\_ID$ once the consensus for the Validation of registration from witnesses gets passed and emits an event of $ExistOrNot$.
\item \textit{SC\_Register}: A unique dedicated cross-chain smart contract for the first time registration, deployed on an \textit{IoT Chain}. It introduces the consensus for the $entrance fee$ transferred to the \textit{IoT Chain}. Its balance is hardcoded and will be run out after the success of registration and commits suicide afterwards by the agreement of Mulsig.
\item \textit{SC\_Inter}: A unique dedicated cross-chain smart contract for the assets transferring back to \textit{Token Chain} as well as the subsequent assets transferred to \textit{IoT Chain} with a consensus process, deployed on \textit{IoT Chain}. It dedicates the assets unlocking to \textit{SC\_Bank} after the success of consensus. It forwards the assets transferred from users' accounts to \textit{SC\_Bank} for assets locking in the case of \textit{IoT-Chain}-to-\textit{Token-Chain}.
\item \textit{SC\_Bank}: A unique dedicated cross-chain smart contract for assets locking/unlocking, deployed on \textit{IoT Chain}. The balance of \textit{SC\_Bank} that equals to the total amount of the upper bound of the potential assets trading on \textit{IoT Chain} is hardcoded in the genesis block. The corresponding amount of token is unlocked and transferred to the destination address with the delegation from \textit{SC\_Inter}. It receives and locks the assets transferred from users' accounts to \textit{SC\_Bank} for assets locking in the case of \textit{IoT-Chain}-to-\textit{Token-Chain}.
\item \textit{SC\_Consensus}: A unique dedicated cross-chain smart contract for both registration and consensus in the context of an \textit{IoT Chain} with the native gas.
\item \textit{SC\_Trading}: A unique dedicated cross-chain smart contract that records the balance for all accounts that have ever been involved in a cross-chain transaction and dynamically maintains a ledger. Both intra-transactions and inter-transactions associated with real business must be manipulated via \textit{SC\_Trading}. In other words, the token value is generated and destroyed with \textit{SC\_Trading} running on an \textit{IoT Chain} with the native gas. The amount of consumable assets is precisely defined even though its native account-balance tends to be infinite.
\item \textit{Witnesses}: A set of nodes that bridge the information exchanged between \textit{Token Chain} and a specific \textit{IoT Chain}. Information exchanged needs to be verified by the consensus on both \textit{Token Chain} and the \textit{IoT Chain}. \textit{Witnesses} retrieve events from:
\begin{enumerate}
\item \textit{SC\_ID}: if the $Chain\_ID$ has been appended in the list;
\item \textit{SC\_A}: if some numbers of cross-chain transactions have arrived;
\item \textit{SC\_Bank}: if the assets has been successfully locked in the context of \textit{IoT-Chain}-to-\textit{Token-Chain};
\item \textit{SC\_Consensus}: if some numbers of cross-chain transactions have arrived.
\end{enumerate}
\end{itemize}
\section{Detailed Description}
A set of pre-defined \textit{Witnesses} exchange the cross chain messages as if a bridge connects \textit{Token Chain} and an \textit{IoT Chain}. Remark that multiple \textit{IoT Chains} can be implemented along with their own pre-defined \textit{Witnesses}.
All \textit{IoT Chains} are pegged to \textit{Token Chain}, which provides an individual channel for each \textit{IoT Chain} to exchange token. \textit{IoT Chains} do not interact with each other directly.
In addition, businesses tend to customized their own \textit{IoT Chains} regardless of a specific type of consensus algorithm or even the existence of native gas, i.e. either of an \textit{IoT Chain} with native gas or an \textit{IoT Chain} without native gas. Note that the native gas refers to the native gas-oriented account-balance. Also note that the native gas-oriented account-balance is firstly implemented in Ethereum platform. It is often used to restrict the malicious transactions flooding among the network. The difference between them is shown as follows,
\begin{itemize}
\item \textbf{\textit{IoT Chain} without the native gas: }an \textit{IoT Chain} has no native gas-oriented account-balance. Nodes on such an \textit{IoT Chain} cannot send transactions without the imported gas transferred from the corresponding \textit{Token Chain}, where the imported gas, in fact, is consuming the main tokens. In other words, the main tokens are responsible for both sending a transaction and dealing with a business issue.
\item \textbf{\textit{IoT Chain} with the native gas: }an \textit{IoT Chain} has the native gas-oriented account-balance. Nodes on such an \textit{IoT Chain} can send transactions (uploading the IoT messages) without the imported gas transferred from the corresponding \textit{Token Chain}. However, any transactions associated with the business issue must come along with the imported gas.
\end{itemize}
\textit{IoT Chains} are free to be customized with low restrictions in terms of the consensus algorithm and the existence of native gas. All tokens on \textit{IoT Chains} are originated from \textit{Token Chain}, referring to the main tokens. In other words, \textit{IoT Chains} have no original token value regardless of the presence of the native gas.
Note that \textit{IoT Chain} is allowed to have its own account-balance mechanism to maintain the whole system running fluently with a necessary access control and a flow control. This account-balance is meant to be a sufficiently large amount of value that can be automatically recharged by administrators, which implies that only the token originated from \textit{Token Chain} that is strictly differentiated from the native account-balance on \textit{IoT Chains}, has the real economical value. Thus users can choose to either registering an \textit{IoT Chain} without the native gas in which only the token originated from \textit{Token Chain} dominates the system, or an \textit{IoT Chain} with the native gas in which its own account-balance mechanism is in charge of its self-driven processing, which are both independent to the consensus algorithm.
For the rest of this section we extend the description to a deeper extent shown as follows,
\begin{itemize}
\item Section 5.1 refers to Figure \ref {cross_1}, \ref {cross_2}, \ref {cross_3}, \ref {cross_4}, \ref {cross_5}, \ref {cross_6} and \ref {cross_7}.
\begin{itemize}
\item Section 5.1.1 refers to Figure \ref {cross_1}.
\item Section 5.1.2 refers to Figure \ref {cross_2} and \ref {cross_3}.
\item Section 5.1.3 refers to Figure \ref {cross_6} and \ref {cross_7}.
\item Section 5.1.4 refers to Figure \ref {cross_4} and \ref {cross_5}.
\end{itemize}
\item Section 5.2 refers to Figure \ref {cross_8};
\item Section 5.3 describes the mechanism of the reverting and resending process of our invention when the consensus fails.
\end{itemize}
Note that the entire flow chart refers to Figure \ref {cross}.
\subsection{Bridging a Side Blockchain without the native gas}
Bridging a side Blockchain without the native gas includes the following steps,
\begin{enumerate}
\item pre-definition of necessary parameters;
\item registration of the chain;
\item subsequent assets transferring:
\begin{itemize}
\item transferring from \textit{IoT Chain} back to \textit{Token Chain};
\item transferring from \textit{Token Chain} to \textit{IoT Chain} after the registration.
\end{itemize}
\end{enumerate}
\subsubsection{Pre-definition}
A set of \textit{Witnesses} run as intermediaries to bridge \textit{Token Chain} and an specific \textit{IoT Chain}. These \textit{Witnesses} have valid accounts on both \textit{Token Chain} and the \textit{IoT Chain}. The interactions between \textit{Token Chain} and the \textit{IoT Chain} are secured by Blockchain. Only interactions verified and recorded in the form of transactions are accepted and performed on both sides. Smart contracts implement the interactions between \textit{Token Chain} and the \textit{IoT Chain}.
The genesis block of the \textit{IoT Chain} has a pre-defined form. The genesis block defines the account states and the interaction parameters of the \textit{IoT Chain}. In the genesis block, all accounts on the \textit{IoT chains} have no balance when the \textit{IoT Chain} initializes, except that two pre-defined smart contracts created by the creator of the \textit{IoT Chain} stating the balance of these smart contracts and promises to lock \textbf{\textit{some fee}} (shown below) from these addresses, respectively. Each of them is verified by the signature of its own key.
\begin{enumerate}
\item the first kind of fee is assigned as the \textit{entrance fee} that is the minimum amount of token required to initialize an interactive \textit{IoT Chain} for the registration phase. \emph{Note that the \textit{entrance fee} should be minimum-bounded to prevent from maliciously frequent registration requests};
\item the second kind of fee is assigned as the \textit{reserved assets} from which the subsequent assets can be unlocked.
\end{enumerate}
The \textit{IoT Chain} predefines $N$ \textit{Witnesses} in the genesis block of the \textit{IoT Chain}. These $N$ \textit{Witnesses} should have valid accounts on \textit{Token Chain}.
In other words, a valid genesis block of an interactive \textit{IoT Chain} without the native gas has the following fields:
\begin{itemize}
\item $Chain\_ID$: The Chain ID of the \textit{IoT Chain};
\item $SC\_Register$: An interactive smart contract on the \textit{IoT Chain} which enables transfer function between the \textit{IoT Chain} and \textit{Token Chain}.
\item $Bal\_Resv$: The amount of token that the creator reserves in $SC\_Register$ for the \textit{IoT Chain};
\item $SC\_Inter$: A smart contract on the \textit{IoT Chain}, used for subsequent transferring from \textit{Token Chain} to the \textit{IoT Chain} and transferring from the \textit{IoT Chain} back to \textit{Token Chain};
\item $SC\_Bank$: A smart contract on the \textit{IoT Chain} which enables the subsequent token locking and unlocking. There is a strict access control upon this smart contract, i.e. the only address who can interact with $SC\_Bank$ is the address of $SC\_Inter$.
\item $Bal\_Bank$: The amount of token that the creator reserves in $SC\_Bank$ for the \textit{IoT Chain}; Note that $Bal\_Resv + Bal\_Bank \equiv$ total amount of token on \textit{Token Chain} including those unmined token;
\item $Wit\_Addr\_List$: The addresses of \textit{Witnesses} for the \textit{IoT Chain}, which is hardcoded in both $SC\_Register$ and $SC\_Inter$.
\end{itemize}
The creator creates and broadcasts the genesis block in the IoT network and await the other nodes being launched and ready for the the new \textit{IoT Chain} within an expected period. Note that any transactions with a source address of one of the \textit{Witnesses} and a destination address of \textit{SC\_Register} or \textit{SC\_Inter} can pay a zero gasprice. Meanwhile, any transactions with a source address of either \textit{SC\_Register}, \textit{SC\_Inter} or \textit{SC\_Bank} can also pay a zero gasprice.
\subsubsection{Registration}
After all \textit{Witnesses} have been ready (All \textit{Witnesses} should then be able to send transactions to and retrieving events from any targeted addresses. Meanwhile, they all are able to fetch the json files of both of the genesis blocks of the \textit{IoT Chain} and \textit{Token Chain}), the creator on \textit{Token Chain} sends a request of registration in the form of a transaction to \textit{SC\_A}, where the following contents are included in the invoked event,
\begin{itemize}
\item all token that is contained in this transaction sent to \textit{SC\_A} subtracting the compensation fee for both \textit{SC\_A} and \textit{Witnesses} providing the transferring service. It guarantees that \textit{SC\_A} and \textit{Witnesses} have sufficient fund to send transactions on \textit{Token Chain};
\item $Chain\_ID$ of the \textit{IoT Chain} that the creator proposes.
\end{itemize}
\textit{Note that \textit{SC\_A} is a smart contract deployed on \textit{Token Chain} with no balance initially, which is a dedicated cross-chain service relay and created by the service provider in the form of \textbf{open source}.}
As such the process of validation on \textit{Witnesses} gets called by retrieving the events invoked by this requesting transaction.
It validates the genesis block according to Alg. \ref{alg: verify IoT chain} from which the output is sent to \textit{SC\_A}.
In other words, once a \textit{Witness} validates the genesis block, it accepts the registration request of the new \textit{IoT Chain} and sends a ``Confirm'' transaction to the smart contract, \textit{SC\_A}, on \textit{Token Chain}.
When \textit{SC\_A} receives ``Confirm'' transactions from more than $N/2$ \textit{Witnesses} (\textit{Note that for any smart contracts on which a consensus process is being conducted, a parameter $round$ must be used to prevent from out-of-order message.
The minimum faulty tolerance is $N/2$}), the smart contract \textit{SC\_A} reserves $Bal\_Resv$ on \textit{Token Chain} from the account of the creator, triggers the smart contract \textit{SC\_ID} to register the \textit{IoT Chain} with its $Chain\_ID$.
Note that Alg. \ref{alg: verify IoT chain} has ensured that the $Chain\_ID$ for each chain is unique in the entire system.
A signal indicating whether the registration has succeeded or not is contained in the event invoked by $Chain\_ID$ that \textit{Witnesses} are listening to.
If and only if a ``True'' is received, a \textit{Witness} sends a ``Transferring'' transaction to the smart contract $SC\_Register$ on the \textit{IoT Chain}.
If $SC\_Register$ receives ``Transferring'' transaction from more than $N/2$ \textit{Witnesses}, the same amount (this amount should now exactly equal to the total balance of $SC\_Register$, ensuring that the balance of $SC\_Register$ will be zero after unlocking) of token are unlocked from $SC\_Register$ on \textit{IoT Chain}.
$SC\_Register$ commits suicide by the owners after the owners both (2-of-2 Mulsigs) have retrieved the event of success of adding the $Chain\_ID$ in $SC\_ID$.
After the \textit{IoT Chain} is created and registered, the \textit{IoT Chain} will be able to work properly and generate blocks based on its consensus protocol. \emph{Note that there is no block rewards on the \textit{IoT Chain}}.
\subsubsection{Token Chain to IoT Chain after Registration}
$SC\_Inter$ is responsible for any subsequent ``Transferring'' transactions sent from \textit{Token Chain} to a specific \textit{IoT Chain}, plus those sent from the \textit{IoT Chain} back to \textit{Token Chain}. If a subsequent ``Transferring'' transaction sent to $SC\_Inter$ is detected, \textbf{$SC\_Inter$ delegates the burden to $SC\_Bank$ along with $to$ and $value$ of each cross-chain transaction after the success of $N/2$ consensus}. This process succeeds if and only if it passes the validation that the current balance of $SC\_Bank$ + the \textit{entrance fee} + the amount of token that is requested for this specific ``Transferring'' transaction $\equiv$ total amount of token on \textit{Token Chain} including those unmined token. The owners of $SC\_Bank$ add the address of $SC\_Inter$ in $SC\_Bank$ and the corresponding amount of token can be unlocked if and only if both owners agree. \emph{Note that it is none of business of $SC\_ID$ for a subsequent cross-chain transferring. \textit{Witnesses} do not listen to the events invoked by $SC\_ID$. Instead, events of $SC\_A$ are listened.}
\subsubsection{IoT Chain to Token Chain after Registration}
Suppose the aforementioned \textit{IoT Chain} is still discussed, there are two types of transactions on the \textit{IoT Chain}: one records the IoT events and does not interact with \textit{Token Chain}; the other one records the interactions with \textit{Token Chain}. Any transaction that is associated with $SC\_Inter$ belongs to Interaction transactions.
Suppose that a user on the \textit{IoT Chain} sends a request of transferring token back to \textit{Token Chain} in the form of a transaction sent to $SC\_Inter$ with an arbitrary amount of token $\le$ (total amount of token on the \textit{Token Chain} including those unmined token $-$ the balance of $SC\_Bank$). $SC\_Inter$ then becomes the relay and transfers the token to $SC\_Bank$ for the further locking. Each of the
\textit{Witnesses} retrieves the events and sends a ``Transferring'' transaction to $SC\_A$ on \textit{Token Chain}. If $SC\_A$ receives a ``Transferring'' transaction from more than $N/2$ \textit{Witnesses}, the same amount (this amount should $\le$ the balance of $SC\_A$) of token is unlocked from $SC\_A$ on \textit{Token Chain}.
\begin{algorithm}
\caption{Verify the Request of Registering an IoT Chain} \label{alg: verify IoT chain}
\LinesNumbered
\KwIn{
$\quad$$Hash(Genesis)$: the hash of the IoT genesis block. \\
$\quad$$Balance_{request}$: the total balance that the creator reserves for the IoT chain. \\
$\quad$$Balance_{json}$: the total balance that the json file of the IoT chain is showing.\\
$\quad$ $ChainID_{request}$: the chain ID that the creator proposes for the IoT chain. \\
$\quad$ $ChainID_{json}$: the chain ID that the json file of the IoT chain is showing.\\
$\quad$ $ChainID_{token}$: the chain ID of the token chain.\\
$\quad$ $SmartContract_{ChainID}$: a list of all chain IDs that have been successfully registered already.}
\KwOut{True or False
\textbf{The witnesses assert the following conditions:}
Confirm the integrity of the Genesis Block:
$\quad$$GetHash(Genesis)$ = $Hash(Genesis)$;
Confirm it is a request to generate a new IoT Chain:
$\quad$$GetHeight()$ = 0;
Verify the balance of the creator:
$\quad$$Balance_{request}$ = $Balance_{json}$
Validate the Chain ID:
$\quad$$ChainID_{request}$ = $ChainID_{json}$
$\quad$$ChainID_{request}$ $\neq$ $ChainID_{token}$
$\quad$$ChainID_{request}$ $\notin$ $SmartContract_{ChainID}$
If ALL above conditions are satisfied, the output is ``TRUE'', which indicates the witness validates the registration request; or else the output is ``FALSE'', which indicates the witness refuses the registration request.
\end{algorithm}
\subsection{Bridging a Side Blockchain with the native gas}
Comparing with bridging an \textit{IoT Chain} without the native gas, there are only a few tiny changes to bridge an \textit{IoT Chain} with the native gas, which is shown as follows,
\begin{enumerate}
\item \textit{SC\_Register} and \textit{SC\_Inter} merge into one smart contract, \textit{SC\_Consensus}, since there is no need to transfer token itself. Instead, the only thing needs to do is to update the balance in \textit{SC\_Trading}. \textit{SC\_Consensus} has two hardcoded owners, offering 2-of-2 MultiSig;
\item A set of \textit{Witnesses} is still needed to be organized and written in \textit{SC\_Consensus}. Note that there is no need to predefine \textit{SC\_Consensus} in the genesis block as long as it provides sufficient native account-balance for self-driven processing. Also, there is no need to set an \textit{entrance fee} in \textit{SC\_Consensus};
\item \textit{SC\_Bank} can be deprecated;
\item It is still required to have a smart contract running on \textit{Token Chain}, \textit{SC\_A} created by the service provider in the form of \textbf{open source}. However, the events fetched by \textit{Witnesses} for the Validation only contain the information of how much is transferred and the \textit{Chain\_ID}. Accordingly, \textit{Witnesses} only validate the \textit{Chain\_ID} part in \textit{Algorithm 1} regardless of the height and balance of the \textit{IoT Chain};
\item As if a trading platform manipulate the financial trading, \textit{SC\_Trading} records the balance for all accounts that have ever been involved in an arbitrary cross-chain transaction and dynamically maintains a ledger based on the historical balance. Both intra-transactions and inter-transactions associated with real business must be manipulated via \textit{SC\_Trading}. In other words, the token value is generated and destroyed with \textit{SC\_Trading} running on an \textit{IoT Chain} with the native gas. In the context of high throughput of the \textit{IoT Chain} along with its native account-balance mechanism, it is believed that \textit{SC\_Trading} is able to handle the transactions in a high volume. Note that the following statement must be satisfied,
$$\left(\sum_{i=0}^{N}Bal_i\right)_{ChainID}=Asset_{ChainID},$$
where the total amount of token that is being locked in \textit{SC\_A} with a certain $Chain\_ID$ must equal to the sum of the balance of all accounts recorded in \textit{SC\_Trading} on an \textit{IoT Chain} with the same $Chain\_ID$. It has two hardcoded owners, offering 2-of-2 MultiSig.
\item Either of \textit{SC\_Consensus} or \textit{SC\_Trading} can be pre-defined in the genesis block without having the owners and offering the Mulsig if the r \textit{IoT Chain} with the native gas is at the height of zero when initializing. As such a valid genesis block of an \textit{IoT Chain} with the native gas has the following fields:
\begin{itemize}
\item $Chain\_ID$: The Chain ID of the \textit{IoT Chain};
\item $SC\_Register$: An interactive smart contract on the \textit{IoT Chain} which enables transfer function between the \textit{IoT Chain} and the \textit{Token Chain};
\item $SC\_Trading$: An smart contract on the \textit{IoT Chain} which records the balance for all accounts that have ever been involved in an arbitrary cross-chain transaction and dynamically maintains a ledger based on the historical balance;
\item $Wit\_Addr\_List$: The addresses of \textit{Witnesses} for the \textit{IoT Chain}, which is hardcoded in both $SC\_Register$ and $SC\_Inter$.
\end{itemize}
\end{enumerate}
\subsection{Process of Reverting and Resending}
Note that there exist multiple solutions to the failure of consensus, such as $Revert$ or $Resend$. In this documentation, we also provide our mechanism shown as follows.
\begin{itemize}
\item \textbf{Registration:} if the consensus of $N/2$ on $SC\_A$ fails to be reached within a period of $t$ (can instead use block height $h$), the token is reverted to the account of the creator on \textit{Token Chain}, with a certain amount of compensation fee paid to $SC\_A$ and \textit{Witnesses}.
\item \textbf{Transferring:} suppose that $h_l$ denotes the height of the current latest block and $\omega$ denotes the unconfirmation window size (a public chain needs such a window size to determine the sufficiently high confidence interval of a block not being reverted by the chain-reorganization). If it is the first time of transferring, each of \textit{Witnesses} waits until $h_l-2\omega+1 \geq 0$, and collects all events associated with the transactions whose destination address is $SC\_A$, $SC\_Register$ or $SC\_Inter$ from block of $h_l-2\omega+1$ to $h_l-\omega$, which is assigned as the data of a ``Transferring'' transaction for the following consensus. If the consensus of $N/2$ in $SC\_Register$, $SC\_Inter$ or $SC\_A$ fails to be reached within a period of $t$ (can instead use block height $h$), each of \textit{Witnesses} sleeps for $T$ (now we let $h_{l2}$ denotes the latest block) and afterwards starts to resend the blocks getting stuck with the height from $h_l-2\omega+1
$ to $h_l-\omega$ (totally $\omega$ blocks). This process of resending is iterated until the consensus of $N/2$ is reached ($h_{l2}$ updates per iteration). Note that a process of resending indicates that it is not the first time of transferring. In this case, each of \textit{Witnesses} finds its $h_{l2}$ and checks whether $h_{l2} \geq h_l-\omega + 2\omega$. If not, each of \textit{Witnesses} sleeps for $T$ and starts over. Otherwise, each of \textit{Witnesses} collects the corresponding events from the blocks with height from $h_l-\omega+1$ to $h_l$ and sends the cross-chain message to $SC\_Register$, $SC\_Inter$ or $SC\_A$ for the following consensus. Also note that each of \textit{Witnesses} listens to the event invoked by the corresponding consensus of $N/2$ in order for the awareness of the result.
\end{itemize}
\section{Claims}
What is claimed is:
\begin{enumerate}
\item A method for transferring assets from a parent chain to multiple side chains and backwards, the method comprising:
\begin{itemize}
\item pre-defining, for a new side chain without the native gas, a set of parameters must be set prior to every other step in order for the later validation;
\item registering, via the Validation and Consensus by \textit{Witnesses}, the information of the side chain gets recorded by an open smart contract on the parent chain;
\item first time sending, which represents the first time transferring from the parent chain to the side chain. A certain amount of token being locked in another open smart contract on the parent chain and the same amount of token is unlocked and released by an open smart contract on the side chain;
\item reversely sending, which represents the transferring from the side chain back to the parent chain. A certain amount of token is locked on the side chain, and the same amount of token is unlocked on the parent chain if the consensus is passes;
\item subsequent sending, which represents the subsequent transferring from the parent chain to the side chain. It differs from the first time transferring in that,
\begin{itemize}
\item \textit{Witnesses} listen to the events from the smart contract that locks the assets but not the smart contract that records the Chain ID of side chains;
\item the smart contract that releases and unlocks the assets on the side chain in the first time transferring now instead delegates this to an open smart contract that locks all the rest of the potential assets.
\end{itemize}
\item reverting and resending, by the consensus whose result \textit{Witnesses} keeps listening to. Revert is implemented in the consensus of registration while Resend is implemented in the other consensuses;
\item trading, by an open smart contract on a new side chain with the native gas. All transactions associated with cross-chain require to be manipulated by this smart contract.
\end{itemize}
\item The method of claim \textbf{1}, wherein users can choose to either registering a side Blockchain without the native gas in which only the token originated from \textit{Token Chain} dominates the system, or a side Blockchain with the native gas in which its own account-balance mechanism is in charge of its self-driven processing. A side Blockchain without the native gas must be at a height of zero when registering, while a side Blockchain without the native gas can be either a new one or an existing one.
\item The method of claim \textbf{1}, wherein the MulSig is not limited to 2-of-2, can be n-of-m instead.
\item The method of claim \textbf{1}, wherein the Mulsig of \textit{SC\_Register} is for suicide. The Mulsig of \textit{SC\_Bank} is for adding the address of \textit{SC\_Inter} and subsequent assets locking and unlocking.
\item The method of claim \textbf{1}, wherein either \textit{SC\_Consensus} or \textit{SC\_Trading} can be pre-defined in the genesis block instead of offering Mulsig if the side chain with the native gas is at the height of zero.
\item The method of claim \textbf{1}, wherein the native gas is not limited to Ethereum platform. All Ethereum-based platforms or other platforms that are self-driven by some kinds of native account-balance are also compatible.
\item The method of claim \textbf{1}, wherein the faulty tolerance of consensus can be set to any number greater than \textit{N/2}.
\item The method of claim \textbf{1}, wherein the \textit{entrance fee} is minimum-bounded to prevent from maliciously frequent registration requests.
\item The method of claim \textbf{1}, wherein the format of data by which \textit{Witnesses} validate the validity of a new side chain is not limited to json file.
\item The method of claim \textbf{1}, wherein there is no need to pay fee for block generator for any cross-chain-related smart contracts and each of \textit{Witnesses} sending a transaction with a destination address of a cross-chain-related smart contract.
\item The method of claim \textbf{1}, wherein the cross-chain-related smart contract initializes with no balance on the parent chain.
\item The method of claim \textbf{1}, wherein for any smart contracts on which a consensus process is being conducted, a parameter \textit{round} is used to prevent from out-of-order messages.
\item The method of claim \textbf{1}, wherein there is no block rewards on the side chain.
\end{enumerate}
\bibliographystyle{IEEEtran}
\section{Introduction}
Blockchain, featured with its decentralized tamper-resistance, has been an active research topic to address the security issue of centralization, and has been widely adopted in many scenarios such as Internet-of-Things (IoT)~\cite{iot-1,iot-2,iot-3,iot-4,iot-5,iot-6}. A public Blockchain, often referred to as \textit{Token Chain}, can accomplish a decentralized tokenomics. Since \textit{Token Chains} are usually public, they require decentralized consensus algorithms, such as Proof of Work (PoW)~\cite{nakamoto2008bitcoin}. However, a decentralized consensus algorithm entails the comparatively low performance and efficiency, implying that those delay-sensitive applications or applications with massive deployment and data, such as IoT applications, cannot operate on \textit{Token Chains} properly. For such applications, consortium chains are proposed with partially decentralized and more efficient consensus algorithms, such as Byzantine Faulty Tolerance (BFT)~\cite{castro1999practical}.
However, \textit{Side Chains} have no economic value, which prevents the trading of IoT data.
We propose a cross-chain architecture consisting of a public \textit{Token Chain} and multiple associated consortium \textit{Side Chains}. In the cross-chain system, the \textit{Token Chain} issues the main tokens and maintains the official token ledger for the cryptocurrency ecosystem, while a number of \textit{Side Chains} import the tokens from the \textit{Token Chain} for the management of the massive records and transactions for applications, such as IoT. The key challenge is the design of cross-chain interactions, including token exchange between the \textit{Token Chain} and its associated \textit{Side Chains}. A set of nodes are defined as \textit{Witnesses} in the genesis block of each \textit{Side Chain}. These \textit{Witnesses} are responsible for the exchange of tokens between the \textit{Token Chain} and the \textit{Side Chains}. \\
The cross-chain system provides the flexibility to meet various capacity demands of heterogeneous applications, including but not limited to IoT applications. It achieves modularity in that the \textit{Side Chains} are decoupled from the \textit{Token Chain} through the design of \textit{Witnesses}. This system is not limited to Ethereum platform, and has no limitations on the number of \textit{IoT Chains}.\footnote{The theoretical maximum number of IoT chains is $16^{40}\approx 1.46\times 10^{48}$ based on the 20-byte address. It is a quite large number to exceed in practice.}
\section{Related Works}
There exist several cross-chain technologies, mainly classified into Relay, Hash-Lock and Notaries.
\begin{itemize}
\item Relay: Chain $A$ can prove the facts and behaviors on Chain $B$. Technologies using Relay are including but not limited to, the following solutions:
\begin{itemize}
\item BTC-Relay~\cite{hash-lock}. It is a technology that can inter-operate between Bitcoin and Ethereum. A light client that introduces Simple Payment Verfication (SPV) is implemented by a smart contract in Ethereum so that facts on Bitcoin can be proved in Ethereum. However, it also implies that Ethereum needs to store all headers of Bitcoin, which is quite expensive.
\item Non-Interactive Proof of PoW~\cite{nipopow}. Combined with Monte Carlo method and Bisection method, this proof tries to calculate the probability that a chosen puzzle with a high difficulty falls into a specific interval. As such the PoW on Chain $B$ can be proved by including this proof of PoW on Chain $A$. However, currently this method is not general and strictly dependent from PoW consensus.
\end{itemize}
\item Hash-Lock~\cite{hash-lock}. It acts as a personal trading platform. Fair and secure assets exchange can be accomplished with the aid of a time-out mechanism. However, it does not support assets portability and it is unfriendly to the interoperation between a public chain and a consortium chain because of the possible different token-scheme on this consortium chain.
\item Notaries~\cite{notary}. It is also known as \textit{Witnesses}. \textbf{Our invention belongs to this type.} In words, there exist some nodes which run clients on both Chain $A$ and $B$. Messages (Assets) exchanged is conducted through these nodes with a necessary consensus process. Notaries is the most frequently used technology so far. There are a number of Blockchain projects announcing their schemes, including but not limited to,
\begin{itemize}
\item Hcash. It is a typical design of Notaries-liked cross-chain technology. It collectively features most of the advantages of Notaries from other similar projects. We believe that comparing our invention with Hcash will be sufficiently significant to differential our invention with other similar designs. In our opinion, notaries act as a set of nodes running on both chains, as if multiple personal trading platforms exist with a necessary consensus process. The On/Off-chain multi-signature wallet takes critical responsibilities when assets exchanging. However, this design does require users to have assets prior to the process of exchanging, which implies that similar to Hash-Lock, it does not support consortium chain very well either. Hcash does not provide the solutions to address the implications of the native gas-oriented account-balance on the consortium chain, and to generate exchangeable token for a consortium chain with the native gas-oriented account-balance.
\item Walton Chain. Notaries are a set of nodes only running on the local chain. The request of generating a new side chain is executed by a smart contract by sending a transaction. In other words, notaries are the bridge to exchange the messages of the most external layer to the internal layer with a necessary consensus process. It implies that a side chain is only an abstractive structure in Walton Chain. Link structure is maintained on smart contracts in which the data storage is simplified as the state-oriented data to be stored in the local database. Such a design is not only strongly restricted by the throughput of the most external chain but also extremely increases the loading of the entire network.
\end{itemize}
\end{itemize}
\emph{There appears an existing national patent (CN 106447309 A) in China that is similar to our invention. Their patent implements the third-party chain that pegs to both the main chain and the side chain with a certain address on each chain as if an accessible API connects the main chain and the side chain. It implies that these APIs have an offline information relaying between the certain address on either the main chain or the side chain, and its third-party chain, which they did not precisely point out. In addition, the existence of the third-party chain exposes the issue of data privacy. It also makes no sense that the third-party is eligible to maintain multiple side chains due to the data privacy, centralization and scalability.} Therefore, we can conclude that our invention is more general and outplays their patent.
Our invention is targeting at such a case that multiple \textit{IoT Chains} with different consensus algorithms, a high throughput and a high volume are anchored with a \textit{Token Chain}. Therein each \textit{IoT Chain} has its own business groups charged with the audition of cross-chain process. One \textit{IoT Chain} can be totally a black-box for the others. Likewise, any behaviors on an \textit{IoT Chain} can be also a black-box for the \textit{Token Chain} apart from the amount of token exchanged. Either an \textit{IoT Chain} without the native gas or an \textit{IoT Chain} withe the native gas can be registered to the \textit{Token Chain} whenever a user requires token involved and afterwards provides the sufficient assets and proofs on the \textit{Token Chain}. In words, we believe that our invention can be a better solution to the separation between the financial-layer and data-layer in order for a more data-friendly and healthy Blockchain system.
\section{Brief Description of the Drawings}
The description of each step and module are shown in the following drawings.
\begin{enumerate}
\item Figure \ref {cross} shows the flow diagram for the entire architecture of our invention. It combines the rest of the Figures into a whole to elaborate the entire flow of the procedure.
\item Figure \ref {cross_1} shows the flow diagram for the pre-defined parameters required to be setup prior to the registration to \textit{Token Chain} for an \textit{IoT Chain} without the native gas.
\item Figure \ref {cross_2} shows the flow diagram for the first time registration to \textit{Token Chain} for an \textit{IoT Chain} without the native gas from the perspective of \textit{Token Chain}.
\item Figure \ref {cross_3} shows the flow diagram for the first time registration to \textit{Token Chain} for an \textit{IoT Chain} without the native gas from the perspective of \textit{IoT Chain}.
\item Figure \ref {cross_4} shows the flow diagram for transferring consumable token from \textit{IoT Chain} back to \textit{Token Chain} from the perspective of \textit{IoT Chain}.
\item Figure \ref {cross_5} shows the flow diagram for transferring consumable token from \textit{IoT Chain} back to \textit{Token Chain} from the perspective of \textit{Token Chain}.
\item Figure \ref {cross_6} shows the flow diagram for subsequently transferring from \textit{Token Chain} to \textit{IoT Chain} that has been registered from the perspective of \textit{Token Chain}.
\item Figure \ref {cross_7} shows the flow diagram for subsequently transferring from \textit{Token Chain} to \textit{IoT Chain} that has been registered from the perspective of \textit{IoT Chain}.
\item Figure \ref {cross_8} shows the flow diagram for transferring between \textit{Token Chain} and an \textit{IoT Chain} with the native gas.
\end{enumerate}
\section{Brief Description of the Smart Contracts and Witnesses}
\begin{itemize}
\item \textit{SC\_A}: A dedicated cross-chain smart contract for an \textit{IoT Chain \{A\}}, deployed on \textit{Token Chain}. It locks/unlocks the assets transferred. It introduces the consensus for the Validation of registration from witnesses, as well as the consensus for the assets transferred back to \textit{Token Chain}.
\item \textit{SC\_ID}: It is deployed on \textit{Token Chain}. A \textit{Token Chain} can only have one \textit{SC\_ID} and it offers 2-of-2 Mulsig (Note that it can be a n-of-m Mulsig that has been claimed in the last section. For the rest of place using Mulsig, 2-of-2 is implemented in order for the simple elaboration). It records the $Chain\_ID$ of \textit{IoT Chains} that have been successfully registered. It $Updates$ the list of $Chain\_ID$ once the consensus for the Validation of registration from witnesses gets passed and emits an event of $ExistOrNot$.
\item \textit{SC\_Register}: A unique dedicated cross-chain smart contract for the first time registration, deployed on an \textit{IoT Chain}. It introduces the consensus for the $entrance fee$ transferred to the \textit{IoT Chain}. Its balance is hardcoded and will be run out after the success of registration and commits suicide afterwards by the agreement of Mulsig.
\item \textit{SC\_Inter}: A unique dedicated cross-chain smart contract for the assets transferring back to \textit{Token Chain} as well as the subsequent assets transferred to \textit{IoT Chain} with a consensus process, deployed on \textit{IoT Chain}. It dedicates the assets unlocking to \textit{SC\_Bank} after the success of consensus. It forwards the assets transferred from users' accounts to \textit{SC\_Bank} for assets locking in the case of \textit{IoT-Chain}-to-\textit{Token-Chain}.
\item \textit{SC\_Bank}: A unique dedicated cross-chain smart contract for assets locking/unlocking, deployed on \textit{IoT Chain}. The balance of \textit{SC\_Bank} that equals to the total amount of the upper bound of the potential assets trading on \textit{IoT Chain} is hardcoded in the genesis block. The corresponding amount of token is unlocked and transferred to the destination address with the delegation from \textit{SC\_Inter}. It receives and locks the assets transferred from users' accounts to \textit{SC\_Bank} for assets locking in the case of \textit{IoT-Chain}-to-\textit{Token-Chain}.
\item \textit{SC\_Consensus}: A unique dedicated cross-chain smart contract for both registration and consensus in the context of an \textit{IoT Chain} with the native gas.
\item \textit{SC\_Trading}: A unique dedicated cross-chain smart contract that records the balance for all accounts that have ever been involved in a cross-chain transaction and dynamically maintains a ledger. Both intra-transactions and inter-transactions associated with real business must be manipulated via \textit{SC\_Trading}. In other words, the token value is generated and destroyed with \textit{SC\_Trading} running on an \textit{IoT Chain} with the native gas. The amount of consumable assets is precisely defined even though its native account-balance tends to be infinite.
\item \textit{Witnesses}: A set of nodes that bridge the information exchanged between \textit{Token Chain} and a specific \textit{IoT Chain}. Information exchanged needs to be verified by the consensus on both \textit{Token Chain} and the \textit{IoT Chain}. \textit{Witnesses} retrieve events from:
\begin{enumerate}
\item \textit{SC\_ID}: if the $Chain\_ID$ has been appended in the list;
\item \textit{SC\_A}: if some numbers of cross-chain transactions have arrived;
\item \textit{SC\_Bank}: if the assets has been successfully locked in the context of \textit{IoT-Chain}-to-\textit{Token-Chain};
\item \textit{SC\_Consensus}: if some numbers of cross-chain transactions have arrived.
\end{enumerate}
\end{itemize}
\section{Detailed Description}
A set of pre-defined \textit{Witnesses} exchange the cross chain messages as if a bridge connects \textit{Token Chain} and an \textit{IoT Chain}. Remark that multiple \textit{IoT Chains} can be implemented along with their own pre-defined \textit{Witnesses}.
All \textit{IoT Chains} are pegged to \textit{Token Chain}, which provides an individual channel for each \textit{IoT Chain} to exchange token. \textit{IoT Chains} do not interact with each other directly.
In addition, businesses tend to customized their own \textit{IoT Chains} regardless of a specific type of consensus algorithm or even the existence of native gas, i.e. either of an \textit{IoT Chain} with native gas or an \textit{IoT Chain} without native gas. Note that the native gas refers to the native gas-oriented account-balance. Also note that the native gas-oriented account-balance is firstly implemented in Ethereum platform. It is often used to restrict the malicious transactions flooding among the network. The difference between them is shown as follows,
\begin{itemize}
\item \textbf{\textit{IoT Chain} without the native gas: }an \textit{IoT Chain} has no native gas-oriented account-balance. Nodes on such an \textit{IoT Chain} cannot send transactions without the imported gas transferred from the corresponding \textit{Token Chain}, where the imported gas, in fact, is consuming the main tokens. In other words, the main tokens are responsible for both sending a transaction and dealing with a business issue.
\item \textbf{\textit{IoT Chain} with the native gas: }an \textit{IoT Chain} has the native gas-oriented account-balance. Nodes on such an \textit{IoT Chain} can send transactions (uploading the IoT messages) without the imported gas transferred from the corresponding \textit{Token Chain}. However, any transactions associated with the business issue must come along with the imported gas.
\end{itemize}
\textit{IoT Chains} are free to be customized with low restrictions in terms of the consensus algorithm and the existence of native gas. All tokens on \textit{IoT Chains} are originated from \textit{Token Chain}, referring to the main tokens. In other words, \textit{IoT Chains} have no original token value regardless of the presence of the native gas.
Note that \textit{IoT Chain} is allowed to have its own account-balance mechanism to maintain the whole system running fluently with a necessary access control and a flow control. This account-balance is meant to be a sufficiently large amount of value that can be automatically recharged by administrators, which implies that only the token originated from \textit{Token Chain} that is strictly differentiated from the native account-balance on \textit{IoT Chains}, has the real economical value. Thus users can choose to either registering an \textit{IoT Chain} without the native gas in which only the token originated from \textit{Token Chain} dominates the system, or an \textit{IoT Chain} with the native gas in which its own account-balance mechanism is in charge of its self-driven processing, which are both independent to the consensus algorithm.
For the rest of this section we extend the description to a deeper extent shown as follows,
\begin{itemize}
\item Section 5.1 refers to Figure \ref {cross_1}, \ref {cross_2}, \ref {cross_3}, \ref {cross_4}, \ref {cross_5}, \ref {cross_6} and \ref {cross_7}.
\begin{itemize}
\item Section 5.1.1 refers to Figure \ref {cross_1}.
\item Section 5.1.2 refers to Figure \ref {cross_2} and \ref {cross_3}.
\item Section 5.1.3 refers to Figure \ref {cross_6} and \ref {cross_7}.
\item Section 5.1.4 refers to Figure \ref {cross_4} and \ref {cross_5}.
\end{itemize}
\item Section 5.2 refers to Figure \ref {cross_8};
\item Section 5.3 describes the mechanism of the reverting and resending process of our invention when the consensus fails.
\end{itemize}
Note that the entire flow chart refers to Figure \ref {cross}.
\subsection{Bridging a Side Blockchain without the native gas}
Bridging a side Blockchain without the native gas includes the following steps,
\begin{enumerate}
\item pre-definition of necessary parameters;
\item registration of the chain;
\item subsequent assets transferring:
\begin{itemize}
\item transferring from \textit{IoT Chain} back to \textit{Token Chain};
\item transferring from \textit{Token Chain} to \textit{IoT Chain} after the registration.
\end{itemize}
\end{enumerate}
\subsubsection{Pre-definition}
A set of \textit{Witnesses} run as intermediaries to bridge \textit{Token Chain} and an specific \textit{IoT Chain}. These \textit{Witnesses} have valid accounts on both \textit{Token Chain} and the \textit{IoT Chain}. The interactions between \textit{Token Chain} and the \textit{IoT Chain} are secured by Blockchain. Only interactions verified and recorded in the form of transactions are accepted and performed on both sides. Smart contracts implement the interactions between \textit{Token Chain} and the \textit{IoT Chain}.
The genesis block of the \textit{IoT Chain} has a pre-defined form. The genesis block defines the account states and the interaction parameters of the \textit{IoT Chain}. In the genesis block, all accounts on the \textit{IoT chains} have no balance when the \textit{IoT Chain} initializes, except that two pre-defined smart contracts created by the creator of the \textit{IoT Chain} stating the balance of these smart contracts and promises to lock \textbf{\textit{some fee}} (shown below) from these addresses, respectively. Each of them is verified by the signature of its own key.
\begin{enumerate}
\item the first kind of fee is assigned as the \textit{entrance fee} that is the minimum amount of token required to initialize an interactive \textit{IoT Chain} for the registration phase. \emph{Note that the \textit{entrance fee} should be minimum-bounded to prevent from maliciously frequent registration requests};
\item the second kind of fee is assigned as the \textit{reserved assets} from which the subsequent assets can be unlocked.
\end{enumerate}
The \textit{IoT Chain} predefines $N$ \textit{Witnesses} in the genesis block of the \textit{IoT Chain}. These $N$ \textit{Witnesses} should have valid accounts on \textit{Token Chain}.
In other words, a valid genesis block of an interactive \textit{IoT Chain} without the native gas has the following fields:
\begin{itemize}
\item $Chain\_ID$: The Chain ID of the \textit{IoT Chain};
\item $SC\_Register$: An interactive smart contract on the \textit{IoT Chain} which enables transfer function between the \textit{IoT Chain} and \textit{Token Chain}.
\item $Bal\_Resv$: The amount of token that the creator reserves in $SC\_Register$ for the \textit{IoT Chain};
\item $SC\_Inter$: A smart contract on the \textit{IoT Chain}, used for subsequent transferring from \textit{Token Chain} to the \textit{IoT Chain} and transferring from the \textit{IoT Chain} back to \textit{Token Chain};
\item $SC\_Bank$: A smart contract on the \textit{IoT Chain} which enables the subsequent token locking and unlocking. There is a strict access control upon this smart contract, i.e. the only address who can interact with $SC\_Bank$ is the address of $SC\_Inter$.
\item $Bal\_Bank$: The amount of token that the creator reserves in $SC\_Bank$ for the \textit{IoT Chain}; Note that $Bal\_Resv + Bal\_Bank \equiv$ total amount of token on \textit{Token Chain} including those unmined token;
\item $Wit\_Addr\_List$: The addresses of \textit{Witnesses} for the \textit{IoT Chain}, which is hardcoded in both $SC\_Register$ and $SC\_Inter$.
\end{itemize}
The creator creates and broadcasts the genesis block in the IoT network and await the other nodes being launched and ready for the the new \textit{IoT Chain} within an expected period. Note that any transactions with a source address of one of the \textit{Witnesses} and a destination address of \textit{SC\_Register} or \textit{SC\_Inter} can pay a zero gasprice. Meanwhile, any transactions with a source address of either \textit{SC\_Register}, \textit{SC\_Inter} or \textit{SC\_Bank} can also pay a zero gasprice.
\subsubsection{Registration}
After all \textit{Witnesses} have been ready (All \textit{Witnesses} should then be able to send transactions to and retrieving events from any targeted addresses. Meanwhile, they all are able to fetch the json files of both of the genesis blocks of the \textit{IoT Chain} and \textit{Token Chain}), the creator on \textit{Token Chain} sends a request of registration in the form of a transaction to \textit{SC\_A}, where the following contents are included in the invoked event,
\begin{itemize}
\item all token that is contained in this transaction sent to \textit{SC\_A} subtracting the compensation fee for both \textit{SC\_A} and \textit{Witnesses} providing the transferring service. It guarantees that \textit{SC\_A} and \textit{Witnesses} have sufficient fund to send transactions on \textit{Token Chain};
\item $Chain\_ID$ of the \textit{IoT Chain} that the creator proposes.
\end{itemize}
\textit{Note that \textit{SC\_A} is a smart contract deployed on \textit{Token Chain} with no balance initially, which is a dedicated cross-chain service relay and created by the service provider in the form of \textbf{open source}.}
As such the process of validation on \textit{Witnesses} gets called by retrieving the events invoked by this requesting transaction.
It validates the genesis block according to Alg. \ref{alg: verify IoT chain} from which the output is sent to \textit{SC\_A}.
In other words, once a \textit{Witness} validates the genesis block, it accepts the registration request of the new \textit{IoT Chain} and sends a ``Confirm'' transaction to the smart contract, \textit{SC\_A}, on \textit{Token Chain}.
When \textit{SC\_A} receives ``Confirm'' transactions from more than $N/2$ \textit{Witnesses} (\textit{Note that for any smart contracts on which a consensus process is being conducted, a parameter $round$ must be used to prevent from out-of-order message.
The minimum faulty tolerance is $N/2$}), the smart contract \textit{SC\_A} reserves $Bal\_Resv$ on \textit{Token Chain} from the account of the creator, triggers the smart contract \textit{SC\_ID} to register the \textit{IoT Chain} with its $Chain\_ID$.
Note that Alg. \ref{alg: verify IoT chain} has ensured that the $Chain\_ID$ for each chain is unique in the entire system.
A signal indicating whether the registration has succeeded or not is contained in the event invoked by $Chain\_ID$ that \textit{Witnesses} are listening to.
If and only if a ``True'' is received, a \textit{Witness} sends a ``Transferring'' transaction to the smart contract $SC\_Register$ on the \textit{IoT Chain}.
If $SC\_Register$ receives ``Transferring'' transaction from more than $N/2$ \textit{Witnesses}, the same amount (this amount should now exactly equal to the total balance of $SC\_Register$, ensuring that the balance of $SC\_Register$ will be zero after unlocking) of token are unlocked from $SC\_Register$ on \textit{IoT Chain}.
$SC\_Register$ commits suicide by the owners after the owners both (2-of-2 Mulsigs) have retrieved the event of success of adding the $Chain\_ID$ in $SC\_ID$.
After the \textit{IoT Chain} is created and registered, the \textit{IoT Chain} will be able to work properly and generate blocks based on its consensus protocol. \emph{Note that there is no block rewards on the \textit{IoT Chain}}.
\subsubsection{Token Chain to IoT Chain after Registration}
$SC\_Inter$ is responsible for any subsequent ``Transferring'' transactions sent from \textit{Token Chain} to a specific \textit{IoT Chain}, plus those sent from the \textit{IoT Chain} back to \textit{Token Chain}. If a subsequent ``Transferring'' transaction sent to $SC\_Inter$ is detected, \textbf{$SC\_Inter$ delegates the burden to $SC\_Bank$ along with $to$ and $value$ of each cross-chain transaction after the success of $N/2$ consensus}. This process succeeds if and only if it passes the validation that the current balance of $SC\_Bank$ + the \textit{entrance fee} + the amount of token that is requested for this specific ``Transferring'' transaction $\equiv$ total amount of token on \textit{Token Chain} including those unmined token. The owners of $SC\_Bank$ add the address of $SC\_Inter$ in $SC\_Bank$ and the corresponding amount of token can be unlocked if and only if both owners agree. \emph{Note that it is none of business of $SC\_ID$ for a subsequent cross-chain transferring. \textit{Witnesses} do not listen to the events invoked by $SC\_ID$. Instead, events of $SC\_A$ are listened.}
\subsubsection{IoT Chain to Token Chain after Registration}
Suppose the aforementioned \textit{IoT Chain} is still discussed, there are two types of transactions on the \textit{IoT Chain}: one records the IoT events and does not interact with \textit{Token Chain}; the other one records the interactions with \textit{Token Chain}. Any transaction that is associated with $SC\_Inter$ belongs to Interaction transactions.
Suppose that a user on the \textit{IoT Chain} sends a request of transferring token back to \textit{Token Chain} in the form of a transaction sent to $SC\_Inter$ with an arbitrary amount of token $\le$ (total amount of token on the \textit{Token Chain} including those unmined token $-$ the balance of $SC\_Bank$). $SC\_Inter$ then becomes the relay and transfers the token to $SC\_Bank$ for the further locking. Each of the
\textit{Witnesses} retrieves the events and sends a ``Transferring'' transaction to $SC\_A$ on \textit{Token Chain}. If $SC\_A$ receives a ``Transferring'' transaction from more than $N/2$ \textit{Witnesses}, the same amount (this amount should $\le$ the balance of $SC\_A$) of token is unlocked from $SC\_A$ on \textit{Token Chain}.
\begin{algorithm}
\caption{Verify the Request of Registering an IoT Chain} \label{alg: verify IoT chain}
\LinesNumbered
\KwIn{
$\quad$$Hash(Genesis)$: the hash of the IoT genesis block. \\
$\quad$$Balance_{request}$: the total balance that the creator reserves for the IoT chain. \\
$\quad$$Balance_{json}$: the total balance that the json file of the IoT chain is showing.\\
$\quad$ $ChainID_{request}$: the chain ID that the creator proposes for the IoT chain. \\
$\quad$ $ChainID_{json}$: the chain ID that the json file of the IoT chain is showing.\\
$\quad$ $ChainID_{token}$: the chain ID of the token chain.\\
$\quad$ $SmartContract_{ChainID}$: a list of all chain IDs that have been successfully registered already.}
\KwOut{True or False
\textbf{The witnesses assert the following conditions:}
Confirm the integrity of the Genesis Block:
$\quad$$GetHash(Genesis)$ = $Hash(Genesis)$;
Confirm it is a request to generate a new IoT Chain:
$\quad$$GetHeight()$ = 0;
Verify the balance of the creator:
$\quad$$Balance_{request}$ = $Balance_{json}$
Validate the Chain ID:
$\quad$$ChainID_{request}$ = $ChainID_{json}$
$\quad$$ChainID_{request}$ $\neq$ $ChainID_{token}$
$\quad$$ChainID_{request}$ $\notin$ $SmartContract_{ChainID}$
If ALL above conditions are satisfied, the output is ``TRUE'', which indicates the witness validates the registration request; or else the output is ``FALSE'', which indicates the witness refuses the registration request.
\end{algorithm}
\subsection{Bridging a Side Blockchain with the native gas}
Comparing with bridging an \textit{IoT Chain} without the native gas, there are only a few tiny changes to bridge an \textit{IoT Chain} with the native gas, which is shown as follows,
\begin{enumerate}
\item \textit{SC\_Register} and \textit{SC\_Inter} merge into one smart contract, \textit{SC\_Consensus}, since there is no need to transfer token itself. Instead, the only thing needs to do is to update the balance in \textit{SC\_Trading}. \textit{SC\_Consensus} has two hardcoded owners, offering 2-of-2 MultiSig;
\item A set of \textit{Witnesses} is still needed to be organized and written in \textit{SC\_Consensus}. Note that there is no need to predefine \textit{SC\_Consensus} in the genesis block as long as it provides sufficient native account-balance for self-driven processing. Also, there is no need to set an \textit{entrance fee} in \textit{SC\_Consensus};
\item \textit{SC\_Bank} can be deprecated;
\item It is still required to have a smart contract running on \textit{Token Chain}, \textit{SC\_A} created by the service provider in the form of \textbf{open source}. However, the events fetched by \textit{Witnesses} for the Validation only contain the information of how much is transferred and the \textit{Chain\_ID}. Accordingly, \textit{Witnesses} only validate the \textit{Chain\_ID} part in \textit{Algorithm 1} regardless of the height and balance of the \textit{IoT Chain};
\item As if a trading platform manipulate the financial trading, \textit{SC\_Trading} records the balance for all accounts that have ever been involved in an arbitrary cross-chain transaction and dynamically maintains a ledger based on the historical balance. Both intra-transactions and inter-transactions associated with real business must be manipulated via \textit{SC\_Trading}. In other words, the token value is generated and destroyed with \textit{SC\_Trading} running on an \textit{IoT Chain} with the native gas. In the context of high throughput of the \textit{IoT Chain} along with its native account-balance mechanism, it is believed that \textit{SC\_Trading} is able to handle the transactions in a high volume. Note that the following statement must be satisfied,
$$\left(\sum_{i=0}^{N}Bal_i\right)_{ChainID}=Asset_{ChainID},$$
where the total amount of token that is being locked in \textit{SC\_A} with a certain $Chain\_ID$ must equal to the sum of the balance of all accounts recorded in \textit{SC\_Trading} on an \textit{IoT Chain} with the same $Chain\_ID$. It has two hardcoded owners, offering 2-of-2 MultiSig.
\item Either of \textit{SC\_Consensus} or \textit{SC\_Trading} can be pre-defined in the genesis block without having the owners and offering the Mulsig if the r \textit{IoT Chain} with the native gas is at the height of zero when initializing. As such a valid genesis block of an \textit{IoT Chain} with the native gas has the following fields:
\begin{itemize}
\item $Chain\_ID$: The Chain ID of the \textit{IoT Chain};
\item $SC\_Register$: An interactive smart contract on the \textit{IoT Chain} which enables transfer function between the \textit{IoT Chain} and the \textit{Token Chain};
\item $SC\_Trading$: An smart contract on the \textit{IoT Chain} which records the balance for all accounts that have ever been involved in an arbitrary cross-chain transaction and dynamically maintains a ledger based on the historical balance;
\item $Wit\_Addr\_List$: The addresses of \textit{Witnesses} for the \textit{IoT Chain}, which is hardcoded in both $SC\_Register$ and $SC\_Inter$.
\end{itemize}
\end{enumerate}
\subsection{Process of Reverting and Resending}
Note that there exist multiple solutions to the failure of consensus, such as $Revert$ or $Resend$. In this documentation, we also provide our mechanism shown as follows.
\begin{itemize}
\item \textbf{Registration:} if the consensus of $N/2$ on $SC\_A$ fails to be reached within a period of $t$ (can instead use block height $h$), the token is reverted to the account of the creator on \textit{Token Chain}, with a certain amount of compensation fee paid to $SC\_A$ and \textit{Witnesses}.
\item \textbf{Transferring:} suppose that $h_l$ denotes the height of the current latest block and $\omega$ denotes the unconfirmation window size (a public chain needs such a window size to determine the sufficiently high confidence interval of a block not being reverted by the chain-reorganization). If it is the first time of transferring, each of \textit{Witnesses} waits until $h_l-2\omega+1 \geq 0$, and collects all events associated with the transactions whose destination address is $SC\_A$, $SC\_Register$ or $SC\_Inter$ from block of $h_l-2\omega+1$ to $h_l-\omega$, which is assigned as the data of a ``Transferring'' transaction for the following consensus. If the consensus of $N/2$ in $SC\_Register$, $SC\_Inter$ or $SC\_A$ fails to be reached within a period of $t$ (can instead use block height $h$), each of \textit{Witnesses} sleeps for $T$ (now we let $h_{l2}$ denotes the latest block) and afterwards starts to resend the blocks getting stuck with the height from $h_l-2\omega+1
$ to $h_l-\omega$ (totally $\omega$ blocks). This process of resending is iterated until the consensus of $N/2$ is reached ($h_{l2}$ updates per iteration). Note that a process of resending indicates that it is not the first time of transferring. In this case, each of \textit{Witnesses} finds its $h_{l2}$ and checks whether $h_{l2} \geq h_l-\omega + 2\omega$. If not, each of \textit{Witnesses} sleeps for $T$ and starts over. Otherwise, each of \textit{Witnesses} collects the corresponding events from the blocks with height from $h_l-\omega+1$ to $h_l$ and sends the cross-chain message to $SC\_Register$, $SC\_Inter$ or $SC\_A$ for the following consensus. Also note that each of \textit{Witnesses} listens to the event invoked by the corresponding consensus of $N/2$ in order for the awareness of the result.
\end{itemize}
\section{Claims}
What is claimed is:
\begin{enumerate}
\item A method for transferring assets from a parent chain to multiple side chains and backwards, the method comprising:
\begin{itemize}
\item pre-defining, for a new side chain without the native gas, a set of parameters must be set prior to every other step in order for the later validation;
\item registering, via the Validation and Consensus by \textit{Witnesses}, the information of the side chain gets recorded by an open smart contract on the parent chain;
\item first time sending, which represents the first time transferring from the parent chain to the side chain. A certain amount of token being locked in another open smart contract on the parent chain and the same amount of token is unlocked and released by an open smart contract on the side chain;
\item reversely sending, which represents the transferring from the side chain back to the parent chain. A certain amount of token is locked on the side chain, and the same amount of token is unlocked on the parent chain if the consensus is passes;
\item subsequent sending, which represents the subsequent transferring from the parent chain to the side chain. It differs from the first time transferring in that,
\begin{itemize}
\item \textit{Witnesses} listen to the events from the smart contract that locks the assets but not the smart contract that records the Chain ID of side chains;
\item the smart contract that releases and unlocks the assets on the side chain in the first time transferring now instead delegates this to an open smart contract that locks all the rest of the potential assets.
\end{itemize}
\item reverting and resending, by the consensus whose result \textit{Witnesses} keeps listening to. Revert is implemented in the consensus of registration while Resend is implemented in the other consensuses;
\item trading, by an open smart contract on a new side chain with the native gas. All transactions associated with cross-chain require to be manipulated by this smart contract.
\end{itemize}
\item The method of claim \textbf{1}, wherein users can choose to either registering a side Blockchain without the native gas in which only the token originated from \textit{Token Chain} dominates the system, or a side Blockchain with the native gas in which its own account-balance mechanism is in charge of its self-driven processing. A side Blockchain without the native gas must be at a height of zero when registering, while a side Blockchain without the native gas can be either a new one or an existing one.
\item The method of claim \textbf{1}, wherein the MulSig is not limited to 2-of-2, can be n-of-m instead.
\item The method of claim \textbf{1}, wherein the Mulsig of \textit{SC\_Register} is for suicide. The Mulsig of \textit{SC\_Bank} is for adding the address of \textit{SC\_Inter} and subsequent assets locking and unlocking.
\item The method of claim \textbf{1}, wherein either \textit{SC\_Consensus} or \textit{SC\_Trading} can be pre-defined in the genesis block instead of offering Mulsig if the side chain with the native gas is at the height of zero.
\item The method of claim \textbf{1}, wherein the native gas is not limited to Ethereum platform. All Ethereum-based platforms or other platforms that are self-driven by some kinds of native account-balance are also compatible.
\item The method of claim \textbf{1}, wherein the faulty tolerance of consensus can be set to any number greater than \textit{N/2}.
\item The method of claim \textbf{1}, wherein the \textit{entrance fee} is minimum-bounded to prevent from maliciously frequent registration requests.
\item The method of claim \textbf{1}, wherein the format of data by which \textit{Witnesses} validate the validity of a new side chain is not limited to json file.
\item The method of claim \textbf{1}, wherein there is no need to pay fee for block generator for any cross-chain-related smart contracts and each of \textit{Witnesses} sending a transaction with a destination address of a cross-chain-related smart contract.
\item The method of claim \textbf{1}, wherein the cross-chain-related smart contract initializes with no balance on the parent chain.
\item The method of claim \textbf{1}, wherein for any smart contracts on which a consensus process is being conducted, a parameter \textit{round} is used to prevent from out-of-order messages.
\item The method of claim \textbf{1}, wherein there is no block rewards on the side chain.
\end{enumerate}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:intro}
Researchers in the visualization community have long viewed interaction as an \textit{analytic discourse} between the analyst and the visualization system~\cite{pike2009science}. Thus, capturing and analyzing the user's passive interactions has been integral to the visual analytics research agenda~\cite{cook2005illuminating,perry2009supporting,pirolli2005sensemaking,ribarsky2009science,crouser2013balancing}. Some believe that this data can provide a \textit{transcript} of the reasoning process, informing more effective visualization encodings, and producing better
intelligent algorithms to assist with data exploration, model refinement, etc.~\cite{ragan2015characterizing, xu2020survey}. Further, many have
seized the opportunity to leverage machine learning techniques to decode the information embedded in the user's interaction log data. This paper adopts the term used by Xu et al.'s~\cite{xu2020survey} recent survey on the analysis of user interactions and refers to the general goal of understanding the user and their sensemaking process as \textit{user modeling}.
Although there has been significant progress in developing algorithms that can reveal valuable information about the user and their analytic process, a unified comparison of the proposed techniques on different datasets, tasks, and analysis scenarios is lacking. Some suggested methods are presented theoretically without in-depth empirical evaluation, while others are validated using controlled user studies, though often with a single dataset. Furthermore, the algorithms are sometimes proprietary, and the community lacks benchmark datasets for easy comparison. These issues point to fundamental problems that can hinder research progress and question the practicality of user modeling techniques outside of academia.
To address this, we present a computational benchmark study, comparing previously proposed user modeling techniques from the visual analytics community. We first narrow the scope to real-time algorithms that learn from low-level interactions with data points that researchers can potentially use for providing real-time support during data exploration. Such user modeling techniques typically fall under two broad categories: (1) \textit{data interaction prediction}, \ie\ inferring data points that the analyst is likely to interact with in the near future and (2) \textit{exploration bias detection}, \ie\ detecting data features that are disproportionately explored by the user. These two categories, however, are not mutually exclusive as the successful detection of exploration bias might could also infer potential data points the user may interact with and vice versa. After surveying the body of work, we selected seven proposed techniques and standardized their input and output specifications to account for a variety of datasets. In addition to the selected models, we developed an ensemble model to predict data interaction and detect exploration bias. We then compared all eight modeling techniques' performance with four different publicly available user study datasets that vary in the number of visual attributes shown and the types of tasks the user study participants performed.
We evaluate the user modeling algorithms based on the success rate of their data interaction prediction, their exploration bias detection, and computational complexity. In addition to the comparison, we highlight the open challenges we discovered. In evaluating these algorithms, we aim to provide a better understanding to researchers and practitioners interested in integrating existing user modeling techniques into their system and to encourage further advances in analyzing user interactions and visualization provenance. The implementations, pre-procesed data files, and links to the manuscripts and datasets used in this study will be freely available on GitHub \footnote{\url{https://github.com/washuvis/vis2022usermodels}} to support reproducibility.
We summarize our contributions as follows:
\begin{itemize}[noitemsep]
\item Using a unified notation, we provide an overview of existing modeling techniques in the visual analytics community for predicting data interaction and detecting exploration bias.
\item We compare and evaluate the performance of eight techniques on two aspects of user modeling with four unique user study interaction log datasets.
\item Based on our evaluation, we provide recommendations and new research directions for analyzing user interactions and visualization provenance.
\end{itemize}
\section{Related Work}
\label{sec:related}
User modeling techniques can fall into broad categories based on shared objective and can exhibit extreme diversity, as presented by the extensive survey from Xu et al.~\cite{xu2020survey}. For example, researchers have proposed techniques to infer data interactions~\cite{monadjemi2020competing, ottley2019follow, healey2012interest}, various forms of bias~\cite{monadjemi2020competing, wall2017warning, gotz2016adaptive}, and even user attributes~\cite{brown2014finding,liu2020survey}. Additionally, other techniques aim to assist the user by data prefetching~\cite{battle2016dynamic, khan2019flux}, recommending visualizations~\cite{gotz2009behavior}, providing interface guidance~\cite{ceneda2016characterizing,ceneda2019review}, and improving data selection~\cite{fan2018fast,gadhave2021predicting}.
In our work, we aim to evaluate existing models within the categories of data interaction prediction and exploration bias detection. We selected these two categories because of their broad applicability to varying datasets and system designs.
\subsection{Exploration Bias Detection}
While exploring datasets and making decisions, humans are susceptible to cognitive limitations and biases that arise naturally from perception and intuition\cite{dimara2018task,ellis2018cognitive}. There are many different types of biases. For example, one form of bias, \emph{confirmation bias}, was described by Nickerson et al.~\cite{nickerson98confirmationbias} as interpreting evidence in ways that are partial to existing beliefs already in mind. Gigerenzer et al.~\cite{gigerenzer2009homo} defined \emph{cognitive bias} as a deviation between human judgment and a rational norm. Lee et al.~\cite{lee2019avoiding} define \textit{drill-down fallacies} as wrongfully attributing a deviation in trend to a local change, while it is in fact a more general phenomena. Cho et al.~\cite{cho2017anchoring} investigate the impact of \emph{anchoring bias} (i.e. focusing too much on one piece of information) while making decisions.
This paper focuses on models for \textit{exploration bias}. Exploration bias is akin to \textit{sample selection bias}, a generic phenomenon in social sciences when the investigator fails to capture a random sample of the target population~\cite{winship1992models}. Much like sample selection bias can lead to biased inferences about a study population, exploration bias can result in false deductions from the data. Using this as motivation, Wall et al.~\cite{wall2017warning} introduced \textit{Attribute Distribution}, which is a metric for quantifying exploration bias by observing user interactions while inspecting individual data points. Their work aimed to draw the connection between cognitive biases and exploration bias (\ie\ cognitive biases can cause exploration bias) , with proposed subsequent solutions for increasing bias awareness during data analysis~\cite{narechania2021lumos}.
While exploration bias in the context of visual analytics often has a negative connotation, it may at times be intentional and be interpreted as a user's data interest \cite{wall2017warning}.
Other works have investigated several techniques to detect and quantify exploration bias a user may manifest while interacting with datasets.
For example, Gotz et al.~\cite{gotz2016adaptive} presented \textit{Adaptive Contextualization}, a statistical approach to detecting exploration bias by observing user interactions while filtering high-dimensional data.
In their recent work, Monadjemi et al.~\cite{monadjemi2020competing} detected exploration bias via Bayesian model selection.
While the motivation behind each of these modeling techniques may be different, they all share a critical characteristic: in essence, they all compare the distribution of the underlying data set with the distribution of data points with which the user has interacted.
\subsection{Data Interaction Prediction}
This paper also examines algorithms that predict users' data interactions. There has been several techniques for inferring which data points the user will likely explore in the future.
For example, Healy et al.~\cite{healey2012interest} dynamically identified and tagged data elements in their visualization that were of potential interest based on the user's actions using Bayesian classification. Battle et al.~\cite{battle2016dynamic} modeled a user's sequence of interactions to pre-fetch data for a map visualization. They maintained a Markov chain model of interactions, and ranked the data for pre-fetching according to the likelihood of the users taking actions corresponding to said data. Similarly, Ottley et al.~\cite{ottley2019follow} observed users' mouse interaction and succeeded in modeling the attention of users during visual data exploration with a hidden Markov model. Their algorithm leverages a user's observed clicks with a visualization system to obtain next click predictions. In addition to bias detection, Monadjemi et al.~\cite{monadjemi2020competing} showed that their framework can predict future interactions and summarize analytic sessions. Zhou et. al~\cite{zhou2021modeling} observed users to develop a model of their analytical focus and use that model to surface relevant medical publications to users during visual analysis of a large corpus of medical records. As evident in this brief overview, there are many variations of techniques that learns data points that may to relevant to a user by observing their data exploration. However, the visual analytics literature lacks a unified comparison among these techniques.
\subsection{Evaluation of User Modeling Techniques}
There are many ways to evaluate an algorithm or technique, including complexity analysis, implementation analysis, and laboratory user study~\cite{munzner2009nested}. The choice of evaluation approach depends on the project goals and contributions. As a result, we observed a diverse set of validation methods ranging from crowdsourced user studies and case studies to interaction simulations (\eg\ \cite{ottley2019follow, zhou2021modeling, monadjemi2020active}). For this paper, we utilize user study datasets to assess how the current modeling approaches might perform under actual user data to gain insight into their potential use in a real-time system. Therefore, we review the prior work on evaluated user modeling techniques with user study datasets.
For example, Ottley et al.\cite{ottley2019follow} validated their model's ability to predict future interactions on one user study. While their study validated the technique successfully, it did not evaluate their technique against any baselines.
Zhou et al.\cite{zhou2021modeling} validated their technique's ability to identify relevant medical concepts via a user study.
They successfully validated their technique by comparing it to the ground truth elicited from the study participants and measured a set of usability metrics via a post-study survey. However, their analysis also did not include any baselines.
Additionally, Gotz et al.\cite{gotz2016adaptive} conducted a formal user study to investigate how their technique (Adaptive Contextualization) can mitigate selection bias. Their user study included a two experimental groups: one with access to bias mitigation features, and one without access to such features (baseline). With their work being one of the earlier work in detecting selection bias, they would not have been able to compare against other techniques that were proposed in later years.
Most similar to our work, Monadjemi et al.~\cite{monadjemi2020competing} evaluated the performance of their technique (Competing Models) with two existing techniques by Ottley et al.~\cite{ottley2019follow} and Wall et al.~\cite{wall2017warning} for data interaction prediction and bias detection, respectively. Their evaluation, while including baselines and multiple user studies, was still limited in that it did not include all existing techniques as baselines and did not consider as many user studies.
We extend upon their work by standardizing and comparing the performance of four additional user modeling techniques with two additional datasets. Moreover, we introduce an ensemble approach that combines the selected modeling techniques for both exploration bias detection and data interaction prediction. We hope that our analysis of these techniques will shed some light for researchers and practitioners who are interested utilizing them.
\input{tables/methods_overview}
\input{sections/scope}
\input{sections/methods2}
\begin{figure*}[!ht]
\centering
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[width=0.87\textwidth]{figures/political_user_session.png}
\end{subfigure}
%
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[width=0.87\textwidth]{figures/stl_agg_session.png}
\end{subfigure}
\caption{Each block represents a certain time $t$ within the observed interactions. The color represents the top-ranking attribute and the transparency level represents the technique's confidence. (\includegraphics[width=.015
\textwidth]{figures/warning.png} indicates that an assumption of the Chi-Square test was not met.)}
\label{fig:bias}
\vspace{-1em}
\end{figure*}
\section{Results}
\subsection{Data Interaction Prediction}
\label{sec:dip_results}
Figure~\ref{fig:inference} shows the models' accuracy across different values of $\kappa \in \{1,5,10,20,50,100\}$, the size of the prediction. We analyzed each algorithms' success rate at predicting the next observed interaction, calculating the overall success rate for all available user interaction traces for each dataset as \mbox{$\sum success \div \sum predictions$}. For simplicity, the results in this section focused on $\kappa = 100$, \ie\ the algorithm picks the 100 points with the highest rankings for the next interaction.
\vspace{.5em}
\noindent
\textbf{STL Crimes:}
The top row of Figure~\ref{fig:inference} shows the three subtasks in the in \textit{STL Crimes} dataset (\textit{location-based}, \textit{type-based}, and \textit{mixed}). We can observe that most of the techniques can predict the next data interaction point with high success rate at $\kappa = 100$, which represents 5\% of the underlying data. For location-based tasks, \hmm\ and \cm\ are the best performing models, with success rates of 95\% and 94\% at $\kappa = 100$, respectively. With the type-based task, \ens\ performs the best with a success rate of 93\% at $\kappa = 100$. For the mixed-based task, \ens\ outperforms all the other models, successfully predicting the next data interaction 98\% of the time.
\vspace{.5em}
\noindent
\textbf{Vastopolis:}
The \textit{STL Crimes} user study was highly directed, asking users to perform specific tasks that lead their exploration. On the other hand, the \textit{Vastopolis} task was less directed, involved a larger data volume, and included textual data (participants interacted with Tweet-like data). The selected models are unable to reach high success rate for data interaction prediction, and success rates were consistently less than 50\% at $\kappa = 100$ points. Nonetheless, \ens\ and \af\ perform the best with this interaction log dataset.
\vspace{.5em}
\noindent
\textbf{Boardrooms:}
The user study sessions with the \textit{Boardrooms} dataset were highly open-ended. Due to the open nature of the task, all of the models had difficulty achieving high success rate for data interaction prediction. \hmm\ outperforms all the other modeling techniques with a success rate of 51\% at $\kappa = 100$.
\vspace{.5em}
\noindent
\textbf{Political Committee:}
With the \textit{Political Committee} interaction logs, we wanted to stress-test the modeling techniques and observe their performance on set of minimal, but \textit{intentional} interactions. However, is it important to note that the dataset set is comparably small, with 180 fictitious politicians. \cm\ and \af\ showed promising performance with success rates of 67\% and 66\% at $\kappa = 100$, respectively.
\subsection{Exploration Bias Detection}
For the evaluation of bias detection, we chose the two following interaction datasets: \textit{STL Crimes} and \textit{Political Committee}. These two datasets were chosen as both studies either aimed to encourage specific biases or elicit the users' inherent biases during data exploration. Although we include the results of \ad\ in this section, we note that its performance may be unreliable as one of the Chi-Square test assumptions was not met. The Chi-Square test assumes that: \textit{the value of the expected cell should be 5 or more in at least 80\% of the cells, and no cell should have an expected of less than one}~\cite{kim2017statistical}.
\vspace{.5em}
\noindent
\textbf{Political Committee:}
Each algorithm outputs a \textit{bias value} for the attributes in the dataset. For example, \textit{Political Committee} has nine attributes including (age, gender, and occupation, \etc) and based on the assigned values, we can rank the attributes by the techniques' bias belief. For simplicity, Figure \ref{fig:bias} and the analysis in this section considers only the attribute with the highest bias belief after each user's observation. As detailed in section~\ref{sec:bias_detection_gt} our qualitative coding revealed no significant differences between the distribution of values explored by the user and the underlying data distribution, meaning no ground-truth bias was detected for every session. Although there is no evidence of exploration bias, the top of ~\autoref{fig:bias} shows the top-ranking bias found by each technique for every user throughout their data exploration.
The absence of a ground-truth bias is potentially depicted by the lack of agreement among the techniques for the top-ranking bias. However, we see some agreement for \textit{Lima}, as the techniques believed their exploration was biased towards the occupation attribute.
\vspace{.5em}
\noindent
\textbf{STL Crimes:}
The \textit{STL Crimes} user study visually displayed three attributes (longitude, latitude, and type of crime) and included three subtasks (\textit{location-based}, \textit{type-based}, and \textit{mixed}). Since users were tasked to specifically interact with subsets of the data, we set these attributes as ground-truth biases. To calculate the bias value of location, we multiplied the bias values for longitude and latitude. To calculate the mixed bias value, we multiplied the location bias with the bias value for type of crime. The bottom row of Figure \ref{fig:bias} shows the aggregated top-ranking bias among the users over time. For both location-based and type-based tasks, all of the modeling techniques were able to accurately determine the attribute that users were most biased towards. However, for the mixed task, we observed that \ad\ and \cm\ were the only modeling techniques that detected that the users were interested in both type and location attributes. For \ens, \ac, and \hmm, the techniques believed the users were focused on only type-based and only location-based attributes, respectively.
\section{Discussion}
This study compared eight algorithms for data interaction prediction and exploration bias detection across four user study datasets. \textit{Perhaps the most salient finding is there is no clear `winner.'} In particular, we found that no single model excelled with all datasets and the differences between the top-ranking techniques are small.
\subsection{Insights from Data Interaction Prediction Results}
The \textit{STL Crimes} data subsets demonstrate that the models generally perform relatively well when the tasks are highly-specified and have limited exploration noise. We use \textit{STL Crimes} to represent a best-case modeling scenario. Equally important, our evaluation included datasets that were selected to test the algorithms' robustness. In particular, \textit{Vastopolis} includes unstructured text, \textit{Boardrooms} involves an open exploration task, and \textit{Political Committee} presents a limited observation window with 10-16 interactions per participant. We can observe that the models' success rates fell drastically with these less directed and open-ended user study datasets. Still, there are a few noteworthy discoveries from this analysis. For example, \af~\cite{zhou2021modeling} (developed initially for textual data exploration) achieved the highest overall success rate with the \textit{Vastopolis} dataset, the only dataset with unstructured text. In addition, \hmm~\cite{ottley2019follow}, surpassed the others on the only dataset for which the task was entirely open-ended. However, in both instances, the overall success rates were relatively low.
Given the high variability in the success rates, we introduced an \ens\ approach~\cite{rokach2010ensemble,dietterich2000ensemble} by averaging the computed probabilities, allowing each model to vote for the data points they believe would best match the users' next data interactions in light of its observations. Then, we rank the data points based on their probability mass and select the \topk\ accordingly for predictions. The \ens\ model was among the top predictors for all but one dataset. We saw that averaging the predictions resulted in \ens\ achieving the second worst overall performance with the \textit{Political Committee} dataset, which we included to represent a limited observation scenario. Although we observe an overall improved performance in data interaction prediction with \ens, especially when $\kappa$ is small, we do note that this approach can be costly to compute and may not be suitable for real-time systems.
\subsection{Insights from Exploration Bias Detection Results}
The \textit{STL Crimes} dataset also provided a convenient means for evaluating exploration bias. In particular, the three data subsets involved tasks in which the study participants either focused on a specified location, data type, or a mixture of the two, allowing us to use these as ground truths. All five tested algorithms successfully identified that participants disproportionally allocated interactions toward location and type-related attributes. However, the most interesting finding relates to the mixed task data subset. Specifically, we can observe that only \cm~\cite{monadjemi2020competing} and \ad~\cite{wall2017warning}, accurately detected that the task elicited a mixed bias.
Furthermore, the exploration bias detection findings are notably insightful for the \cm\ and \hmm\ techniques -- the only two algorithms capable of both data interaction prediction and exploration bias detection. Although both algorithms achieved more than a 90\% success rate in predicting the next data interaction at $\kappa = 100$, the exploration bias detection results provide further insight into the models' beliefs and \textit{how} they produced their data relevance rankings. Markedly, correctly identifying the mixed nature of the task meant that \cm\ was one of the best data interaction predictors, second to only \ens\ for this data subset. In contrast, misclassifying the task as location-based resulted in an overall lower prediction performance for the \hmm\ approach.
Although our qualitative coding failed to detect exploration bias in the committee picks from participants in the \textit{Political Committee} dataset, we included the evaluation results to assess the algorithms' predictions in light of unbiased interaction traces. We observed that all the models presented false positives, often with high certainty, as indicated by the opacity of the blocks in Figure \ref{fig:bias}. It is noteworthy that the one underlying assumption of some of the models is that the user is always biased. For example, the \hmm\ algorithm explicitly includes bias in the algorithm's specification. Specifically, it maintains a \textit{bias probability} for each feature in the dataset, allowing the algorithm to capture the likelihood of bias toward one or more data features. \cm\ is the only algorithm that directly encodes the potential for no bias. The algorithm tracks bias by building a model for all possible feature subsets, including the null set, representing no potential exploration bias. However, the null set is only one of the numerous model choices considered in \cm, making it an unlikely pick.
Altogether, this underlying assumption that the user's interaction is always biased has both positives and negatives for real-world applications. On the one hand, we believe that identifying the features that the user disproportionately explores drives the algorithms' success in data interaction prediction. Also, it is potentially a reasonable assumption for many exploration scenarios. On the other hand, the false positives can erode the user's trust in the algorithms, especially in proposed applications related to helping the user recognize their own exploration bias or recommender systems~\cite{gotz2016adaptive, wall2017warning, wall2022left}. That said, a potential solution exists within two models, \ac~\cite{gotz2016adaptive} and \ad~\cite{wall2017warning}, which leverage statistical methods for hypothesis testing in their bias detection. If used as intended, a failure to reject the null hypothesis provides evidence of no bias. However, practitioners and researchers should use the appropriate tests for the given data and verify that the data meet their assumptions.
\subsection{Challenges and Limitations}
There were several challenges we faced when preparing to perform a unified comparison of these models. For example, implementations of the techniques were not readily available. We invested time gaining an understanding of all the techniques and standardizing the implementations to handle various datasets. We note that we are not able to completely replicate the performance of \cm\ and \hmm\ on data interaction prediction with the \textit{STL Crimes} dataset that we see in Monadjemi et al~\cite{monadjemi2020competing}. We believe that the differences in performance are likely due to our hyperparameters selections.
When deploying algorithms in real-world settings, it is important to gather data and tune hyperparameters. However, we excluded this step in our work as the datasets were small, presenting a high risk for overfitting. Instead, we used our domain knowledge to make conscientious decisions about each model's hyperparameters, which are all available and adjustable in our codebase.
Although we selected a diverse collection of datasets, we acknowledge that they do not represent many visual analytics scenarios. Therefore, we take care in reporting our observations and hesitate to make broad statements. Further, we were limited by the small number of interaction datasets that are publicly available, especially those with ground truths. This highlights important considerations for balancing ecological validity where the ground truth is not known or may not exist, and having some measure of success for evaluation. Determining ground truth when evaluating bias in open-ended tasks is not trivial. A call for more guided user studies with built-in biases may be needed so that researchers can further evaluate bias detection algorithms.
\section{Future Directions}
Despite our limitations, our work is a step forward in evaluating user modeling techniques proposed in the visual analytics community and opens possibilities for future work in this area. For example, our evaluations only involved datasets with a single interaction type, even though some of the modeling techniques are able to or have the potential to handle multiple types of interactions~\cite{zhou2021modeling, ottley2019follow, monadjemi2020competing}. Although our implementations did not include this feature, future work could extend this benchmark study with the added ability to learn from multiple types of interactions with data points. Further, we primarily focused on quantitative measures, but there are other factors beyond accuracy that might affect the appropriateness of a technique for a given user modeling scenario. Factors such as speed, trustworthiness, and flexibility can be important considerations to explore in future evaluations of user modeling techniques.
\section{Conclusion}
In this paper, we present a computational benchmark study that provides a unified comparison of user modeling techniques for data interaction prediction and exploration bias detection. We implemented standardized versions of seven previously proposed user modeling techniques that learns from a user's low-level interactions in real-time. Additionally, we developed an ensemble approach by averaging the models' predictions. We then evaluated the performance of all these techniques across four different interaction logs for data interaction prediction/data ranking and exploration bias detection. We found that there is no clear `winner' among the modeling techniques, but our analysis highlights the tasks and datasets that elicited the best performance for each technique. Finally, we discuss the open challenges for user modeling and evaluations. This work is a step towards gaining an understanding of user modeling techniques within the visual analytics community and we hope that it encourages further advances in analyzing user interactions and visualization provenance.
\acknowledgments{
This work is supported in part by the National Science Foundation under Grant No. OAC-2118201 and IIS-2142977.}
\clearpage
\bibliographystyle{abbrv-doi}
\section{Methods and Materials}
\label{sec:methods}
We compare a total of eight user models, based on a variety of theoretical approaches and summarized in \autoref{tab:methods-overview}. This section describes our notation, the algorithms, datasets, and the performance measures.
\subsection{User Modeling Methods}
To form our collection of algorithms discussed in this paper, we started with the pool of papers that were discussed in Xu et al.'s~\cite{xu2020survey} 2020 survey on the analysis of user interactions and provenance in visual analytics. We accounted for the papers published around the same time and after the survey by performing a forward citation search on the manuscripts in the survey and the survey itself. We narrowed the pool to the final selection of seven papers\cite{ottley2019follow, wall2017warning, monadjemi2020competing, gotz2016adaptive, monadjemi2020active, healey2012interest, zhou2021modeling} based on the criteria described in \autoref{sec:scope}, each including algorithms for either learning information relevance or bias detection from observing data interactions. Of the algorithms we selected, three can only infer information relevance (\textit{k}-Nearest Neighbors\cite{monadjemi2020active}, Boosted Naive Bayes\cite{healey2012interest}, and Analytic Focus \cite{gotz2016adaptive} ), two can only detect bias (Attribute Distribution Bias Metric\cite{wall2017warning} and Adaptive Contextualization\cite{gotz2016adaptive}), and two can infer both (Hidden Markov Model\cite{ottley2019follow} and Competing Models\cite{monadjemi2020competing}). In addition to the selected algorithms, we also introduce and provide the performance of an ensemble model which was developed with all the modeling techniques mentioned above that infer information relevance. We provide their algorithmic details below.
\subsubsection{Inferring Information Relevance}
This category of algorithms aims to learn users' latent data interest and rank data points based on this interest. Given this ranking, a common applications are attention tracking and next click prediction\cite{monadjemi2020active,monadjemi2020competing,ottley2019follow}. We define the \emph{relevance function} as follows:
\begin{equation}
\label{eq:f_rank}
f_{r} \colon \mathcal{D} \mapsto [0, 1],
\end{equation}
where the set of data points, $\mathcal{D}$, is mapped to a numerical value between 0 and 1. Higher values correspond data points being more relevant to the task at hand.
Given two data points $x_i, x_j \in \mathcal{D}$, we say $x_i$ is more relevant to the task at hand, $x_i \succ x_j$, if it has a higher relevance value, $f_{r}(x_i) \textgreater f_{r}(x_j)$.
\vspace{1em}
\noindent
\textbf{\textit{k}-Nearest Neighbors Classifier} as seen in \cite{monadjemi2020active}:
\vspace{0.5em}
\noindent
A straightforward method for learning users' data interest by observing interactions is to train a $k$-nearest neighbors (\knn) binary classifier.
\textbf{This approach assumes that proximity drives a user's exploration patterns.}
Monadjemi et al.\cite{monadjemi2020active} built a \knn classifier that relies on a notion of distance between data points and computing the matrix of $k$ nearest neighbors for every data point given the defined distance formula as a pre-processing step. The complexity for this pre-processing step is $O(n^{2}d)$. Refer to \autoref{tab:methods-overview} for more information and complexity variables. This model assumes that every data point in in the dataset, $x_i \in \mathcal{D}$, has a binary label $y_i \in \{0, 1\}$.
A label of $y_i=1$ means the data point $x_i$ is relevant to the task at hand, whereas a label of $y_i=0$ means $x_i$ is irrelevant. By considering the labels of the nearest neighbors, this model is able to provide us with the probability of any given point being relevant in light of past observations: $\Pr(y_i=1 \mid x_i, \mathcal{I})$. Hence, we use this posterior belief as our relevance function:
\begin{equation}
f_r(x_i) = \Pr(y_i=1 \mid x_i, \mathcal{I}).
\end{equation}
\vspace{1em}
\noindent
\textbf{Boosted Naive Bayes Classifier} as seen in \cite{healey2012interest}:
\vspace{0.5em}
\noindent
Healey et al.\cite{healey2012interest} propose a naive Bayes classifier for maintaining a belief over users' interest in data points.
This approach tracks the frequencies of attributes explored and \textbf{assumes that the user's latent data interest is related to the dataset attributes' occurrence rate.}
In a similar formulation as the \knn\ approach, each data point $x_i \in \mathcal{D}$ has a binary label $y_i \in \{0, 1\}$. A label of $y_i = 1$ means the point $x_i$ is relevant to the task at hand and a label of $y_i = 0$ means it is not relevant. The goal of this approach is to reason about the unknown labels in light of frequencies of observations calculated via the Bayes' law:
\begin{equation}
\Pr(y_i=1 \mid x_i) \propto \Pr(x_i \mid y_i = 1)\Pr(y_i=1).
\end{equation}
Extending $x_i$ to each of its individual $d$ dimensions and assuming conditional independence among attributes, we get:
\begin{equation}
\begin{split}
&\Pr(y_i=1 \mid x_{i, 1}, x_{i, 2}, ..., x_{i, d})\\
\propto &\Pr( x_{i, 1}, x_{i, 2}, ..., x_{i, d} \mid y_i = 1)\Pr(y_i=1) \\
= & \Pr(y_i=1) \prod_{j=1}^{d} \Pr(x_{i,j} \mid y_i=1),
\end{split}
\end{equation}
where $x_{i, j}$ denotes the $j$th attribute of the $i$th data point. Therefore, we get the relevance function as:
\begin{equation}
f_r(x_i) = \Pr(y_i=1) \prod_{j=1}^{d} \Pr(x_{i,j} \mid y_i=1).
\end{equation}
While Naive Bayes often performs as well as more sophisticated models in practice \cite{langley1992analysis}, its performance can further be improved by \emph{boosting}. As recommended by Healey et al.\cite{healey2012interest}, we used AdaBoost to train a stronger classifier.
\vspace{1em}
\noindent
\textbf{Analytic Focus Modeling} as seen in \cite{zhou2021modeling}:
\vspace{0.5em}
\noindent
Zhou et al.\cite{zhou2021modeling} propose a technique for modeling analytic focus (\af) during a session. Their technique is based on an abstract notion of \emph{concepts} defined as \emph{``meaningful data attribute[s] in the problem domain."} This technique tracks user focus on each of the concepts by observing their interactions and maintain an importance score for each concept. \textbf{The model assumes that the user's latent interest is related to the occurrence rate of discrete \textit{concepts} observed in their data interactions.}
With concepts being left open for practitioners to define in their particular domain, here, we let each unique value appearing in the discrete attributes of the dataset to represent a concept, performing a pre-processing step on continuous attributes to convert them into discrete bins.
Each interaction $\alpha \in \Alpha$ is defined by two values: an initial importance score $I_\alpha(0)$ and a persistence score $P_\alpha$. In a sense, $I_\alpha(0)$ refers to how intentional an action is and how informative it is in revealing analytic focus while $P_\alpha$ is a measure of how long the visual changes resulting from action $\alpha$ remain.
Using these parameters and the Ebbinghaus forgetting curve \cite{ebbinghaus1880urmanuskript,ebbinghaus1885gedachtnis,murre2015replication}, the authors define the per-action importance score function $t$ time steps after an action is taken as:
\begin{equation}
I_{\alpha}(t) = I_{\alpha}(0) \times e^{-\frac{t}{P_{\alpha}}}.
\end{equation}
Using the equation above, the authors propose an additive model for calculating the importance score for each concept $c$ at time $\tau$:
\begin{equation}
I^c(\tau) = \sum_{(x_{i, t}, \alpha) \in \mathcal{I}_{\tau,c}}I_{\alpha} (\tau - t),
\end{equation}
where $\mathcal{I}_{\tau, c}$ denotes the set of interactions with points involving concept $c$ up until time step $\tau$. Considering that each data point $x_i \in \mathcal{D}$ may subscribe to multiple concepts (one concept per dimension in our case), we extend this method so that the importance of a data point is the product of the importance of its individual concepts. In other words, given a data point $x_i \in \mathcal{D}$ and the current time step $\tau$,
\begin{equation}
f_r(x_i) = \prod_{c \in C(x_i)}{I^c(\tau)},
\end{equation}
where $C(x_i)$ is the list of concepts with which $x_i$ is associated.
\subsubsection{Detecting Exploration Bias}
The goal of bias detection is to determine if user interactions are biased towards a certain attribute in the dataset. Detected biases can be used to bring awareness and mitigate users from focusing on a subset of the data~\cite{gotz2016adaptive,wall2017warning,wall2022left}. Additionally, bias detection can be leveraged to infer the user's interest and incorporated into intelligent systems to better aid the user during data exploration. Using a standard notation for every approach considered in this work, we define the \emph{bias function}, $f_{b}$, as follows:
\begin{equation}
\label{eq:f_bias}
f_{b} \colon \mathcal{A} \mapsto [0, 1],
\end{equation}
where the set of attributes, $\mathcal{A}$, is mapped to a numerical value between 0 and 1. Higher values correspond to more severe bias towards a certain attribute.
\vspace{1em}
\noindent
\textbf{Attribute Distribution Bias Metric} as seen in \cite{wall2017warning}:
\vspace{0.5em}
\noindent
Wall et al.\cite{wall2017warning} propose a set of metrics for quantitatively detecting various types of cognitive biases. The most relevant to our work is the metric they call \emph{attribute distribution} bias.
\textbf{Their approach assumes that in an unbiased exploratory session, the distribution of values explored by the user will resemble the distribution of the underlying data.} It relies on well-known statistical tests, Chi-Square\cite{mchugh2013chi} and Kolmogorov–Smirnov\cite{massey1951kolmogorov}, which are used for discrete and continuous attributes respectively. In each case, the test takes the underlying data as well as the set of points with which the user has interacted as inputs and returns a $p$-value indicating whether or not the two samples could be reasonably drawn from the same distribution. Then, they define their attribute distribution metric for a given attribute $a_i \in \mathcal{A}$ to be:
\begin{equation}
f_{b}(a_i) = 1 - p\text{-value}_{a_{i}}.
\end{equation}
\vspace{1em}
\noindent
\textbf{Adaptive Contextualization} as seen in \cite{gotz2016adaptive}:
\vspace{0.5em}
\noindent
Gotz et al.\cite{gotz2016adaptive} propose a metric to measure selection bias in real-time. This methods assumes that you can detect exploration bias by comparing the distributing
In order to detect such biases, they propose measuring the Hellinger distance between the datasets before and after filtering. Hellinger distance is a statistical measure designed to quantify the similarity between two distributions. Thus, \textbf{this approach assumes that we can detect bias by comparing the observed distribution to the expected data distribution}. The authors propose using the Hellinger distance defined for discrete probability distributions with a pre-processing step to discretize the continuous attribute into discrete bins.
Consider a particular dimension of interest, $a_i$, represented as a set of $m$ discrete values $\{v_1, v_2, ..., v_m\}$. The distribution of these values in the underlying data and interaction data are $\{p_1, p_2, ..., p_m\}$ and $\{q_1, q_2, ..., q_m\}$ respectively. We compute the Hellinger distance as:
\begin{equation}
f_b(a_i) = H(\mathcal{D}_{a_i}, \mathcal{I}_{a_i}) = \sqrt{\frac{1}{2} \sum_{j=1}^{m}{(\sqrt{p_j} + \sqrt{q_j})^2} },
\end{equation}
where $\mathcal{D}_{a_i}$ and $\mathcal{I}_{a_i}$ denote only the $a_i$ attribute of the underlying and interaction datasets respectively. While this technique was intended to detect selection bias (i.e.\ before vs.\ after filtering), we adopt it to quantify exploration bias.
\subsubsection{Inferring Both Information Relevance and Exploration Bias}
\label{sec:hmm-cm}
\vspace{1em}
\noindent
\textbf{Hidden Markov Model} as seen in \cite{ottley2019follow}:
\vspace{0.5em}
\noindent
Ottley et al.\cite{ottley2019follow} propose a Hidden Markov Model (\hmm) approach for modeling user attention during visual exploratory analysis. \textbf{Their approach assumes that the visual features of the interface will drive the user's attention (it considers all possible combinations)} and their interactions such that attention at time $t+1$ will be similar to the observation at time $t$ . They constructed a hidden Markov model, presuming the user's behavior evolves under a Markov process (that is, the behavior at a particular time only depends on their behavior at the previous time step), and interaction events are generated conditionally independently given this sequence of attention shifts. The set of visible visual features (\eg\ color, position, size, etc.) creates the \textit{mark space} and the latent states. Further, their algorithm uses particle filtering~\cite{gordon1993novel,doucet2000sequential} to sample from the model's posterior distribution and predict the user's attention for the next time step. Integral to this technique is a \textit{bias vector} that captures the relative importance of the mark space components. We use the particle filter output to predict information relevance and the bias vector for bias detection.
\begin{figure*}[!ht]
\centering
\includegraphics[width=\linewidth]{figures/datasets.png}
\caption{A summary of the characteristics of the selected user interaction datasets and the performance of the user modeling techniques with said datasets. We represent success using each technique's performance at top-100 points for information relevance.}
\label{fig:datasets}
\end{figure*}
\vspace{1em}
\noindent
\textbf{Competing Models} as seen in \cite{monadjemi2020competing}:
\vspace{0.5em}
\noindent
Monadjemi et al.\cite{monadjemi2020competing} propose a Bayesian model selection approach for detecting exploration bias and inferring information relevance during visual exploratory analysis. This approach, called \emph{Competing Models} (\cm), enumerates a set of probabilistic models each of which represent exploration based on a subset of data attributes.
Given a dataset with $d$ attributes, they enumerate the model space ${\mathbf{M} = \{\mathcal{M}_1, \mathcal{M}_2, ..., \mathcal{M}_{2^d}\}}$ where each model represents exploration based on one subset of attributes. \textbf{Thus, it assumes and one or more data attributes (it considers all possible combinations) are sufficient to represent a users data interest.} As interactions arrive, they maintain a belief ${\Pr(\mathcal{M}_i \mid \mathcal{I})}$ which is interpreted as the viability of model $\mathcal{M}_i$ in explaining user interactions.
To infer the relevance of data points to the task at hand in light of past interactions, they use Bayesian model averaging to get:
\begin{equation}
f_r(x_i) = \Pr(x_i \mid \mathcal{I}) = \sum_{j=0}^{2^d}\Pr(x_i \mid \mathcal{I}, \mathcal{M}_j).
\end{equation}
To quantify bias towards each attribute, they use the law of total probability as follows:
\begin{equation}
f_b(a_i) = \sum_{\mathcal{M}_j \in \mathbf{M}_{a_i}} \Pr(\mathcal{M}_j \mid \mathcal{I}),
\end{equation}
where $\mathbf{M}_{a_i} \subset \mathbf{M}$ denotes the set of models involving attribute $a_i$.
\subsubsection{Inferring Information Relevance via an Ensemble Approach}
\vspace{1em}
\noindent
\textbf{Ensemble Model}
\vspace{0.5em}
\noindent
We anticipate that no model will be able to correctly recognize relevant points in every system, task, or dataset. This phenomena was also observed by the machine learning community which led to the development of \emph{ensemble methods} \cite{rokach2010ensemble,dietterich2000ensemble}. Ensembles are a collection of classification models combined to construct a model with higher predictive power. Using this approach, we consider a final technique which combines predictions from all five techniques that infer information relevance reviewed in this paper. For a given data point, $x_i \in \mathcal{D}$, we compute the relevance function, $f_r(x_i)$, using each of the modeling techniques and average the relevance values to get the ensemble prediction. Then, we proceed with ranking the data points according to their relevance, a process identical to the one in the preceding modeling techniques.
\subsection{User Study Datasets}
\label{sec:datasets}
We used four publicly available datasets from user study experiments with visualization systems. These studies contained varying interfaces as well as task designs. We aimed to evaluate and compare the performance of the user modeling techniques with a variety of interaction logs, ranging from studies with very directed and specific tasks to the polar opposite, open data exploration. Refer to ~\autoref{fig:datasets} for a detailed summary of the characteristics of each dataset.
\subsubsection{Directed User Studies}
Although the user studies in this section are directed, they differ in the intensity of direction a user was provided while performing the task. We describe the user studies in more detail below.
\vspace{1em}
\noindent
\textbf{STL Crimes} as seen in \cite{ottley2019follow}:
\vspace{0.5em}
\noindent
The \textit{STL Crimes} dataset \textbf{involved a highly specified task thus capturing less exploration noise. We selected this dataset as it represents the best-case observational scenario for user modeling.}
Curated by Ottley et al.~\cite{ottley2019follow},
this dataset consists of 1,951 crime instances reported in the month of March 2017 with a total of eighteen attributes. Each instance of crime on the map was displayed as a dot with a position and color, indicating the location and type of the crime respectively. Eight types of crime exist in the database: homicide, theft-related, assault, arson, fraud, vandalism, weapons, and vagrancy. 30 participants were recruited from Mechanical Turk and they completed three categories of directed search tasks that were designed to encourage them to either click on: (1) \textit{Type-Based} data points with similar crime categories (e.g. Homicide, Assault, etc.), (2) \textit{Location-Based} data points that are within in the same vicinity, or (3) \textit{Mixed} data points that are of the same crime category and in the same vicinity. Each participant completed a total of six tasks, two from each category. Since the authors specified to the participants target attributes to interact with, this provides us with a notion of ground-truth bias. In their user study, Ottley et al.~\cite{ottley2019follow} captured mouse click data as participants interacted the visualization. Consistent with Ottley et al.~\cite{ottley2019follow}, we filtered the data to hold only the interactions of participants who successfully answered the task questions and created three data subsets: 28 user sessions for location-based tasks, 23 for typed-based, and 27 for mixed.
\vspace{1em}
\noindent
\textbf{Vastopolis} as seen in \cite{monadjemi2020active}:
\vspace{0.5em}
\noindent
We selected the \textit{Vastopolis} dataset because of the attributes in the dataset. \textbf{It is the only user interaction dataset in this study that includes unstructured text.}
Inspired by the 2011 VAST Challenge, Monadjemi et al.'s~\cite{monadjemi2020active} study describes a major epidemic that started in the fictitious city of Vastopolis. Participant's interacted with a map of Vastopolis with 3000 geolocated tweet-like data. In addition to the social media posts, the map display major roadways, waterways, and landmarks in the city. They recruited 130 participants from Amazon Mechanical Turk and tasked each participant to search through the dataset of microblogs via an interactive map and bookmark as many posts containing illness-related information as possible. Following the same pre-processing step as the authors, 74 user interaction logs remained from the control group of the study. The only interaction type we considered in the interaction logs were intentional ``bookmarks" of the microblogs.
\vspace{1em}
\noindent
\textbf{Political Committee} as seen in \cite{wall2022left}:
\vspace{0.5em}
The \textit{Political Committee} user study \textbf{examined cognitive and exploration bias as participants selected their picks for a hypothetical committee. Thus we selected this dataset as a candidate for examining exploration bias.}
\noindent
Wall et al.~\cite{wall2022left} generated a dataset of 180 fictitious politicians, representing the composition of the Georgia General Assembly. The dataset contains three discrete and six continuous attributes that characterizes each politician. In the study conducted by Wall et al.~\cite{wall2022left}, the authors had users interact with an visualization system that supports the exploration of the fictitious political committee data. There were two versions of the visualization system: a Control version of the interface, and an Intervention version of the interface, which was modified to visualize traces of the user’s interactions with the data in real-time. To assess the effectiveness of visualizing the user's interaction traces, the authors performed an in-lab study with 24 participants. They were tasked with selecting a committee of 10 candidates to review public opinion in Georgia on the controversial bill that bans abortion after 6 weeks. In order to observe the raw, inherent biases of the users, we selected the interaction logs of those in the Control group (total of 12 participants). The authors recorded several types of interactions the user can make with the interface, however, we filtered the interaction log for each user to contain only the interactions when the user ``bookmarked'' a politician to be a part of their political committee.
\subsubsection{Open-Ended User Study}
\vspace{1em}
\noindent
\textbf{Boardrooms} as seen in \cite{feng2018effects}:
\vspace{0.5em}
\noindent
The \textit{Boardrooms} study did not involve a particular tasks and instructed the participants to freely interact with the visualisation. \textbf{We expect this dataset to include the highest levels of exploration noise. }
Feng et al.~\cite{feng2018effects,feng2018patterns} utilized the visualization, ``Inside America’s Boardroom," which was published by the \textit{Wall Street Journal}~\cite{wsjboardrooms} in 2016. This point-based visualization displays seven different attributes to the user: market capitalization, ratio of unrelated board members, ratio of female board members, average age of board members, average tenure of board members, and median pay of board members. The study adopted an open-ended approach and observed the user's interactions as they explored the dataset through the visualization. The interaction log contains a succession of hovers each user made on different data points. To account for unintentional hovers, we followed the same approach as Monadjemi et al.~\cite{monadjemi2020competing}, each user's session was filtered only include hovers that lasted for over one second. Additionally, we only considered sessions that consisted of more than three hovers, resulting in a total of 39 user sessions.
\subsection{Evaluation Measures}
In this section, we describe the performances measures we used to evaluate the modeling technique on information relevance and bias detection.
\subsubsection{Information Relevance}
We developed a standardized probabilistic output for each algorithm to directly compare the models to each other. Specifically, techniques treat every data point as a potential candidate for being most relevant to the user data interest. Thus, each method outputs the relevance probability for each data point given a set of interaction observations. This output represent the technique's unique interpretation of information relevance. For example, \knn\ assumes that the data points most relevant to the user will minimize some measure of distance while \af\ examines the frequency on \textit{concepts} from the past interactions.
In our evaluation, ground truth relevance is unobservable. Still, we used the next interaction as a proxy, presuming that the user will interact with the data points they perceive to be most relevant to their task.
Given the model's probabilities, we can rank each data point based on algorithm's information relevance belief.
We report \topk\ data retrieval for the algorithms for $\kappa \in \{1,5,10,20,50,100\}$ which we will refer to as the \textit{prediction sets}. We considered a variety of set sizes. For example, an alternative is to use percentages instead of a fixed values. However, for large datasets, retrieving a small percentage to the dataset can still be overwhelming to the user if we consider the goal of provided real-time analysis support. This standardized output enables us to evaluate and compare each technique's ability to infer information relevance.
We record the following measures for every observed interaction the datasets:
\vspace{-.5em}
\begin{itemize}[noitemsep]
\item \textbf{\success:} $\in \{0,1\}$, binary value for whether the next interaction was included in the prediction set for all values of $\kappa$.
\item \textbf{\rank:} the ranked position of the next observed interaction such that lower is better.
\end{itemize}
\begin{figure}[t!]
\centering
\includegraphics[width=\linewidth]{figures/Lima.png}
\caption{Comparing participant Lima's interactions (foreground) to the full dataset distribution (background). Both coders reported no significant difference between the two distributions, thus indicating that no bias detected. We observed similar result for all twelve study participants from \cite{wall2022left}.}
\label{fig:exampleuser}
\end{figure}
\begin{figure*}[!ht]
\centering
\begin{subfigure}[b]{.95\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/ncp_results.png}
\vspace*{3mm}
\end{subfigure}
%
\begin{subfigure}[b]{.95\textwidth}
\vspace{-1em}
\centering
\includegraphics[width=\textwidth]{figures/ranking_results.png}
\end{subfigure}
\caption{We evaluated each model's performance with two metrics: (top) Average \textbf{success rate} across all the tasks in each user study dataset for varying values of the prediction set size and (bottom) Average \textbf{rank} of the next interaction point of the user modeling techniques.}
\label{fig:inference}
\end{figure*}
\subsubsection{Exploration Bias Detection}
\label{sec:bias_detection_gt}
Evaluating each technique's ability to detect exploration bias poses a unique challenge, in that \textit{ground truth} is not always readily available. When evaluating information relevance, we had access to a notion of \emph{success}: we wanted to predict next data points with which the user interacts, and we were able to verify our ability to do so by observing the next interaction point. For bias selection, however, such ground truth is not as easy to access, especially for user study datasets with open-ended tasks. With the \textit{Political Committee} dataset collected by Wall et al.~\cite{wall2022left}, for example, each user can exhibit their own set of biases as they explore the data.
For this set of evaluations, we use two datasets: \textit{STL Crimes} separated into \textit{location-based}, \textit{type-based}, and \textit{mixed} subsets and \textit{Political Committee}. The \textit{STL Crimes} user study involved a directed task, asking participant to interaction with specific portions of the data. We used the specific task as the ground truth. In our evaluation, an algorithm successfully detects exploration bias if its outputs indicate data attributes related to crime location, type, or a mixture model of type and location for \textit{location-based}, \textit{type-based}, and \textit{mixed} respectively.
For the \textit{Political Committee}, two researchers performed qualitative coding on each study participants' committee selections, following best practices~\cite{brod2009qualitative}. We used an iterative approach. First, the coders reviewed a single participant's selections and density plots showing the attribute distributions for a given set of interaction logs and the original data distribution. Their goal was to identify evidence of data bias in the observed selections. Next, they established a codebook from this first round and independently coded all twelve participants' interaction logs. After inspecting each of the twelve study participants, both coders agreed the was no observable bias in the interaction study. For example, \autoref{fig:exampleuser} compares study participant Lima's data selection to the individual attribute distribution. Although the coding sessions detected no bias, we included this analysis to assess the algorithms' performance in detecting bias when none exists.
\section{User Modeling Methods}
\label{sec:methods}
In this section, we provide a brief overview on the existing modeling techniques considered in our benchmark study. For most of the selected modeling techniques, their implementations were not readily available. For ones that were readily available, the implementations were hard-coded to work with specific datasets. In order to easily evaluate the modeling techniques across multiple datasets, we implemented every technique with a standardized syntax and input/output format.
The \nexticon\ icon indicates that the technique is capable of data interaction prediction. The \biasicon\ icon indicates that the technique is capable of exploration bias detection. We compare a total of eight user models, based on a variety of theoretical approaches and summarized in \autoref{tab:methods-overview}. Of the algorithms we selected, three can only infer next data interaction (\textit{k}-Nearest Neighbors\cite{monadjemi2020active}, Boosted Naive Bayes\cite{healey2012interest}, and Analytic Focus \cite{gotz2016adaptive}), two can only detect exploration bias (Attribute Distribution~\cite{wall2017warning} and Adaptive Contextualization\cite{gotz2016adaptive}), and two can infer both (Hidden Markov Model\cite{ottley2019follow} and Competing Models\cite{monadjemi2020competing}). In addition to the selected algorithms, we also introduce and evaluate an ensemble model, developed by averaging the output of all modeling techniques.
Each model assumes there is a visualization of a dataset ${\mathcal{D} = \{x_1, x_2, ..., x_n\}}$, where each data point in $\mathcal{D}$ corresponds to a visual element. Each data point $x_i \in \mathcal{D}$ has $d$ attributes which may be continuous, categorical, or ordinal. Users may take different actions while interacting with data points using the visualization (i.e. click, hover, etc.). Thus, the models also assume a non-empty set of interactions supported by the system as $C = \{c_1, c_2, ...\}$. As users interact the visualized data points, they maintain a set of observation ${\mathcal{I} = \{(x_{i, t}, c) \mid t = 1, 2, ...}\}$, where $x_{i, t}$ denotes an interaction of type $c$ with data point $x_i$ at time step $t$.
\subsection{\textit{k}-Nearest Neighbors Classifier as seen in \cite{monadjemi2020active} \hfill \nexticon\ }
\noindent
A straightforward method for learning users' data interest by observing interactions is to train a $k$-nearest neighbors (\knn) binary classifier.
\textbf{This approach assumes that proximity drives a user's exploration patterns.}
Monadjemi et al.\cite{monadjemi2020active} built a \knn classifier that relies on a notion of distance between data points and computing the matrix of $k$ nearest neighbors for every data point given the defined distance formula as a pre-processing step. The complexity for this pre-processing step is $O(n^{2}d)$. Refer to \autoref{tab:methods-overview} for more information and complexity variables. This model assumes that every data point in in the dataset, $x_i \in \mathcal{D}$, has a binary label $y_i \in \{0, 1\}$.
A label of $y_i=1$ means the data point $x_i$ is relevant to the task at hand, whereas a label of $y_i=0$ means $x_i$ is irrelevant. By considering the labels of the nearest neighbors, this model is able to provide us with the probability of any given point being relevant in light of past observations: $\Pr(y_i=1 \mid x_i, \mathcal{I})$. Hence, we use this posterior belief as our ranking function:
\begin{equation}
f_r(x_i) = \Pr(y_i=1 \mid x_i, \mathcal{I}).
\end{equation}
\subsection{Boosted Naive Bayes Classifier as seen in \cite{healey2012interest} \hfill \nexticon\ }
\noindent
Healey et al.\cite{healey2012interest} propose a naive Bayes classifier for maintaining a belief over users' interest in data points.
This approach tracks the frequencies of attributes explored and \textbf{assumes that the user's latent data interest is related to the dataset attributes' occurrence rate.}
In a similar formulation as the \knn\ approach, each data point $x_i \in \mathcal{D}$ has a binary label $y_i \in \{0, 1\}$. A label of $y_i = 1$ means the point $x_i$ is relevant to the task at hand and a label of $y_i = 0$ means it is not relevant. The goal of this approach is to reason about the unknown labels in light of frequencies of observations calculated via the Bayes' law:
\begin{equation}
\Pr(y_i=1 \mid x_i) \propto \Pr(x_i \mid y_i = 1)\Pr(y_i=1).
\end{equation}
Extending $x_i$ to each of its individual $d$ dimensions and assuming conditional independence among attributes, we get:
\begin{equation}
\begin{split}
&\Pr(y_i=1 \mid x_{i, 1}, x_{i, 2}, ..., x_{i, d})\\
\propto &\Pr( x_{i, 1}, x_{i, 2}, ..., x_{i, d} \mid y_i = 1)\Pr(y_i=1) \\
= & \Pr(y_i=1) \prod_{j=1}^{d} \Pr(x_{i,j} \mid y_i=1),
\end{split}
\end{equation}
where $x_{i, j}$ denotes the $j$th attribute of the $i$th data point. Therefore, we get the ranking function as:
\begin{equation}
f_r(x_i) = \Pr(y_i=1) \prod_{j=1}^{d} \Pr(x_{i,j} \mid y_i=1).
\end{equation}
While Naive Bayes often performs as well as more sophisticated models in practice \cite{langley1992analysis}, its performance can further be improved by \emph{boosting}. As recommended by Healey et al.\cite{healey2012interest}, we used AdaBoost to train a stronger classifier.
\subsection{Analytic Focus Modeling as seen in \cite{zhou2021modeling} \hfill \nexticon\ }
\noindent
Zhou et al.\cite{zhou2021modeling} propose a technique for modeling \textit{analytic focus} (\af) during a session. Their technique is based on an abstract notion of \emph{concepts} defined as \emph{``meaningful data attribute[s] in the problem domain."} This technique tracks user focus on each of the concepts by observing their interactions and maintain an importance score for each concept. \textbf{The model assumes that the user's latent data interest is related to the occurrence rate of discrete \textit{concepts} observed in their data interactions.}
With concepts being left open for practitioners to define in their particular domain, here, we let each unique value appearing in the discrete attributes of the dataset to represent a concept, performing a pre-processing step on continuous attributes to convert them into discrete bins.
Each interaction $\alpha$ is defined by two values: an initial importance score $I_\alpha(0)$ and a persistence score $P_\alpha$. In a sense, $I_\alpha(0)$ refers to how intentional an action is and how informative it is in revealing analytic focus while $P_\alpha$ is a measure of how long the visual changes resulting from action $\alpha$ remain.
Using these parameters and the Ebbinghaus forgetting curve \cite{ebbinghaus1880urmanuskript,ebbinghaus1885gedachtnis,murre2015replication}, the authors define the per-action importance score function $t$ time steps after an action is taken as:
\begin{equation}
I_{\alpha}(t) = I_{\alpha}(0) \times e^{-\frac{t}{P_{\alpha}}}.
\end{equation}
Using the equation above, the authors propose an additive model for calculating the importance score for each concept $c$ at time $\tau$:
\begin{equation}
I^c(\tau) = \sum_{(x_{i, t}, \alpha) \in \mathcal{I}_{\tau,c}}I_{\alpha} (\tau - t),
\end{equation}
where $\mathcal{I}_{\tau, c}$ denotes the set of interactions with points involving concept $c$ up until time step $\tau$. Considering that each data point $x_i \in \mathcal{D}$ may subscribe to multiple concepts (one concept per dimension in our case), we extend this method so that the importance of a data point is the product of the importance of its individual concepts. In other words, given a data point $x_i \in \mathcal{D}$ and the current time step $\tau$,
\begin{equation}
f_r(x_i) = \prod_{c \in C(x_i)}{I^c(\tau)},
\end{equation}
where $C(x_i)$ is the list of concepts with which $x_i$ is associated.
\subsection{Hidden Markov Model as seen in \cite{ottley2019follow} \hfill \nexticon\ \biasicon\ }
\noindent
Ottley et al.\cite{ottley2019follow} propose a \textit{Hidden Markov Model} (\hmm) approach for modeling user attention during visual exploratory analysis. \textbf{Their approach assumes that the visual features of the interface will drive the user's attention (it considers all possible combinations)} and their interactions such that attention at time $t+1$ will be similar to the observation at time $t$ . They constructed a hidden Markov model, presuming the user's behavior evolves under a Markov process (that is, the behavior at a particular time only depends on their behavior at the previous time step), and interaction events are generated conditionally independently given this sequence of attention shifts. The set of visible visual features (\eg\ color, position, size, etc.) creates the \textit{mark space} and the latent states. Further, their algorithm uses particle filtering~\cite{gordon1993novel,doucet2000sequential} to sample from the model's posterior distribution and predict the user's attention for the next time step. Integral to this technique is a \textit{bias vector} that captures the relative importance of the mark space components. We use the particle filter output to predict next data interaction and the bias vector for bias detection.
\begin{figure*}[!ht]
\centering
\includegraphics[width=0.95\linewidth]{figures/datasets.png}
\caption{A summary of the characteristics of the selected user interaction datasets and the performance of the user modeling techniques with said datasets. We represent success using each technique's performance at top-100 points for next data interaction prediction.}
\label{fig:datasets}
\end{figure*}
\subsection{Competing Models as seen in \cite{monadjemi2020competing} \hfill \nexticon\ \biasicon\ }
\noindent
Monadjemi et al.\cite{monadjemi2020competing} propose a Bayesian model selection approach for detecting exploration bias and inferring next data interaction during visual exploratory analysis. This approach, called \emph{Competing Models} (\cm), enumerates a set of probabilistic models each of which represent exploration based on a subset of data attributes.
Given a dataset with $d$ attributes, they enumerate the model space ${\mathbf{M} = \{\mathcal{M}_1, \mathcal{M}_2, ..., \mathcal{M}_{2^d}\}}$ where each model represents exploration based on one subset of attributes. \textbf{Thus, it assumes and one or more data attributes (it considers all possible combinations) are sufficient to represent a users data interest.} As interactions arrive, they maintain a belief ${\Pr(\mathcal{M}_i \mid \mathcal{I})}$ which is interpreted as the viability of model $\mathcal{M}_i$ in explaining user interactions.
To rank the data points to the task at hand in light of past interactions, they use Bayesian model averaging to get:
\begin{equation}
f_r(x_i) = \Pr(x_i \mid \mathcal{I}) = \sum_{j=0}^{2^d}\Pr(x_i \mid \mathcal{I}, \mathcal{M}_j).
\end{equation}
To quantify bias towards each attribute, they use the law of total probability as follows:
\begin{equation}
f_b(a_i) = \sum_{\mathcal{M}_j \in \mathbf{M}_{a_i}} \Pr(\mathcal{M}_j \mid \mathcal{I}),
\end{equation}
where $\mathbf{M}_{a_i} \subset \mathbf{M}$ denotes the set of models involving attribute $a_i$.
\subsection{Attribute Distribution as seen in \cite{wall2017warning} \hfill \biasicon\ }
\noindent
Wall et al.\cite{wall2017warning} propose a set of metrics for quantitatively detecting various types of cognitive biases. The most relevant to our work is the metric they call \emph{attribute distribution}.
\textbf{Their approach assumes that in an unbiased exploratory session, the distribution of values explored by the user will resemble the distribution of the underlying data.} It relies on well-known statistical tests, Chi-Square\cite{mchugh2013chi} and Kolmogorov–Smirnov\cite{massey1951kolmogorov}, which are used for discrete and continuous attributes respectively. In each case, the test takes the underlying data as well as the set of points with which the user has interacted as inputs and returns a $p$-value indicating whether or not the two samples could be reasonably drawn from the same distribution. Then, they define their attribute distribution metric for a given attribute $a_i \in \mathcal{A}$ to be:
\begin{equation}
f_{b}(a_i) = 1 - p\text{-value}_{a_{i}}.
\end{equation}
\subsection{Adaptive Contextualization as seen in \cite{gotz2016adaptive} \hfill \biasicon\ }
\noindent
Gotz et al.\cite{gotz2016adaptive} propose \textit{Adaptive Contextualization}, a metric to measure selection bias in real-time. In order to detect such biases, they propose measuring the Hellinger distance between the datasets before and after filtering. The Hellinger distance is a statistical measure designed to quantify the similarity between two distributions. Thus, \textbf{this approach assumes that we can detect bias by comparing the observed distribution to the expected data distribution}. The authors propose using the Hellinger distance defined for discrete probability distributions with a pre-processing step to discretize the continuous attribute into discrete bins.
Consider a particular dimension of interest, $a_i$, represented as a set of $m$ discrete values $\{v_1, v_2, ..., v_m\}$. The distribution of these values in the underlying data and interaction data are $\{p_1, p_2, ..., p_m\}$ and $\{q_1, q_2, ..., q_m\}$ respectively. We compute the Hellinger distance as:
\begin{equation}
f_b(a_i) = H(\mathcal{D}_{a_i}, \mathcal{I}_{a_i}) = \sqrt{\frac{1}{2} \sum_{j=1}^{m}{(\sqrt{p_j} + \sqrt{q_j})^2} },
\end{equation}
where $\mathcal{D}_{a_i}$ and $\mathcal{I}_{a_i}$ denote only the $a_i$ attribute of the underlying and interaction datasets respectively. While this technique was intended to detect selection bias (i.e.\ before vs.\ after filtering), we adopt it to quantify exploration bias.
\subsection{Ensemble Approach \hfill \nexticon\ \biasicon\ }
We anticipate that no model will be able to correctly recognize relevant points in every system, task, or dataset. This phenomena was also observed by the machine learning community which led to the development of \emph{ensemble methods} \cite{rokach2010ensemble,dietterich2000ensemble}. Ensembles are a collection of classification models combined to construct a model with higher predictive power. Using this approach, we consider a final technique which combines predictions from all five techniques that predict data interactions and all four techniques that detect exploration bias reviewed in this paper. To predict next data interaction, for a given data point, $x_i \in \mathcal{D}$, we compute the ranking function, $f_r(x_i)$, using each of the modeling techniques and average the ranking values to get the ensemble prediction. Then, we proceed with ranking the data points, a process identical to the one in the preceding modeling techniques. Similarly, we average the bias values towards each data attribute from each modeling technique to get the ensemble prediction for exploration bias.
\section{User Study Interaction Logs}
\label{sec:datasets}
We used four publicly available interaction logs from user study experiments with visualization systems. These studies contained varying interfaces as well as task designs. We aimed to evaluate and compare the performance of the user modeling techniques with a variety of interaction logs, ranging from studies with very directed and specific tasks to the polar opposite, open data exploration. Refer to ~\autoref{fig:datasets} for a detailed summary of the characteristics of each dataset.
\subsection{\textbf{STL Crimes} as seen in \cite{ottley2019follow} \hfill \disticon\ \histicon\ }
The \textit{STL Crimes} dataset \textbf{involved a highly specified task thus capturing less exploration noise. We selected this dataset as it represents the best-case observational scenario for user modeling.}
Curated by Ottley et al.~\cite{ottley2019follow},
this dataset consists of 1,951 crime instances reported in the month of March 2017 with a total of eighteen attributes. Each instance of crime on the map was displayed as a dot with a position and color, indicating the location and type of the crime respectively. Eight types of crime exist in the database: homicide, theft-related, assault, arson, fraud, vandalism, weapons, and vagrancy. 30 participants were recruited from Mechanical Turk and they completed three categories of directed search tasks that were designed to encourage them to either click on: (1) \textit{Type-Based} data points with similar crime categories (e.g. Homicide, Assault, etc.), (2) \textit{Location-Based} data points that are within in the same vicinity, or (3) \textit{Mixed} data points that are of the same crime category and in the same vicinity. Each participant completed a total of six tasks, two from each category. Since the authors specified to the participants target attributes to interact with, this provides us with a notion of ground-truth bias. In their user study, Ottley et al.~\cite{ottley2019follow} captured mouse click data as participants interacted the visualization. Consistent with Ottley et al.~\cite{ottley2019follow}, we filtered the data to hold only the interactions of participants who successfully answered the task questions and created three data subsets: 28 user sessions for location-based tasks, 23 for typed-based, and 27 for mixed.
\subsection{\textbf{Vastopolis} as seen in \cite{monadjemi2020active} \hfill \texticon\ \disticon\ }
We selected the \textit{Vastopolis} dataset because of the attributes in the dataset. \textbf{It is the only user interaction dataset in this study that includes unstructured text.}
Inspired by the 2011 VAST Challenge, Monadjemi et al.'s~\cite{monadjemi2020active} study describes a major epidemic that started in the fictitious city of Vastopolis. Participant's interacted with a map of Vastopolis with 3000 geolocated tweet-like data. In addition to the social media posts, the map display major roadways, waterways, and landmarks in the city. They recruited 130 participants from Amazon Mechanical Turk and tasked each participant to search through the dataset of microblogs via an interactive map and bookmark as many posts containing illness-related information as possible. Following the same pre-processing step as the authors, 74 user interaction logs remained from the control group of the study. The only interaction type we considered in the interaction logs were intentional ``bookmarks" of the microblogs.
\subsection{\textbf{Political Committee} as seen in \cite{wall2022left}\hfill \ordinalicon\ \disticon\ \histicon\ }
The \textit{Political Committee} user study \textbf{examined cognitive and exploration bias as participants selected their picks for a hypothetical committee. Thus we selected this dataset as a candidate for examining exploration bias.}
Wall et al.~\cite{wall2022left} generated a dataset of 180 fictitious politicians, representing the composition of the Georgia General Assembly. The dataset contains three discrete and six continuous attributes that characterizes each politician. In the study conducted by Wall et al.~\cite{wall2022left}, the authors had users interact with an visualization system that supports the exploration of the fictitious political committee data. There were two versions of the visualization system: a Control version of the interface, and an Intervention version of the interface, which was modified to visualize traces of the user’s interactions with the data in real-time. To assess the effectiveness of visualizing the user's interaction traces, the authors performed an in-lab study with 24 participants. They were tasked with selecting a committee of 10 candidates to review public opinion in Georgia on the controversial bill that bans abortion after 6 weeks. In order to observe the raw, inherent biases of the users, we selected the interaction logs of those in the Control group (total of 12 participants). The authors recorded several types of interactions the user can make with the interface such as hovers, changing of axes on the scatterplot, committee selection etc. We filtered the interaction log for each user to contain only committee selection interactions with the data points.
\subsection{\textbf{Boardrooms} as seen in \cite{feng2018effects}\hfill \disticon\ \histicon\ }
The \textit{Boardrooms} study did not involve a particular task, but instructed the participants to freely interact with the visualisation. \textbf{We expect this dataset to include the highest levels of exploration noise. }
Feng et al.~\cite{feng2018effects,feng2018patterns} utilized the visualization, ``Inside America’s Boardroom," which was published by the \textit{Wall Street Journal}~\cite{wsjboardrooms} in 2016. This point-based visualization displays seven different attributes to the user: market capitalization, ratio of unrelated board members, ratio of female board members, average age of board members, average tenure of board members, and median pay of board members. The study adopted an open-ended approach and observed the user's interactions as they explored the dataset through the visualization. The interaction log contains a succession of hovers each user made on different data points. To account for unintentional hovers, we followed the same approach as Monadjemi et al.~\cite{monadjemi2020competing}, each user's session was filtered only include hovers that lasted for over one second. Additionally, we only considered sessions that consisted of more than three hovers, resulting in a total of 39 user sessions.
\section{Evaluation Measures}
In this section, we describe the performances measures we used to evaluate the modeling techniques' accuracy in predicting data interactions and detecting exploration bias with the four user study logs.
We note that 2 out of 4 interaction logs included a single interaction type. In particular, the \textit{STL Crimes} captured only mouse clicks and \textit{Boardrooms} recorded mouse hovers alone. \textit{Vastopolis} included bookmarks and hovers, but recording inconsistencies in the logging made it impossible to integrate the two interaction modalities. Thus, our evaluation focused solely on predictions based on the more intentional ``bookmark'' interaction. Finally, the \textit{Political Committee} dataset included the most diverse interface and data interactions ( \eg\ hovers, reconfiguring, committee selections). For consistency, we select a single interaction observation (committee selections), and the algorithms’ predictions are limited to their observations. Suppose, for example, the algorithm’s observations are limited to mouse clicks. Then the algorithm can only predict the next click or detect data exploration bias among the clicked data points. In the case of \textit{Political Committee}, this means that the algorithms can only observe when the user selects a committee member and must make predictions about the next committee selection or estimate selection biases solely from the limited observations.
\subsection{Evaluating Data Interaction Prediction}
We developed a standardized probabilistic output for each algorithm to compare the models to each other. Specifically, the techniques treat every data point as a potential candidate for being most relevant to the user's data interest. Thus, each method outputs the ranking probability for each data point given its observations. This output represents the technique's unique interpretation of the data points that the user will likely interact with. For example, \knn\ assumes that the data points most relevant to the user will minimize some measure of distance while \af\ examines the frequency on \textit{concepts} from the past interactions.
Given the model's probabilities, we can rank each data point based on the algorithm's belief.
We report \topk\ data retrieval for the algorithms for $\kappa \in \{1,5,10,20,50,100\}$ which we will refer to as the \textit{prediction sets}. We considered a variety of set sizes. For example, an alternative is to use percentages instead of a fixed values. However, for large datasets, retrieving a small percentage to the dataset can still be overwhelming to the user if we consider the goal of provided real-time analysis support. This standardized output enables us to evaluate and compare each technique's ability to predict the next data interaction. We record the following measures for every interaction the datasets:
\begin{itemize}[noitemsep, topsep=0pt]
\item \textbf{\success:} $\in \{0,1\}$, binary value for whether the next data interaction was included in the prediction set for all values of $\kappa$.
\item \textbf{\rank:} the ranked position of the next observed interaction such that lower is better.
\end{itemize}
\begin{figure}[b!]
\centering
\includegraphics[width=0.9\linewidth]{figures/Lima.png}
\caption{Comparing participant \textit{Lima's} interactions (foreground) to the full dataset distribution (background). Both coders reported no significant difference between the two distributions, indicating no bias detected. The result were similar for all 12 study participants from \cite{wall2022left}.}
\label{fig:exampleuser}
\end{figure}
\begin{figure*}[!ht]
\centering
\begin{subfigure}[b]{0.87\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/ncp_results.png}
\vspace*{3mm}
\end{subfigure}
%
\begin{subfigure}[b]{0.87\textwidth}
\vspace{-1em}
\centering
\includegraphics[width=\textwidth]{figures/ranking_results.png}
\end{subfigure}
\caption{We evaluated each model's performance with two metrics: (top) average \textbf{success rate} across all the tasks in each user study dataset for varying values of the prediction set size and (bottom) average \textbf{rank} of the next data interaction of the user modeling techniques.}
\label{fig:inference}
\vspace{-2mm}
\end{figure*}
\subsection{Evaluating Exploration Bias Detection}
\label{sec:bias_detection_gt}
Evaluating each technique's ability to detect exploration bias poses a unique challenge, in that \textit{ground truth} is not always readily available. When evaluating data interaction prediction, we had access to a notion of \emph{success}: we wanted to predict next data points with which the user interacts, and we were able to verify our ability to do so by observing the data interaction point. For bias selection, however, such ground truth is not as easy to access, especially for user study datasets with open-ended tasks. With the \textit{Political Committee} dataset collected by Wall et al.~\cite{wall2022left}, for example, each user can exhibit their own set of biases as they explore the data.
For this set of evaluations, we use two datasets: \textit{STL Crimes} separated into \textit{location-based}, \textit{type-based}, and \textit{mixed} subsets and \textit{Political Committee}. The \textit{STL Crimes} user study involved a directed task, asking participant to interaction with specific portions of the data, inherently resulting in biased data exploration. For example, the type-based task instructed participants to inspect all ``Arsons'' and note the time of day they occurred. We used the specific task as the ground truth. In our evaluation, an algorithm successfully detects exploration bias if its outputs indicate data attributes related to crime location, type, or a mixture model of type and location for \textit{location-based}, \textit{type-based}, and \textit{mixed} respectively.
For \textit{Political Committee}, two researchers performed qualitative coding on each study participants' committee selections, following best practices~\cite{brod2009qualitative}. We used an iterative approach. First, the coders reviewed a single participant's selections and density plots showing the attribute distributions for a given set of interaction logs and the original data distribution. Their goal was to identify evidence of data bias in the observed selections. Evidence of data bias was determined if the attribute distributions for a given set of interaction logs was different than original data distributions. Next, the coders established a codebook from this first round and independently coded all twelve participants' interaction logs. After inspecting each of the twelve study participants, both coders agreed that there was no observable bias in the any of the interaction logs. For example, \autoref{fig:exampleuser} compares study participant Lima's data selection to the individual attribute distribution. No significant difference was observed between the two distributions for each attribute. Although the coding sessions detected no bias, we included this analysis to assess the algorithms' performance in detecting bias when none exists.
\section{Algorithms to compare}
Future interactions
\begin{enumerate}
\item HMM (Ottley et al, 2019) \tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;
\item Competing models (Monadjemi et al, 2020) \tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;
\item Naive Bayes/AdaBoost (Healey and Dennis, 2012) {\color{blue}{(Jen)}}
\item $k$-NN {\color{blue}{(Shayan)}} \tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;
\end{enumerate}
\noindent
Bias
\begin{enumerate}
\item Emily Wall's algorithm for detecting bias \tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;
\item Competing models (Monadjemi et al, 2020) \tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;
\item Adaptive Contextualization (Gotz et al. 2016) {\color{blue}{(Jen)}} \tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;
\item Future Clicks: Naive Bayes/AdaBoost (Healey and Dennis, 2012) {\color{blue}{(Jen)}}
\item Analytic Focus (Gotz Group) {\color{blue}{(Shayan???)}}
\item $k$-NN {\color{blue}{(Shayan)}} \tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;
\end{enumerate}
\section{Comparison metrics}
\begin{itemize}
\item precision at top-K
\item average ranking of the next click
\item probability for the next click
\item
\item User's interest (bias)
\item User's future interaction point
\item accuracy vs accuracy over time (how much interaction is needed for the algorithm)
\item Runtime
\item Complexity
\item Can we use Bayesian Model selection to compare in realtime?
\end{itemize}
\section{Datasets to use}
We want to have datasets that will work well in addition to datasets that won't work as well.
\begin{itemize}
\item STL crimes
\end{itemize}
\subsection{Input Data format}
\begin{itemize}
\item Underlying data: a Pandas table (can be saved as CSV or something) with a valid notion of index. Each row would be a data point and each column would be an attribute.
\item User interaction data: a Pandas table (again, saved as CSV or pickle or whatever), where each row corresponds to a participant. We can have a column that contains the sequence of data indices they interacted with as a list, consistent with the indices in the underlying data.
\end{itemize}
\subsection{Output Data format}
Output format for each algorithm will be a list of objects such that $i$ denotes the number of participants and $j$ represents that number of interactions for each participant.
\[
\mathcal{O}_{i_j} = \{ p, t, b, b', k \}
\]
\noindent
where:
\begin{itemize}
\item $p$: the participant id,
\item $t$: time index/stamp?,
\item $b$: bias towards each attribute,
\item $b'$: bias towards combinations of attributes,
\item $k$: top k prediction set for next click where $k \geq 100$
\end{itemize}
\end{document}
\section{Selected User Study Datasets}
\label{sec:datasets}
\ao{I like the structure of this section. Let's do something similar for the algorithms. See \autoref{sec:methods}}
To compare the performance of the selected user modeling techniques, we chose four user study datasets. The selection of datasets we chose vary in the number of visual attributes shown to the user and the degree of openness of the task(s) the user was asked to perform. In this section, we provide an overview of each of the underlying data as well as details of its corresponding user study interaction log. See to Table~\ref{tab:user_study_data} for detailed statistics characterizing the selected user interaction log datasets.
\subsection{St. Louis Crimes}
\subsubsection{Description of the Underlying Data}
The St. Louis Crime dataset was pulled from the database of crimes maintained by the Metropolitan Police Department of St. Louis \cite{Crime}. This dataset consists of 1,951 crime instances committed in the month of March 2017 with a total of eighteen attributes.
\subsubsection{Description of the Interaction Data}
In their user study, Ottley et al.~\cite{ottley2019follow} captured mouse click data as participants interacted with a crime map visualization of St. Louis. Each instance of crime on the map was displayed as a dot with a position and color, indicating the location (denoted by \textit{longitude} and \textit{latitude}) and \textit{type} of the crime respectively. Eight types of crime exist in the database: homicide, theft-related, assault, arson, fraud, vandalism, weapons, and vagrancy. 30 participants were recruited from Mechanical Turk and they completed three categories of directed search tasks that were designed to encourage them to either click on: (1) \textit{Type-Based} data points with similar crime categories (e.g. Homicide, Assault, etc.), (2) \textit{Location-Based} data points that are within in the same vicinity, or (3) \textit{Mixed} data points that are of the same crime category and in the same vicinity. Their participants completed two tasks for each category, resulting in a total of six tasks. The tasks presented to the participants were:
\begin{enumerate}[noitemsep]
\item Type-Based: examine data points with similar crime categories
\begin{enumerate}[noitemsep]
\item Out of all the cases of Homicide, one case differs from the other cases with regard to time. What is the time of that case?
\item How many cases of arson occur during PM?
\end{enumerate}
\item Location-Based: examine data points that are within in the same vicinity
\begin{enumerate}[noitemsep]
\item There are four types Theft-Related crime in the red shaded region: Larceny, Burglary, Robbery and Motor Vehicle Theft. Count the number of cases of Robbery in the red shaded region.
\item There are two types of Assault: Aggravated and Non-Aggravated assault. Count the number of Non-Aggravated Assault in the red shaded region.
\end{enumerate}
\item Mixed: examine data points that are of the same crime category and in the same vicinity
\begin{enumerate}[noitemsep]
\item Count the number of crimes that occur during 7:00 AM - 12:30 PM in the red shaded region.
\item Count the number of crimes during AM in the red shaded region.
\end{enumerate}
\end{enumerate}
Questions (1) and (2) were type-based tasks, (3) and (4) were location-based tasks and (5) and (6) were mixed tasks. Consistent with Ottley et al.~\cite{ottley2019follow}, we filtered the data to hold only the interactions of participants who successfully answered the task questions, resulting in 28 user sessions for location-based tasks, 23 for typed-based, and 27 for mixed.
\subsection{S\&P 500 Boardrooms}
\subsubsection{Description of the Underlying Data}
The S\&P 500 Boardrooms dataset comes from the Wall Street Journal~\cite{wsjboardrooms} and it holds various information about the companies and their board of directors in the S\&P 500. The data contains one discrete and six continuous attributes.
\subsubsection{Description of the Interaction Data}
Feng et al.~\cite{feng2018patterns} utilized the visualization, ``Inside America’s Boardroom," which was published by the Wall Street Journal~\cite{wsjboardrooms} in 2016. This point-based visualization displays seven different attributes to the user: \textit{market capitalization, ratio of unrelated board members, ratio of female board members, average age of board members, average tenure of board members, and median pay of board members}. The authors took an open-ended approach to their study and observed the user's interactions as they explored the dataset through the visualization. The interaction log contains a succession of hovers each user made on different data points. To account for unintentional hovers, we followed the same approach as Monadjemi et al.~\cite{monadjemi2020competing}, each user's session was filtered only include hovers that lasted for over one second. Additionally, we only considered sessions that consisted of more than three hovers, resulting in a total of 39 user sessions.
\subsection{Vastopolis - VAST Challenge 2011}
\subsubsection{Description of the Underlying Data}
In 2011, the Visual Analytics Science and Technology (VAST) community published a fictitious epidemic dataset for their annual challenge. A truck accident over a major river contaminates the water and air with harmful chemicals in the town of Vastopolis. The water flow and wind transport these chemicals to two distinct parts of the city, causing citizens to exhibit symptoms of an illness. This dataset originally contains 1,023,077 microblogs posted on social media from locations around the town during a 21-day period. Monadjemi et al.~\cite{monadjemi2020active} created a subset of this dataset to contain data from the first two days of the epidemic with 33\% of points being related to illness. The authors then picked a random set of 3000 points from this subset to display to users in their study. Each microblog in this dataset is classified to a certain topic keyword based with a labeling heuristic. The data contains one discrete and two continuous attributes.
\subsubsection{Description of the Interaction Data}
The user study presented in Monadjemi et al.~\cite{monadjemi2020active} involved users in identifying the parts of Vastopolis impacted by the epidemic via analyzing social media posts and their posting location, denoted by \textit{longitude} and \textit{latitude}. The authors recruited 130 participants from Amazon Mechanical Turk and tasked each participant to search through the dataset of microblogs via an interactive map and bookmarking as many posts containing illness-related information as possible. Following the same pre-processing step as the authors, we ended up with a total of 74 user interaction logs from the control group of the study. The only interaction type we considered in the interaction dataset were hovers that lasted at least 500 milliseconds.
\subsection{Political Committee}
\subsubsection{Description of the Underlying Data}
Wall et al.~\cite{wall2022left} generated a dataset of 180 fictitious politicians, representing the composition of the Georgia General Assembly. The dataset contains three discrete and six continuous attributes that characterizes each politician.
\subsubsection{Description of the Interaction Data}
In the study conducted by Wall et al.~\cite{wall2022left}, the authors had users interact with an visualization system that supports the exploration of the fictitious political committee data. There were two versions of the visualization system: a Control version of the interface, and an Intervention version of the interface, which was modified to visualize traces of the user’s interactions with the data in real-time. The goal of interaction traces are to promote reflection and awareness during visual data analysis. To assess the effectiveness of visualizing the user's interaction traces, the authors performed an in-lab study with 24 participants and gave them the following open-ended task: \textit{Imagine you are engaged in political decision-making in the state of Georgia. The debate about abortion is ongoing, with variations of these bills cropping up across other states in the United States, which can potentially learn from the ongoing debate in Georgia. Select a committee of 10 candidates that you feel should review public opinion in Georgia on the controversial Georgia HB481.} For context, the Georgia HB481 bans abortion after 6 weeks (earlier than the previous state law of 20 weeks. In order to observe the raw, inherent biases of the users, we selected the interaction logs of those in the control group (total of 12 participants). The authors recorded several types of interactions the user can make with the interface, however, we filtered the interaction log for each user to contain only the interactions when the user "bookmarked" a politician to be part of their political committee.
\subsection{Hidden Markov Model}
Ottley et al.\cite{ottley2019follow} propose a Hidden Markov Model (\acro{HMM}) approach for modeling user attention during visual exploratory analysis. From a high level, their approach is based on the assumption that users interact with nearby points. They construct a probability belief over the set of data points $x_i \in \mathcal{D}$ being involved in the next interaction in light of past observations. Using the Markov property, this probabilistic belief only depends on the last interaction observed. \shayan{Key ideas missing here ares (1) the model is based on visual attributes, (2) the $\pi$ vector for biases, (3) our extension to go beyond visual channels and use every attribute in the data.}
\subsection{Attribute Distribution Bias Metric}
Wall et al.\cite{wall2017warning} propose a set of metrics for quantitatively detecting various types of cognitive biases. The most relevant to our work is the metric they call \emph{attribute distribution} bias. Their approach is based on the assumption that in an unbiased exploratory session, the distribution of values considered by the user resembles the distribution of the underlying data.
This approach is simple in that it relies on well-known statistical tests. Chi-Square and Kolmogorov–Smirnov tests are used for discrete and continuous attributes respectively. In each case, the test takes the underlying data as well as the set of points with which the user has interacted as inputs and returns a $p$-value indicating whether or not the two samples could be reasonably drawn from the same distribution. Then, they define their attribute distribution metric for a given attribute $a_i \in \mathcal{A}$ to be:
\begin{equation}
f_{b}(a_i) = 1 - p\text{-value}_{a_{i}}.
\end{equation}
\shayan{Missing ideas: the underlying assumptions for KS and ChiSquare tests, and alternatives when the assumptions are not met.}
\subsection{Competing Models}
Monadjemi et al.\cite{monadjemi2020competing} propose a Bayesian model selection approach for detecting exploration bias and inferring information relevance during visual exploratory analysis. This approach, called \emph{Competing Models}, enumerates a set of probabilistic models each of which represent exploration based on a subset of data attributes. Given a dataset with $d$ attributes, they enumerate the model space ${\mathbf{M} = \{\mathcal{M}_1, \mathcal{M}_2, ..., \mathcal{M}_{2^d}\}}$ where each model represents exploration based on one subset of attributes. As interactions arrive, they maintain a belief ${\Pr(\mathcal{M}_i \mid \mathcal{I})}$ which is interpreted as the viability of model $\mathcal{M}_i$ in explaining user interactions.
\shayan{Missing ideas: based on exact inference; not always easy to have; assumptions of the exact inference models. We omit the inner workings of the models in this discussion.}
To infer the relevance of data points to the task at hand in light of past interactions, they use Bayesian model averaging to get:
\begin{equation}
f_r(x_i) = \Pr(x_i \mid \mathcal{I}) = \sum_{j=0}^{2^d}\Pr(x_i \mid \mathcal{I}, \mathcal{M}_j).
\end{equation}
To quantify bias towards each attribute, they use the law of total probability as follows:
\begin{equation}
f_b(a_i) = \sum_{\mathcal{M}_j \in \mathbf{M}_{a_i}} \Pr(\mathcal{M}_j \mid \mathcal{I}),
\end{equation}
where $\mathbf{M}_{a_i} \subset \mathbf{M}$ denotes the set of models involving attribute $a_i$.
\subsection{Adaptive Contextualization}
Gotz et al.\cite{gotz2016adaptive} propose a metric to measure selection bias in real-time. Their work was motivated by the fact that, in some scenarios, filtering the dataset by some attributes may introduce unintentional bias towards other attributes. In order to detect such biases, they propose measuring the Hellinger distance between the datasets before and after filtering. Hellinger distance is a statistical measure designed to quantify the similarity between two distributions. The authors propose using the Hellinger distance defined for discrete probability distributions. When working with continuous attributes, they propose a pre-processing step to discretize the continuous domain into discrete bins. Consider a particular dimension of interest, $a_i$, represented as a set of $m$ discrete values $\{v_1, v_2, ..., v_m\}$. The distribution of these values in the underlying data and interaction data are $\{p_1, p_2, ..., p_m\}$ and $\{q_1, q_2, ..., q_m\}$ respectively. We compute the Hellinger distance as:
\begin{equation}
f_b(a_i) = H(\mathcal{D}_{a_i}, \mathcal{I}_{a_i}) = \sqrt{\frac{1}{2} \sum_{j=1}^{m}{(\sqrt{p_j} + \sqrt{q_j})^2} },
\end{equation}
where $\mathcal{D}_{a_i}$ and $\mathcal{I}_{a_i}$ denote only the $a_i$ attribute of the underlying and interaction datasets respectively. While this technique was intended to detect selection bias (i.e.\ before vs.\ after filtering), we adopt it to quantify exploration bias.
\subsection{Nearest Neighbors Classifier}
One simple method for learning users' data interest by observing interactions is to train a $k$-nearest neighbors ($k$-\acro{NN}) binary classifier. This method has the benefit of being non-parametric (hence flexible) and computationally fast to update in light of observations. Building a $k$-\acro{NN} classifier relies on a notion of distance between data points and computing the matrix of $k$ nearest neighbors for every data point given the defined distance formula. This model assumes that every data point $x_i \in \mathcal{D}$ has a binary label $y_i \in \{0, 1\}$. A label of $y_i=1$ means the data point $x_i$ is relevant to the task at hand, whereas a label of $y_i=0$ means $x_i$ is irrelevant. By considering the labels of the nearest neighbors, this model is able to provide us with the probability of any given point being relevant in light of past observations: $\Pr(y_i=1 \mid x_i, \mathcal{I})$. Hence, we use this posterior belief as our relevance function:
\begin{equation}
f_r(x_i) = \Pr(y_i=1 \mid x_i, \mathcal{I}).
\end{equation}
Monadjemi et al.\cite{monadjemi2020active} used a variation of this modeling technique to maintain a belief over users' latent interest for each attribute. Using a maximum likelihood estimation (\acro{MLE}) approach to computing weights, they merged predictions for each of the attributes to compute the overall relevance value. While their approach performed well in their application, the optimization step for computing the weights becomes prohibitive as the number of attributes grow.
\subsection{Boosted Naive Bayes Classifier}
Healey et al.\cite{healey2012interest} propose a naive Bayes classifier for maintaining a belief over users' interest in data points. In a similar formulation as the $k$-\acro{NN} approach, each data point $x_i \in \mathcal{D}$ has a binary label $y_i \in \{0, 1\}$. A label of $y_i = 1$ means the point $x_i$ is relevant to the task at hand and a label of $y_i = 0$ means it is not relevant. The goal of this approach is to reason about the unknown labels in light of frequencies of observations calculated via the Bayes' law:
\begin{equation}
\Pr(y_i=1 \mid x_i) \propto \Pr(x_i \mid y_i = 1)\Pr(y_i=1).
\end{equation}
Extending $x_i$ to each of its individual $d$ dimensions and assuming conditional independence among attributes, we get:
\begin{equation}
\begin{split}
&\Pr(y_i=1 \mid x_{i, 1}, x_{i, 2}, ..., x_{i, d})\\
\propto &\Pr( x_{i, 1}, x_{i, 2}, ..., x_{i, d} \mid y_i = 1)\Pr(y_i=1) \\
= & \Pr(y_i=1) \prod_{j=1}^{d} \Pr(x_{i,j} \mid y_i=1),
\end{split}
\end{equation}
where $x_{i, j}$ denotes the $j$th attribute of the $i$th data point. Therefore, we get the relevance function as:
\begin{equation}
f_r(x_i) = \Pr(y_i=1) \prod_{j=1}^{d} \Pr(x_{i,j} \mid y_i=1).
\end{equation}
This model is called \emph{naive} because of the conditional independence assumption among attributes. In more sophisticated models, the dependencies among variable are leaned via techniques called Bayesian structure learning. While Naive Bayes often performs as well as more sophisticated models in practice \cite{langley1992analysis}, its performance can further be improved by \emph{boosting}. As recommended by Healey et al.\cite{healey2012interest}, we also use AdaBoost to train a stronger classifier.
\subsection{Analytic Focus Modeling}
Zhou et al.\cite{zhou2021modeling} propose a technique for modeling analytic focus (\acro{AF}) during a session. Their technique is based on an abstract notion of \emph{concepts} defined as \emph{``meaningful data attribute[s] in the problem domain."} For example, when working with medical literature, concepts can be medical terms appearing in a document. With concepts being left open for practitioners to define in their particular domain, we let each unique value appearing in discrete attributes of the dataset to represent a concept. We perform a pre-processing step on continuous attributes to convert them into discrete bins. The goal of this technique is to track user focus on each of the concepts by observing their interactions and maintain an importance score for each concept.
As opposed to preceding techniques, \acro{AF} supports different types of interactions. Each interaction $\alpha \in \Alpha$ is defined by two values: an initial importance score $I_\alpha(0)$ and a persistence score $P_\alpha$. In a sense, $I_\alpha(0)$ refers to how intentional an action is and how informative it is in revealing analytic focus. For example, $I_{filter}(0)$ is typically higher than $I_{hover}(0)$ because filtering is often an intentional action taken to focus on a subset of the data whereas hovering can sometimes be unintentional. Moreover, $P_\alpha$ is a measure of how long the visual changes resulting from action $\alpha$ remain. Once again, we expect $P_{filter}$ to be greater than $P_{hover}$ because filtering makes changes to the visualized data until the user reverts the changes, whereas hovering may result in a tooltip opening which will close when the user stops hovering on the point. Using these parameters and the Ebbinghaus forgetting curve [cite], the authors define the per-action importance score function $t$ time steps after an action is taken as:
\begin{equation}
I_{\alpha}(t) = I_{\alpha}(0) \times e^{-\frac{t}{P_{\alpha}}}.
\end{equation}
Using the equation above, the authors propose an additive model for calculating the importance score for each concept $c$ at time $\tau$:
\begin{equation}
I^c(\tau) = \sum_{(x_{i, t}, \alpha) \in \mathcal{I}_{\tau,c}}I_{\alpha} (\tau - t),
\end{equation}
where $\mathcal{I}_{\tau, c}$ denotes the set of interactions with points involving concept $c$ up until time step $\tau$. Considering that each data point $x_i \in \mathcal{D}$ may subscribe to multiple concepts (one concept per dimension in our case), we extent this method so that the importance of a data point is the product of the importance of its individual concepts. In other words, given a data point $x_i \in \mathcal{D}$ and the current time step $\tau$,
\begin{equation}
f_r(x_i) = \prod_{c \in C(x_i)}{I^c(\tau)},
\end{equation}
where $C(x_i)$ is the list of concepts with which $x_i$ is associated.
\section{Preliminaries}
\label{sec:prelims}
\ao{I don't think this section is necessary}
In this section, we narrow the scope of user modeling to learning users' data interest during exploratory analysis. We assume there is a dot-based visualization of a dataset ${\mathcal{D} = \{x_1, x_2, ..., x_n\}}$, where each data point in $\mathcal{D}$ corresponds to a visual element (i.e. a dot) on the visualization. Each data point $x_i \in \mathcal{D}$ has $d$ dimensions (or attributes) which may be continuous, categorical, or ordinal. We denote the set of attributes as ${\mathcal{A} = \{a_1, a_2, ..., a_d\}}$. Due to limitations in visualizing high-dimensional data, only a subset of these attributes may be represented by visual channels on the visualization.
Users may take different actions while interacting with data points using the visualization. For example, they can filer, zoom, hover, or click. We denote the set of interactions supported by the system as $\Alpha = \{\alpha_1, \alpha_2, ...\}$.
As users interact the visualized data points, we maintain a set of interactions ${\mathcal{I} = \{(x_{i, t}, \alpha)\}}$, where $x_{i, t}$ denotes an interaction of type $\alpha$ with data point $x_i$ at time step $t$.
Consider the following as a running example throughout this section. A user is utilizing Google Maps to determine where to have dinner with a friend. Each restaurant is displayed on the map with visual data attributes such as the star ratings, the type of food, as well as its location (ordinal, categorical, and continuous respectively). We give concrete definitions of two aspects of user modeling related to sessions similar to our motivating example: (i) bias detection and (ii) data ranking.
\subsection{Bias Detection}
The goal of bias detection is to determine if user interactions are biased towards a certain attribute in the dataset. For example, our user may be biased towards location (searching for restaurants in a particular area), food type (looking for Mexican food), ratings (looking for popular and well reviewed places), or any combination of the attributes. Wall et al.\cite{wall2017warning} refer to this type of bias as \emph{attribute distribution} bias. Using a standard notation for every approach considered in this work, we define the \emph{bias function}, $f_{b}$, as follows:
\begin{equation}
\label{eq:f_bias}
f_{b} \colon \mathcal{A} \mapsto [0, 1],
\end{equation}
where the set of attributes, $\mathcal{A}$, is mapped to a numerical value between 0 and 1. Higher values correspond to more severe bias towards a certain attribute.
\subsection{Data Ranking}
The goal of data ranking is to learn users' latent interest and rank data points accordingly. Using the example from above, as the user interacts with restaurant points on the map, the data is ranked based on what the algorithm thinks is relevant to the user. We define the \emph{relevance function} as follows:
\begin{equation}
\label{eq:f_rank}
f_{r} \colon \mathcal{D} \mapsto [0, 1],
\end{equation}
where the set of data points, $\mathcal{D}$, is mapped to a numerical value between 0 and 1. Higher values correspond data points being more relevant to the task at hand.
Given two data points $x_i, x_j \in \mathcal{D}$, we say $x_i$ is more relevant to the task at hand, $x_i \succ x_j$, if it has a higher relevance value, $f_{r}(x_i) \textgreater f_{r}(x_j)$.
\section{Purpose and Scope}
\label{sec:scope}
This paper presents a benchmark study to compare the user modeling methods proposed in the visual analytics literature, using publicly available user interaction datasets. As demonstrated in \autoref{sec:related}, the body of work on analyzing user interactions is diverse.
However, instead of a comprehensive comparison of all existing user models, this paper focuses on techniques for predicting data interaction and detecting exploration bias. These categories of user modeling algorithms are complementary. For example, inferring the next data interaction could aid exploration bias detection, and we can use exploration bias inferences to identify data points relevant to the user's current tasks. Furthermore, we can observe from the literature that researchers have already developed algorithms that can already perform both tasks (\eg\ \cite{monadjemi2020competing,ottley2019follow}).
Additionally, we selected these two user modeling categories because of their potential application in providing real-time support for data exploration. We pose the following research questions to guide our evaluations:
\begin{itemize}[noitemsep, topsep=0pt]
\item []\textbf{RQ1:} Which algorithm is most accurate at predicting next data interactions for dataset with (i) goal-driven tasks and (ii) open-ended tasks?
\item []\textbf{RQ2:} Which algorithm is most accurate at detecting exploration bias for datasets with (i) goal-driven tasks and (ii) open-ended tasks?
\end{itemize}
\subsection{Algorithm and Dataset Selection}
To form our collection of algorithms and datasets discussed in this paper, we started with the pool of papers that were discussed in Xu et al.'s~\cite{xu2020survey} 2020 survey on the analysis of user interactions and provenance in visual analytics. The authors built a corpus of 105 manuscripts with seed papers collected from provenance-related prior work, supplemented by manually scanning all issues of major journals and all proceedings in the Visualization and HCI communities from 2009 to 2019. We accounted for the papers published around the same time and after the survey by performing a forward citation search on the manuscripts in the survey and the survey itself, resulting in a corpus of 110 manuscripts.
We narrowed the pool to the final selection of seven papers with algorithms within scope~\cite{ottley2019follow, wall2017warning, monadjemi2020competing, gotz2016adaptive, monadjemi2020active, healey2012interest, zhou2021modeling}.
To the best of our knowledge, this paper includes all real-time algorithms for detecting data interaction prediction and exploration bias that are (i) freely available, (ii) have precise code, or pseudocode descriptions, or (iii) can successfully be implemented following a reasonable amount of troubleshooting and debugging. A few edge cases could theoretically be included but are excluded from this analysis. For example, the \textit{Forecache} system proposed by Battle et al.~\cite{battle2016dynamic} captures movement data as a user navigates a large image, predicts where they will likely explore next, and uses this prediction to drive a prefetching algorithm. We can theoretically reformulate this problem and use the expected next move to infer the user's interaction, assuming the tasks is navigation-related. However, we did not include this or any other algorithm that required such reformulation.
We used a similar process for our dataset selection. We manually examined our corpus of 110 papers to identify manuscripts with user studies and publicly available datasets. Four user study datasets logged participants' interactions with data elements throughout the study session. We included all four datasets in our evaluations.
|
\section{Introduction}
Survival Analysis, or Time-to-Event Analysis, is an important subject in Statistics, where statistical and machine learning models are proposed to analyze survival data, whose outcome of interest is the time until one event happens (\textit{event time}), such as the time when one patient dies during the clinical study. In general two reasons make survival analysis necessary for analyzing survival data. Firstly, researchers are interested in the \textit{distribution} of the time for the event, not only the status of the event \textit{at one given time point}. Secondly, in most times, we cannot record the event time for one individual due to the missing follow-up time or the limited time range for one study. We can only record a \textit{censored} time (such as the last follow-up or diagnosing time) for the individual. These two reasons motivate many modern algorithms to deal with survival data, and methodologies proposed in survival analysis have been extended to other fields, such as fraud detection used in Finance.
Discrete time model is one branch of survival analysis, where the observed time is recorded as a discrete value. It is widely used when the event time is recorded based on a given time unit, such as day or month, or we want to model the grouped time, where each discrete time value can be understood as the index of a continuous time interval. Note that although our model focuses on discrete time model, we borrow ideas a lot from literature researching on continuous time model, \cite{kalbfleisch2011statistical} provides a unified form of continuous and discrete time model, we refer that book for readers who are interested in the connection.
Statisticians have developed various statistical models in survival analysis with rich theoretical properties to analyze survival data, such as Kaplan-Meier estimator \cite{kaplan1958nonparametric} and Cox model \cite{cox1972regression}. However, the assumptions for these models usually do not hold in many real datasets, especially when the size of data is large. For instance, in SUPPORT (stands for Study to Understand Prognoses Preferences Outcomes and Risks of Treatment, \cite{knaus1995support}) data, most of the covariates are biological levels that have non-linear relationship with risk score, while Cox model assumes the relationship to be linear. Additionally, \cite{gensheimer2019scalable} uses a graph and one categorical feature (whether the patient has cancer or not) to show that proportional hazard assumption does not hold in SUPPORT, either, which is another important assumption made in Cox model. Since in real cases either one or both of the assumptions (linear, proportional hazard) will be violated, more flexible models are required for survival data analysis, such as machine learning algorithms, which emphasize not on \textit{the form of the model}, but on \textit{what to learn from the data}.
In machine learning field, deep learning (also known as neural network training) is renowned for its strong performance on several supervised learning problems, such as image classification. Deep learning methods have been proposed in Survival Analysis recently in the sense that neural network possesses high flexibility to model data and do feature transformations. Although the theoretical principles of deep learning are still unclear, many deep learning experiments and works show the superiority of deep learning compared with other traditional statistical or machine learning algorithms when the size of survival data is large. Notably, several works have been proposed to apply deep learning into survival analysis (\cite{faraggi1995neural, katzman2018deepsurv, ching2018cox, lee2018deephit, tang2022soden, gensheimer2019scalable, kvamme2019time, kvamme2019continuous, tarkhan2021survival, purushotham2018benchmarking, yousefi2017predicting}) and most of the works have shown that deep learning can achieve better performances compared with other statistical survival models or machine learning methods designed for Survival Analysis and we provide an overview of most these methods in Section 2.
Although deep learning is famous for its ability to achieve an extraordinary performance when the size of data is large, this may not be the case in many real datasets. For example, due to the privacy, we may not be able to obtain a large amount of data for a clinical study, but only obtain partial information such as the parameters of some estimated statistical models or those of a pre-trained deep learning model. However, the strict assumptions required by many statistical models still do not hold for some datasets even though their size of data is short. Take SUPPORT as an example, it is obvious that the non-linear property of the covariates is independent with the size of data. In these cases, we have a small size of data, but we still require a flexible and satisfactory model.
In this paper, we provide two approaches for training neural network on a small size of data. One approach is still training a model on a small size of data, but exploits information from other sources, such as an estimated model (\textit{prior model}). To integrate other information, we propose a model that combines the prior information into our loss function for deep learning. Generally, we use Kullback-Leibler (KL) divergence \cite{kullback1951information} to calculate the similarity between \textit{local model} (model that is trained with the accessible small dataset and the information from prior data) and \textit{prior model} (model that has been estimated before, such as published statistical model or trained model from other data sources) and select a hyperparameter to determine the weight for local information and prior information. High weight for local data means the model tends not to use much prior information, while low weight means the model tends to believe that prior information is good enough for the model. We also extend this method to competing risk problem with a slight (but useful) modification.
As for the other approach, we follow \cite{gensheimer2019scalable}, but specially design the neural network structure, with two structures that used for modeling proportional and non-proportional data, and also commonly-seen three link functions that used in discrete time model \cite{kalbfleisch2011statistical} (altogether six options). For proportional version, the new design is useful in data suitable for using these 3 link functions, but also reduces the number of nodes in the neural network, which will relieve the overparameterization problem \cite{arora2018optimization}, meaning the number of parameters is too large compared with the size of data. For non-proportional version, we do not restrict the activation function used in the output layer to be only sigmoid function (simulating logit link), but also the log and log-log link functions.
Based on our experiences in dealing with survival data with deep learning, we also provide with a comprehensive software combining all the contributions proposed in this paper and other necessary components for deep learning in survival analysis with the aid of prior model. This software is written by \texttt{Python} and deep learning process is completed by \texttt{Pytorch}, a well-designed package for training neural network. We design this software to not only show the novelty and comprehensiveness of our work, but also make it super convenient for users who are interested in this topic but would like to apply these methods into their own problems, such as combining a Cox model with local model a grouped relative risk model to model a clinical dataset with time values grouped by a series of time intervals \cite{wang2021kullback}.
As a recap, we summarize our work as follows:
\begin{itemize}
\item We provide two approaches to alleviate small data problems in deep learning. One is applying KL divergence to integrate data or model from other sources. The other one is specially design six neural network structures to adapt to datasets with various statistical properties.
\item We extend our likelihood function and KL divergence to competing risk setting, which provides a novel approach solving competing risk problems.
\item We provide a software which is comprehensive and user-friendly.
\end{itemize}
The organization of this paper is as follows. Section 2 recaps related works and their connections with our paper. Section 3 provides the details of discrete time model and the loss function used in our deep learning model, we have also provided details on how to parameterize the loss function by modeling hazard function with deep learning and how to extend it into competing risk problems. Section 4 provides some simulation studies, showing the proposed two approaches in our paper are useful when data has desired properties. Section 5 shows the results for real data analysis. We show the trend of our results compared with other models when we increase the size of local data and also run our model with hyperparameter tuning on a complicated EHR dataset. Section 6 concludes our work and proposes some future extensions for this paper.
\section{Related Work}
There has been a large body of models dealing with survival data. Cox proportional hazard model \cite{cox1972regression} may be the most commonly used model in current survival analysis problems and also is a starting point of the future works. The formula of Cox model can be written as
\begin{equation}
\lambda(t; \mathbf Z) = \lambda_0(t) \exp(\mathbf Z^\top \beta)
\end{equation}
where $\mathbf Z = [Z_1, \ldots, Z_p]^\top$ is the derived covariate vector obtained from the basic covariates $\mathbf x$\footnote{$\mathbf Z$ can be understood as a feature transformation of $\mathbf x$, which means the covariates used in the model can be raw features ($\mathbf x$) or processed features (such as adding on interaction terms or higher-order terms for $\mathbf x$).}. $\beta = (\beta_1, \ldots, \beta_p)^\top$ is a vector of unknown parameters requiring estimation. It has two assumptions, where the first one is linear assumption, meaning the log risk of the hazard function is a linear function of derived features ($\mathbf Z^\top \beta$), while the second one is proportional hazard assumption, which means the ratio of hazard function is constant over time ($\frac{\lambda(t_0; \mathbf x)}{\lambda_0(t_0)} = \exp(\mathbf Z^\top \beta)$ is constant for any $t_0$). Since either one or both of the assumptions will be violated in real cases, many alternative models are proposed, such as \cite{aalen1980interaction}, \cite{gray1994spline} and \cite{yu2011learning}, but still suffer from restrictions brought by improper assumptions.
As mentioned in Section 1, machine learning algorithms are favored with researchers due to free of strict statistical assumptions. Random Survival Forest (RSF)\cite{ishwaran2008random} uses only observed time and status to create a forest which ensures each node of the tree to be populated by individuals with similar survival. It achieves improvement in several survival datasets and is a benchmark for survival analysis in machine learning algorithms, but it does not consider the information of covariates for each individual, which may lead to inaccurate results.
Deep learning is shown to be empirically promising in many previous works compared with other machine learning algorithms. To the best of our knowledge, \cite{faraggi1995neural} is the first paper proposing neural network to tackle with survival analysis problems and Deepsurv \cite{katzman2018deepsurv} is the first paper that shows empirically that deep learning can provide improvement in results for survival analysis. For Deepsurv, it relieves the log risk term of the Cox proportional hazard model, from $\mathbf Z^T \beta$ to a general function $g(\theta; \mathbf Z)$ and this function is modeled by a neural network, with $\theta$ the set of unknown parameters required estimation. However, Deepsurv still keeps the proportional hazard assumption even if it provides opportunity to model non-linear relation between covariates and log risk. Additionally, it chooses partial likelihood \cite{cox1975partial} as the loss function, which prevents the usage of stochastic gradient descent (SGD) or mini-batch gradient descent due to the existence of at-risk set. Although \cite{kvamme2019time} proposes a case-control sampling method to approximate partial likelihood empirically by restricting the size of at-risk set into a small number (and even $1$ in that paper), there is still no theoretical guarantee that restricting the size of at-risk set is reasonable.
Nnet-survival \cite{gensheimer2019scalable}, LogisticHazard \footnote{we refer LogisticHazard as a name of \cite{kvamme2019continuous}, but LogisticHazard is actually one (but the most important) of the models proposed in that paper.} \cite{kvamme2019continuous} and Deephit \cite{lee2018deephit} are three main papers that regard loss function the discrete likelihood function. Nnet-survival models the hazard value at different time points for one individual using neural network, Deephit models the probability mass functions (PMF), and LogisticHazard experiments both ways. Nnet-survival provides two versions of model structures, with proportional version same as using log-log link function in the last hidden layer. It also exploits the fact well known in survival analysis that discrete time survival data can be expressed as a Bernoulli log-likelihood of the hazard values (\textit{data expansion}). Nevertheless it neither provides justification on the design, nor gives thorough consideration on other extensions, which are complemented in our paper.
Deephit focuses on modeling PMF with competing risk setting, and it adds a penalization term, making the model focuses more on the discrimination loss and provides advantages to have satisfactory result when the metric is chosen to be Concordance-index. For competing risk setting, our paper instead models cause-specific hazard function (the hazard function for a given time point and also a given cause leading to the event) instead of PMF, and we are the first group using neural network to model hazard function for competing risk problems.
LogisticHazard provides a numerical stable implementation of flexible version model in Nnet-survival, it also provides comparison experiments on modeling PMF and hazard function and two ways of interpolation, making it possible to provide with a smooth survival function for discrete time model. In our paper, we take these two ways of interpolation as two tuning options and we can offer fair and comprehensive experiment result on a data with continuous time value, unifying both continuous and discrete time model. Moreover, our software follows \texttt{pycox}, an existing Python software for deep learning in survival analysis and LogisticHazard, we also show in our paper that due to the similarity of old and new loss functions, we can smoothly transfer all the techniques that used in \texttt{pycox} into our software, but not coding from scratch.
\section{Methodology}
This section is divided into 2 parts, in the first part we provide with a comprehensive introduction of discrete time model. In the second part we show the details to parameterize our model with a neural network.
\subsection{Discrete Time Model: An Introduction}
\subsubsection{Probability Definitions and Data Structure}
To capture the random behavior of the event and censored time, we define the probability density functions of them and further use likelihood function to estimate the parameters based on the given data. Define $T^*$ and $C$ to be event time and censored time (we will define $T$ as the observed time for each individual, so here we use $T^*$ instead), and define $\tau_1 < \tau_2 < \ldots < \tau_K$ for $K$ values that $T^*$ and $C$ take, then for event time variable $T^*$, we define
\[
f(\tau_k |\mathbf Z ) = P(T^* = \tau_k \mid \mathbf Z), k = 1, 2, \ldots, K
\]
as the probability of $T^* = \tau_k$ given the derived covariates $\mathbf Z$. Define $F(t| \mathbf Z) = P(T^* \le t \mid \mathbf Z)$ as the cumulative density function of $T^*$. $S(t| \mathbf Z) = P(T > t \mid \mathbf Z) = \sum_{j | \tau_j > t} f(\tau_j) = 1 - F(t| \mathbf Z)$ as the \textit{survival function}. We also define a \textit{hazard function}
\begin{equation}
\lambda(\tau_k| \mathbf Z) = P(T^* = \tau_k \mid T^* \ge \tau_k, \mathbf Z), k = 1, \ldots, K
\end{equation}
which describes the instant death rate at time $t$ given the individual still survives at time $t$.
One interesting property in discrete time model is that we can uniformly express the survival function and density function by hazard function. Note that
\begin{equation*}
\lambda(\tau_k| \mathbf Z) = P(T^* = \tau_k \mid T^* \ge \tau_k, \mathbf Z) = \frac{P(T^* = \tau_k \mid \mathbf Z)}{P(T^* \ge \tau_k \mid \mathbf Z)} = \frac{f(\tau_k|\mathbf Z)}{S(\tau_k^{-}|\mathbf Z)} = \frac{S(\tau_{k - 1}|\mathbf Z) - S(\tau_k|\mathbf Z)}{S(\tau_{k - 1}|\mathbf Z)}
\end{equation*}
We can get $S(\tau_k|\mathbf Z) = [1 - \lambda(\tau_k|\mathbf Z)] S(\tau_{k - 1}|\mathbf Z)$ and $f(\tau_k|\mathbf Z) = \lambda(\tau_k|\mathbf Z) S(\tau_{k - 1}|\mathbf Z)$. So the final expression is
\begin{equation}
S(\tau_k|\mathbf Z) = \prod_{i = 1}^k[1 - \lambda(\tau_i|\mathbf Z)]
\label{hazard S}
\end{equation}
and
\begin{equation}
f(\tau_k|\mathbf Z) = \prod_{i = 1}^{k - 1}[1 - \lambda(\tau_i|\mathbf Z)]\lambda(\tau_k|\mathbf Z)
\label{hazard f}
\end{equation}
For censoring time $C$, we can similarly define the probabilistic functions as
\[
f_C(\tau_k| \mathbf Z) = P(C = \tau_k \mid \mathbf Z)
\]
\[
F_C(t | \mathbf Z) = P(C \le t \mid \mathbf Z)
\]
\[
S_C(t | \mathbf Z) = P(C > t \mid \mathbf Z) = \sum_{j | \tau_j > t} f_C(\tau_j) = 1 - F_C(t \mid \mathbf Z)
\]
And finally we define the observed time and status as
\begin{equation}
T = \min(T^*, C), \Delta = I(T^* \le C)
\end{equation}
where $T$ is the random value for observed time, which is the minimum of event and censored time. $\Delta$ is the status indicator, where $\Delta = 1$ means the event time for individual is observed and known, $\Delta = 0$ means it is missing and censored time is used as the alternative. $I(\cdot)$ is the indicator function. Note that we also require $T^*$ and $C$ and independent conditional on the derived covariates $\mathbf Z$, which is a common assumption in survival analysis.
Given the necessary probabilistic definitions, we can understand the data as the samples drawn from the population described by the distributions for each random variable. Suppose we have gotten the data $\{\mathbf{Z}_i, t_i, \delta_i\}_{i = 1}^n$ with $n$ the number of observations. Then $\mathbf{Z}_i$ is the derived covariates vector for individual $i$, $t_i \in \{\tau_1, \tau_2, \ldots, \tau_K\}$ is the observed discrete time and $\delta_i$ is the status indicator. $\delta_i = 1$ means the event time for this patient is known, then $t_i$ is the event time. $\delta_i = 0$ means the event time for this patient is unknown, then $t_i$ is the censored time.
\subsubsection{Likelihood Function}
Note that given the observed time $T = \tau_k$ with $k \in \{1, \ldots, K\}$ and the status $\Delta = \delta$ with $\delta \in \{0, 1\}$, we have
\[
\begin{aligned}
P(T = \tau_k, \Delta = \delta) & = P(T^* = \tau_k, C \ge \tau_k)^\delta P(T^* > \tau_k, C = \tau_k)^{1- \delta} \\
& = P(T^* = \tau_k)^\delta P(C \ge \tau_k)^\delta P(T^* > \tau_k)^{1 - \delta}P(C = \tau_k)^{1- \delta} \\
& = f(\tau_k)^\delta S(\tau_k)^{1 - \delta} P(C \ge \tau_k)^\delta P(C = \tau_k)^{1- \delta}
\end{aligned}
\]
(We omit the covariates $\mathbf Z$ for simplicity) We want to model the event time distribution, so we delete the term $P(C \ge \tau_k)^\delta P(C = \tau_k)^{1- \delta}$ (due to random censoring) and we can write out the likelihood contribution for one individual as $f(\tau_k)^\delta S(\tau_k)^{1 - \delta}$. More rigorously, suppose $\{\mathbf Z_i, t_i, \delta_i\}$ is the information for the $i$-th observation, then we can get the likelihood contribution
\begin{equation*}
L_i = f(t_i| \mathbf Z_i)^{\delta_i }S(t_i| \mathbf Z_i)^{1 - \delta_i }
\end{equation*}
And the final likelihood function for the observed data can be written as
\begin{equation}
L = \prod_{i = 1}^n L_i = \prod_{i = 1}^n f(t_i| \mathbf Z_i)^{\delta_i }S(t_i| \mathbf Z_i)^{1 - \delta_i}
\label{l full}
\end{equation}
and the log-likelihood is
\begin{equation}
l = \log L = \sum_{i = 1}^n \delta_i \log f(t_i | \mathbf Z_i) + (1 - \delta_i) \log S(t_i | \mathbf Z_i)
\label{log l original}
\end{equation}
Following survival literatures, we note (\ref{l full}) as \textit{discrete time log-likelihood}.
Moreover, combining (\ref{hazard f}) and (\ref{hazard S}) into (\ref{log l original}), define $T_i = \tau_j$, we can get
\begin{equation}
\begin{aligned}
l & = \sum_{i = 1}^n \delta_i (\sum_{k = 1}^{j - 1} \log (1 - \lambda(\tau_k|\mathbf Z_i)) + \log \lambda(\tau_j | \mathbf Z_i)) + (1 - \delta_i) \sum_{k = 1}^{j}\log (1 - \lambda(\tau_k | \mathbf Z_i))
\\ & = \sum_{i = 1}^n \left\{ [\sum_{k = 1}^{j} \log (1 - \lambda(\tau_k|\mathbf Z_i))] + \delta_i \log \frac{\lambda (\tau_j | \mathbf Z_i)}{1 - \lambda (\tau_j | \mathbf Z_i)} \right\}
\\ & = \sum_{i = 1}^n \sum_{k = 1}^K I(T_i \ge \tau_k)\left[ \log(1 - \lambda(\tau_k | \mathbf Z_i)) + I(T_i = \tau_j, \Delta_i = 1) \log \frac{\lambda (\tau_j | \mathbf Z_i)}{1 - \lambda (\tau_j | \mathbf Z_i)}\right]
\end{aligned}
\end{equation}
where $I(T_i = \tau_j, \Delta_i = 1) = 1$ if $T_i = \tau_j$ and $\Delta_i = 1$ both hold. In other words, this term equals to zero at any time point as long as this individual is censored (only has censored time but not event time).
At last, define $Y_i(k) = I(T_i \ge \tau_k)$ and $\delta_i(j) = I(T_i = \tau_j, \Delta_i = 1)$, we have the final log-likelihood function
\begin{equation}
l = \sum_{i = 1}^n \sum_{k = 1}^K Y_i(k)\left[ \log(1 - \lambda(\tau_k | \mathbf Z_i)) + \delta_i(j) \log \frac{\lambda (\tau_j | \mathbf Z_i)}{1 - \lambda (\tau_j | \mathbf Z_i)}\right]
\label{log l hazard}
\end{equation}
\subsubsection{Likelihood Function with Prior}
The likelihood function introduced above paves the way for exploiting deep learning methods to estimate unknown parameters. However, they suffer the shortage of data in many cases, which motivates us to use prior information, the information from external sources, to improve the performance of the model. In this section we introduce the likelihood function with prior that firstly proposed by \cite{wang2021kullback}, which is also used as the main loss function in our deep learning model. For the second way of relieving small data problem (six special designs of neural networks), see Section
Kullback-Leibler (KL) divergence is one common metric used for measuring the similarity between two probabilistic distributions. \cite{cox1972regression} mentioned firstly the connection between linear models and life tables and \cite{jiang2016variable} applies the KL divergence in the Generalized Linear Model. Note that hazard function is a conditional probability in discrete time survival model, we can exploit it to compute the similarity between hazard functions for local and prior models for a given individual at a given time point. After that we can summarize over time points and observations to get an overall KL divergence. Note that we only show required formulas for support of our deep learning method and put additional explanations in the appendix, for people who are interested in more theoretical details, we refer \cite{wang2021kullback} as the reference. Also, we will frequently exchange word \textit{local} with \textit{internal} and \textit{prior} with \textit{external} in this paper, we mention it here for unambiguity.
Let $\boldsymbol {\tilde \lambda} (\cdot) \in \mathbb R^{n \times K}$ be the hazard function for external model and $\boldsymbol \lambda(\cdot) \in \mathbb R^{n \times K}$ be that for internal model. They define
\[
d_{k}\left( \boldsymbol {\tilde \lambda}, \boldsymbol \lambda ; \mathbf{Z}_{i}\right)=\tilde \lambda\left(\tau_k | \mathbf{Z}_{i}\right) \log \left\{\frac{\tilde \lambda\left(\tau_k | \mathbf{Z}_{i}\right)}{\lambda\left(\tau_k | \mathbf{Z}_{i}\right)}\right\}+\left\{1-\tilde \lambda\left(\tau_k | \mathbf{Z}_{i}\right)\right\} \log \left\{\frac{1-\tilde \lambda\left(\tau_k | \mathbf{Z}_{i}\right)}{1-\lambda\left(\tau_k | \mathbf{Z}_{i}\right)}\right\}
\]
to be the KL divergence between two hazard functions for individual $i$ at time $\tau_k$, and they define the overall KL divergence to be
\[
D_{KL}(\boldsymbol {\tilde \lambda}, \boldsymbol \lambda) = \sum_{i = 1}^n \sum_{k = 1}^K Y_i(k) d_k(\boldsymbol {\tilde \lambda}, \boldsymbol \lambda; \mathbf Z_i)
\]
By simple calculations, we can obtain
\[
D_{K L}\left(\boldsymbol {\tilde \lambda}, \boldsymbol \lambda \right)=-\sum_{i=1}^{n} \sum_{k=1}^{K} Y_{i}\left(k \right)\left[\tilde{\lambda}\left(\tau_k | \mathbf Z_{i}\right) \log \left\{\frac{\lambda(\tau_k | \mathbf Z_i)}{1-\lambda(\tau_k | \mathbf Z_i)}\right\}+\log \left\{1-\lambda(\tau_k | \mathbf Z_i)\right\}\right] + C
\]
To be more clear, $ \lambda(\tau_k | \mathbf Z_i)$ is the estimated hazard function with internal dataset based on external prior model at time $\tau_k$ for individual $i$. Here $C$ is the term not related to $\lambda(\tau_k | \mathbf Z_i), \forall i, k$, we do not need to write out this part since only $\boldsymbol\lambda(\cdot)$ is unknown and need to be estimated by training a neural network.
To balance the trade-off between internal and external model, we can use a hyperparameter $\eta \ge 0$ to balance the weights. More specifically, we define a penalized log-likelihood as
\begin{equation}
l_{\eta} = \frac{l - \eta D_{KL} (\boldsymbol {\tilde \lambda}, \boldsymbol \lambda)}{1 + \eta}
\label{hyper 2}
\end{equation}
And our goal is to maximize it. Larger $\eta$ means more weights on minimizing $D_{KL}(\boldsymbol {\tilde \lambda}, \boldsymbol \lambda)$, and lower value means less difference between $\boldsymbol {\tilde \lambda}$ and $\boldsymbol \lambda$. In that case model will tend to believe more on prior information and similarly, smaller $\eta$ means model will put more weights on the discrete time log-likelihood.
Note that (\ref{hyper 2}) can be rewritten as
\begin{equation}
l_\eta =\sum_{i=1}^{n} \sum_{k=1}^{K} Y_{i}\left(k \right)\left[ \frac{\delta_i(j) + \eta \tilde \lambda(\tau_k | \mathbf Z_i)}{1 + \eta} \log \left\{\frac{\lambda(\tau_k | \mathbf Z_i)}{1-\lambda(\tau_k | \mathbf Z_i)}\right\}+\log \left\{1-\lambda(\tau_k | \mathbf Z_i)\right\}\right],
\label{log l prior hazard}
\end{equation}
which is the final loss function used for training our deep learning model. Note that $\lambda(\tau_k|\mathbf Z_i)$, the hazard value at time $k$ for the $i$-th individual, is the parameter that required estimation. So there are altogether $nK$ unknown values in this function.
In our model, the negative likelihood function ($-l$) is taken as the loss function and our goal is to optimize the loss function with deep learning. More details will be introduced later.
\subsubsection{Computational Advantage in Implementation}
Although our model training procedure described above seems hard to implement in practice due to the requirement of using chain rule to compute the gradient, we can apply one computational advantage, which makes it much easier for us to implement this algorithm. In details, we can use the same form of log-likelihood function to unify formula (\ref{log l hazard}) and (\ref{log l prior hazard}). Define
\[
l_{all}(\boldsymbol{\theta}, \mathbf p) = \sum_{i=1}^{n} \sum_{k=1}^{K} Y_{i}\left({k}\right)\left[ \left\{ p(\tau_k| \mathbf Z_i) \log \left\{\lambda(\tau_k| \mathbf Z_i)\right\}\right\}+ (1 - p(\tau_k| \mathbf Z_i) ) \log \left\{1-\lambda(\tau_k| \mathbf Z_i)\right\}\right]
\]
be the form of the log-likelihood functions introduced before with $\boldsymbol \theta$ all unknown parameters in the model and $\mathbf p = \{p(\tau_k|\mathbf Z_i)\}_{i, k} \in \mathbb R^{n \times K}$. It is easy to see that $ l(\boldsymbol{\theta}) = l_{all}(\boldsymbol{\theta}, \boldsymbol \delta)$ and $ l_{\eta} (\boldsymbol{\theta}) = l_{all} (\boldsymbol{\theta}, \frac{\boldsymbol \delta + \eta \boldsymbol {\tilde\lambda}}{1 + \eta})$, where $\boldsymbol \delta = \{\delta_i(k)\}_{i, k} \in \mathbb R^{n \times K}$. Since $\mathbf p$ has no unknown values in both loss functions, for each batch-training step, we can pre-compute $\mathbf p$ and pass the results into the loss function. Due to this finding, we only need to add one new class for our loss function and pass one additional input parameter matrix ($p(k| \mathbf x_i), \forall i, k$) into the loss function method, holding other modules unchanged. In this way all computational advantages mentioned in LogisticHazard \cite{kvamme2019continuous} (\texttt{pycox}) can be smoothly transplanted into our model instead of coding from scratch, with a minor additional memory occupation.
\subsubsection{Discussion: What is Prior Information?}
\subsection{Deep Learning: What and How?}
\subsubsection{What is Deep Learning?}
Although the idea for combining prior information in loss function comes from \cite{wang2021kullback}, it still assumes a statistical assumption for the hazard function, which is $\lambda(t_k; \mathbf Z_i) = g(\gamma_k + \mathbf Z_i^\top \boldsymbol \beta)$ with $g(u) = h^{-1}(u)$ and $h(u)$ being our interest, a monotonically increasing and twice differentiable link function. The linear form of $\beta, \gamma$ and $\mathbf Z$ can be too restrictive in many cases, and deep learning is a natural way of modeling non-linear relationship between the covariates and the outcome of interest.
Deep learning is also named as neural network training, which is renowned for its ability to approximate a large amount of functions. Mathematically, a neural network is to find a functional approximation between input and output. The inner architecture of a neural network (also known as \textit{computational graph}) can simulate basic mathematical operations such as add, multiply, nesting of functions and non-linear transformations, leading to really complex functions by combining these elements. For example:
\begin{itemize}
\item A one-layer neural network with an identity activation function can be used to simulate a linear model $y = \beta_0 + \beta_1 x_1 + \ldots + \beta_p x_p$.
\item A one-layer neural network with a sigmoid activation function can be used to simulate a GLM with logic link: $\operatorname{logit}(y) = \beta_0 + \beta_1 x_1 + \ldots + \beta_p x_p$.
\item A three-layer neural network is equivalent to defining $\lambda(\cdot; \boldsymbol \phi) = f_3(f_2(f_1(\cdot)))$ with $f_1, f_2, f_3$ three arbitrary and learnable functions \cite{lecun2015deep}.
\end{itemize} The first two examples show that linear models and GLM with logic link function can be covered by a one-layer fully-connected feedforward neural network with or without sigmoid activation function added into the final layer, respectively. For the structure of neural network, the first layer takes the input value, and the output of each layer will be used as the input of the next layer. Mathematically, suppose that we have a $l$-layer fully-connected feed-forward neural network with $\{h_1, h_2, \ldots, h_{l - 1}\}$ pre-defined numbers of nodes for each hidden layer (The number of nodes for input layer and output layer are fixed and they depend on the data and the value of interest, we define them as $h_0$ and $h_l$, respectively, for unification) and $\{s^{(1)}, s^{(2)}, \ldots, s^{(l)}\}$ activation functions for each layer, then the formula for the $i$-th individual can be written as
\begin{equation}
\begin{aligned}
& \mathbf Z^{(1)}_i = f^{(1)}(\mathbf Z_i) = s^{(1)}(W^{(1)} \mathbf Z_i + b^{(1)}) \\
& \mathbf Z^{(2)}_i = f^{(2)}(\mathbf Z^{(1)}_i) = s^{(2)}(W^{(2)} \mathbf Z^{(1)}_i + b^{(2)}) \\
& \mathbf Z^{(3)}_i = f^{(3)}(\mathbf Z^{(2)}_i) = s^{(3)}(W^{(3)} \mathbf Z^{(2)}_i + b^{(3)}) \\
& \cdots \\
& \mathbf Z_i^{(l)} = f^{(l)}(\mathbf Z^{(l - 1)}_i) = s^{(l)}(W^{(l)} \mathbf Z^{(l - 1)}_i + b^{(l)}) \\
& l_i = r(\mathbf Z_i^{(l)})
\end{aligned}
\label{neural network}
\end{equation}
where $\mathbf Z_i^{(m)} \in \mathbb R^{h_m}, m = 0, 1, \ldots, l$ is the input of the $m$-th layer with $h_m$ defined as the number of dimensions for the input of the $m$-th layer. We also assume $\mathbf Z^{(0)}_i = \mathbf Z_i$ and $h_0 = p$ the dimension of $\mathbf Z_i$ without loss of generality. $f^{(m)}: \mathbb R^{h_{m - 1}} \to \mathbb R^{h_{m}}$ is the transformation function for the $m$-th layer. $W^{(m)} \mathbf Z^{(m)} + b^{(m)}$ is the linear combination term with $W^{(m)} \in \mathbb R^{h_{m} \times h_{m - 1}}, b^{(m)} \in \mathbb R^{h_{m}}$ as the unknown parameters require learning, $s^{(m)}$ is the activation function such as identity, sigmoid or ReLU function operating element-wise (the function will be operated on each element of the input vector and give an output vector with the same size, same for below). $r(\mathbf Z_i^{(l)})$ is a function of $\mathbf Z_i^{(l)}$ and $l_i$ is the final value and usually the loss function. For this model, $\{W^{(1)}, b^{(1)}, W^{(2)}, b^{(2)}, \ldots, W^{(l)}, b^{(l)}\}$ are unknown parameters requiring estimation and there are altogether $\sum_{k = 1}^l h_k(h_{k - 1} + 1)$ unknown values, a linear function with respect to $h_i, i = 0, 1, \ldots, l$.
In our paper, we model the hazard function $\boldsymbol \lambda_i, i = 1, 2, \ldots, n$ by a neural network, which means for the $i$-th observation, we will take its derived covariates $\mathbf Z_i$ as input, and take the output vector of the neural network as the estimated hazard value. Mathematically, we want to use neural network to model a function $g(\boldsymbol \theta): \mathbb R^p \to \mathbb R^K$ where $p$ is the dimension of $\mathbf Z_i$ and $K$ means the number of time intervals, $\boldsymbol \theta$ is the set of unknown parameters $\{W^{(1)}, b^{(1)}, W^{(2)}, b^{(2)}, \ldots\}$ require learning. If we want to use fully-connected neural network, then the formula can be written as (\ref{neural network}) but with the last two equations changing as
\[
\begin{aligned}
& \boldsymbol { \lambda_i} = f^{(l)}(\mathbf Z^{(l - 1)}_i) = s^{(l)}(W^{(l)} \mathbf Z^{(l - 1)}_i + b^{(l)}) \\
& l_i = r(\boldsymbol { \lambda_i})
\end{aligned}
\]
and it has a one-to-one correlation with the series of pre-defined hyperparameters, such as number of layers, number of hidden units for each layer, activation function and so on.
\paragraph{Remark} As a side note, fully-connected feedforward neural network is flexible enough to model structured data such as clinical data. Nevertheless, it is inconvenient to model unstructured data (image, text) using this structure. In that case, the linear combination term $W^{(l)} \mathbf Z^{(l)}_i + b^{(l)}$ can also be changed to adapt to different data structures. For example, in Convolutional Neural Network (CNN), the connection between input and weight is not fully-connected, but focusing on a 'local patch' of the data. This is mainly used in image data, where the 'local patch' structure is extremely significant for data analysis. Although the combination of $W, \mathbf Z_i$ and $b$ will be slightly different for other structures such as CNN, fully-connected feed-forward neural network is flexible enough to tackle with the survival data. Moreover, as we will also see later, this has little effect on the training process of neural network, so we will assume our model of interested to be fully-connected feed-forward neural network except for extra mentioning.
\subsubsection{How to Use Deep Learning}
The way for training a neural network is the same as machine learning, which lies in the field of optimization, minimizing/maximizing a loss function computed by data by iteratively computing the value and gradient (for some algorithms, Hessian is also required) of the loss function and then updating the unknown parameters by optimization algorithms such as gradient descent or Newton method until convergence. In Deep Learning, computing the value and gradient can be done by forward and backward propagation efficiently and exactly \cite{lecun2015deep}, especially at the time when Graphical Processing Unit (GPU) can be used for dramatically accelerating of the deep learning model by parallel computing. We only introduce the main ideas of forward and backward propagation in the main part, but leave some technical details into Appendix.
For forward propagation, it means passing the input vector of each individual ($\mathbf Z_i$) to get the values of each layer ($\{\mathbf Z_i^{(1)}, \mathbf Z_i^{(2)}, \ldots, \mathbf Z_i^{(l)}\}$) and the value of loss function $l_i$ following (\ref{neural network}) given the values of $\{W^{(1)}, b^{(1)}, \ldots, W^{(l)}, b^{(l)}\}$. In this step we can get all necessary values for each layer. For backward propagation, we use the values computed by forward propagation to obtain the gradients with respect to unknown parameters $\{W^{(1)}, b^{(1)}, \ldots, W^{(l)}, b^{(l)}\}$. Note that for forward propagation, we will firstly use $W^{(1)}, b^{(1)}$ for calculation and then use $W^{(2)}, b^{(2)}$ until $W^{(l)}, b^{(l)}$. But for backward propagation, the gradient with respect to $W^{(l)}, b^{(l)}$ will be firstly obtained, and then $W^{(l - 1)}, b^{(l - 1)}$ until $W^{(1)}, b^{(1)}$. On a computational graph, the backward propagation behaves like computing from the output layer to the input layer, just the opposite direction of the forward propagation. Therefore for the calculation of gradients in a neural network, it is also called "backward" propagation.
After obtaining the function values and gradients, neural network can obtain the next update for each unknown parameter by gradient descent based optimization algorithms. In our model, we would like to use Mini-batch method as the optimization algorithm, which means we only need to compute the function values (obtained by forward propagation) and the gradients of each parameter (obtained by backward propagation) for a small batch of training set. Mathematically, suppose that $\alpha$ is the parameter we would like to update, $b$ is the batch size, $\{l_1, l_2, \ldots, l_b\}$ are the loss functions obtained by $b$ individuals in that batch \footnote{To be more clear, we give another set of indices for the observations of that batch sampled from the original dataset, which means they do not necessarily need to be the first $b$ samples.}. Then for each step, we can write the update formula as
\[
\alpha' = \text{Optim}(\alpha, \frac 1 b \sum_{k = 1}^b \frac{\partial l_k}{\partial \alpha})
\]
and usually $b$ is the hyperparameter requires tuning. Optim can be SGD \cite{ruder2016overview}, Adam, AdamW \cite{kingma2014adam} and RMSProp \cite{tieleman2012lecture}. In implementation, we only provide the last 3 advanced optimizers for options in hyperparameter tuning since SGD will provide much inferior result compared with the other three optimizers.
The reason for using Mini-batch training is twofold: Firstly, it is difficult to fit the large dataset into the memory, making the normal gradient descent infeasible. Even if the data can be fit into the memory, this method can also fall into local minima easily \cite{bottou1991stochastic} for deep learning. Secondly, SGD suffers from a high variance, which leads to a high fluctuation especially when the algorithm is close to the convergence. Mini-batch training can reduce the variance and relieve this situation.
In summary, the training procedure for our model can be described as follows:
\begin{algorithm}
\caption{Training Procedure of Neural Network}\label{Training}
\begin{algorithmic}
\Require Dataset ($\{\mathbf Z_1, \ldots, \mathbf Z_n\}$), Neural Network with Initial Value $\{W^{(1)}, b^{(1)}, \ldots, W^{(l)}, b^{(l)}\}$, Batch Size $b$
\While{Stopping criteria meets}
\State Sample out $b$ individuals.
\State \textbf{Forward Propagation}: Compute $\{l_1, \ldots, l_b\}$ with $\{W^{(1)}, b^{(1)}, \ldots, W^{(l)}, b^{(l)}\}$ following (\ref{neural network}).
\State \textbf{Backward Propagation}: Compute $\frac{\partial l_i}{\partial W^{(j)}}$ and $\frac{\partial l_i}{\partial b^{(j)}}$ for each $i = 1, 2, \ldots, b$ and $j = 1, 2, \ldots, l$.
\State Update the parameters $W_j$ with $W_j$ and approximated gradient $\frac 1b \sum_{i = 1}^b \frac{\partial l_i}{\partial W_j}$, similarly for $b_j$. This means
\[
W_j' = \text{Optim}(W_j, \frac 1b \sum_{i = 1}^b \frac{\partial l_i}{\partial W_j}), b_j' = \text{Optim}(b_j, \frac 1b \sum_{i = 1}^b \frac{\partial l_i}{\partial b_j})
\]
with Optim be Adam, AdamW or RMSProp.
\State $W_j = W_j', b_j = b_j', \forall j = 1, \ldots, l$
\EndWhile
\end{algorithmic}
\end{algorithm}
Note that in our simulation and real data experiments, we stop the algorithm based on early stopping, which means we will split the training data into 'true' training data used for model training and validation data. After each iteration, we will compute the loss based on validation data and if the loss does not decrease after several iterations, we stop the algorithm. It is called early stopping since the model may not meet the convergence criteria, but it is one really common trick in machine learning community and its goal is to prevent overfitting.
Last but not least, hyperparameter tuning is required in deep learning since each set of hyperparameters uniquely define a structure of neural network and there is no theoretical guarantee that which set of hyperparameters is the best for different tasks (similar as 'no free lunch theorem' \cite{shalev2014understanding}). For saving time we narrow down the range of tuning for each hyperparameter based on our experiments and we put the main technical details and results in the Appendix. Here are some take-away recommendations.
\begin{enumerate}
\item Batch size $b$ should be tuned in the range of $0.1\%$ to $5\%$ of the data size.
\item No significant difference in three optimizers (RMSProp, Adam, AdamW) with default values of parameters recommended in Pytorch.
\item Dropout Rate can be chosen between 0 and 0.25.
\item Learning Rate can be chosen from $0.0001$ to $0.05$.
\end{enumerate}
We also put them into our software for the sake of users.
\subsubsection{Six Designs for Proportional Hazards Assumption and Three Link Functions}
Deep learning is renowned for its extraordinary results in many tasks, including survival data analysis. However, this may lead to over-parameterization problem, which means the number of unknown parameters is too high compared with the size of data. In that case, training with neural network with few restrictions may provide in contrast worse results compared with other statistical or machine learning algorithms. Also, when the data is observed to have some desired statistical properties, it is also not wise to directly use deep learning for model training, but not consider mature statistical models.
In this part, we design 6 neural networks adapting to different demands. We provide options for modeling proportional and non-proportional data\footnote{Actually, the proportional design will not simulate a model that follows proportional hazard assumption when the derived covariates for one patient are time-dependent.} and also three link functions, combining them to get altogether 6 designs.
In general, discrete time model can be unified as
\begin{equation}
h[\lambda(t|\mathbf Z)] = h[\lambda_0 (t)] + \mathbf Z' \beta \cite{kalbfleisch2011statistical}
\label{Jack}
\end{equation}
where $h$ is a monotone-increasing and twice-differentiable function mapping $[0, 1]$ into $[-\infty, \infty]$ with $h(0) = - \infty$. Then if $h(u) = \log (- \log (1 - u))$, we have the grouped relative risk model \cite{kalbfleisch1973marginal}. If $h(u) = \log u$, we have the discrete relative risk model \cite{prentice2003mixed}. If $h(u) = \log \frac u {1 - u}$, we have the discrete logistic model \cite{cox1972regression}. Thanks to (\ref{Jack}) we can specially define the last layer and simply change the form of function $h(\cdot)$. To be more clear, for proportional version, the second last hidden layer will be fully connected with the final hidden layer, which is a scalar and shows the total effect of derived covariates. Then the final hidden layer will be fully connected with the output layer, with edges the baseline hazard values for each time point and the connection for the last two layer is specially defined, depending on $h(\cdot)$. The formula of neural network for proportional data follows (\ref{neural network}) but with the last two equations changing as
\[
\begin{aligned}
& Z_i^{(l)} = f^{(l)}(\mathbf Z^{(l - 1)}_i) = w^{(l)^\top} \mathbf Z^{(l - 1)}_i \\
& h(\boldsymbol {\lambda_i}) = h(\sigma(w^{(l + 1)})) + \mathbf Z_i^{(l)} \\
& l_i = r(\boldsymbol { \lambda_i})
\end{aligned}
\]
and $w^{(l)} \in \mathbb R^{h_{l - 1}}$, $w^{(l + 1)} \in \mathbb R^K$ and $\sigma(x) = \frac 1 { 1 + e^{-x}}$ is the sigmoid activation function, $\mathbf Z_i^{(l)} \in \mathbb R^K$ is the vector with all values the same as $Z_i^{(l)}$. Note that $\sigma(w^{(l + 1)})$ simulates the baseline hazard value at each time point. We put on a sigmoid function on the initial weights $w^{(l + 1)}$ since the baseline hazard value belongs in the range $[0, 1]$ and sigmoid function is a one-to-one mapping, which means each distinct value of $w^{(l + 1)}$ will lead to one unique value of baseline hazard for each time point.
For modeling non-proportional data, the structure does not require sepcial design. Figure \ref{Prop Non Prop} provides a comparison. For the normal design (which is also model designed for non-proportional model), the edges between the final hidden layer and the output layer are different for hazard values at different time points, which means the weights between the last two layers are time-dependent. Mathematically, we can write the formula for the last two layers as
\[
\lambda_{i}(\tau_k) = w(\tau_k)^{(l)^\top} \mathbf Z_i^{(l - 1)}, k = 1, \ldots, K
\]
where $\boldsymbol \lambda_i = (\lambda_{i}(\tau_1), \ldots, \lambda_i(\tau_K))^\top$ and $W^{(l)} = (w(\tau_1), \ldots, w(\tau_K))^\top $. It is obvious that $w(\tau_k)$ is a time-dependent coefficients operating on $\mathbf Z_i^{(l - 1)}$ and this ensures that the model itself will be non-proportional.
\begin{figure}[htbp!]
\begin{minipage}{0.48\linewidth}
\centering
\includegraphics[width=4in]{Proportional_Version.png}
\end{minipage}
\hfill
\begin{minipage}{0.48\linewidth}
\centering
\includegraphics[width=4in]{NonProportional_Version.png}
\end{minipage}
\caption{Comparison between proportional and non-proportional designs}
\label{Prop Non Prop}
\end{figure}
\subsection{Extension: Competing Risk Setting}
Competing Risk Problem focuses on the case when we not only want to model the event time, but also the cause that lead to the event. For example, people suffering from cancers usually have co-morbidities, such as cardiovascular disease. Although the competing risk problem is regarded really challenging in survival analysis, our model can solve the discrete competing risk problem just by a simple extension of the activation function. We noticed that \cite{lee2018deephit} is the first paper using deep learning to solve competing risk problem in discrete time model to the best of our knowledge. However, it models the PMF for each cause at each time point, but not the hazard function. In other words, our model is the first to model hazard function with neural network for solving competing risk problem in discrete time model. We also included the option for solving competing risk problem in our software.
In this section, we will introduce the theoretical guarantees for our extension. Since some of the formulas can be transferred directly from previous sections, we only introduce new notations and refer readers to the previous equations for old ones.
\subsubsection{Notations}
Compared with single-risk (simpler case that introduced in Section 3.1) case, here we define the \textit{cause-specific hazard function} resulting from cause $r$ as
\begin{equation}
\lambda_r(\tau_k|\mathbf Z) = P(T^* = \tau_k, R = r \mid T^* \ge \tau_k, \mathbf Z)
\end{equation}
where $r \in \{1, 2, \ldots, q\}$ means the number of causes for the event and we assume there are altogether $q$ causes. The \textit{overall hazard function} is defined as
\begin{equation}
\lambda(\tau_k|\mathbf Z) = \sum_{r = 1}^q \lambda_r(\tau_k|\mathbf Z) = P(T^* = \tau_k \mid T^* \ge \tau_k, \mathbf Z)
\end{equation}
and the formula for the survival function $S(t|\mathbf Z)$ and the probability density function $f(\tau_k|\mathbf Z)$ can be obtained with the same expression but different meaning of $\lambda(\tau_k|\mathbf Z)$. One difference is
\begin{equation}
f_r(\tau_k|\mathbf Z) = P(T^*=\tau_k, R=r \mid \mathbf{Z})=\lambda_{r}(\tau_k| \mathbf{Z}) \prod_{i=1}^{k - 1}(1-\lambda(\tau_i|\mathbf Z))
\end{equation}
As for the likelihood function, suppose for the $i$-th individual, we have $\{\mathbf Z_i, t_i, r_i, \delta_i\}$ as the information and $t_i = \tau_k$, then similarly we can define the likelihood contribution from individual $i$ as
\begin{equation}
\begin{aligned}
L_{i} & =P\left(T^*_{i}=t_{i}, R_{i}=r_{i}\right)^{\delta_{i}} P\left(T^*_{i}>t_{i}\right)^{1-\delta_{i}} P\left(C_{i} \geq t_{i}\right)^{\delta_{i}} P\left(C_{i}=t_{i}\right)^{1-\delta_{i}} \\
& \propto P\left(T^*_{i}=t_{i}, R_{i}=r_{i}\right)^{\delta_{i}} P\left(T^*_{i}>t_{i}\right)^{1-\delta_{i}} \\
& = f_{r_i}(t_i)^{\delta_i} S(t_i)^{1- \delta_i} = \lambda_{r_{i}}\left(\tau_k | \mathbf{Z}_{i}\right)^{\delta_{i}}\left(1-\lambda\left(\tau_k | \mathbf{Z}_{i}\right)\right)^{1-\delta_{i}} \prod_{j=1}^{k - 1}\left(1-\lambda\left(\tau_j | \mathbf{Z}_{i}\right)\right)
\end{aligned}
\end{equation}
Take the logarithm and combining the log-likelihood for each observation, define $t_i = \tau_j$ we can get the final log-likelihood function
\begin{equation}
\begin{aligned}
l & = \sum_{i = 1}^n \sum_{k = 1}^K Y_i(k)\left[ \log(1 - \lambda(\tau_k | \mathbf Z_i)) + \delta_{i}(j) \log \frac{\lambda_{r_i} (\tau_j | \mathbf Z_i)}{1 - \lambda (\tau_j | \mathbf Z_i)}\right] \\
& = \sum_{i = 1}^n \sum_{k = 1}^K Y_i(k)\left[ \log(1 - \lambda(\tau_k | \mathbf Z_i)) + \sum_{r = 1}^q \delta_{i}^r(j) \log \frac{\lambda_{r} (\tau_j | \mathbf Z_i)}{1 - \lambda (\tau_j | \mathbf Z_i)}\right]
\end{aligned}
\label{log l hazard competing}
\end{equation}
where we define $\delta_i^r(j) = I(T_i = j, R_i = r, \Delta_i = 1)$.
For the KL divergence, we follow \cite{wang2021kullback} but make the extension from single-risk to competing risk, we define \[
d_{k}\left( \boldsymbol {\tilde \lambda}, \boldsymbol \lambda ; \mathbf{Z}_{i}\right)= \sum_{r = 1}^q \tilde \lambda_r\left(\tau_k | \mathbf{Z}_{i}\right) \log \left\{\frac{\tilde \lambda_r\left(\tau_k | \mathbf{Z}_{i}\right)}{\lambda_r\left(\tau_k | \mathbf{Z}_{i}\right)}\right\}+\left\{1-\tilde \lambda\left(\tau_k | \mathbf{Z}_{i}\right)\right\} \log \left\{\frac{1-\tilde \lambda\left(\tau_k | \mathbf{Z}_{i}\right)}{1-\lambda\left(\tau_k | \mathbf{Z}_{i}\right)}\right\}
\]
Then we can get
\[
\begin{aligned}
D_{KL}(\tilde{\boldsymbol{\lambda}}, \boldsymbol{\lambda}) & =-\sum_{i=1}^{n} \sum_{k=1}^{K} Y_{i}(k)\left[\sum_{r = 1}^q \tilde \lambda_r \left(\tau_k \mid \mathbf{Z}_{i}\right) \log \left\{ \lambda_r(\tau_k | \mathbf Z_i) \right\}+ (1 - \tilde \lambda(\tau_k|\mathbf Z_i) )\log \left\{1-\lambda\left(\tau_k \mid \mathbf{Z}_{i}\right)\right\}\right]+C \\
& = -\sum_{i=1}^{n} \sum_{k=1}^{K} Y_{i}(k)\left[\sum_{r = 1}^q \tilde \lambda_r \left(\tau_k \mid \mathbf{Z}_{i}\right) \log \left\{ \frac{\lambda_r(\tau_k | \mathbf Z_i)}{1 - \lambda(\tau_k|\mathbf Z_i)} \right\}+ \log \left\{1-\lambda\left(\tau_k \mid \mathbf{Z}_{i}\right)\right\} )\right]+C
\end{aligned}
\]
by noting that $\sum_{r = 1}^q \tilde \lambda_r(\tau_k|\mathbf Z_i) = \tilde \lambda(\tau_k|\mathbf Z_i)$. Define $l_\eta = \frac{l - \eta D_{KL}(\tilde{ \boldsymbol \lambda}, \boldsymbol \lambda)}{1 + \eta} $, we have
\begin{equation}
l_\eta = \sum_{i = 1}^n \sum_{k = 1}^K Y_i(k)\left[ \log(1 - \lambda(\tau_k | \mathbf Z_i)) + \sum_{r = 1}^q \frac{\delta_{i}^r(j) + \eta \tilde \lambda_r(\tau_k|\mathbf Z_i)}{1 + \eta} \log \frac{\lambda_{r} (\tau_j | \mathbf Z_i)}{1 - \lambda (\tau_j | \mathbf Z_i)}\right]
\end{equation}
One question is to get $\tilde \lambda_r(\tau_k|\mathbf Z_i), \forall r$ at time point $k$ for the individual $i$. However this can be implemented easily by setting the activation function of the output layer as softmax function. More detailedly, instead of providing one node for each time point $k$ and individual $i$, we provide $q$ nodes and each node value is an estimation of cause-specific hazard value. The formula can also be written as $(\ref{neural network})$ but with the last two equations changing as
\[
\begin{aligned}
& \boldsymbol { \lambda_i} = f^{(l)}(\mathbf Z^{(l - 1)}_i) = \text{MarginalSoftmax}(W^{(l)} \mathbf Z^{(l - 1)}_i + b^{(l)}) \\
& l_i = r(\boldsymbol { \lambda_i})
\end{aligned}
\]
where the dimension of $\boldsymbol \lambda_i$ will be changed from $K$ to $qK$ (which means $W^{(l)} \in \mathbb R^{qK \times h_{l - 1}}$ and $b^{(l)} \in \mathbb R^{qK}$), and MarginalSoftmax will provide a transformation from $W^{(l)} \mathbf Z^{(l - 1)}_i + b^{(l)} \in \mathbb R^{q K}$ to $ \boldsymbol \lambda_i' \in R ^{K \times q}$ with each row of the new matrix the values for cause-specific hazard function. Then a softmax function will be operated on each row to get the final results. Mathematically, suppose $\boldsymbol \lambda_{ki}' \in \mathbb R^q$ is the $k$-th row for $\boldsymbol \lambda_i'$, then
\[
\boldsymbol \lambda_{kij} = \frac{e^{\boldsymbol \lambda_{kij}'}}{\sum_{j = 1}^q e^{\boldsymbol \lambda_{kij}'}}, j = 1, 2, \ldots, q
\]
which is the formula of softmax function and $\boldsymbol \lambda_{kij}$ and $\boldsymbol \lambda_{kij}'$ mean the $j$-th element of $\boldsymbol \lambda_{ki}$, $\boldsymbol \lambda_{ki}'$, respectively. After that the matrix will be changed back to a vector with dimension $qK$ without changing of meaning for each coordinate.
In summary, changing from sigmoid to softmax function means a generalization from binomial model to multinomial model, and softmax function corresponds to a multinomial logit function.
\section{Simulations}
In this section, we design various simulation settings for different goals. Firstly, we want to show that our model can perform better than other models when the quality of prior information is high, while when the quality of prior information is low, our model can at least obtain similar results compared with other models, since in this case, prior information can barely help improve the performance and robustness of the model. Secondly, we want to show that properties of deep learning still preserve and our model can obtain better results with the aid of suitable prior information. This can be done by simulating a non-linear and non-proportional dataset for the experiment. Our third goal is an extension, we observed that in real clinical cases, images are really important for experts to find out potential diseases. Based on this observation, we use MNIST \cite{lecun1998gradient}, a fundamental imaging dataset in Artificial Intelligence, to design a clinical trial dataset for training, our goal is to show that Convolutional Neural Network (CNN) can be successfully applied in our model, demonstrating that our model has a huge potential to apply on unstructured data.
\subsection{Simulation Settings}
\subsubsection{Simulation 1}
In this experiment, we generate data with non-linear relationship between covariates and targets. Moreover, Part of the covariates are function of time, which means the proportional hazard assumption does not hold in this dataset. The goal for this simulation is to show that the ability for modeling non-linear and non-proportional data still exists in our model.
The form of the hazard function in this setting is given as
\begin{equation}
h(t | x) = h_0 (x) \exp (g (t, x))
\end{equation}
where $g(t, x) = a(x) + b(x) t$ and
\begin{equation}
\begin{aligned}
& a(x) = \operatorname{sign} (x_2) + 0.44 x_0 + 0.66 x_1 + 0.88 x_2 + \frac 23 (x_0^2 + x_2^2 + x_0 x_1 + 2 x_1 x_2) \\
& b(x) = |0.2 (x_0 + x_1) + 0.5 x_0 x_1|
\end{aligned}
\end{equation}
In this simulation, we experiment with 3 models ({\color{blue} Maybe 4 will be better}): Cox model with proportional hazard assumption, LogisticHazard and our model. The event time is grouped with 20 time intervals and We want to show that LogisticHazard can achieve much better result compared with Cox model due to the flexibility of deep learning model, and our model can still get a better result compared with LogisticHazard when using prior information.
Additionally, we will generate 10000 observations with given model assumptions, where 95\% of the data will be used as external data and the remaining 5\% will be used as internal data. For the existing models, only internal data will be available. While for our models, we will use LogisticHazard to train prior model (i.e. to get $\tilde \lambda(t_k, \mathbf Z_i), \forall i, k$) and apply the prior information into the local data in our model.
\subsubsection{Simulation 3}
In this experiment, we use MNIST, a well-known imaging dataset, to simulate the time and event status. We want to show that more advanced structures, such as Convolutional Neural Network (CNN), can also be used in our model for extracting information from imaging data. This is meaningful since there exist many real applications relying on imaging data, such as using using deep learning for tumor detection and so on.
MNIST is an imaging dataset with each image a $28 \times 28$ handwritten-digit, which has a training set of 60000 examples and a test set of 10000 examples. Each image represents a digit and the event time will be generated based on the digit. More specifically, we define the event time to follow an exponential distribution $t \sim \exp(\alpha)$ (which means the density function is $f(x; \alpha) = \frac 1 \alpha e ^{- x / \alpha}$) with $\alpha$ given as
\begin{equation}
\alpha (\mathrm{digit}) = \frac{365 \cdot \exp (-0.6 \cdot \mathrm{digit})}{\log (1. 2)}
\end{equation}
And the data is right-censored with a manually set threshold, which means the individual whose survival time is greater than the given threshold will be marked as censored and the event of time will be changed to that threshold.
In this setting, we sample 30000 images from the training dataset for saving time. The model training procedure is the same as what in simulation 2-1.
In summary, for each individual, the covariates are pixels of the image, the event time is generated by an exponential distribution defined above, and observations will be marked as censored when the time is greater than a given threshold.
\subsection{Evaluation Metrics}
Three metrics are used for evaluating discrete time survival model and each of them focuses on slightly different aspects of the model.
\paragraph{Time Dependent Concordance Index} Concordance Index (C-index) \cite{harrell1984regression} is a discriminative evaluation metric used in survival analysis and it comes from the concept of Area Under Curve (AUC), the metric used in machine learning for measuring discriminative ability of the model. Since deep learning is mainly proposed for modeling non-proportional data, C-index measured at a given time point may not provide an overall picture of the discriminative ability of the survival model, so in this experiment, we use the version of C-index in \cite{antolini2005time}, to use the inverse probability weighted time dependent C-index combining the C-index measured at all possible event time points, which is
\begin{equation}
C^{t d}=\frac{\sum_{i: \Delta_{i}=1} \sum_{j: t_{i}<t_{j}} I\left(\hat{S}_{\mathbf Z_{i}}\left(t_{i}\right)<\hat{S}_{\mathbf Z_{j}}\left(t_{i}\right)\right) / \hat{G}^{2}\left(t_{i}\right)}{\sum_{i: \delta_{i}=1} \sum_{j: t_{i}<t_{j}} 1 / \hat{G}^{2}\left(t_{i}\right)}
\label{4}
\end{equation}
where $\mathbf Z_i, t_i, \delta_i$ are covariates, observed time and the indicator of event/censoring for observation $i$. $\hat S_{\mathbf Z_i}(t_i)$ is the predicted survival function and $\hat G(t)$ is the Kaplan-Meier (KM) estimator for the survival function of the censoring time. Note that under the independence assumption of censoring and event time, we have $C^{td} \overset p {\to} P(S_{\mathbf Z_i}(T) < S_{\mathbf Z_j}(T) \mid T_i < T_j )$ and higher $C^{td}$ means better model.
\paragraph{Integrated Brier Score} Brier Score (BS) measures calibration and discriminative performance simultaneously since it can be decomposed into two terms measuring these two aspects respectively. It follows the idea of Mean Squared Error (MSE) and the formula for BS is defined as
\begin{equation}
\mathrm{BS}(t)=\frac{1}{n} \sum_{i=1}^{n}\left\{\frac{\left(\hat{S}_{\mathbf Z_{i}}(t)\right)^{2} I\left(t_{i} \leq t, \delta_{i}=1\right)}{\hat{G}\left(t_{i}\right)}+\frac{\left(1-\hat{S}_{\mathbf Z_{i}}(t)\right)^{2} I\left(t_{i}>t\right)}{\hat{G}(t)}\right\}
\end{equation}
where $n$ is the number of observations and the other notations are the same as (\ref{4}). The Integrated Brier Score (IBS) \cite{graf1999assessment} is defined as an integration of the Brier Score in the time interval, which is defined as
\begin{equation}
\mathrm{IBS} = \frac 1 {t_{\max} - t_{\min}} \int_{t_{\min}}^{t_{\max}} \operatorname{BS}(t) dt
\end{equation}
And for IBS, we will firstly use the two interpolation schemes (CDI, CHI \cite{kvamme2019continuous}) to transfer the survival function trained for discrete time model to continuous time and then define a fine grid of time points to compute the IBS. The integral will be approximated by numerical integration. Lower IBS means better result.
\paragraph{Integrated Binomial Log Likelihood} This metric is also proposed in \cite{graf1999assessment} and is very similar to IBS, which also focuses both calibration and discriminative performance of the model. The definition of Binomial Log Likelihood (BLL) is
\begin{equation}
\operatorname{BLL}(t)=\frac{1}{n} \sum_{i=1}^{n}\left\{\frac{\log \left(1-\hat{S}_{\mathbf Z_{i}}(t)\right) I\left(t_{i} \leq t, \delta_{i}=1\right)}{\hat{G}\left(t_{i}\right)}+\frac{\log \left(\hat{S}_{\mathbf Z_{i}}(t)\right) I\left(t_{i}>t\right)}{\hat{G}(t)}\right\}
\end{equation}
and Integrated Negative Binomial Log Likelihood (INBLL) is defined as
\begin{equation}
\mathrm{INBLL} = - \frac 1 {t_{\max} - t_{\min}} \int_{t_{\min}}^{t_{\max}} \operatorname{BLL}(t) dt
\end{equation}
and the notations are the same as what in IBS. Note that compared with IBS, it penalizes the error with the scale $-\log (1 -x)$, which will lead to more penalization for larger magnitude of error. Lower INBLL means better model.
\subsection{Results}
\subsubsection{Simulation 1}
In our project, we compete with LogisticHazard \cite{gensheimer2019scalable} \cite{kvamme2019continuous}. 50 sets of internal data and 1 set of external data are generated in our model, which means all sets of internal data share the same prior information. The hyperparameter $\eta$ for our model is tuned by 5-fold cross validation, where for each fold, 80\% of the data will be sampled out for training and 20\% for validation, with IBS as the selection criteria. After $\eta$ is determined, we compute the prior information and both models will be trained with only local data, where 20\% of the data is sampled as test data, and in the remaining 80\% of the data, 80\% will be used for training model and 20\% is used for validation. Early Stopping is adopted with patience 5 and the stopping criteria is the loss function (\ref{3})
For neural network structures, we set a fully-connected neural network with two hidden layers. Each layer has 32 hidden nodes equipped with ReLU \cite{agarap2018deep} activation function to prevent gradient vanishing, the dropout rate is set to be 0.1 and batch normalization \cite{ioffe2015batch} is used in each layer. The optimizer is chosen to be Adam \cite{kingma2014adam} with two parameters set for two momentum values default as $0.9$ and $0.999$. The learning rate is set to be 0.01 and the batch size for training is set to be 20 due to the small size of training data. Note that for both models, the parameters are set to be the same for fairness.
Figure \ref{"fig1"} to \ref{"fig5"} have shown the results for different schemes when using 3 different metrics, where for IBS and IBLL, we divide the time as 100 time intervals with equal length. Significant improvement in average performance and robustness is shown in experiment results. Note that for figure \ref{"fig3"} and figure 5 (scheme 3 and scheme 5), our model does not perform better than existing model, it is reasonable because in these 2 cases, the external data has information that far different from local and true model (with correct day effects and 10 covariate effects), leading to $\eta$ be near 0. Note that for figure \ref{"fig4"} (scheme 4), our model performs really well, this is because the magnitude of loss does not differ a lot compared with scheme 5, which means the noise does not corrupt the prior data too much. In real cases this is significant since this shows that our model can not only behave better with high quality of prior information, but also can be used to identify the quality of prior data when we would like to apply prior information with unknown quality.
\begin{figure}[htbp!]
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.5in]{simulation_1_graph_1-1.png}
\end{minipage}
\hfill
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.5in]{simulation_1_graph_1-2.png}
\end{minipage}
\hfill
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.5in]{simulation_1_graph_1-3.png}
\end{minipage}
\caption{Results for Scheme 1}
\label{"fig1"}
\end{figure}
\begin{figure}[htbp!]
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.5in]{simulation_1_graph_3-1.png}
\end{minipage}
\hfill
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.5in]{simulation_1_graph_3-2.png}
\end{minipage}
\hfill
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.5in]{simulation_1_graph_3-3.png}
\end{minipage}
\caption{Results for Scheme 2}
\label{"fig2"}
\end{figure}
\begin{figure}[htbp!]
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.5in]{simulation_1_graph_2-1.png}
\end{minipage}
\hfill
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.5in]{simulation_1_graph_2-2.png}
\end{minipage}
\hfill
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.5in]{simulation_1_graph_2-3.png}
\end{minipage}
\caption{Results for Scheme 3}
\label{"fig3"}
\end{figure}
\begin{figure}[htbp!]
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.5in]{simulation_1_graph_4-1.png}
\end{minipage}
\hfill
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.5in]{simulation_1_graph_4-2.png}
\end{minipage}
\hfill
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.5in]{simulation_1_graph_4-3.png}
\end{minipage}
\caption{Results for Scheme 4}
\label{"fig4"}
\end{figure}
\begin{figure}[htbp!]
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.5in]{simulation_1_graph_5-1.png}
\end{minipage}
\hfill
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.5in]{Simulation_1_graph_5-2.png}
\end{minipage}
\hfill
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.5in]{Simulation_1_graph_5-3.png}
\end{minipage}
\caption{Results for Scheme 5}
\label{"fig5"}
\end{figure}
\subsubsection{Simulation 2-1}
For this experiment, we compete our model with Cox model and LogisticHazard. For our model, the hyperparameter tuning procedure, the model parameters and the data splitting rules are the same as what in Simulation 1.
Figure \ref{"fig 2-1"} has shown the comparison results with respect to 3 different metrics. For Cox model, it does not perform really well compared with the other two models since Cox model cannot deal with non-linear and non-proportional data really well, which is one advantage of Deep Learning. For the other two models, our model performs slightly better than LogisticHazard, which makes sense due to the aid of prior information in our model.
\begin{figure}[htbp!]
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.8in]{simulation_2-1_1.png}
\end{minipage}
\hfill
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.8in]{Simulation_2-1_2.png}
\end{minipage}
\hfill
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1.8in]{Simulation_2-1_3.png}
\end{minipage}
\caption{Results for Simulation 2-1}
\label{"fig 2-1"}
\end{figure}
\subsubsection{Simulation 2}
In this experiment, we compare our models with LogisticHazard, but with varied ratio of local data. We try 3 ratios, $\{2\%, 5\%, 10\%\}$ (which means a given ratio of data will be sampled out as internal data) and run 50 experiments with hyperparameter tuning and data splitting the same as what applied in Simulation 1. The model structures are summarized in figure \ref{Nerual Network}. We train with Adam optimizer with two parameters for moments fixed as default values ($0.9$ and $0.999$ described before), the learning rate is set as constant 0.005. We also apply early stopping with patience 10. The maximum number of training epoch is set to be 100 for saving time. We also use GPU in this model for accelerating training of Convolutional Neural Network, which can be easily done with \texttt{torchtuple}.
\begin{table}[htbp!]
\centering
\begin{tabular}{ |p{1.5cm}|p{2.5cm}|p{2cm}|p{1.5cm}|p{2.5cm}|p{2cm}|p{2cm}| }
\hline
\multicolumn{7}{|c|}{Architecture} \\
\hline
Layer & Name & Kernel Size & Stride & \# of Channels & Input Size & Output Size\\
\hline
1 & Conv2d &$5 \times 5$& $ 1 \times 1 $ & $16$ & $1 \times 28 \times 28$ & $16 \times 24 \times 24$\\
\hline
2 & MaxPool2d & $2 \times 2$ & $2$ & & $16 \times 24 \times 24$ & $16 \times 12 \times 12$\\
\hline
3 &Conv2d & $5 \times 5$& $1 \times 1$ & $16$ & $16 \times 12 \times 12$ & $16 \times 8 \times 8$\\
\hline
4 &AdaAvgPool2d & $8 \times 8$& $8$ & & $16 \times 8 \times 8$ & $16 \times 1 \times 1$\\
\hline
& Flatten & & & & $16 \times 1 \times 1$& $16$ \\
\hline
5& Linear & & & &$16$ & $16$\\
\hline
6& Linear & & & & $16$ & $20$\\
\hline
\end{tabular}
\caption{Neural Network Structure Used in Simulation 3}
\label{Nerual Network}
\end{table}
The results are shown below and the stability is significantly improved with the amount of internal data increasing. Meanwhile, our model performs better when IBS is used as the evaluation criteria.
\begin{figure}[htbp!]
\begin{minipage}{0.48\linewidth}
\centering
\includegraphics[width=2.8in]{simulation_3_graph_1.png}
\end{minipage}
\hfill
\begin{minipage}{0.48\linewidth}
\centering
\includegraphics[width=2.8in]{simulation_3_graph_2.png}
\end{minipage}
\caption{Results for Simulation 3}
\label{"fig 2-2"}
\end{figure}
\section{Real Data Analysis}
In this section, we apply our method on 3 real datasets, which are the Study to Understand Prognoses Preferences Outcomes and Risks of Treatment (SUPPORT), the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) and Medical Information Mart for Intensive Care \Rmnum{3} (MIMIC-\Rmnum{3}). Note that the covariates of each dataset have different properties that require processing by neural network, which provides with space for us to demonstrate some advanced usages of neural network architectures in our model.
\subsection{Information of the datasets}
\paragraph{SUPPORT} SUPPORT dataset is a large study focusing on the survival time of seriously illed hospitalized patients \cite{knaus1995support}. Around 68.1\% people died during the study with median time 58 days. The original data consists of 9105 patients and 14 features, for which all features are indices observed or diagnosed by patients (age, sex, race, number of comorbidities, presence of diabetes, presence of dementia, presence of cancer, mean arterial blood pressure, heart rate, respiration rate, temperature, white blood cell count, serum’s sodium, and serum’s creatinine). The data we use in our experiment follows \cite{katzman2018deepsurv}, where missing data are dropped and altogether 8873 patients are kept in the dataset.
\begin{figure}
\centering
\includegraphics[scale = 0.5]{Graph_Illustration_of_Embedding.jpg}
\caption{Illustration of Embedding. Suppose the dimensionality of all nominal features is $m$ and the desired embedding layer is $n$, essentially this is a one-layer fully connected neural network and this can be understood as a linear projection from $\mathbb R^m$ to $\mathbb R^n$. After getting the embedding layer with $n$ dimensions, we combine the values with other features (suppose the dimensionality of all other features is $p$) to get the input layer with dimensionality $n + p$. This is the input layer used for model training. }
\label{Graph Illustration of Embedding}
\end{figure}
Note that in this dataset, we have continuous, binary and nominal features. Therefore we apply embedding techniques to specially do data-preprocessing for nominal features, which is a common trick used in deep learning for nominal features. Figure \ref{Graph Illustration of Embedding} shows the meaning of embedding. This shows the easy-to-use essence of deep learning since if we would like to use traditional statistical models such as GLM, we need to carefully learn the knowledge in the dataset, which is not necessary for deep learning.
\paragraph{METABRIC} METABRIC \cite{curtis2012genomic} is one dataset that focuses on studying the relationship between gene and clinical features and breast cancer. Following \cite{katzman2018deepsurv}, we drop some missing data and 1904 patients are preserved with 9 features, where 4 are gene features and 5 are clinical features. This dataset has only binary and continuous features, so no other additional data-preprocessing steps are required.
\paragraph{MIMIC-\Rmnum{3}} MIMIC-\Rmnum{3} \cite{physiobank2000physionet}, \cite{johnson2016mimic}, is a deidentified dataset associated with over 40k patients who stayed in critical care units of the Beth Israel Deaconess Medical Center between 2001 and 2012. Compared with other datasets, this is an electronic health record (EHR) data, which means real-time features that observed in a routine basis can be accessible. In other words, the size of this dataset is much larger and features are more complicated than the other two datasets. The goal for using this dataset is to show that with proper tuning of some common parameters for deep learning, our model can achieve a satisfactory result with only part of this dataset accessible. We do not use SUPPORT or METABRIC for this aim due to the reason that the size of these two datasets is not so large, which means other models may not be able to train a reasonable result.
\subsection{Results}
\paragraph{SUPPORT} For this dataset, we have subsampled 5\%, 10\% and 20\% of the data as internal data, and then for each ratio, we have 50 groups of data for experiments and $\eta$ is tuned with 5-fold cross validation, where the data splitting procedure and evaluation metrics are the same as what in Simulation 1. Due to the existence of categorical covariates in SUPPORT, we create an additional embedding layer to deal with them and then combine the embedded features with other features to be the input of the 3-layer fully-connected neural network. ReLU is used as the activation function for each layer, Batch Normalization is also equipped for better training of the neural network in each layer with momentum 0.1. Dropout rate is set to be 0.2 for each layer. For the optimization, AdamWR \cite{loshchilov2017decoupled} is applied with decoupled weight decay rate to be 0.01 and learning rate is set to be the constant 0.01, holding other parameters default. Early Stopping is used with patience 10.
\begin{figure}[htbp!]
\begin{minipage}{0.48\linewidth}
\centering
\includegraphics[width=3in]{Real_Data_graph_1.png}
\end{minipage}
\hfill
\begin{minipage}{0.48\linewidth}
\centering
\includegraphics[width=3in]{Real_Data_graph_2.png}
\end{minipage}
\caption{Results (left: SUPPORT, right: METABRIC)}
\label{"Real Data"}
\end{figure}
The left graph in Figure \ref{"Real Data"} shows the trend of changing sample size. Similar as what has shown in Simulation 3, the stability of the model improves when the sample size of local data increases. Additionally, for LogisticHazard, the results improve with more local data, while for our model, the improvement is not so straightforward. We guess this is because when the size of training data is large enough, deep learning can model well enough and in such case, prior information is not a significant factor for prediction accuracy.
\paragraph{METABRIC} For this dataset, we follow similar data splitting rules as SUPPORT, except for the change of ratio of internal data. We sampled out $10\%, 20\%$ and $30\%$ of the data as internal since the size of data for METABRIC is much smaller than SUPPORT. Since we require no data-preprocessing tricks for this dataset, we use the same model parameters as what used in simulation 1, as the data structures for these two experiments are essentially equivalent. The goal for this experiment is the same as SUPPORT, which means we want to show the trend of our comparison result with the size of internal data increasing.
The right graph in Figure \ref{"Real Data"} shows the results for METABRIC, similar trends appear and the conclusion is the same, we do not repeat that due to the length limit.
\paragraph{MIMIC-\Rmnum{3}} The
\section{Conclusion and Future Work}
\bibliographystyle{plain}
|
\section{Introduction}
Metaverse, which is considered as the next generation of the Internet, has attracted researchers' attention recently~\cite{xu2022full, mystakidis2022metaverse}. In Metaverse, people can interact with the virtual world through technologies like virtual reality and augmented reality. Currently, people typically access the servers of Metaverse providers through Metaverse local devices, such as head-mounted glasses, to get the Metaverse services. These devices are required to assist users in accessing and experiencing Metaverse services, as well as to process resource-intensive computing tasks locally.
However, using Metaverse local devices for computing faces severe challenges: 1) the computing and communication resources of the devices are limited; 2) the locations of devices performing local computing are dispersed and the devices may be constantly moving.
Fortunately, there is one existing solution for addressing these challenges, namely mobile edge computing (MEC)~\cite{abbas2017mobile, huda2022survey}. Specifically, local devices can offload their computing tasks to proxy MEC servers, e.g., base stations; then, the MEC servers finish those tasks and return the results to local devices. Since the MEC servers are located close to local devices and usually have enough computing resources, their involvement can reduce the latency of communication and computing, thus providing low-latency and high-quality Metaverse services.
In practice, the MEC servers in Metaverse are usually responsible for computing multiple tasks from different users. Since their resources are not infinite, one single MEC server may not be able to handle those offloading tasks in time, leading to low quality of Metaverse services. One possible solution is to involve other MEC servers with extra resources to work together for offloading computing in Metaverse, which makes it necessary to establish a secure resource trading platform
for edge servers. To that aim, we propose a blockchain system running on MEC servers to form a distributed computing framework, named blockchain-based MEC platform,
which enables resource integration and optimal utilization among MEC servers in a trustless environment.
Currently, there is no existing study focusing on implementing blockchain-based MEC for resource sharing and optimization in Metaverse.
Although there exist several studies about MEC in Metaverse, they focus on the latency analysis~\cite{dhelim2022edge} and incentive mechanism design~\cite{xu2021wireless}. While for research about blockchain-based MEC~\cite{sheng2020near, zhaofeng2019blockchain}, no one has considered solving the resource sharing problem in MEC.
To fill this gap, the proposed blockchain-based MEC platform aims to assist resource sharing and optimization in Metaverse via trading offloading tasks in a transparent but secure way, so that more offloading tasks from Metaverse users can be finished timely. This platform comprises multiple Metaverse users, MEC servers, and a consortium blockchain system running the practical Byzantine fault tolerance (PBFT) consensus protocol~\cite{castro1999practical, xu2021concurrent}. And there are four main procedures in the proposed system, i.e., \textit{data submission, task allocation, offloading computing}, and \textit{payment of tasks}. As the pivotal step, task allocation faces a critical challenge brought by the heterogeneity of multi-task requests from users. Specifically, given the limited computing and communication resources of each MEC server and offloading tasks with different price policies, data sizes, and completion time requirements, our proposed system needs to assign the requested multiple tasks to multiple servers under various constraints. In addition, since the offloading tasks are time-sensitive, the task allocation step is expected to make the (near) optimal decisions as fast as possible so as to reduce the latency of the whole system.
To address these challenges, we design a learning-based task allocation mechanism. Specifically, we first formulate the task allocation issue as an integer-programming problem; then, we analyze the utilities of different decisions by considering the resource constraints of servers and the time requirements of tasks; according to our analysis, the multiple tasks allocation (MTA) problem is NP-complete, and the time complexity is too high, so we design a learning-based algorithm to find its approximate optimal value in polynomial time.
Since the objective function and constraints of the MTA problem are conditional, which are affected by the servers uploading the tasks to blockchain, in our proposed learning-based solution, we first transform the MTA problem into a reinforcement learning problem and then calculate the rewards for each state and action considering the influences of the server source of tasks.
To the best of our knowledge, we are the first to implement blockchain-based MEC in Metaverse. The main contributions of this paper are summarized as below:
\begin{itemize}
\item We are the first to propose a practical blockchain-based MEC platform for resource sharing and optimal utilization in Metaverse. Our proposed system can satisfy the demands of low latency, multiple requests, energy efficiency, incentive compatibility, and data privacy protection for Metaverse users.
\item We design a task allocation mechanism for edge resource sharing. Under the heterogeneous constraints, we first formulate an MTA problem for reasonably distributing offloading tasks among multiple MEC servers
\item We propose a learning-based solution to find the approximate optimal solution for the MTA problem with polynomial time complexity. We speed up the learning process via identifying available actions
for each state.
\item We conduct extensive experiments to verify the effectiveness, efficiency, and validity of our proposed system, mechanisms, and algorithms.
\end{itemize}
\label{related}
\section{System Model}
\label{system}
\subsection{System Overview}
Our proposed blockchain-based edge resource sharing platform is illustrated in Fig.~\ref{fig_sys} for supporting the Metaverse applications, consisting of mobile devices as Metaverse users, MEC servers, and the consortium blockchain running with practical Byzantine fault tolerance (PBFT) consensus. Here we assume that users in our considered system are devices with offloading requests to MEC servers. Based on the arrival time of offloading tasks in the blockchain network, we assign a specific task number to each of them.
Specifically, we define $\mathcal{T}=\{t_1,\cdots,t_j,\cdots,t_m\}$ as the set of offloading tasks from users
with $m$ denoting the number of all offloading tasks.
Let $\mathcal{S}=\{s_1,\cdots,s_i,\cdots,s_n\}$ denote the set of MEC servers in the Metaverse
with $n$ being the total number of servers.
\begin{figure}[h!t]
\centering
\includegraphics[width=\linewidth]{bc_mec-Page-2_2.pdf}
\caption{The illustration of our proposed system.}
\label{fig_sys}
\end{figure}
The workflow of our proposed system can be described as below: 1) \textit{Data Submission}: The Metaverse users upload their raw data and the descriptions of tasks to their nearest MEC servers. We use $R_j<p_j, D_j, \tau_{e,j}>$ to denote the description of task $t_j$,
where $p_j$ is the unit price of per CPU cycle of computing for finishing $t_j$, $D_j$ is the data size of $t_j$, and $\tau_{e,j}$ is the time requirement for finishing $t_j$. 2) \textit{Task Allocation}: Once server $s_i$ receives the data and $R_j$ from users, the data is stored locally and $R_j$ is submitted to the blockchain for task allocation which will be elaborated in the following sections. 3) \textit{Offloading Computing}: According to the task allocation results, server $s_i$ will either send the data of tasks it received to appropriate servers or finish the computing of the tasks locally. Then, the servers start to work on the allocated tasks and will broadcast the computing results to the blockchain network when the tasks are completed. 4) \textit{Payment of Tasks}: The users first pay offloading computing fees through the blockchain network where they can also get the computing results later. Each server can get its own payment through the coordination of the blockchain.
To protect the privacy of the computing results, we can employ the asymmetric encryption technique, such as Rivest–Shamir–Adleman (RSA) algorithm~\cite{izdemir2021rivest}, in this system. Specifically, at the beginning of each round, the users submit the public keys generated by RSA along with the offloading task requests while keep their private keys secret. After the tasks are finished by the MEC servers, the computing results will be encrypted with the public keys of the corresponding users. Then, the users can download the encrypted results via accessing the blockchain
and decrypt the results with their private keys.
Since the amount of tasks received by each server can be different, we define the tasks submitted by server $s_i$ as $\mathcal{T}_{s,i}=\{t_1,t_2,\cdots,t_{m'}\}$, where $m'\leq m$ is the number of tasks from $s_i$. For example, if $t_1$, $t_3$, and $t_{20}$ are submitted by $s_1$, then we have $\mathcal{T}_{s, 1}=\left\{t_1,t_3,t_{{20}}\right\}$.
\subsection{Consortium Blockchain}
The consortium blockchain is running on the MEC servers,
where all servers are authorized nodes and thus can be trusted. Besides, as mentioned earlier, we implement PBFT, a lightweight consensus protocol in blockchain, to assist servers in reaching consensus for generating blocks to help disseminate task descriptions and allocation results, as well as enforce payment distribution
in a round-by-round manner.
Here is the workflow of the blockchain network in our system: 1) At the beginning of each round, each server publishes its available resources for finishing tasks, including both computing and communication resources (see Section~\ref{model} for details), and uploads received task descriptions $R_j$ to the blockchain network. 2) All of the above-mentioned information will be broadcast in the blockchain network so that the leader node can allocate the received tasks to appropriate servers with the help of the task allocation mechanism detailed in the next section. Then the allocation decisions, resource information of the MEC servers, and task descriptions will be packaged into a new block. 3) If $s_i$ is assigned to process the tasks submitted by itself, it can start processing right away; otherwise, it has to transfer the received data to other servers for finishing the offloading tasks. 4) After tasks are finished, the results will be broadcast so as to be recorded on the blockchain. 5) The users can get the results from the blockchain, and the servers will be paid accordingly.
Overall, the blockchain offers a decentralized and trusted platform to conduct resource sharing for MEC empowered Metaverse applications, thus enabling offloading computing for Metaverse users to overcome the challenges posed by the insufficient computing power of individual server and device mobility. With the combination of MEC and blockchain, our proposed resource sharing platform can handle multiple offloading tasks in a efficient and distributed manner. However, regarding the pivotal step, we need to design a task allocation mechanism to assign offloading tasks to proper MEC servers so as to achieve the optimal resource utilization while satisfying user requests, which will be discussed in the next section.
\label{mechanism}
\section{Design of Task Allocation Mechanism }
\label{model}
In this section, we detail the design of task allocation mechanism by modeling all possible computing and communication costs for $s_i$
\subsection{Computing Cost Model}
According to~\cite{wang2022incentive}, if task $t_j$ is allocated to server $s_i$,
its energy cost can be calculated as $E_{i,j}^\text{comp}=\alpha_i \mu_{i,j} f_{i}^2$,
where $\alpha_i$ is the parameter related to the architecture of CPU, $\mu_{i,j}=D_j \theta_{i}$ is the total CPU cycles required to finish task $t_j$ on the MEC server $s_i$ with $\theta_{i}$ being the CPU cycles required to process unit data sample, and $f_i$ is the CPU frequency used to compute the offloading tasks. Besides, the time consumption of computing $t_j$ on $s_i$ can be calculated by $T_{i,j}^\text{comp}=\frac{\mu_{i,j}}{f_{i}}$.
Let $\mu_i$ be the maximum available computing capacity of $s_i$. Then $\alpha_i, \theta_i, f_i$ and $\mu_i$ will be submitted to the blockchain as the available computing resource of $s_i$ for task allocation.
\subsection{Communication Cost Model}
If $t_j$ is submitted by $s_i$ but not finally assigned to $s_i$, $s_i$ would only be responsible for transmitting the data of $t_j$ to the determined server.
We define $B_{i}$ as the allocated bandwidth of $s_i$ for data transmission; let $H_{i}$ and $G_{i}$ be the transmission power and channel power gain, respectively.
Based on Shannon Bound, we can get the data transmission rate as $r_{i}=B_{i}\log_2 (1+\frac{H_{i} G_{i}}{\delta^2})$,
where $0\leq \delta \leq 1$ is the Gaussian noise during the transmission. Then we can calculate the transmission time as $T_{i,j}^\text{comm}=\frac{D_j}{r_{i}}$.
And the energy consumption of data transmission can be calculated by $E_{i,j}^\text{comm}=H_{i} T_{i,j}^\text{comm}$.
Note that $B_i, H_i$, and $G_i$ are uploaded to the blockchain network as the available communication resource of $s_i$.
\subsection{Utility Model}
After $t_j$ is completed, server $s_i$ will receive the payment from users, denoted by $p_j\mu_{i,j}$ for task $t_j$, which is the product of unit price per CPU cycle and the total number of consumed CPU cycles. Then, the utility of $s_i$ regarding finishing $t_j$ in the system is the difference between the received payment and the local cost.
Since there are two possible task allocation results for any server $s_i$ with respect to processing task $t_j$, i.e., computing $t_j$ locally or not, which will significantly affect the utility of $s_i$, we define an indicator $\mathds{1}_{i,j}$ to capture this: if $t_j$ is assigned to $s_i$, then $\mathds{1}_{i,j}=1$; otherwise, $\mathds{1}_{i,j}=0$.
If $t_j \in \mathcal{T}_{s,i}$, the utility can be expressed as $U_{i,j}^{I}= (p_j \mu_{i,j}-E_{i,j}^\text{comp}) \mathds{1}_{i,j}+(\lambda p_j \mu_{i,j}-E_{i,j}^\text{comm})(1-\mathds{1}_{i,j})$,
where $\lambda p_j \mu_{i,j}$ is the intermediary fee paid by the server finishing $t_j$, with $\lambda$ being a predefined constant parameter that can be known globally;
and the total time consumption is $T_{i,j}^I=T_{i,j}^\text{comp}\mathds{1}_{i,j}+T_{i,j}^\text{comm} (1-\mathds{1}_{i,j})$. While if $t_j \notin \mathcal{T}_{s,i}$, the utility is defined as $U_{i,j}^N=[(1-\lambda)p_j \mu_{i,j}-E_{i,j}^\text{comp}] \mathds{1}_{i,j}$, and the total time consumption is $T_{i,j}^N=(T_{i,j}^\text{comp}+T_{i^*,j}^\text{comm})\mathds{1}_{i,j}$, where $T_{i^*,j}^\text{comm}$ is the transmission time of server submitting $t_j$.
Thus, we can get the utility function and time consumption as: if $t_j \in \mathcal{T}_{s,i}$, $U_{i,j}=U_{i,j}^I$ and $T_{i,j}=T_{i,j}^I$; otherwise, $U_{i,j}=U_{i,j}^N$ and $T_{i,j}=T_{i,j}^N$.
\subsection{Problem Formulation}
\label{problem_formulation}
Recall the goal of task allocation for achieving resource integration and optimal utilization in our system, we would like to make sure that more tasks can be proceed in time and thus servers can get more rewards. So we can formulate it into a multi-task allocation (MTA) problem as follows:
\begin{align*}
\mathbf{MTA:} &\mathop{\arg\max}_{\mathds{1}_{i,j}} \sum_{i=1}^n \sum_{j=1}^m U_{i,j} \\
\text{s.t.}:
\text{C1}&:\sum_{j=1}^{m} \mu_{i,j} \mathds{1}_{i,j} \leq \mu_i, \\
\text{C2}&: T_{i,j}\leq \tau_{e,j}, \\
\text{C3}&:\sum_{j=1}^m \mathds{1}_{i,j}\leq 1, \\
\text{C4}&: i\in\left\{1,2,\cdots,n\right\},j\in\left\{1,2,\cdots,m\right\}, 0< n \leq m.
\end{align*}
In the above MTA problem, the optimization objective is to maximize the total utility of all servers given requested tasks from Metaverse users;
C1 is the computing capacity constraint to make sure that the selected tasks can be proceeded by server $s_i$; C2 is the time constraint, consisting of computing time and transmission time to ensure that $t_j$ can be completed in time; C3 guarantees that one task can only be assigned to one server, and it is possible that some servers are not assigned with tasks; C4 defines the domain of this optimization problem.
\begin{theorem}
MTA is an NP-complete problem.
\end{theorem}
\begin{proof}
The MTA problem can be described as a complex multiple knapsacks problem (MKP), which has been proved to be NP-complete in~\cite{li2019multi}. Specifically, MKP is defined as: let $\mathcal{T'}=\left\{t_1,t_2,\cdots,t_{m'}\right\}$ denote the set of $m$ items, and let $\mathcal{S'}=\left\{s_1,s_2,\cdots,s_{n'}\right\}$ as the set of $n$ knapsacks, and $m'>n'$. Each item $t_i'$ has its weight $D_j'$ and value $p_j'$ and each knapsack has its maximum weight $D_i'$. The goal is to decide where should each item be placed in the knapsacks so that the total utility is maximized. In other words, it aims to solve the following problem: $\max_{\mathds{1}_{i,j}}\sum_j^m \sum_i^n p_j'\mathds{1}_{i,j} $, where $\mathds{1}_{i,j}$ is an indicator: when $\mathds{1}_{i,j} = 1$, item $t_j'$ will be assigned to knapsack $s_i'$, and otherwise, $\mathds{1}_{i,j} = 0$.
The time complexity is $O(n^{m})$, which means MKP cannot be solved in polynomial time. MKP can be reduced to the simplified MTA problem.
In the original MTA problem, we need to allocate multiple tasks to multiple servers with computing resource and time consumption constraints. The objective of MTA is to maximize the total utility, i.e., $\max_{\mathds{1}_{i,j}}\sum_j^n \sum_i^m U_{i,j}$. If we remove the time constraints and assume that every server offers the same computing and communication resources for each task, we get the simplified MTA problem which is equivalent to MKP. As we can see, the simplified MTA also has the time complexity of $O(n^{m})$, so MTA is an NP-complete problem.
\end{proof}
\section{Learning-based Solution for MTA}
\label{solution}
Based on the analysis in Section~\ref{problem_formulation}, we know that the MTA problem is NP-complete, so we need to design a computationally efficient algorithm to solve it. In this section, we design a learning-based algorithm with $\epsilon$-greedy strategy
\subsection{Problem Reformulation}
To begin with, we need to reformulate the MTA problem based on the Q-learning algorithm~\cite{jiang2020q, chien2020q}. There are three main components in Q-learning, i.e., state space, action space, and reward function, which are detailed as follows.
\subsubsection{State Space}
We denote the state space of the agent, i.e., the MEC server executing the Q-learning algorithm to allocate tasks, as $\Omega <\mathcal{T},\mathcal{P}, \mathcal{D}, \mathcal{T}_{e} >$. Specifically, $\mathcal{T}=\left\{t_1,t_2,\cdots,t_m\right\}$ is the set of tasks,
$\mathcal{P}$ is the set of the unit prices of tasks, i.e., $\mathcal{P}=\left\{p_1,p_2,\cdots,p_m\right\}$, $\mathcal{D}=\left\{D_1,D_2,\cdots,D_m\right\}$ is the set of data sizes of tasks, and the set of execution time is defined as $\mathcal{T}_{e}=\left\{\tau_{e,1},\tau_{e,2},\cdots,\tau_{e,m}\right\}$. In other words, the state space is composed of all tasks with the corresponding descriptions, including prices, data sizes, and execution time requirements. Thus, $\Omega$ is a matrix with $m$ rows and 4 columns. The agent selects an action for each state based on the current task requirements and resource conditions of all servers. Once the action is chosen at a state, the agent will turn to the next state to conduct action selection.
\subsubsection{Action Space}
Since servers have different amount of available resources, such as total CPU cycles, CPU cycle frequencies, and communication bandwidth, we can denote the action space as $\mathcal{A}<\mathcal{S},\mathcal{M},\mathcal{F},\mathcal{B},\mathcal{H},\mathcal{G},\alpha,\Theta >$. In detail, $\mathcal{S}$ is the set of servers, and $\mathcal{M}=\left\{\mu_1,\mu_2,\cdots,\mu_n\right\}$ is the set of total available CPU cycles, and $\mathcal{F}=\left\{f_1,f_2,\cdots,f_n\right\}$ is the set of CPU frequencies, and $\mathcal{B}=\left\{B_1,B_2,\cdots,B_n\right\}$ is the set of communication bandwidth; as for $\mathcal{H}=\left\{H_1,H_2,\cdots,H_n\right\}$ and $\mathcal{G}=\left\{G_1,G_2,\cdots,G_n\right\}$, they are the sets of transmission power and channel power gain, respectively; and $\alpha=\left\{\alpha_1,\alpha_2,\cdots,\alpha_n\right\}$ is the set of the parameter correlated to the CPU architectures and $\Theta=\left\{\theta_1,\theta_2,\cdots,\theta_n\right\}$ is the set of CPU cycles required for processing one data sample.
Specifically, there are two statuses for each action, i.e., \textit{selected} and \textit{not selected}, and the agent can choose only one action in one state while one action can be selected multiple times in all the states. This is to ensure that one task can only be assigned to one server; however, one server could process multiple tasks if it has sufficient resources. In this way, we can know that the action space is an $n\times8$ matrix.
\subsubsection{Reward Function}
\label{reward_function}
The objective of MTA problem is to maximize the utility by allocating tasks to proper servers, so the rewards here are defined by the utilities.
In other words, the rewards are determined by the allocation decision, resource conditions, and time constraints. Besides, the rewards would also be affected by where the task is submitted from to the blockchain network. Thus, in the design of the reward function, we need to consider all of these aspects.
We evaluate whether our system can process $t_j$ by two criteria: 1) server $s_i$ has enough computing power to be devoted to the computation, i.e., $\mu_{i,j}\leq\mu_i$; and 2) $s_i$ is able to process the task $t_j$ within the required time constraint, i.e., $T_{i,j}\leq \tau_{e,j}$. The second criteria is C2 while the first criteria is a weak C1. And C1 can only be used when selecting actions and updating Q-value, which will be discussed in Section~\ref{learning process}.
If $t_j\in T_{s_i}$, the reward function can be expressed as:
\begin{align}
U_{i,j}=
\begin{cases}
p_j \mu_{i,j}-E_{i,j}^\text{comp},& \mu_{i,j}\leq\mu_i\ \text{and}\ T_{i,j} \leq \tau_{e,j},\\
\lambda p_j \mu_{i,j}-E_{i,j}^\text{comm},& \text{otherwise}.
\end{cases}
\label{reward_1}
\end{align}
When $t_j$ is from $s_i$, if $s_i$ is not assigned to process $t_j$,
it has to transmit $t_j$ to another server and obtain some intermediary fee, so the reward is $\lambda p_j \mu_{i,j}-E_{i,j}^\text{comm}$; but if $s_i$ is able to process $t_j$, it can get the payment $p_i \mu_{i,j}$ at the cost of the computing energy consumption $E_{i,j}^\text{comp}$, and thus, the reward is $U_{i,j}=p_i \mu_{i,j}-E_{i,j}^\text{comp}$.
If $t_j\notin T_{s_i}$, the reward function is as below:
\begin{align}
U_{i,j}=\begin{cases}
(1-\lambda)p_j \mu_{i,j}-E_{i,j}^\text{comp},&\mu_{i,j}\leq\mu_i\ \text{and}\ T_{i,j} \leq \tau_{e,j},\\
0,& \text{otherwise}.
\end{cases}
\label{reward_2}
\end{align}
The logic of (\ref{reward_2}) is similar to (\ref{reward_1}): when $t_j$ is not from $s_i$, if $s_i$ has the capability to process $t_j$, it can get the reward the same as in (\ref{reward_1}) but has to pay the intermediary fee; and it will get nothing if it cannot process this task, which means that $A_i$ as one action cannot be chosen in state $R_j$.
Based on the above analysis, we know that the reward functions are the transformation of the objective function under certain constraints. For simplicity, we summarize the calculation of reward functions in Algorithm~\ref{al_11}. To get an $n \times m$ matrix $\mathcal{U}=\left\{U_{1,1},U_{1,2},\cdots,U_{i,j},\cdots,U_{n,m}\right\}$ containing the reward value for each state and each action, we need to calculate the energy consumption and time cost for both computing and communication processes (Lines 3-6). And then we can calculate and return the rewards based on (\ref{reward_1}) and (\ref{reward_2}) (Lines 7-25).
\begin{algorithm}
\caption{StateActionReward}
\label{al_1}
\begin{algorithmic}[1]
\REQUIRE $\Omega$, $\mathcal{A}$
\ENSURE $\mathcal{U}$
\FOR{$i\in \{1,\cdots,n\}$}
\FOR{$j \in \{1,\cdots,m\}$}
\STATE $E_{i,j}^\text{comp}$ $\leftarrow$ $\alpha_i \mu_{i,j} f_{i}^2$
\STATE $T_{i,j}^\text{comp}$ $\leftarrow$ $\frac{\mu_{i,j}}{f_{i}}$
\STATE $T_{i,j}^\text{comm}$ $\leftarrow$ $\frac{D_j}{r_{i}}$
\STATE $E_{i,j}^\text{comm}$ $\leftarrow$ $H_{i} T_{i,j}^\text{comm}$
\IF{$s_j \in T_{s_i}$}
\STATE $T_{i,j} \leftarrow T_{i,j}^\text{comp}\mathds{1}_{i,j}+T_{i,j}^\text{comm} (1-\mathds{1}_{i,j})$
\IF{($\mu_{i,j}\leq \mu_i$) and ($T_{i,j}$ $\leq$ $\tau_{e,j}$)}
\STATE $U_{i,j} \leftarrow p_j \mu_{i,j}-E_{i,j}^\text{comp}$
\ELSE
\STATE$U_{i,j} \leftarrow \lambda p_j \mu_{i,j}-E_{i,j}^\text{comm}$
\ENDIF
\ENDIF
\IF {$s_j \notin T_{s_i}$}
\STATE $T_{i,j}=(T_{i,j}^\text{comp}+T_{i^*,j}^\text{comm})\mathds{1}_{i,j}$
\IF{($\mu_{i,j}\leq \mu_i$) and ($T_{i,j}$ $\leq$ $\tau_{e,j}$)}
\STATE $U_{i,j}\leftarrow (1-\lambda)p_j \mu_{i,j}-E_{i,j}^\text{comp}$
\ELSE
\STATE$U_{i,j} \leftarrow 0$
\ENDIF
\ENDIF
\ENDFOR
\ENDFOR
\RETURN $\mathcal{U}$
\end{algorithmic}
\label{al_11}
\end{algorithm}
\subsection{Learning Process}
\label{learning process}
\subsubsection{Available Actions}
To avoid selecting unmatched servers that cannot process such tasks as actions,
and to improve the learning efficiency by reducing the time complexity, we need to clarify which set of actions are available in each state before selecting one as the action. Thus, we define the concept of available actions as below.
\begin{definition}[Available Actions]
\label{ava_actions}
Assume the agent is at state $\Omega_j$, $\mathcal{A}_{j}^{ava}=\left\{s_1,s_2,\cdots,s_{n'}\right\}$ is the set of available actions, i.e., servers, that can process $t_j$ under the constraints of $\mathrm{C1}-\mathrm{C4}$, with $n'\leq n$ being the number of available actions.
\end{definition}
Since the transfer of states is a dynamic process and the states will affect each other, one of the most direct effects is that once an action is selected, its computing power will be reduced and therefore will constrain the action selection of the subsequent states. Therefore, we evaluate the available actions of each state one by one.
Recall the discussions in Section~\ref{reward_function}, we get an $n\times m$ rewards table $\mathcal{U}$, and we can use $\mathcal{U}_j=\left\{U_{1,j},U_{2,j},\cdots,U_{n,j}\right\}$ to denote all the possible rewards of $\Omega_j$.
Generally speaking, $\mathcal{U}_j$ has three types of value: positive, zero, and negative. We can select those actions with non-zero values as the available actions. However, this naive method can lead to some severe consequences. For example, one server is selected multiple times due to its non-zero value (this could happen in Q-learning, especially when the number of servers is small), but its capacity of available computing resources is not enough to handle all of these tasks even if the reward for each task is positive or optimal.
To address this challenge, we design a dynamic table
to record the accumulative $\mu_{i,j}$, which can be denoted by $\mu_{i,j}^{acc}$. This table shares the same dimensions with the rewards table $\mathcal{U}$ generated in Algorithm~\ref{al_11}. Hence, we can design the following mechanism to get the available actions for each state.
First, naively select the actions with non-zero values from $\mathcal{U}$ and initialize the values as zero to reduce the computational cost by avoiding going through the whole action space again. Let $\mathcal{A}_j^{nz}=\left\{A_{1,j}^{nz},A_{2,j}^{nz},\cdots,A_{n',j}^{nz}\right\}$ be the set to contain the actions with non-zero values at state $\Omega_j$. In other words, the agent only needs to check the actions in $\mathcal{A}_j^{nz}$ at state $\Omega_j$.
Then, based on the selected actions, we can get a value of $\mu_{i,j}$. In this way, after multiple rounds, we can get the accumulative value $\mu_{i,j}^{acc}$. This is a dynamic process, which means that only those selected actions can contribute to their corresponding accumulative values. Besides, we need to consider the time constraint, ensuring that one server can compute the allocated tasks in time. We use $\tau_{e,i}^{ava}$ to demonstrate the available computing time for the actions in $\mathcal{A}_{j}^{nz}$, which is calculated via $\tau_{e,i}^{ava}=(\mu_{i}-\mu_{i,j}^{acc})/f_{i}$.
Next, we can decide whether an action ${A}_{i,j}^{nz}$ is available by the following rules: if $\mu_{i,j}^{acc}<\mu_i$ and $\tau_{e,j}^{ava}>\tau_{e,j}$, then $\mathcal{A}_{i,j}^{nz}$ is considered as available action; otherwise, ${A}_{i,j}^{nz}$ will be discarded from $\mathcal{A}_j^{nz}$. In this way, we can get a set of available actions $\mathcal{A}_j^{ava}$ at state $\Omega_j$, and the final action ${A}_j$ should be chosen from it according to the selection policy which will be discussed in Section~\ref{action_selection}.
\begin{algorithm}[b!]
\caption{StateAvailableActions}
\begin{algorithmic}[1]
\REQUIRE $\mathcal{A}$, $\Omega_j$, $\mathcal{U}$
\ENSURE ${A}_j$
\STATE Initialize $\mathcal{A}$, $\mathcal{A}_j^{ava}$, $\mu_{i,j}^{acc}$
\FOR {$i \in \{1,\cdots,n\}$}
\IF {$U_{i,j} \neq 0$}
\STATE Append $A_i$ into $\mathcal{A}_j^{nz}$
\ENDIF
\ENDFOR
\STATE Choose $A_{i,j}^{nz}$ from $\mathcal{A}_j^{nz}$
\STATE $\tau_{e,i}^{ava}\leftarrow (\mu_i-\mu_{i,j}^{acc})/f_{i}$
\IF {$A_{i,j}^{nz} \in \mathcal{A}_j^{ava}$}
\IF {($\mu_{i,j}^{acc} > \mu_i$) or ($t_{i,j}^{ava} < \tau_{e,j}$)}
\STATE Remove $A_{i,j}^{nz}$ from $ \mathcal{A}_j^{ava}$
\STATE $A_j$ $\leftarrow$ choose another action from $ \mathcal{A}_j^{nz}$
\ENDIF
\ELSE
\STATE Append $A_{i,j}^{nz}$ into $\mathcal{A}_j^{ava}$
\STATE ${A}_j \leftarrow A_{i,j}^{nz}$
\ENDIF
\STATE $\mu_{i,j}^{acc}$ $\leftarrow$ $\mu_{i,j}^{acc}+\mu_{i,j}$
\RETURN ${A}_j$
\end{algorithmic}
\label{al_2}
\end{algorithm}
The whole process of choosing available actions for a specific state is summarized in Algorithm 2. At the beginning, we initialize $\mathcal{A}_j^{nz}$, $\mathcal{A}_j^{ava}$, and $\mu_{i,j}^{acc}$ (Line 1), and we get the non-zero values from $\mathcal{U}$ (Lines 2-6). Next, we choose an action from $\mathcal{A}_j^{nz}$ and calculate its available computing time (Lines 7-8).
Then, we can determine whether the selected action is available or not (Line 9-17), and we update $\mu_{i,j}^{acc}$ (Line 18).
Finally, we get the available action ${A}_j$ at state $\Omega_j$. Please note that the selection method in Lines 7 and 12 will be further discussed in Section~\ref{action_selection}. In general, Algorithm~\ref{al_2} is the transformation of $\mathrm{C1}$ and $\mathrm{C2}$ in the MTA problem.
\subsubsection{The Update of Q-table}
\label{action_selection}
The Q-table is applied during the learning process, which is used to facilitate the action selection. First, we generate an $m\times n$ matrix as the Q-table $Q(\Omega, \mathcal{A})$, and initialize its value with 0. Then, we need to update the Q-value at each state. Here we use the following equation to update Q-value:
\begin{align*}
&Q(\Omega_j, A_i)\\
&=Q(\Omega_j, A_i)+\alpha[U_{i,j}+\gamma \max Q(\Omega'_j, A_j^{ava})-Q(\Omega_j, A_i)],
\end{align*}
where $Q(\Omega'_j, A_j^{ava})$ means all the possible Q-value at next stage; $0\leq \alpha \leq 1$ is the learning rate and $0\leq \gamma \leq 1$ is the discount factor.
\begin{algorithm}[b!]
\caption{Learning-based Algorithm for the MTA Problem}
\begin{algorithmic}[1]
\REQUIRE $\Omega, \mathcal{A}, K$
\ENSURE $TS^*$
\STATE Initialize $Q(\Omega,\mathcal{A})$, $k$
\STATE $\mathcal{U} \leftarrow$ StateActionReward($\Omega, \mathcal{A}$)
\STATE $ A_{j}^{nz} \leftarrow$ actions with non-zero values in $\mathcal{U}$
\STATE Generate a random value $x$
\IF{$x \leq \epsilon$}
\STATE $A_j \leftarrow$ randomly selected from $\mathcal{A}_{j}^{nz}$
\ELSE
\STATE $A_j$ $\leftarrow$ $\max Q(\Omega_j,.)$
\ENDIF
\FOR{$k \in \{1,\cdots,K\}$}
\FOR{$j \in \{1,\cdots,m\}$}
\STATE $A_j \leftarrow$ StateAvailableActions($\mathcal{A}$, $\Omega_j$, $\mathcal{U}$)
\STATE $Q(\Omega_j,A_i) \leftarrow Q(\Omega_j,A_i)+\alpha[U_{i,j}+\gamma \max Q(\Omega'_j, \mathcal{A}_j^{ava})-Q(\Omega_j, A_i)]$
\ENDFOR
\STATE $TS \leftarrow $ the sum of all rewards in this episode
\ENDFOR
\STATE $TS^* \leftarrow \max TS$
\RETURN $TS^*$
\end{algorithmic}
\label{al_3}
\end{algorithm}
The agent selects the action based on the Q-value, and we adopt the $\epsilon$-greedy algorithm as the policy of action selection. The $\epsilon$-greedy strategy selects the action with the largest expected rewards most of the time, and the parameter $\epsilon$ balances exploration and exploitation. We can use a larger value of $\epsilon$ to allow the agent to exploit what have been learned, and the agent will explore more actions that have not been learned with a smaller $\epsilon$.
The learning process requires multiple rounds so that the agent can learn more about the rewards and find the best solution. For each episode $k\in \left\{1,2,\cdots,K\right\}$ where $K$ is the total episodes of learning, we let the agent go through all the states and find a solution, and then we compare all the solutions $TS$ and choose the one $TS^*$ with the highest sum of rewards as the optimal solution.
Algorithm 3 is the detailed process of the learning-based solution for the MTA problem. We first initialize $Q(\Omega,\mathcal{A})$, $k$, and get $U_{i,j}$ via Algorithm~\ref{al_11} (Lines 1-2). The $\epsilon$-greedy strategy is implemented to select the action (Lines 3-9). Then, the learning process will last until episode $k$ reaches the predefined total number of episodes $K$ (Lines 10-16), and the optimal solution will be obtained and returned (Line 17-18).
\subsection{Complexity Analysis}
The learning-based algorithm comprises three sub-algorithms, so the complexity analysis needs to take all of them into account. First, the computational complexity of Algorithm~\ref{al_11} is $O(m\times n)$, and the complexity of Algorithm~\ref{al_2} is $O(n)$. As for the time complexity of Algorithm~\ref{al_3}, it is $O(k\times n)$. Thus, in general, the time complexity of the learning-based solution should be $O(m\times n)=O(k\times n)+O(n)+O(m\times n)$, which means that we can solve the MTA problem in polynomial time.
\section{Experimental Evaluation}
\label{experiments}
\subsection{Experimental Setting}
In our experiments\footnote{The code is available in: {https://github.com/wzljerry/Blockchain-based-Edge-Resource-Sharing-for-Metaverse}}, we consider a blockchain-based MEC system in Metaverse with 20 servers and 50 offloading tasks as the primary setting. We change relevant parameters to analyze the impacts of the numbers of servers and tasks on the total rewards and the time complexity, as well as the influence of Q-learning parameters on the convergence. For clarity, we summarize the basic parameter settings in Table~\ref{tb_1}
\begin{table}[h!t]
\centering
\caption{Basic Parameter Settings}
\begin{tabular}{|l|l|l|l|}
\hline
$n=20$ & $m=50$ & $D_i=[200,400]$ & $\theta=0.01 $ \\ \hline
$p_j=[1,10]$ & $D_j=[10,20]$ & $\tau_{e,j}=[1,100]$ & $\gamma=0.9$ \\ \hline
$f_i=[1,10]$ & $H_i=[5,10]$ & $G_i=[5,10]$ &$\alpha=0.01$ \\ \hline
$\delta=0.01$ & $B_i=[5,10]$ & $\epsilon=0.9$ & $K=500$ \\ \hline
\end{tabular}
\label{tb_1}
\end{table}
\subsection{Experimental Results}
First, we design experiments to prove the effectiveness of our proposed learning-based algorithm for the MTA problem.
To that aim, we compare the learning-based solution with two benchmark algorithms, i.e., the random allocation algorithm and the greedy-based algorithm. Specifically, The random allocation algorithm allocates each task to one server randomly; and the greedy-based algorithm selects the server with the maximum reward for each task
We run these three algorithms with different numbers of servers and tasks to obtain
the comparison results as shown in Fig.~\ref{fig_1}. From Figs.~\ref{fig_1}(a) and~\ref{fig_1}(b), we can see that the learning-based algorithm can always outperform the other two algorithms with higher rewards. The random strategy fluctuates a lot due to the randomness of each selection. The greedy and learning algorithms, in contrast, perform more smoothly. Intuitively, more servers will not result in higher total rewards; however, more tasks would increase the overall revenue. But from our experimental results, this trend does not hold. There are at least three reasons for it. First, C1--C4 constrain the allocation decisions, resulting in dynamic decisions as the number of servers or tasks changes, which does not have a cumulative effect on rewards; second, although more tasks result in more rewards, they also increase the cost of computation and communication; and third, servers and tasks are heterogeneous with different parameters, exacerbating the results from the first two reasons.
\begin{figure}[h]
\centering
\subfigure[The Number of the MEC Servers]{
\includegraphics[width=0.465\linewidth]{test2_3.pdf}}\hfill
\subfigure[The Number of Offloading Tasks]{
\includegraphics[width=0.465\linewidth]{test5_1.pdf}}
\caption{The influences of the numbers of MEC servers and offloading tasks on the total rewards.}
\label{fig_1}
\end{figure}
Then, we explore the influences of unit prices and data sizes of offloading tasks on the total rewards. We increase $p_j$ and $D_j$ with the percentage from 10\% to 100\%, and the results are shown in Fig.~\ref{fig_5}. It is clear that both the increase of $p_j$ and $D_j$ will affect the total rewards. When the unit prices are larger, the users will pay more to the system, so the total rewards will increase. And, if the data size of $t_j$ is larger, then the CPU cycles required to process $t_j$ will also be increased, and thus the total rewards will be higher.
\begin{figure}[h]
\centering
\subfigure[The Increase Percentage of Unit Price]{
\includegraphics[width=0.465\linewidth]{test_para_1.pdf}}\hfill
\subfigure[The Increase Percentage of Data Size]{
\includegraphics[width=0.465\linewidth]{test_para_2.pdf}}
\caption{The influences of unit prices and data sizes of offloading tasks on the total rewards.}
\label{fig_5}
\end{figure}
Last but not least, we mainly examine the convergence performance of the learning algorithm. Q-learning is determined by two main variables, i.e., $\alpha$ and $\gamma$.
From Fig.~\ref{fig_2}(a), we can see that when $\alpha$ is smaller, the current choice is more influenced by experience and lacks further exploration, so although it achieves a high payoff at the beginning, the convergence rate does not improve effectively.
From Fig.~\ref{fig_2}(b), we can see that the larger the value of $\gamma$, the better the convergence. This is because $\alpha$ takes into account the effect of future rewards on current choices, so as $\gamma$ becomes larger, it emphasizes rewards more and thus learns the optimal combinations more efficiently to speed up the convergence.
\begin{figure}[h!t]
\centering
\subfigure[The Learning Rate $\alpha$.]{
\includegraphics[width=0.465\linewidth]{test7_1.pdf}}\hfill
\subfigure[The Discount Factor $\gamma$.]{
\includegraphics[width=0.465\linewidth]{test_8_1.pdf}}
\caption{The convergence of the learning-based algorithm.}
\label{fig_2}
\end{figure}
\section{Conclusion}
\label{conclusion}
In this paper, we establish a blockchain-based MEC platform for resource sharing and optimization to facilitate Metaverse applications. In particular, we design a task allocation scheme to assist our proposed system. To that aim, a learning-based algorithm is proposed to help the system make task allocation decisions in polynomial time. Numerous experiments prove that our proposed system and algorithms are efficient.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Deep inelastic scattering off a nuclear target at high energies constitutes the best environment to probe the small x tail of the hadron wave function and to confirm experimentally the existence of gluon saturation.
There is a trivial reason why one is interested in the smallest probe possible: to minimize the effect of final state interaction and make experimental data interpretable from the perspective of the hadron wave function. At high energy, the coherence length of the virtual photon is significantly larger than the proton radius. Thus the larger the nuclear target the denser the gluon field which the projectiles probes. While the Electron Ion Collider will be the best discovery machine for the saturation physics, the ultra-peripheral collisions~\cite{Bertulani:2005ru,Baltz:2007kq} at RHIC and the LHC may also be a very useful probe of the small x hadron wave function as a nearly real photon likely has a smaller partonic content than the proton, and consequently the influence of final state effects should be smaller as well.
In this paper we focus on correlations between produced particles in Ultraperipheral Collisions (UPC). Traditionally, hydrodynamics has been successful in dealing with correlations produced in heavy ion collisions. However, the discovery of such correlation in small systems brought to the fore the question of their origin. Following the high multiplicity pp measurement by CMS~\cite{CMS:2010ifv}, a set of measurements was done in the past decade, specifically, p-Pb at LHC~\cite{ALICE:2012eyl,ATLAS:2012cix,CMS:2012qk} and p-Au, d-Au, He-Au at RHIC~\cite{PHENIX:2016cfs,PHENIX:2013ktj,STAR:2014qsy}, see more details in review articles~\cite{Schenke:2021mxx,Nagle:2018nvi}. In addition similar correlations have been recently observed in UPC \cite{ATLAS:2021jhn}. Similar measurement was also performed in $e^+e^-$ collision \cite{Badea:2019vey}, and ep Deep-inelastic-scattering \cite{ZEUS:2019jya}, and no ridge correlation was found in this data.
Our focus in this paper is on the possibility that particle correlations originate from the quantum correlation in the initial state hadronic wave function.
From theoretical perspective, the wave-function of a nearly real photon is poorly understood due to uncontrollably large non-perturbative effects. Thus theoretical results describing scattering and particle production in UPC are unavoidably model dependent. In this paper, in order to at least provide a sense of the systematic uncertainty, we consider two models for the photon wave functions. The first one is based on a naive perturbative process of photon splitting into a virtual quark-antiquark pair. The second assumes that the real photon can be represented as, in general non-perturbative, hadronic state, e.g. $\rho$-meson~\cite{Sakurai:1960ju}. A $\rho$-meson wave-function evolved to asymptotically high energy/small x would be similar to the wave function of a nucleus or a proton and thus can be approximated by the McLerran-Venugopalan (MV) model. For the energies available at the LHC and especially at RHIC, this approximation can be expected to be wanting. Nonetheless, two diametrically opposite approximations may help to understand the influence and the importance of the associated systematics.
Within the Color Glass Condensate approach, the scattering and central rapidity particle production in photon-nucleus collisions pose yet another challenge (see Ref.~\cite{Zhao:2022ayk} for hydrodynamic approach to UPC).
At the moment, the two - particle production was only studied in the forward, photon going direction~\cite{Shi:2020djm}, which is not appropriate to describe the central rapidity production for two reasons. First, theoretically, forward production is a significantly different process than the production at central rapidity and as such it has qualitatively different systematics as a function of the control parameters. Second, the experimental data shows extremely strong rapidity dependence, see Ref.~\cite{ATLAS:2021jhn}. The lack of boost-invariance~\cite{Bozek:2015swa,Ke:2016jrd,Schenke:2016ksl,Shen:2020jwv,Wu:2021hkv} invalidates approximating of the mid-rapidity physics by the physics of the forward region. In particular only a small fraction of particles is produced in the forward region. Thus it is hard to imagine that particle correlations integrated over a sizable rapidity interval can be reasonably described in terms of these hadrons alone as in \cite{Shi:2020djm}. Thus addressing mid-rapidity production is absolutely crucial in order for the CGC based theoretical analysis to be taken seriously.
Here we undertake such a calculation, albeit within the constraints mentioned above.
As in the rest of the CGC based calculations, we consider inclusive two-gluon production from the projectile wave function in the eikonal approximation for the emission vertex. This approximation is justified if the gluons have sufficiently different rapidities. On the other hand when the rapidity difference becomes too large, $\sim \frac{1}{\alpha_s}$ this approximation has to be revised as one would be required to account for the evolution between the gluon emissions~\cite{Jalilian-Marian:2004vhw}. However experimentally one is not dealing with such large rapidity differences, and so our simple approximation should be adequate.
The calculation of particle production in the CGC framework necessitates averaging of products of Wilson lines over the color field ensemble of the target.
This in general is a rather complex endeavor since to study correlations we are forced to go beyond the large $N_c$ limit which then requires us to calculate correlators of a large number of Wilson lines. In order to make the calculations tractable and not overly numerically demanding, in this paper we use the so-called {\it factorized dipole approximation} (FDA) discussed in Ref.~\cite{Kovner:2017ssr,Kovner:2018vec} and successfully applied to study particle production and correlations in Refs.~\cite{Altinoluk:2018ogz,Kovner:2018fxj,Kovner:2018azs,Li:2021ntt,Altinoluk:2020psk}. This approximation is well suited for a dense nuclear target in and close to saturation regime. We will explain this in more detail in the following.
The paper is organized as follows.
We start from summarizing the machinery used to calculate particle production in dilute-dense scattering in Sec.~\ref{sec2}. In Sec.~\ref{sec3}, we calculate a general expression for double gluon production cross section. In Sec.~\ref{sec4} and Sec.~\ref{sec5}, we discuss projectile ensemble average for the dipole and MV models respectively. Sec.~\ref{Sec:Num} lays out the details of our numerical procedure. We conclude with the discussion of our results in Sec.~\ref{sec7}.
\section{Particle production in dilute-dense scattering}\label{sec2}
In this section we outline the most important ingredients of the formalism and define the observables.
The formalism of our choice is the wave function approach of Refs.~\cite{Kovner:2001vi,Baier:2005dv}. Alternatively one can use diagrammatic methods of Ref.~\cite{Kovchegov:2012mbw}. While the final expressions are the same, the representation we use is convenient for organizing calculations.
Consider a projectile at high energy before a collision with a hadronic/nuclear target. The projectile's wave function can be written in the general form
\begin{align}
| \Psi_{\rm in} \rangle =
\sum_{\{\v{x}, k^+, \alpha\}}
\psi (\{\v{x}, k^+, \alpha\})
\prod_i|{\v{x}_i, k^+_i, \alpha_i}\rangle
\end{align}
where $\v{x}_i$ is the $i$-th parton's transverse position, $k_i^+$ longitudinal momentum and $\alpha_i$ the color index. Subject to the physical process, the latter can belong to either (anti)fundamental
or adjoint representation describing a (anti)quark or a gluon in the incoming wave function.
At high energy, the separation of relevant timescales allows for significant simplifications of the analysis of the scattering process. The propagation time through the target in high energy collisions can be arbitrary small compared to effective inter-parton interaction time. One thus can neglect the interactions between the partons and approximate the scattering process by independent propagation of partons.
For a similar reason,
at high energy
the partons do not change their transverse positions during the scattering process. The high energy scattering in QCD is non-trivial due to the different eikonal phases acquired by different components of the wave function. This difference in phases is responsible for decoherence of the components of the wave function and leads to particle production in the final state. The wave function of the outgoing projectile right after scattering is
\begin{align}\label{out}
| \Psi_{\rm out} \rangle = S |\Psi_{\rm in}\rangle = \sum_{\{\v{x}, k^+, \alpha\}}
\psi (\{\v{x}, k^+, \alpha\})
\prod_i U_{\alpha_i \beta_i}( \v{x}_i) |\v{x}_i, k^+_i, \beta_i\rangle
\end{align}
where $U$ is the Wilson line defined in the representation of $SU(N_c)$ group appropriate to the color charge of a given parton:
\begin{align}
U_{\v{x}} = {\cal P} e^{ i \int d x^- T^a A^+_a(\v{x},x_-) }.
\end{align}
Here $A^+$ is the $+$ component of the gauge field of the target.
To calculate the expectation value of a gluon observable
in the final state, one has to allow for the propagation of the projectile state \eqref{out} to asymptotic time $t\rightarrow\infty$.
This evolution results in emission of additional gluons which dress the bare partons by their WW fields as well as in recombination of some of the outgoing soft gluons into the WW field of the outgoing fast partons.
It is straightforward to account for the evolution after scattering (or equivalently to change the basis from free to dressed partons) by introducing the coherent operator
\begin{align}\label{c}
C = {\cal P} e^{i \sqrt{2} \int d^2 x d\xi\,
\hat b_a^i(\xi, \v{x})
\left[
a^\dagger_{i,a} (\xi, \v{x})
+ a_{i,a} (\xi, \v{x})
\right]
}
\end{align}
where the WW field operator is defined through the total charge density operator $\hat{\rho}_{\rm P}^a(\xi,\v{y})$ integrated from the rapidity of the projectile to $\xi$
\begin{align}
\hat{b}^i_a(\xi,\v{x})=\frac{g}{2\pi}\int d^2y \frac{(\v{x}-\v{y})^i}{|\v{x}-\v{y}|^2} \hat{\rho}_{\rm P}^a(\xi,\v{y})+...
\end{align}
Here we have explicitly written out only the leading order in $\alpha_s$ expression. Higher order corrections can be included, but we will not take them into account in our calculations. The expectation value of a
gluon observable $O(a,a^\dagger)$ in the final state is then given by
$ \langle \Psi_{\rm out} | C \,O(a,a^\dagger) \, C^\dagger | \Psi_{\rm out} \rangle $.
For example the differential single inclusive gluon production is given by
\begin{align}
\frac{ dN } { dy d^2 \v q} = \langle \Psi_{\rm out} | C \,
a_{i,b}^\dagger(y, \v q)
a_{i,b}(y, \v q) \,
C^\dagger | \Psi_{\rm out} \rangle
\end{align}
This is nothing but the high energy counterpart of the Faddeev-Kulish construction of dressed states in theories with massless gauge bosons.
In this paper we will consider double gluon production, which is a straightforward generalization of the previous equation
\begin{align}
\frac{dN}{d\eta d^2\v{q}_1 d\xi d^2\v{q}_2} = \langle \Psi_{\rm out} | C \,
a_{i,b}^\dagger(\eta, \v{q}_1)
a_{i,b}^\dagger(\xi, \v{q}_2)
a_{i,b}(\eta, \v{q}_1) \,
a_{i,b}(\xi, \v{q}_2) \,
C^\dagger | \Psi_{\rm out} \rangle
\end{align}
To perform the actual calculations, we need of course a model for the wave function of the projectile. In general, in the high energy CGC approach,
this wave function has the form of a wave function for valence charges dressed by the WW field, that is \begin{align}
| \Psi_{\rm in} \rangle = C | v \rangle,
\end{align}
where $C$ is the coherent operator Eq.(\ref{c}) and $| v \rangle$ describes a state in the free Fock space of a small number of valence (large rapidity) partons. As mentioned in the introduction, we use two models for the valence wave function $|v \rangle $ - the dipole and MV model, which we describe in detail below.
Finally, the target is treated as an ensemble of classical gluon field configurations which have to be averaged over.
We now proceed with the detailed calculations within the framework just described.
\subsection{Modelling the Projectile and the Target}
As stated earlier, we explore two models for the wave function of a nearly on shell photon.
\subsection{Dilute dipole projectile}
For a nearly on-shell photon, the longitudinal polarization is suppressed by a square of the virtuality. Hence only the transverse polarization is of importance.
The leading order perturbative expression for the photon state is given by ~\cite {Beuf:2016wdz}
\begin{equation}
\begin{split}
|\gamma_{\lambda}^*\rangle \simeq &
\sum_{s_1,s_2}\int_0^{\infty} \frac{dz}{4(2\pi) z(1-z)} \int \frac{d^2 \v{k}_1 }{(2\pi)^2} \Psi^{T}_{\lambda}(z, \v{k}_1,s_1,s_2)b^{\dagger}_{\alpha,s_1}(\v{k}_1^+, k_{1})d^{\dagger}_{\alpha,s_2}(\v{k}_2^+, -\v{k}_{1})|0\rangle\\
=& \sum_{s_1,s_2}\int_0^{\infty} \frac{dz}{4(2\pi) z(1-z)} \int d^2 \bm{z}_1 d^2 \bm{z}_2 \Psi^{T}_{\lambda}(z, \v r,s_1,s_2)b^{\dagger}_{\alpha,s_1}(\v{k}_1^+, \v{z_1})d^{\dagger}_{\alpha,s_2}(\v{k}_2^+, \v{z_2})|0\rangle\\
\end{split}
\end{equation}
where $s_{1,2}$ are polarizations of quark and anti-quark, $\alpha$ is the color index in the fundamental representation, $\lambda=\pm 1$ is the photon polarization, $r=z_1-z_2$. The longitudinal momentum of the (anti)quark is ${k}_1^+ = z p^+$ (${k}_2^+ = (1-z) p^+$). The above expression is written in the reference frame where the photon has zero transverse momentum. The photon splitting functions in the momentum and coordinate space are
\begin{align}
\Psi^{T}_{\lambda}(z,\v k_{1},s_1,s_2)&=-2ee_f\delta_{s_1,-s_2}(2z-1+2\lambda s_1)\sqrt{z(1-z)}\frac{\v k_{1}\cdot \v{\epsilon_{\lambda}}}{\v k_{1}^2+\varepsilon_f^2}
\\
\Psi^{T}_{\lambda}(z, \v r ,s_1,s_2)&=-i\frac{2ee_f}{2\pi}\delta_{s_1,-s_2}(2z-1+2\lambda s_1)\sqrt{z(1-z)}\frac{\v r \cdot \v{\epsilon_{\lambda} }}{|\v r|}\varepsilon_fK_1(\varepsilon_f|\v r|),
\end{align}
where $\varepsilon_f^2 = z(1-z)Q^2$. In numerical calculations we use a small but nonzero value of $Q\sim 200$ MeV, see Sec.~\ref{Sec:Num} for details.
\subsubsection{Correlators of the charge density.}
For the calculation we will need to know correlators of up to four color charge density operators. Here we calculate those in the dipole model of the photon.
The charge density operator in terms of the quark and anti quark creation and annihilation operators is defined as
\begin{align}
\hat{\rho}^a(x^-,\v x)= b^{\dagger}_{\alpha \sigma}(x^-, \v{x}) t^a_{\alpha \beta} b_{\beta \sigma}(x^-,\v{x})-d^{\dagger}_{\alpha \sigma}(x^-, \v{x}) t^a_{\beta \alpha } d_{\beta \sigma}(x^-,\v{x})
\end{align}
For the purpose of a CGC calculation we need the correlators of the integrated quantity
\begin{equation}
\hat\rho^a(\v x)\equiv \int dx^-\hat{\rho}^a(x^-,\v x)
\end{equation}
We now evaluate the correlators of $\hat\rho$ in a dipole state
$|q\bar{q}\rangle = b^{\dagger}_{\alpha \sigma}(\v{z_1})d^{\dagger}_{\alpha \sigma}(\v{z_2})|0\rangle$,
where we have suppressed indices and arguments irrelevant for this computation. \\[10pt]
Using
\begin{equation}
\begin{split}
&\hat{\rho}^a(\v x) b^{\dagger}_{h}(\v {z_1})d^{\dagger}_{h}(\v{z_2})|0\rangle\\
=& \left[ t^a_{\alpha h}b^{\dagger}_{\alpha}(\v{z_1})d^{\dagger}_{h}(\v{z_2}) \delta^{(2)}(\v{x}-\v{z_1})
-t^a_{ h \alpha}b^{\dagger}_{h}(\v{z_1})d^{\dagger}_{\alpha}(\v{z_2})\delta^{(2)}(\v{x}-\v{z_2})\right]|0\rangle
\end{split}
\end{equation}
and
\begin{equation}
\begin{split}
& \langle 0|d_{l}(\v{z_2})b_{l}(\v{z_1})\hat{\rho}^a(\v{x})\\
=&\langle 0| \left[ t^a_{l \beta}d_{\beta}(\v{z_2})b_{l}(\v{z_1}) \delta^{(2)}(\v{x}-\v{z_1})-t^a_{ \beta l}d_{\beta}(\v{z_2})b_{l}(\v{z_1})\delta^{(2)}(\v{x}-\v{z_2})\right]
\end{split}
\end{equation}
we obtain
\begin{equation}
\langle q\bar{q} | \hat \rho^{a}(\v{x_1}) \hat\rho^{b}(\v{x_2}) | q\bar{q}\rangle
=\frac{\delta^{ab}}{2}
\prod_{i=1,2}
\left( \delta^{(2)}(\v{x_i}-\v{z_1}) - \delta^{(2)}(\v{x_i}-\v{z_2}) \right)\,.
\end{equation}
For higher correlators, to simplify expressions, we account explicitly for the symmetry of the photon splitting function
$\v{z_1} \leftrightarrow \v{z_2}$ (and $z \leftrightarrow 1 -z$)\footnote{That is we
consider $|q\bar{q}\rangle = \frac12 \left[ b^{\dagger}_{\alpha \sigma}(\v{z_1},z)d^{\dagger}_{\alpha \sigma}(\v{z_2},1-z) +
b^{\dagger}_{\alpha \sigma}(\v{z_2},1-z)d^{\dagger}_{\alpha \sigma}(\v{z_1},z)\right]
|0\rangle$
}
to obtain
\begin{equation}
\begin{split}
& \langle q\bar{q} | \hat\rho^{a}(\v{x_1}) \hat\rho^{b}(\v{x_2}) \hat\rho^{c}(\v{x_3}) | q\bar{q}\rangle \\
=& \frac{if_{abc}}{4} \left( \delta^{(2)}(\v{x_2}-\v{z_1}) + \delta^{(2)}(\v{x_2}-\v{z_2}) \right) \prod_{i=1,3} \left( \delta^{(2)}(\v{x_i}-\v{z_1}) - \delta^{(2)}(\v{x_i}-\v{z_2}) \right)
\end{split}
\end{equation}
and
\begin{equation}
\begin{split}
& \langle q\bar q| \hat\rho^{a}(\v{x_1}) \hat\rho^{b}(\v{x_2}) \hat\rho^{c}(\v{x_3}) \hat\rho^{d}(\v{x_4})|q\bar q\rangle
= \Tr(t^at^bt^ct^d)\prod_{i=1}^4\left[ \delta^{(2)}(\v{x_i}-\v{z_1}) - \delta^{(2)}(\v{x_i}-\v{z_2}) \right]\\
-&\frac{f_{cde}f_{abe}}{8} \left[\delta^{(2)}(\v{x_2}-\v{z_1})+\delta^{(2)}(\v{x_2}-\v{z_2})\right] \left[\delta^{(2)}(\v{x_3}-\v{z_1}) + \delta^{(2)}(\v{x_3}-\v{z_2})\right]
\\ &\quad \quad \quad \times
\prod_{i=1,4}\left[ \delta^{(2)}(\v{x_i}-\v{z_1}) - \delta^{(2)}(\v{x_i}-\v{z_2}) \right]
\end{split}
\end{equation}
where $\Tr(t^a t^b t^c t^d)=\frac{1}{4N_c}\delta^{ab}\delta^{cd}+\frac{1}{8}\left(d^{abe}+if^{abe}\right)\left(d^{cde}+if^{cde}\right)$.
In the following we will need a representation of this trace in terms its fully symmetric and the remaining parts:
$\Tr(t^a t^b t^c t^d) = \Tr(t^a t^b t^c t^d)_{\rm sym}
+ \Tr(t^a t^b t^c t^d)_{\rm ns}$
with
\begin{equation}
\begin{split}
\Tr(t^a t^b t^c t^d)_{\rm sym}&=\frac{1}{12}\Big[\frac{1}{N_c}(\delta^{ab}\delta^{cd}+\delta^{ac}\delta^{bd}+\delta^{ad}\delta^{bc})+\frac{1}{2}(d_{abe}d_{cde}+d_{ace}d_{bde}+d_{ade}d_{bce}) \Big]
\end{split}
\end{equation}
and
\begin{equation}
\begin{split}
\Tr(t^a t^b t^c t^d)_{\rm ns}&=
\frac{1}{12} (f_{ade} f_{bce} - f_{abe} f_{cde})
+ \frac{1}{8} i (d_{abe} f_{cde} + d_{cde} f_{abe} )\,.
\end{split}
\end{equation}
For $N_c=3$, one can use the following identity (the factor of two was missing in Refs.~\cite{haber2016useful}; the validity of our identity can be easily checked by summing with respect to $a=b$)
\begin{align}
d_{abe}d_{cde}+d_{ace}d_{bde}+d_{ade}d_{bce}=\frac{2}{3}(\delta^{ab}\delta^{cd}+\delta^{ac}\delta^{bd}+\delta^{ad}\delta^{bc})
\end{align}
to obtain
\begin{equation}
\label{Eq:TrSym}
\begin{split}
\Tr(t^a t^b t^c t^d)^{N_c=3}_{\rm sym}
&=\frac{1}{18} (\delta^{ab}\delta^{cd}+\delta^{ac}\delta^{bd}+\delta^{ad}\delta^{bc})\,.
\end{split}
\end{equation}
We stress again that this expression is only true for $N_c=3$.
\subsection{The MV model for the projectile}
The other model we use to describe the photon wave function is the McLerran-Venugopalan model. It
posits that the averaging over the ``classical'' color charge density of the projectile should be performed with Gaussian weight, so that all correlators are expressed in terms of Wick contractions of the basic "propagator"
\begin{align}
\label{Eq:rho2mv1}
%
\langle \rho_a(\v x) \rho_b(\v y) \rangle_{\rm MV}
= \mu^2(x) \delta^{(2)}(\v x- \v y) \delta_{ab}\,.
\end{align}
Note that \eqref{Eq:rho2mv1} goes beyond the conventional MV model in that it includes the dependence of the color charge density on the impact parameter $\v{b} = \frac{\v{x}+\v{y}}{2}$. In principle one can go even further and account for the color neutrality by substituting a suitable function instead of $\delta^{(2)}(\v x- \v y)$, see e.g. Ref.~\cite{McLerran:2016ivs}. We will not do this in the current paper.
One has to be aware that the ``classical'' charge density in \eqref{Eq:rho2mv1} is not identical to the color charge density operator $\hat \rho$. To calculate the correlator of color charge density operators, one has to perform the operator symmetrization first, and use the MV for the completely symmetric parts. This was shown in \cite{PhysRevD.72.074023} and explained from the point of view of quantum-classical correspondence in \cite{Li:2020bys}.
For bilinears, the symmetrization is straightforward:
\begin{align}
\hat \rho_a(\v x) \hat \rho_b(\v y)
&= \left\{ \hat \rho_a(\v x), \hat \rho_b(\v y) \right\} + \left[ \hat \rho_a(\v x), \hat \rho_b(\v y) \right] \notag \\
&= \rho_a(\v x) \rho_b(\v y) - \frac{1}{2} \delta^{(2)}(\v x- \v y) T^{c}_{ab}
\rho_c (\v x)
\end{align}
where we have replaced the fully symmetric combinations by the classical color density $\left\{ \hat \rho_a(\v x), \hat \rho_b(\v y) \right\} \to \rho_a(\v x) \rho_b(\v y)$ and took into account that the commutator $\left[ \hat \rho_a(\v x), \hat \rho_b(\v y) \right] = - \frac{1}{2} \delta^{(2)}(\v x-\v y) T^{c}_{ab} \rho_c (\v x) $ as follows from the definition of $\hat \rho$. We thus see that
\begin{align}
\label{Eq:rho2mv}
%
\langle \hat \rho_a(\v x) \hat \rho_b(\v y) \rangle_{\rm MV}
= \langle \rho_a(\v x) \rho_b(\v y) \rangle_{\rm MV}
= \mu^2(x) \delta^{(2)}(\v x- \v y) \delta_{ab}\,.
\end{align}
This procedure can be extended to any number of operators $\hat \rho$.
For three operators we obtain:
\begin{align}
\notag
&\hat \rho_a(\v x) \hat \rho_b(\v y) \hat
\rho_c(\v z)
= \rho_a(\v x) \rho_b(\v y)
\rho_c(\v z) \\ &-
\frac12 \big( \DDelta{\v y}{\v z} T^e_{ b c} \rho_a(\v x) \rho_e (\v y) +
\DDelta{\v x}{\v z} T^e_{ a c} \rho_b (\v y) \rho_e(\v x) +
\DDelta{\v x}{\v y} T^e_{a b} \rho_e(\v x) \rho_c(\v z)\big)\,,
\end{align}
with $T^{a}_{bc} = - i f_{abc}$.
Here the terms linear in $\hat \rho$ were omitted as they do not contribute to the expectation value over the color invariant ensemble~\footnote{Note that this expression also reproduces the result of the previous section. Indeed, taking into account the symmetry of the dipole wave function we need to account only for the term quadratic in $\rho$
\begin{align}
\langle \rho_a(\v x) \rho_b(\v y) \rangle
= \langle \{ \rho_a(\v x) , \rho_b(\v y) \} \rangle
= \frac{\delta_{ab}}{ 2 } (\delta(\v x-\v {z_1})-\delta(\v{x}-\v{z_2}))(\delta(\v{y}-\v{z_1})-\delta(\v{y}-\v{z_2})) \,.
\notag
\end{align}
Therefore
\begin{align}
\langle \rho_a(\v x) \rho_b(\v y) \rho_c(\v z) \rangle
= &-
\frac1{4} T^{a}_{bc}
\Big[ \DDelta{\v y}{\v z} (\delta(\v x-\v{z_1})-\delta(\v{x}-\v{z_2}))(\delta(\v{y}-\v{z_1})-\delta(\v{y}-\v{z_2}))
\notag \\ & -
\DDelta{\v x}{\v z}
(\delta(\v {x} -\v{z_1})-\delta(\v{x}-\v{z_2}))(\delta(\v{y}-\v{z_1})-\delta(\v{y}-\v{z_2}))
\notag \\ & +
\DDelta{\v x}{\v y} (\delta(\v x- \v{z_1})-\delta(\v x-\v{z_2}))(\delta(\v{z}-\v{z_1})-\delta(\v{z}-\v{z_2})) \Big] \notag \\
& = -
\frac1 {4} T^{a}_{bc}
(\delta(\v{x}-\v{z_1})-\delta(\v{x}-\v{z_2}))
(\delta(\v{z}-\v{z_1})-\delta(\v{z}-\v{z_2}))
(\delta(\v{y}-\v{z_1})+\delta(\v{y}-\v{z_2}))\,.
\notag
\end{align}
}.
Thus in the MV model :
\begin{align}
\label{Eq:rho3mv}
\langle
\hat \rho_a(\v x) \hat \rho_b(\v y) \hat
\rho_c(\v z)
\rangle_{\rm MV}
&=
- \frac12 \DDelta{\v x}{\v y} \DDelta {\v y} {\v z} T^a_{bc} \, \mu^2 (\v x)
\end{align}
Similarly for four operators, we obtain
\begin{align}
\notag
&\langle
\hat \rho_a(\v x) \hat \rho_b(\v y) \hat
\rho_c(\v z) \hat\rho_d(\v u)
\rangle_{\rm MV}
\\ & =
\mu^2(\v x) \mu^2(\v z) \delta^{ab}\delta^{cd} \DDelta{\v x}{\v y} \DDelta{\v z}{\v u}
+ \mu^2(\v x) \mu^2(\v y) \delta^{ac}\delta^{bd} \DDelta{\v x}{\v z} \DDelta{\v y}{\v u}
\notag \\ &
+ \mu^2(\v x) \mu^2(\v z) \delta^{ad}\delta^{bc} \DDelta{\v x}{\v u} \DDelta{\v z}{\v y}
\notag \\ & +
\frac{\mu^2(x)}{6} \DDelta{\v x}{\v y} \DDelta {\v x} {\v z} \DDelta {\v x} {\v u} \left(
f_{a d e} f_{b c e} - f_{a b e} f_{c d e}
\right)
\label{Eq:rho4mv}
\end{align}
The last term here is of lower order in $\mu^2$ and is thus subleading for a projectile with high charge density. We do not neglect it here, but as will become clear later, it does play a somewhat different role than the first two terms even for a dilute projectile.
\subsection{Small $x$ gluon component of the projectile wave function}
\label{Sec:Coherent}
At the leading order, the dressing of the color charge by gluons of rapidity $\eta$ is given by the coherent operator:
\begin{align}
C_{\eta}=\exp{i\sqrt{2} \int d^2 \v x \hat{b}^i_a(\v x)\left[ a^{i\dagger}_a(\eta, \v x)+a^i_a(\eta, \v x) \right]}
\end{align}
where the creation/annihilation operators are introduced as the decomposition of the transverse
gluon field
\begin{equation}
\begin{split}
A^i_a(x^+,\v x)=\int_{0}^{\infty}\frac{dk^+}{2k^+(2\pi)}\left(a^i_a(k^+,\v x)e^{-ik^-x^+} +a^{i\dagger}_a(k^+,\v x)e^{ik^-x^+} \right)\,.
\end{split}
\end{equation}
The rapidity variable is defined here as
\begin{align}
d\eta=\frac{dk^+}{k^+}
\end{align}
and the corresponding
\begin{align}
a^i_a(\eta, \v x)={\sqrt{\frac{k^+}{4\pi}}} a^i_a(k^+, \v x).
\end{align}
The commutation relations for the creation/annihilation operators are
\begin{align}
\left[a^i_a(\eta, \v x), a^{j\dagger}_b(\eta', \v y)\right]=\delta^{ij}\delta^{ab} \delta^2(\eta-\eta') \delta(\v x - \v y)\,.
\end{align}
The Weizs\"acker-Williams field at the leading order is proportional to the color charge density operator
\begin{align}
\hat{b}^i_a(\v x)=\frac{g}{2\pi}\int d^2\v y \frac{(\v x-\v y)^i}{(\v x- \v y)^2} \hat{\rho}^a(\v y)\, .
\end{align}
The distribution (wave function) of $\hat \rho$ can be modelled by the previously discussed dipole or MV models.
Our calculation requires the knowledge of the incoming wave function at the rapidities of both observed gluons: $\eta$ and $\xi<\eta$. As we alluded to in the introduction, we restrict our calculation to the case when the difference between the rapidites of the observed gluons $\eta-\xi$ is large enough to use the eikonal vertex, that is $\eta-\xi \ll 1$, but small enough to neglect corrections due to evolution, that is $\eta-\xi \ll 1/\alpha_s$.
The gluon with rapidity $\eta$ itself contributes to the source for emission of gluons at lower rapidity. This can be accounted for by an appropriate redefinition of the Weizs\"acker-Williams field at lower rapidity $\hat{b}^i_a(\v x) \to \hat{b}^i_a(\v x) + \delta \hat{b}^i_a(\v x)$.
Thus at rapidity $\xi$ we have
\begin{align}
C_{\xi}=\exp{i\sqrt{2} \int d^2\v x \left(\hat{b}^i_a(\v x)+\delta \hat{b}^i_a(\v x)\right)\left[ a^{i\dagger}_a(\xi, \v x)+a^i_a(\xi, \v x) \right]}
\end{align}
where the extra contribution due to the presence of the gluon in the dipole wave function is given by
\begin{equation}
\begin{split}
\delta \hat{b}^i_a(\v x)&=\frac{g}{2\pi} \int_{\xi}^{\eta} d\zeta \int d^2y \frac{(\v x-\v y)^i}{(\v x- \v y)^2} \hat{\rho}_g^a(\zeta, \v y)
\end{split}
\end{equation}
where
\begin{align}
\hat{\rho}^a_g(\zeta, \v x)= a^{i\dagger}_b(\zeta, \v x) T^a_{bc} a_c(\zeta, \v x)\,.
\end{align}
Thus to the required order, the valence state dressed by the gluon cloud is
\begin{equation} |\Psi_{in}\rangle=C_{\xi}C_{\eta} |0\rangle |v\rangle
\end{equation}
\subsection{Scattering off the dense target}
\label{Sec:Scattering}
To describe a large nucleus, we use the classic field approximation for the gluon field of the target. In the eikonal approximation we have,
\begin{align}
\label{Eq:S_g}
\hat{S}^{\dagger}a^i_{a}(\zeta, \v{\bar{u}})\hat{S}&=U^{aa'}_{\v{\bar{u}}}a^i_{a'}(\zeta, \v{\bar{u}})
\end{align}
for gluon scattering, where $U$ is the infinite Wilson line in the adjoint representation:
\begin{equation}
U_{\v{x}}={\cal P}\exp{ig\int_{-\infty}^\infty dx^{+}T^aA_a^-(x^+,\v{x})}\,.
\end{equation}
Similarly for quark and antiquark
\begin{equation}
\begin{split}
\hat{S}^{\dagger}b_{\alpha}(\v{x})\hat{S}&=V_{\alpha \beta}(\v{x})b_{\beta}(\v{x})\\
\hat{S}^{\dagger}d_{\alpha}(\v{x})\hat{S}&=V^{\dagger}_{\beta \alpha }(\v{x})d_{\beta}(\v{x})
\end{split}
\end{equation}
where the Wilson line in the fundamental representation is given by
\begin{equation}
V(\v{x})={\cal P}\exp{ig\int_{-\infty}^\infty dx^{+}t^aA_a^-(x^+,\v{x})}\,.
\end{equation}
The consequence of these equations and the definition of the color charge density is that
\begin{align}
\hat{S}^{\dagger}\hat{\rho}_a(\v{\bar{u}})\hat{S}&=U^{aa'}_{\v{\bar{u}}}\hat{\rho}_{a'}(\v{\bar{u}})
\end{align}
and similarly
\begin{equation}
\hat{S}^{\dagger}\hat{\rho}_a(\zeta,\v{ \bar{u}})\hat{S}=a^{i\dagger}_{b}(\zeta, \v{\bar{u}}) a_{c}(\zeta, \v{\bar{u}})\left[U^{\dagger}_{\v{\bar{u}}} T^a U_{\v{\bar{u}}} \right]_{bc}
= U^{aa'}_{\v{\bar{u}}}
\hat{\rho}_{a'}(\zeta,\v{ \bar{u}})
\,.
\end{equation}
A calculation of a physical observable requires averaging over the ensemble of target fields which determines various correlators of the Wilson lines.
For a large nucleus at small x, the ensemble of the color fields of the target is frequently taken as an MV model with a relatively large value of the saturation momentum. This procedure in practice works well for calculation of a correlator of a small number of Wilson lines, but becomes increasingly complicated and cumbersome for more complicated correlators.
As we will see, the observable we work with leads to the expectations values of four Wilson lines in the adjoint representation. In the limit of large number of colors, this can be reduced to a product of up to two dipoles and a quadrupole in the fundamental representation in the coordinate space.
In order to get particle production cross section, this combination has to be Fourier-transformed with respect to four two-dimensional transverse momenta. Although, there are analytical expressions for the dipole and quadrupole in the McLerran-Venugopalan model, the Fourier transforms render the problem numerically prohibitively complex, as fast Fourier transform is not feasible in this multidimensional space while Monte-Carlo integration is extremely ineffective due to the presence of the oscillatory phases.
The two particle correlations at large $N_c$
only arise at order $1/N_c^2$~\cite{Kovner:2010xk,Kovchegov:2012nd,Kovchegov:2013ewa}. Partly this can be attributed to the nature of the correlations arising either from gluon HBT or gluon Bose-Einstein correlations~\cite{Altinoluk:2018ogz}. Therefore the large $N_c$ limit does not capture the physics we want to study.
Going beyond the large $N_c$ limit provides additional level of complication, as the cross section also involves sixtupoles and octupoles for which explicit analytical expressions, if exist, are too complex for a numerical simulation.
Instead we will use another approach applicable for a well evolved, near black disc, target. The approximation in question is the so-called {\it factorized dipole approximation} (FDA) discussed in Ref.~\cite{Kovner:2017ssr,Kovner:2018vec} and successfully applied to study particle production and correlations in Refs.~\cite{Altinoluk:2018ogz,Kovner:2018fxj,Kovner:2018azs,Li:2021ntt,Altinoluk:2020psk}.
The idea behind this approximation is extremely intuitive. Consider an arbitrary combination of even number of Wilson lines
$\langle U_{x_1} U_{x_2} \ldots U_{x_{2n}} \rangle$ which is multiplied by a non-trivial function of coordinates and integrated with respect to all two-dimensional vectors $x_i$, $i=1, \ldots, 2n$ over some finite, but large compared to the inverse saturation scale of the target squared, transverse area $S_\perp$. If a single Wilson line were to have a nonzero expectation value the largest contribution to the integral would be of order $S^{2n}_\perp$ arising from the integration over the part of the phase space where the coordinates of all Wilson lines are far away from each other. However for a dense target $\langle U_{ab} \rangle = 0$, and so the largest power of the transverse area possible is $S^{n}_\perp$, with corrections of order $S^{n-1}_\perp$.
The leading order contribution includes terms which contain smallest color singlets in the projectile propagating through the target. On one hand, any non singlet state that in the transverse plane separated by more than $1/Q_s$ from other propagating partons must have a vanishing S-matrix on the dense target due to color neutralization on the scale of $1/Q_s$, see e.g. \cite{Iancu:2001md,Ferreiro:2002kv,Iancu:2002xk,Mueller:2002pi}. On the other hand, if the singlet state contains more than two partons, one looses a power of the area when integrating over the coordinates of the partons. Thus the leading contribution in the dense target regime is the one where only dipole contribution to the S-matrix should be accounted for.
For illustration lets consider a four Wilson line observable
\begin{align}
&\int \prod_{i=1}^{4} d^2 \v x_i f(\v x_1, \ldots , \v x_4)
\langle {\rm tr} [U_{\v x_1} U_{\v x_2} U_{\v x_3} U_{\v x_4} ]\rangle
\notag \\ &=
\int \prod_{i=1}^{4} d^2 \v x_i f(\v x_1, \ldots ,\v x_4)\Big[
\langle U^{ab}_{\v x_1} U^{bc}_{\v x_2} \rangle
\langle U^{cd}_{\v x_3} U^{da}_{\v x_4} \rangle
\notag \\ & +
\langle U^{da}_{\v x_4} U^{ab}_{\v x_1}\rangle
\langle U^{bc}_{\v x_2} U^{cd}_{\v x_3}\rangle
+
\langle U^{ab}_{\v x_1} U^{cd}_{\v x_3} \rangle
\langle U^{bc}_{\v x_2} U^{da}_{\v x_4} \rangle
\Big]
+ {\cal O}(S_\perp) \notag
\end{align}
where the function $f$ is assumed to have support for all $x$ within the area $S_\perp$. The correlator $\langle U^{ab}_{\v x} U^{cd}_{\v y} \rangle$ is non zero only for a color singlet state, in this case -- the dipole:
\begin{align}
\label{Eq:KovnersHammer}
\langle U^{ab}_{\v x} U^{cd}_{\v y} \rangle = \frac{1}{N_c^2-1} \delta_{ac} \delta_{bd} D(|\v x-\v y|)\,,
\end{align}
where $D(r) = \frac{1}{N_c^2-1} {\rm tr} [ U^{\dagger}_{\v r} U_{0} ]$.
The quality of this approximation was explicitly checked for a number of Wilson line combinations in Ref.~\cite{Li:2021ntt}. The key necessary (albeit not sufficient) condition for its applicability is $Q^2_s S_{{\rm proj}, \perp} \gg 1$, where $Q_s$ is the saturation momentum of the nucleus and $S_{{\rm proj}, \perp}$ is the effective transverse area of the projectile. For LHC energies
this condition is satisfied if we use the natural $ S_{{\rm proj}, \perp} = \frac{1}{Q_{\rm eff}^2}$ where $Q_{\rm eff}$ is the larger between the QCD non-perturbative scale $\Lambda_{\rm QCD}$ and the virtuality of the photon. For the nearly real photon we thus have $S_{{\rm proj}, \perp} \sim \Lambda_{\rm QCD}^{-2}$. One has to be aware that when large momentum particle production is considered the effective area of the projectile might be significantly smaller than $\Lambda_{\rm QCD}^{-2}$. Although this is a process-dependent statement, in general it means that at large momenta $k\gg Q_s$, the FDA is not uniformly applicable; it however still can be justified for certain momentum configurations.
For calculations in this paper we apply the FDA to evaluate the correlators of more than
two Wilson lines. As our calculations are mostly in momentum space, the basic correlator used in the FDA is the Fourier transform of Eq.~\eqref{Eq:KovnersHammer}:
\begin{align}
\langle U^{ab}_{\v{p}} U^{cd}_{\v{q}} \rangle_T=\frac{(2\pi)^2}{N_c^2-1}\delta_{ac}\delta_{bd}\delta^{(2)}( \v{p}+ \v{q})D(\v{p})
\end{align}
where $D(\v{p})$ is
\begin{align}
D(\v{p})&=\frac{1}{(N^2_c-1) S_{\perp}}\langle \Tr( U^{\dagger}_{\v{p}}U_{\v{p}})\rangle_T\,.
\end{align}
We will also assume spatial isotropy of the target field ensemble, so that $D( \v{p}) = D(|\v p|)$.
We will use the MV model for the dipole. That is in the coordinate space, we have
\begin{align}
D(\v r) = \exp\left[-\frac{1}{4}Q_s^2r^2\ln(\frac{1}{\Lambda^2r^2}) \right]\,
\end{align}
with $D(|\v p|) = \int d^2 r e^{i \v p\cdot \v r} D(\v r)$.
\section{Double inclusive gluon production - general expressions}\label{sec3}
The differential cross section for double inclusive gluon production with rapidities and the transverse momenta $(\eta, \v{q}_1)$ and $(\xi, \v{q}_2)$ is given by
\begin{equation}
\label{Eq:Double}
\begin{split}
\frac{d{\cal N}}{d\eta d \v{q}_1^2 d\xi d \v{q}_2^2}=&\frac{1}{(2\pi)^4} \int
d^2 \v u_1 d^2\v u_2
d^2 \bar{\v u}_1 d^2 \bar{\v u}_2
e^{-i\v{q}_1(\v{u}_1-\v{\bar{u}}_1)}e^{-i\v{q}_2(\v{u}_2-\v{\bar{u}}_2)}
\\ &\times
\langle \gamma^*|C^{\dagger}\hat{S}^{\dagger}C a_{i,a}^{\dagger}(\eta, \v{u}_1)a_{j,b}^{\dagger}(\xi, \v{u}_2)
a_{i,a}(\eta, \v{\bar{u}}_1)a_{j,b}(\xi, \v{\bar{u}}_2)C^{\dagger}\hat{S}C| \gamma^* \rangle
\end{split}
\end{equation}
The coherent operator $C$ is given by the product of $C = C_\xi C_\eta$, see Sec.~\ref{Sec:Coherent}. For double-inclusive gluon production, it is sufficient to expand the coherent operator to the leading order in the argument:
\begin{equation}
\begin{split}
C_{\eta}\simeq& 1+i\sqrt{2} \int d^2\v v_1 \hat{b}^i_{a}(\v{v}_1)\left[ a^{i\dagger}_a(\eta, \v{v}_1)+a^i_a(\eta, \v{v}_1) \right]\,,\\
C_{\xi}\simeq& 1+i\sqrt{2} \int d^2\v v_2 \left(\hat{b}^j_{b}(\v{v}_2)+\delta \hat{b}^j_b(\eta, \v{v}_2)\right)\left[ a^{j\dagger}_b(\xi, \v{v}_2)+a^j_b(\xi, \v{v}_2)\right] \,.
\end{split}
\end{equation}
Hence to the required order, the coherent operator can be approximated by $C_{\rm eff}$
\begin{equation}
\begin{split}
C_{\rm eff} =&1+i\sqrt{2} \int d^2\v v_1 \hat{b}^i_{a}(\v v_1)\left[ a^{i\dagger}_a(\eta, \v{v}_1)+a^i_a(\eta, \v{v}_1) \right]\\
&+i\sqrt{2} \int d^2\v v_2 \left(\hat{b}^j_{b}(\v{v}_2)+\delta \hat{b}^j_b(\eta, \v{v}_2)\right)\left[ a^{j\dagger}_b(\xi, \v{v}_2)+a^j_b(\xi, \v{v}_2)\right] \\
&-2 \int d^2\v v_1d^2\v v_2 \left(\hat{b}^j_{b}(\v{v}_2)+\delta \hat{b}^j_b(\eta, \v{v}_2)\right)\left[ a^{j\dagger}_b(\xi, \v{v}_2)+a^j_b(\xi, \v{v}_2)\right] \notag \\ &\quad \quad \quad \quad \quad \quad \times \hat{b}^i_{a}(\v{v}_1)\left[ a^{i\dagger}_a(\eta, \v{v}_1)+a^i_a(\eta, \v{v}_1) \right]\,.
\end{split}
\end{equation}
The matrix element in Eq.~\eqref{Eq:Double}, can be decomposed into three different contributions classified by the number of gluons emitted directly from the photon state $\mathcal{M}=\Sigma_2+\Sigma_3+\Sigma_4$, or equivalently, the powers of color charge density, schematically,
\begin{align}
b^2(b+\delta b)^2=b^4+2b^3 \delta b+ b^2(\delta b)^2\,.
\end{align}
Here every factor of $b$ corresponds to a gluon emitted directly from the photon, while a factor of $\delta b$ corresponds to a gluon emitted from another gluon with larger rapidity.
We will denote the corresponding contribution to the matrix element by $\Sigma_n$, where $n$ is the number of gluons emitted directly from the projectile. To elaborate on this: $\Sigma_2$ contributes to the probability of the process where a gluon with rapidity $\eta$ is emitted from the photon, and subsequently this gluon splits into a pair of gluons which are then observed in the final state. The splitting can occur either before of after the scattering on the target field. $\Sigma_4$ contributes to the probability of the process where both gluons are emitted directly from the photon state. Finally $\Sigma_3$ represents the interference of the previous two contributions.
We now turn to evaluating the production amplitude.
\begin{figure}[t!]
\centering
\includegraphics[width=0.4\textwidth]{Sigma2.pdf} \quad \quad
\includegraphics[width=0.4\textwidth]{Sigma4.pdf}
\caption{An illustration of two types of contribution to the two gluon production amplitude. Note that the number of actual diagrams is significantly larger, as the position of the shock wave can vary. For example, in total, there are six diagrams for the topology of the left diagram. }
\end{figure}
\subsection{Production amplitude}
Our goal is to evaluate the production amplitude in the coordinate space defined by
\begin{align}
{\cal A}(\bar{\v u}_1,\bar{\v u}_2) =
\hat S^\dagger a_{i,a}(\eta, \v{\bar{u}}_1)a_{j,b}(\xi, \v{\bar{u}}_2)C^{\dagger}\hat{S}C| \gamma^* \rangle\,.
\end{align}
We have introduced the factor $\hat S^\dagger$ for convenience, although it cancels in the calculation in the matrix element ${\cal M} = {\cal A}^\dagger {\cal A}$.
Using Eq.~\eqref{Eq:S_g}, we arrive at
\begin{align}
{\cal A}(\bar{\v u}_1,\bar{\v u}_2) =
U^{aa'}_{\v{\bar{u}}_1}U^{bb'}_{\v{\bar{u}}_2}
a_{ i,a}(\eta, \v{\bar{u}}_1)a_{j,b}(\xi, \v{\bar{u}}_2)
\hat S^\dagger C^{\dagger}\hat{S}C| \gamma^* \rangle \equiv {\cal A}_1+{\cal A}_2\,.
\end{align}
The operator $\hat S^\dagger C^{\dagger}\hat{S}C$ has terms proportional to $ b$ and to $b^2$, and these terms have been separated in the previous equation into the two amplitudes ${\cal A}_1$ and ${\cal A}_2$.
These amplitudes are directly related to the aforementioned $\Sigma_n$, $\Sigma_2 = {\cal A}_1{\cal A}^*_1$, $\Sigma_3 = {\cal A}_1{\cal A}^*_2+{\cal A}^*_1{\cal A}_2$ and $\Sigma_4 = {\cal A}_2{\cal A}^*_2$.
\subsubsection{Amplitude ${\cal A}_1$}
Expanding the coherent operators and collecting the terms of first order in $b$, we obtain
the production amplitude for two gluons with colors $a,b$, polarizations $i,j$ at positions $\v{\bar{u}}_1$ and $\v{\bar{u}}_2$:
\begin{equation}
\begin{split}
&{\cal A}_1(\bar{\v u}_1,\bar{\v u}_2)\\
&= 2\int d^2\v v_1d^2\v v_2
U^{aa'}_{\v{\bar{u}}_1} U^{bb'}_{\v{\bar{u}}_2}a^i_{a'}(\eta, \v{\bar{u}}_1)
a^j_{b'}(\xi, \v{\bar{u}}_2)
\left[ \hat S^{\dagger} \delta \hat{b}^k_c(\eta, \v{v}_2)a^{k\dagger}_c(\xi, \v{v}_2)\hat S\hat{b}^l_{d}(\v{v}_1)a^{l\dagger}_d(\eta, \v{v}_1) \right.\\
&\left.- \hat S^{\dagger} \hat{b}^l_{d}(\v{v}_1)a^{l\dagger}_d(\eta, \v{v}_1)\delta \hat{b}^k_c(\eta, \v{v}_2)a^{k\dagger}_c(\xi, \v{v}_2)\hat S- \delta \hat{b}^k_c(\eta, \v{v}_2)a^{k\dagger}_c(\xi, \v{v}_2)\hat{b}^l_{d}(\v{v}_1)a^{l\dagger}_d(\eta, \v{v}_1) \right]|\gamma^*\rangle
\end{split}\label{Eq:am1}
\end{equation}
The second term in this expression gives zero when the operator $\delta \hat{b}^k_c(\eta, \v{v}_2)$ acts on the gluon vacuum at rapidity $\eta$ (we always assume that $\xi < \eta$).
We thus have two non-trivial terms which can be readily evaluated. The first one gives
\begin{equation}
\begin{split}
&U^{aa'}_{\v{\bar{u}}_1}U^{bb'}_{\v{\bar{u}}_2}\int d^2\v v_1d^2\v v_2 a^i_{a'}(\eta, \v{\bar{u}}_1)a^j_{b'}(\xi, \v{\bar{u}}_2) \hat S^{\dagger} \delta \hat{b}^k_c(\eta, \v{v}_2)a^{k\dagger}_c(\xi, \v{v}_2)\hat S\hat{b}^l_{d}(\v{v}_1)a^{l\dagger}_d(\eta, \v{v}_1)|\gamma^*\rangle\\
&= \int d^2 \v x f^j({\bar{\v u}}_2-{\bar{\v u}}_1)f^i(\v{\bar{u}}_1-\v{x})\left[ T^b U_{\v{\bar{u}}_1} \right]_{af}\hat\rho_f(\v{x})|\gamma^*\rangle
\end{split}
\end{equation}
where
\begin{align}
f^i(\v{x}) = \frac{ g } { 2\pi} \frac{ \v x_i} {\v x^2} \,.
\end{align}
The last term in Eq.~\eqref{Eq:am1} simplifies into
\begin{equation}
\begin{split}
&U^{aa'}_{\v{\bar{u}}_1}U^{bb'}_{\v{\bar{u}}_2}\int d^2\v v_1d^2\v v_2 a^i_{a'}(\eta, \v{\bar{u}}_1)a^j_{b'}(\xi, \v{\bar{u}}_2) \delta \hat{b}^k_c(\eta, \v{v}_2)a^{k\dagger}_c(\xi, \v{v}_2)\hat{b}^l_{d}(\v{v}_1)a^{l\dagger}_d(\eta, \v{v}_1)|\gamma^*\rangle\\
=&\int d^2 \v x U^{aa'}_{\v{\bar{u}}_1}U^{bb'}_{\v{\bar{u}}_2}T^{a'}_{db'} f^j(\v{\bar{u}}_2-\v{\bar{u}}_1)f^i(\v{\bar{u}}_1-\v{x})\hat\rho_d(\v{x})|\gamma^*\rangle\,.
\end{split}
\end{equation}
Combining these two contributions we complete ${\cal A}_1$
\begin{equation}
\begin{split}
{\cal A}_1(\bar{\v u}_1,\bar{\v u}_2)=&2\int d^2 \v x f^j(\v{\bar{u}}_2-\v{\bar{u}}_1)f^i(\v{\bar{u}}_1-\v{x}) (U_{\v{\bar{u}}_2}-U_{\v{\bar{u}}_1})^{bb'}[U^{\dagger}_{\v{\bar{u}}_1}T^{a}U_{\v{\bar{u}}_1}]_{b'd} \hat\rho_d(\v{x})
\end{split}
\end{equation}
where to obtain the final expression we used the identity
\begin{align}
U^{aa'}_{\v{\bar{u}}_1}T^{a'}_{db'}= [U^{\dagger}_{\v{\bar{u}}_1}T^{a}U_{\v{\bar{u}}_1}]_{db'} \,.
\end{align}
\subsubsection{Amplitude ${\cal A}_2$}
Following similar calculations, we obtain:
\begin{equation}
\begin{split}
&{\cal A}_2(\bar{\v u}_1,\bar{\v u}_2)\\
&=2\int d^2\bm{x}d^2\bm{y}f^i(\bar{\v u}_1-\bm{x})f^j(\bar{\v u}_2-\bm{y})\left[U^{ae}_{\bm{x}}U^{bc}_{\bar{\v u}_2}\hat\rho_e(\bm{x})\hat\rho_{c}(\bm{y})+U^{bc}_{\bm{y}}U^{ae}_{\bar{\v u}_1}\hat\rho_{c}(\bm{y})\hat\rho_{e}(\bm{x})\right.\\
&\left.-U^{ae}_{\bm{x}}U^{bc}_{\bm{y}}\hat\rho_e(\bm{x})\hat\rho_c(\bm{y})-U^{ae}_{\bar{\v u}_1}U^{bc}_{\bar{\v u}_2}\hat{\rho}_{c}(\bm{y})\hat{\rho}_{e}(\bm{x})\right]
\\&= 2\int d^2\bm{x}d^2\bm{y}f^i(\bar{\v u}_1-\bm{x})f^j(\bar{\v u}_2-\bm{y})
\\ &\times \left[U^{ae}_{\bm{x}} \left( U^{bc}_{\bar{\v u}_2}
- U^{bc}_{\bm{y}}
\right)
\hat\rho_e(\bm{x})\hat\rho_{c}(\bm{y})+
U^{ae}_{\bar{\v u}_1}
\left(U^{bc}_{\bm{y}}-U^{bc}_{\bar{\v u}_2}\right)\hat\rho_{c}(\bm{y})\hat\rho_{e}(\bm{x})\right]\,.
\end{split}
\end{equation}
If we were to forget about the quantum nature of the color charge density operators (neglect their commutation relations) the later expression would be identical to the frequently used expression for two gluon production in the dilute-dense limit, see e.g. Appendix A of Ref.~\cite{Kovner:2012jm}.
\subsection{Matrix elements $\Sigma$}
We now compute the matrix elements.
The matrix element proportional to the second power of the projectile charge density is:
\begin{equation}
\begin{split}
&\Sigma_2(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )= {\cal A}_{1'}^*(\v u_1,\v u_2){\cal A}_1(\bar{\v u}_1,\bar{\v u}_2) \\
&=4\int d^2\bm{x}\int d^2\bar{\bm{x}}f^i(\bar{\v u}_1-\bm{x})f^i(\v u_1-\bar{\bm{x}})f^j(\bar{\v u}_2-\bar{\v u}_1)f^j(\v u_2-\v u_1)\left<\hat \rho_{d'}(\bar{\bm{x}})\hat \rho_{d}(\bm{x})\right>_{P}\\
&\left<\left[ [U^{\dagger}_{\v u_1}T^{a}U_{\ vu_1}][U^{\dagger}_{\v u_2}-U^{\dagger}_{\v u_1}] [U_{\bar{\v u}_2}-U_{\bar{\v u}_1}][U^{\dagger}_{\bar{\v u}_1}T^{a}U_{\bar{\v u}_1}]\right]_{d'd} \right>_{T}\\
&=4\int d^2\bm{x}\int d^2\bar{\bm{x}}f^i(\bar{\v u}_1-\bm{x})f^i(\v u_1-\bar{\bm{x}})f^j(\bar{\v u}_2-\bar{\v u}_1)f^j(\v u_2-\v u_1)\left<\hat \rho_{d'}(\bar{\bm{x}})\hat \rho_{d}(\bm{x})\right>_{P}\\
&\left<\Tr(U_{\v u_1}T^{e}[U^{\dagger}_{\v u_2}-U^{\dagger}_{\v u_1}][U_{\bar{\v u}_2}-U_{\bar{\v u}_1}]T^eU^{\dagger}_{\bar{\v u}_1})\right>_T
\end{split}
\end{equation}
The interference terms involving the cubic power of the projectile charge density:
\begin{equation}
\begin{split}
&\Sigma_3(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )= {\cal A}_{2'}^*(\v u_1,\v u_2){\cal A}_1(\bar{\v u}_1,\bar{\v u}_2)+{\cal A}_{1'}^*(\v u_1,\v u_2){\cal A}_{2}(\bar{\v u}_1,\bar{\v u}_2) \\
&=4 \int d^2\bar{\bm{x}}d^2\bar{\bm{y}} d^2\bm{x}f^j(\bar{\v u}_2-\bar{\v u}_1)f^i(\bar{\v u}_1-\bm{x})f^i(\v u_1-\bar{\bm{x}})f^j(\v u_2-\bar{\bm{y}})\\
& \times \left(\left<\hat{\rho}_{c'}(\bar{\bm{y}})\hat{\rho}_{e'}(\bar{\bm{x}}) \hat{\rho}_{d}(\bm{x})\right>_{P} \right.\\
& \times \left< U^{\dagger e'a}(\bar{\bm{x}}) \left[ [U^{\dagger}_{\v u_2}-U^{\dagger}_{\bar{\bm{y}}}] [U_{\bar{\v u}_2}-U_{\bar{\v u}_1}][U^{\dagger}_{\bar{\v u}_1}T^{a}U_{\bar{\v u}_1}]\right]_{c'd} \right>_T\\
&\left. + \left<\hat{\rho}_{e'}(\bar{\bm{x}})\hat{\rho}_{c'}(\bar{\bm{y}}) \hat{\rho}_{d}(\bm{x})\right>_{P} \right.\\
& \left. \times \left< U^{\dagger e'a}(\v u_1) \left[ [U^{\dagger}_{\bar{\bm{y}}}-U^{\dagger}_{\v u_2}] [U_{\bar{\v u}_2}-U_{\bar{\v u}_1}][U^{\dagger}_{\bar{\v u}_1}T^{a}U_{\bar{\v u}_1}]\right]_{c'd} \right>_T \right) \\
&+4 \int d^2\bar{\bm{x}}d^2\bm{y} d^2\bm{x}f^j(\v u_2-\v u_1)f^i(\v u_1-\bar{\bm{x}})f^i(\bar{\v u}_1-\bm{x})f^j(\bar{\v u}_2-\bm{y})\\
& \times \left(\left< \hat{\rho}_{d'}(\bar{\bm{x}})\hat{\rho}_{e}(\bm{x})\hat{\rho}_{c}(\bm{y})\right>_{P} \right.\\
& \times \left< \left[ [U^{\dagger}_{\v u_1}T^{a}U_{\v u_1}] [U^{\dagger}_{\v u_2}-U^{\dagger}_{\v u_1}][U_{\bar{\v u}_2}-U_{\bm{y}}]\right]_{d'c} U^{ ae}(\bm{x}) \right>_T\\
&\left. + \left< \hat{\rho}_{d'}(\bar{\bm{x}})\hat{\rho}_{c}(\bm{y})\hat{\rho}_{e}(\bm{x})\right>_{P} \right.\\
& \left. \times \left< \left[ [U^{\dagger}_{\v u_1}T^{a}U_{\v u_1}] [U^{\dagger}_{\v u_2}-U^{\dagger}_{\v u_1}][U_{\bm{y}}-U_{\bar{\v u}_2}]\right]_{d'c} U^{ ae}(\bar{\v u}_1) \right>_T \right)
\end{split}
\end{equation}
Finally the term with the highest power of the projectile density is
\begin{equation}
\begin{split}
&\quad\Sigma_4(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )\\&= {\cal A}_{2'}^*(\v u_1,\v u_2){\cal A}_2(\bar{\v u}_1,\bar{\v u}_2) \\
&=4\int d^2\bm{x}d^2\bm{y}\int d^2\bar{\bm{x}}d^2\bar{\bm{y}}f^i(\bar{\v u}_1-\bm{x})f^j(\bar{\v u}_2-\bm{y})f^i(\v u_1-\bar{\bm{x}})f^j(\v u_2-\bar{\bm{y}})\\
& \times \left(\left<\hat{\rho}_{c'}(\bar{\bm{y}})\hat{\rho}_{e'}(\bar{\bm{x}}) \hat{\rho}_{e}(\bm{x})\hat{\rho}_{c}(\bm{y})\right>_{P} \left< [U^{\dagger}_{\bar{\bm{x}}}U_{\bm{x}}]_{e'e} \left[ [U^{\dagger}_{\v u_2}-U^{\dagger}_{\bar{\bm{y}}}][U_{\bar{\v u}_2}-U_{\bm{y}}]\right]_{c'c} \right>_T \right.
\\
&\left.+\left<\hat{\rho}_{e'}(\bar{\bm{x}})\hat{\rho}_{c'}(\bar{\bm{y}}) \hat{\rho}_{e}(\bm{x})\hat{\rho}_{c}(\bm{y})\right>_{P} \left< [U^{\dagger}_{\v u_1}U_{\bm{x}}]_{e'e} \left[ [U^{\dagger}_{\bar{\bm{y}}}-U^{\dagger}_{\v u_2}][U_{\bar{\v u}_2}-U_{\bm{y}}]\right]_{c'c} \right>_T \right.
\\
&\left.+\left<\hat{\rho}_{e'}(\bar{\bm{x}})\hat{\rho}_{c'}(\bar{\bm{y}}) \hat{\rho}_{c}(\bm{y})\hat{\rho}_{e}(\bm{x})\right>_{P} \left< [U^{\dagger}_{\v u_1}U_{\bar{\v u}_1}]_{e'e} \left[ [U^{\dagger}_{\bar{\bm{y}}}-U^{\dagger}_{\v u_2}][U_{\bm{y}}-U_{\bar{\v u}_2}]\right]_{c'c} \right>_T \right.\\
&\left.+\left<\hat{\rho}_{c'}(\bar{\bm{y}})\hat{\rho}_{e'}(\bar{\bm{x}}) \hat{\rho}_{c}(\bm{y})\hat{\rho}_{e}(\bm{x})\right>_{P} \left< [U^{\dagger}_{\bar{\bm{x}}}U_{\bar{\v u}_1}]_{e'e} \left[ [U^{\dagger}_{\v u_2}-U^{\dagger}_{\bar{\bm{y}}}][U_{\bm{y}}-U_{\bar{\v u}_2}]\right]_{c'c} \right>_T \right)
\end{split}
\end{equation}
\section{Inclusive two gluon production: the dipole model}\label{sec4}
\label{Sect:2glDPM}
The two particle inclusive cross section in terms of the matrix element $\Sigma = \Sigma_2 + \Sigma_3 + \Sigma_4$ computed previously, reads
\begin{equation}
\begin{split}
\frac{dN}{d\eta d \v{q}_1^2 d\xi d \v{q}_2^2}&=\frac{1}{4} \sum_{s_1,s_2,\lambda}\int_{0}^{1}\frac{dz}{2\pi z(1-z)} \int_{\v z_1,\v z_2} \Psi^{T*}_{\lambda}(z,\v z_1-\v z_2,s_1)\Psi^{T}_{\lambda}(z,\v z_1-\v z_2,s_1)\\ & \times \frac{1}{(2\pi)^4} \int_{\v u_1, \bar{\v u}_1,\v u_2, \bar{\v u}_2} e^{-i \v{q}_1(\v u_1-\bar{\v u}_1)}e^{-i \v{q}_2(\v u_2-\bar{\v u}_2)}\Sigma(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )\,.
\end{split}
\end{equation}
Summing over the quark and (transverse) photon polarizations we obtain
\begin{align}
\sum_{s_1,s_2,\lambda}|\Psi^{T}_{\lambda}(z,\v z_1-\v z_2,s_1)|^2
=&\frac{32e^2e_f^2}{(2\pi)^2}(z^2+ {\bar z}^2)z \bar z \varepsilon^2_fK_1(\varepsilon_f|\v r|)^2
\end{align}
where we introduced $\bar z=1-z$ to simplify notation.
In principle $\varepsilon_f$ is a function of $z$; however for small $\varepsilon_f \propto Q$, the combination $\varepsilon_f^2 K_1^2(\varepsilon_f r)$ is approximately $\varepsilon_f$ and therefore $z$ independent. We thus set $z$ to $1/2$ in $\varepsilon_f^2 K_1^2(\varepsilon_f r)$ and integrate the rest of $z$ dependence explicitly:
\begin{align}
\int^1_0 dz \, (z^2+{\bar z}^2)=\frac{2}{3}\,.
\end{align}
Hence the cross section reads
\begin{equation}
\begin{split}
\frac{dN}{d\eta d \v{q}_1^2 d\xi d \v{q}_2^2}&=\frac{16e^2e_f^2}{3 (2\pi)^3}\int_{\bm{z}_1,\bm{z}_2}\varepsilon^2_fK_1(\varepsilon_f|\bm{z}_1-\bm{z}_2|)^2 \\
&\times \frac{1}{(2\pi)^4} \int_{u_1, \bar{u}_1,u_2, \bar{u}_2} e^{-i \v{q}_1(\v u_1-\bar{\v u}_1)}e^{-i \v{q}_2(\v u_2-\bar{\v u}_2)}\Sigma(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )
\end{split}
\end{equation}
where $\varepsilon_f = Q/2$.
Numerical calculations are easier performed in momentum space, therefore in the following we present all expressions in terms of momentum integrals.
It is convenient to split the production cross section into three terms corresponding to $\Sigma_2$, $\Sigma_3$ and $\Sigma_4$.
Using the correlators of the color charge density in the dipole, calculated in Section 2, we obtain for the contribution of $\Sigma_2$:
\begin{equation}
\label{Eq:N2}
\begin{split}
\frac{dN^{(2)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}&=\frac{16 e^2e_f^2}{3 \pi^4} \alpha_s^2\int_{P, \v{k} ,\v{\bar{k}} } \frac{\mathcal{I}(\varepsilon_f,P)}{P^2}\, \Gamma(\v{q}_2,\v k,\v{\bar{k}})
\\
&\Tr(U_{\v{q}_1+\v{k}-\v{P}}T^{e}U^{\dagger}_{\v{k}-\v{q}_2}U_{\v{\bar{k}}-\v{q}_2}T^eU^{\dagger}_{\v{q}_1+\v{\bar{k}} -\v{P}})\,.
\end{split}
\end{equation}
In Eq.~\eqref{Eq:N2} we introduced
\begin{align}
\mathcal{I}_{\varepsilon_f}(|\v P|) &=
\int_0^{\infty} dr r \epsilon_f^2 K^2_1(\epsilon_f |\v r|)(1-J_0(|\v r||\v P|))\notag \\ &=
-\frac{1}{2}+\frac{(\v P^2+2\epsilon_f^2)\ln(\frac{|\v P|}{2\epsilon_f}+\sqrt{1+\frac{\v P^2}{4 \epsilon_f^2}}) }{|\v P|\sqrt{\v P^2+4\epsilon_f^2}} ,
\\
\Gamma (\v{q}_2,\v{k},\v{\bar{k}}) &= \left(\frac{\v q^i_2}{\v{q}_2^2}-\frac{ \v k^i}{ \v k^2}\right) \left(\frac{\v q^i_2}{\v{q}_2^2}-\frac{\v{\bar{k}}^i}{\v{\bar{k}}^2}\right)\,.
\end{align}
Note that the ratio $\mathcal{I}(\varepsilon_f,P) / P^2$ as a function of $P^2$ approaches a finite constant in the IR($P/\varepsilon_f\to0$).
The contributions involving $\Sigma_3$ cannot be simplified into one homogeneous expression.
Instead we have
\begin{equation}
\frac{dN^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}=\frac{dN_1^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}+\frac{dN_2^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}+\frac{dN_3^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}+\frac{dN_4^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}
\end{equation}
We list all four terms separately.
\begin{equation}
\begin{split}
\frac{dN_1^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}
=&
- \frac{8 e^2e_f^2}{3 \pi^4} \alpha_s^2
\int_{\v{k}_1,\v{k}_2,\v{k}_3} \left\langle \Tr \left[T^d U^{\dagger}_{\v{k}_1}U_{\v{k}_2}T^dU^{\dagger}_{\v{k}_3}U_{\v{k}_1-\v{k}_2+\v{k}_3}\right]\right\rangle_T \\
&\times\frac{\v{q}_1 \cdot (\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_3)}{\v{q}_1^2(\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_3)^2}\Gamma(\v{q}_2,\v{q}_2+\v{k}_1,\v{q}_2+\v{k}_2)\\
&\times\Big[\mathcal{I}_{\varepsilon_f}( |\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_3|)+\mathcal{I}_{\varepsilon_f}( |\v{q}_2+\v{k}_1|)-\mathcal{I}_{\varepsilon_f}( |\v{q}_1-\v{k}_1+\v{k}_2-\v{k}_3|)\Big]\,.
\end{split}
\end{equation}
Note that this expression does not have poles at $\v{q}_2+\v{k}=0$ and $\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_3=0$ as the combination of $\mathcal{I}$ functions vanishes at these points.
Similarly for the remaining three contributions we have
\begin{equation}
\begin{split}
\frac{dN_2^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}&=-
\frac{8 e^2e_f^2}{3 \pi^4} \alpha_s^2
\int_{\v{k}_1,\v{k}_2,\v{k}_3}\left\langle \Tr \left[T^d U^{\dagger}_{\v{k}_1}U_{\v{k}_2}T^dU^{\dagger}_{\v{k}_3}U_{\v{k}_1-\v{k}_2+\v{k}_3} \right]\right\rangle_T \\
&\times\frac{(\v{q}_1-\v{k}_1+\v{k}_2-\v{k}_3) \cdot
(\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_3)}{(\v{q}_1-\v{k}_1+\v{k}_2-\v{k}_3)^2(\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_3)^2} \Gamma(\v{q}_2,\v{q}_2+\v{k}_1,\v{q}_2+\v{k}_2) \\
&\times \Big[\mathcal{I}_{\varepsilon_f}( |\v{q}_1-\v{k}_1+\v{k}_2-\v{k}_3|)+\mathcal{I}_{\varepsilon_f}( |\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_3|)-\mathcal{I}_{\varepsilon_f}( |\v{q}_2+\v{k}_1|)\Big]\,,
\end{split}
\end{equation}
\begin{equation}
\begin{split}
\frac{dN_3^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}&=
- \frac{8 e^2e_f^2}{3 \pi^4} \alpha_s^2
\int_{\v{k}_1,\v{k}_2,\v{k}_3}
\left \langle \Tr \left [T^{d} U^{\dagger}_{\v{k}_1}U_{\v{k}_2}T^{d}U^{\dagger}_{\v{k}_3}U_{\v{k}_1-\v{k}_2+\v{k}_3} \right] \right\rangle_T \\
&\times\frac{\v{q}_1\cdot (\v{q}_1+\v{q}_2-\v{k}_2+\v{k}_3)}{\v{q}_1^2(\v{q}_1+\v{q}_2-\v{k}_2+\v{k}_3)^2}\Gamma(\v{q}_2,\v{q}_2+\v{k}_3,\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3)\\
&\times \Big[\mathcal{I}_{\varepsilon_f}( |\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)+\mathcal{I}_{\varepsilon_f}( |\v{q}_1+\v{q}_2-\v{k}_2+\v{k}_3|)-\mathcal{I}_{\varepsilon_f}( |\v{q}_1-\v{k}_1|)\Big]\,,
\end{split}
\end{equation}
and, finally,
\begin{equation}
\begin{split}
\frac{dN_4^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}&=
- \frac{8 e^2e_f^2}{3 \pi^4} \alpha_s^2
\int_{\v{k}_1,\v{k}_2,\v{k}_3}\left\langle \Tr \left[T^d U^{\dagger}_{\v{k}_1}U_{\v{k}_2}T^{d}U^{\dagger}_{\v{k}_3}U_{\v{k}_1-\v{k}_2+\v{k}_3}\right]\right\rangle_T \\
&\times\frac{(\v{q}_1-\v{k}_1)\cdot (\v{q}_1+\v{q}_2-\v{k}_2+\v{k}_3)}{(\v{q}_1-\v{k}_1)^2(\v{q}_1+\v{q}_2-\v{k}_2+\v{k}_3)^2}\Gamma(\v{q}_2,\v{q}_2+\v{k}_3,\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3)\\
&\times \Big[\mathcal{I}_{\varepsilon_f}( |\v{q}_1-\v{k}_1|)+\mathcal{I}_{\varepsilon_f}( |\v{q}_1+\v{q}_2-\v{k}_2+\v{k}_3|)-\mathcal{I}_{\varepsilon_f}( |\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)\Big]\,.
\end{split}
\end{equation}
Switching to the contribution from $\Sigma_4$ involving four $\rho(x)$ correlators, we again obtain four terms that cannot be obviously combined:
\begin{equation}
\frac{dN^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}=\frac{dN_1^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}+\frac{dN_2^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}+\frac{dN_3^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}+\frac{dN_4^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}
\end{equation}
The explicit expressions are:
\begin{equation}
\begin{split}
&\frac{dN_1^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}\\&= \frac{32 e^2e_f^2}{3 \pi^4} \alpha_s^2 \int_{\v{k}_1,\v{k}_2,\v{k}_3}\frac{1}{\v{q}_1^2}\Gamma(\v{q}_2,\v{q}_2+\v{k}_3,\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3)
\\
&\times[U^{\dagger}_{\v{k}_1}U_{\v{k}_2}]_{e'e} \left[U^{\dagger}_{\v{k}_3}U_{\v{k}_1+\v{k}_3-\v{k}_2}\right]_{c'c}
\\
&\times\Big[\Tr(t^{c'}t^{e'}t^{e}t^{c})\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)
\\
+&\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}_1+\v{k}_3|)\Big)
\\
+&\frac{T_{ec}^aT_{c'e'}^a}{8}
\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)
\\
-&\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}_1+\v{k}_3|)\Big)
\Big]
\end{split}
\end{equation}
One in principle can substitute an explicit expression for the trace of four Gell-Mann matrices, in practice however, this does not help to simplify the expression.
\begin{equation}
\begin{split}
&\frac{dN_2^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}\\=&-\frac{32 e^2e_f^2}{3 \pi^4} \alpha_s^2 \int_{\v{k}_1,\v{k}_2,\v{k}_3}\frac{\v{q}_1 \cdot(\v{q}_1+\v{k}_1)}{\v{q}_1^2(\v{q}_1+\v{k}_1)^2}\Gamma(\v{q}_2,\v{q}_2+\v{k}_3,\v{q}_2+\v{k}_1+\v{k}_3-\v{k}_2)
\\
&\times[U^{\dagger}_{\v{k}_1}U_{\v{k}_2}]_{e'e} \left[U^{\dagger}_{\v{k}_3}U_{\v{k}_1+\v{k}_3-\v{k}_2}\right]_{c'c}
\\
&\times\Big[\Tr(t^{e'}t^{c'}t^{e}t^{c})\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)
\\
+&\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2 + \v{k}_1+\v{k}_3|)\Big)
\\
+&\frac{T_{ec}^aT_{e'c'}^a}{8}
\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)
\\
-&\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)+\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2 + \v{k}_1+\v{k}_3|)\Big)\Big]
\end{split}
\end{equation}
\begin{equation}
\begin{split}
&\frac{dN_3^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}\\=&
\frac{32 e^2e_f^2}{3 \pi^4} \alpha_s^2
\int_{\v{k}_1,\v{k}_2,\v{k}_3}\frac{(\v{q}_1+\v{k}_2) \cdot(\v{q}_1+\v{k}_1)}{(\v{q}_1+\v{k}_2)^2(\v{q}_1+\v{k}_1)^2}\Gamma(\v{q}_2,\v{q}_2+\v{k}_3,\v{q}_2+\v{k}_1+\v{k}_3-\v{k}_2)
\\
&\times[U^{\dagger}_{\v{k}_1}U_{\v{k}_2}]_{e'e} \left[U^{\dagger}_{\v{k}_3}U_{\v{k}_1+\v{k}_3-\v{k}_2}\right]_{c'c}
\\
&\times\Big[\Tr(t^{e'}t^{c'}t^{c}t^{e})\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)
\\
+&\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2 + \v{k}_1+\v{k}_3|)\Big)
\\
&\frac{T_{ce}^aT_{e'c'}^a}{8}\Big(-\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)
\\
&+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2 + \v{k}_1+\v{k}_3|)\Big)\Big]
\end{split}
\end{equation}
\begin{equation}
\begin{split}
&\frac{dN_4^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}\\=&-\frac{32 e^2e_f^2}{3 \pi^4} \alpha_s^2 \int_{\v{k}_1,\v{k}_2,\v{k}_3}\frac{(\v{q}_1+\v{k}_2) \cdot \v{q}_1}{(\v{q}_1+\v{k}_2)^2\v{q}_1^2}\Gamma(\v{q}_2,\v{q}_2+\v{k}_3,\v{q}_2+\v{k}_1+\v{k}_3-\v{k}_2)
\\
&\times[U^{\dagger}_{\v{k}_1}U_{\v{k}_2}]_{e'e} \left[U^{\dagger}_{\v{k}_3}U_{\v{k}_1+\v{k}_3-\v{k}_2}\right]_{c'c}
\\
&\times\Big[\Tr(t^{c'}t^{e'}t^{c}t^{e})\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)
\\
+&\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2 + \v{k}_1+\v{k}_3|)\Big)
\\
+&\frac{T_{ce}^aT_{c'e'}^a}{8}\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_2 + \v{k}_1-\v{k}_2 + \v{k}_3|)
\\
&+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)+\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}_1+\v{k}_3|)\Big)\Big]
\end{split}
\end{equation}
Note that the combinations of the functions $\mathcal{I}$ multiplying the trace of the four Gell-Mann matrices are identical, while those multiplying the product of two adjoint matrices are different in different terms due to distinct patterns of signs.
It is interesting to separate out the contribution which is analogous to calculations performed in the ``classical'' MV model. As we have discussed above, it corresponds to keeping only those terms in $\Sigma_4$ which are completely symmetric in the color charge density operators. Here this corresponds to symmetrizing the four generators in the trace with respect to the color indices.
For this symmetric part of the trace, see Eq.~\eqref{Eq:TrSym}, and $N_c=3$ we obtain
\begin{equation}
\begin{split}
\frac{dN_{\rm sym}^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}
=&\frac{16 e^2e_f^2}{27 \pi^4} \alpha_s^2 \int_{\v{k}_1,\v{k}_2,\v{k}_3}\Gamma(\v{q}_1,\v{q}_1+\v{k}_1,\v{q}_1+\v{k}_2)\Gamma(\v{q}_2,\v{q}_2+\v{k}_3,\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3)
\\
&\times\Big[\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)
\\
&\quad \quad +\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)
\\
&\quad \quad -
\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}_1+\v{k}_3|)\Big)
\Big]\\
&\Big( \Tr[U^{\dagger}_{\v{k}_1}U_{\v{k}_2}]\Tr\left[U^{\dagger}_{\v{k}_3}U_{\v{k}_1+\v{k}_3-\v{k}_2}\right]
\\
&\quad +
\Tr\left[U^{\dagger}_{\v{k}_1}U_{\v{k}_2}U^{\dagger}_{\v{k}_3}U_{\v{k}_1+\v{k}_3-\v{k}_2}\right]
\\
&\quad +
\Tr\left[U^{\dagger}_{\v{k}_1}U_{\v{k}_2}U^{\dagger}_{-\v{k}_1+\v{k}_2-\v{k}_3}U_{-\v{k}_3}\right] \Big)\,.
\end{split}
\end{equation}
This expression has a few physically transparent properties. First, it vanishes if any of the four interactions with the target is set to zero, $U_{\v{k}} \to \delta^{(2)}(\v{k})$, due to $\Gamma(\v q,\v k,\v q) = \Gamma(\v q,\v q,\v k)=0$. Second, the color neutrality of the dipole suggests that a zero momentum gluon will not be able to resolve the dipole and thus lead to a trivial contribution. Gluons originating from the dipole have the momenta $\v{q}_1+\v{k}_1$, $\v{q}_1+\v{k}_2$, $\v{q}_2+\v{k}_3$, and $\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3$ (this can be read off the arguments of $\Gamma$ functions). Setting any of this to zero leads to vanishing contribution in the integral due to the combination of $\mathcal{I}$ functions.
Finally, one can trivially show that the cross section is symmetric under $\v{q}_2\to -\v{q}_2$.
\subsection{Averaging over the target in FDA}
We now need to average over the target field ensemble using the FDA.
We present here the detailed computation only for the term $\Sigma_2$. We have
\begin{equation}
\begin{split}
&\langle \Tr(U_{\v{q}_1+\v{k}-\v{P}}T^{e}U^{\dagger}_{\v{k}-\v{q}_2}U_{\v{\bar{k}}-\v{q}_2}T^eU^{\dagger}_{\v{q}_1+\v{\bar{k}} -\v{P}})\rangle_T\\
=&T^e_{bc}T^e_{fg} \langle U^{ab}_{\v{q}_1+\v{k}-\v{P}} U^{\dagger cd}_{\v{k}-\v{q}_2} \rangle_T \langle U^{df}_{\v{\bar{k}}-\v{q}_2} U^{\dagger ga}_{\v{q}_1+\v{\bar{k}} -\v{P}})\rangle_T\\
+&T^e_{bc}T^e_{fg} \langle U^{ab}_{\v{q}_1+\v{k}-\v{P}}U^{df}_{\v{\bar{k}}-\v{q}_2} \rangle_T \langle U^{\dagger cd}_{\v{k}-\v{q}_2} U^{\dagger ga}_{\v{q}_1+\v{\bar{k}} -\v{P}})\rangle_T\\
+&T^e_{bc}T^e_{fg} \langle U^{ab}_{\v{q}_1+\v{k}-\v{P}} U^{\dagger ga}_{\v{q}_1+\v{\bar{k}} -\v{P}}) \rangle_T \langle U^{\dagger cd}_{\v{k}-\v{q}_2} U^{df}_{\v{\bar{k}}-\v{q}_2} \rangle_T\\
\end{split}
\end{equation}
The first term vanishes since $T^e_{bc} \delta_{bc}=0$. The second term
\begin{equation}
\begin{split}
&T^e_{bc}T^e_{fg} \langle U^{ab}_{\v{q}_1+\v{k}-\v{P}}U^{df}_{\v{\bar{k}}-\v{q}_2} \rangle_T \langle U^{\dagger cd}_{\v{k}-\v{q}_2} U^{\dagger ga}_{\v{q}_1+\v{\bar{k}} -\v{P}})\rangle_T\\
&=T^e_{bc}T^e_{b c}\frac{(2\pi)^4}{N_c^2-1} \delta^{(2)}(0) \delta^{(2)}(\v{q}_1+\v{k}+\v{\bar{k}}-\v{P}-\v{q}_2)D(\v{q}_2-\v{\bar{k}})D(\v{q}_2- \v k)
\\
& =
- (2\pi)^2 N_c S_\perp \delta^{(2)}(\v{q}_1+\v{k}+\v{\bar{k}}-\v{P}-\v{q}_2)D(\v{q}_2-\v{\bar{k}})D(\v{q}_2- \v k) ,
\end{split}
\end{equation}
where we took into account that in the momentum space $\delta^{(2)}(0) = S_\perp/(2\pi)^2$.
The third term
\begin{equation}
\begin{split}
&T^e_{bc}T^e_{fg} \langle U^{ab}_{\v{q}_1+\v{k}-\v{P}} U^{\dagger ga}_{\v{q}_1+\v{\bar{k}} -\v{P}}) \rangle_T \langle U^{\dagger cd}_{\v{k}-\v{q}_2} U^{df}_{\v{\bar{k}}-\v{q}_2} \rangle_T\\
&= (2\pi)^2 \Tr (T^e T^e) S_\perp
\delta^{(2)}(\v{k}-\v{\bar{k}})D(\v{q}_1+\v{k}-\v{P})D(\v{q}_2-\v{k})\\
&= (2\pi)^2 N_c (N_c^2-1) S_\perp
\delta^{(2)}(\v{k}-\v{\bar{k}})D(\v{q}_1+\v{k}-\v{P})D(\v{q}_2-\v{k})\,.
\end{split}
\end{equation}
Hence the cross section for the contribution involving $\Sigma_2$ is
\begin{equation}
\begin{split}
\frac{dN^{(2)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}=&\frac{64e^2e_f^2 }{3\pi^2} N_c \alpha_s^2 S_{\perp} \int_{k,\v{\bar{k}} } \frac{\mathcal{I}(\varepsilon,|\v{q}_1+\v{q}_2+\v{k}+\v{\bar{k}}|)}{(\v{q}_1+\v{q}_2+\v{k}+\v{\bar{k}})^2}\\
&\Big((N_c^2-1)\Gamma(\v{q}_2,\v{q}_2+\v{k},\v{q}_2+\v{k})- \Gamma(\v{q}_2,\v{q}_2+ k,\v{q}_2+ \v{\bar{k}})\Big) D(\v{\bar{k}})D(\v{k})\,.
\end{split}
\end{equation}
Similarly we obtain
\begin{equation}
\begin{split}
\frac{dN^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}
&=\frac{32 e^2e_f^2}{3 \pi^2} N_c\alpha_s^2 S_{\perp} \int_{k,\v{\bar{k}}}
\Big[\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}-\v{\bar{k}}|) L(\v{q}_1, \v{\bar{k}}-\v{q}_1)\\
&+\Big(\mathcal{I}_{\varepsilon_f}(|\v{k}+\v{q}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{\bar{k}}-\v{q}_1|)\Big) L(\v{q}_1, \v{q}_1-\v{\bar{k}})\Big] \cdot \frac{
(\v{\bar{k}}-\v{k}-\v{q}_1-\v{q}_2)}{(\v{\bar{k}}-\v{k}_1-\v{q}_1-\v{q}_2)^2}
\\ &\times
\Big((N_c^2-1)\Gamma(\v{q}_2,\v{q}_2+\v{k},\v{q}_2+\v{k})-\Gamma(\v{q}_2,\v{q}_2+\v{k},\v{q}_2-\v{\bar{k}})\Big)
D(\v{k})D(\v{\bar{k}} )
\end{split}
\end{equation}
For the symmetric part of $\Sigma_4$ at $N_c=3$ and using FDA, we obtain
\begin{equation}
\begin{split}
\frac{dN^{(4)}_{\rm sym}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}
&
=\frac{16e^2e_f^2S_{\perp}}{27\pi^4} \alpha^2_s (N_c^2+1)
(\mathcal{W}_1+\mathcal{W}_2+\mathcal{W}_3)
\end{split}
\end{equation}
where $\mathcal{W}_i$ are the following integrals (although our equations are valid for $N_c=3$ only, we keep explicit variable $N_c$ in order to distinguish different contributions)
\begin{equation}
\label{Eq:W1}
\begin{split}
\mathcal{W}_1&=(N_c^2-1) \int_{\v k,\v{\bar{k}}}\Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1+\v{k})\Gamma(\v{q}_2,\v{q}_2+\v{\bar{k}},\v{q}_2+\v{\bar{k}}) D(\v{k})D(\v{\bar{k}})
\\
&\times\Big(2\mathcal{I}_{\varepsilon_f}(| \v{q}_2 +\v{\bar{k}}|)+2\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}|)
- \\& \quad -\mathcal{I}(\varepsilon_f,|\v{q}_2-\v{q}_1
+\v{\bar{k}}-\v{k}|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}+\v{\bar{k}}|)\Big) \,,
\end{split}
\end{equation}
\begin{equation}
\label{Eq:W2}
\begin{split}
\mathcal{W}_2&= \int_{\v k,\v{\bar{k}}}\Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1-\v{\bar{k}})\Gamma(\v{q}_2,\v{q}_2-\v{k},\v{q}_2+\v{\bar{k}}) D(\v{k})D(\v{\bar{k}})
\\
&
\times\Big({\mathcal{I}_{\varepsilon_f}(|\v{q}_2-\v{k}|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{\bar{k}}|)+
\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{\bar{k}}|)}\\&-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2|)
-\mathcal{I}_{\varepsilon_f}(|\v{k}+\v{\bar{k}}|)
-\mathcal{I}_{\varepsilon_f}(|\v{q}_2-\v{q}_1 + \v{\bar{k}}-\v{k}|)
\Big)\,,
\end{split}
\end{equation}
\begin{equation}
\label{Eq:W3}
\begin{split}
\mathcal{W}_3&= \int_{\v k,\v{\bar{k}}}\Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1+\v{\bar{k}})\Gamma(\v{q}_2,\v{q}_2+\v{\bar{k}},\v{q}_2+ k) D(\v{k})D(\v{\bar{k}})
\\
&
\times\Big({\mathcal{I}_{\varepsilon_f}(|\v{q}_2 + \v{\bar{k}}|)
+\mathcal{I}_{\varepsilon_f}(|\v{q}_1 +\v k|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1 + \v{\bar{k}}|)+
\mathcal{I}(\varepsilon_f,|\v{q}_2 +\v k|}\\
&-\mathcal{I}_{\varepsilon_f}(|\v{q}_2-\v{q}_1|)
-\mathcal{I}_{\varepsilon_f}(|\v{k}-\v{\bar{k}}|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2 + \v{k}+\v{\bar{k}}|)\Big)\,.
\end{split}
\end{equation}
In these contributions one can identify the underlying physical origin:
\begin{itemize}
\item Uncorrelated production:
the terms appearing in the second line of Eq.~\eqref{Eq:W1} and proportional to
$\mathcal{I}(\varepsilon_f,|\v{q}_1 +\v k|$ and $\mathcal{I}_{\varepsilon_f}(|\v{q}_2 +\v{\bar{k}}|)$
represent {uncorrelated production}. This is
manifested by factorization of the momenta $\v{q}_1$ and $\v{q}_2$ as well as the integral with respect to $\v k$ and $\v{\bar{k}}$.
\item Bose Enhancement in the projectile:
the last line of Eq.~\eqref{Eq:W1} is due to Bose Enhancement in the projectile wave function. The two gluons from the projectile scatter on the target independently.
This contribution is peaked at the momenta when the emitted gluons from the projectile have
either colinear or anti-colinear momenta, that is
when the arguments of the $\mathcal{I}$ function vanish. There is also an $N_c^2$-suppressed corrections to this contribution in the last terms of Eqs.~\eqref{Eq:W2} and ~\eqref{Eq:W3}.
\item Hanbury Brown and Twiss effect: the terms $\mathcal{I}(\varepsilon_f,|\v{q}_1 \pm \v{q}_2| )$ originate from the HBT effect.
\item Bose Enhancement in the target:
Gluons originating from the target and having
either colinear or anti-colinear momenta lead to the Bose Enhancement in target wave function. This is the origin of the rest of terms in the above expression. This is obvious for the contribution is proportional to $\mathcal{I}_{\varepsilon_f}(|\v k\pm\v{\bar{k}}|)$. The rest of the contributions are also due to Bose enhancement of the target: they correspond to the situation where the two gluons from the projectile scatter in the correlated manner from the target-by exchanging either the same or opposite transverse momenta. The somewhat unusual momentum combinations in factors of ${\cal I}$, like $\mathcal{I}_{\varepsilon_f}(|\v{\bar{k}}-\v{q}_1|)$ express the fact that gluons with momentum of order of inverse dipole size are emitted from the dipole in coherent fashion.
\end{itemize}
The remaining, non-symmetric part has kinematic factors that cannot be combined to Lipatov vertices.
We again write it as a sum of three contributions
\begin{equation}
\begin{split}
\frac{dN_{\rm ns}^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}\approx
& \frac{32 e^2e_f^2}{3 \pi^4} \alpha_s^2 S_\perp N_c\Big( W_1+ W_2+ W_3\Big)
\end{split}
\end{equation}
with
\begin{equation}
\begin{split}
W_1
=&\frac{(N_c^2-1)}{12}\int_{\v k,\v{\bar{k}}}\Big[\Gamma(\v{q}_1,-(\v{q}_1+\v{k}),-(\v{q}_1+\v{k}))+2 \frac{(\v{q}_1+ \v k) \cdot \v{q}_1}{(\v{q}_1+ \v k)^2\v{q}_1^2} \Big]\Gamma(\v{q}_2,\v{q}_2+\v{\bar{k}},\v{q}_2+\v{\bar{k}})\\
&\times D(\v{k})D(\v{\bar{k}})
\Big(2\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{\bar{k}}|)+2\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}|)
-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}-\v{\bar{k}}|)\\&\quad\quad\quad\quad-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}+\v{\bar{k}}|)\Big)\\
&
+
\frac{(N_c^2-1)}{8}\int_{\v k,\v{\bar{k}}}\Gamma(\v{q}_2,\v{q}_2+\v{\bar{k}},\v{q}_2+\v{\bar{k}})\Gamma(\v{q}_1,-(\v{q}_1+\v{k}),-(\v{q}_1+\v{k}))D(\v{k})D(\v{\bar{k}})
\\
&\times \Big(2\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{\bar{k}}|)-2\mathcal{I}_{\varepsilon_f}(|\v{q}_1+ \v k|)
+\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}-\v{\bar{k}}|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}+ \v{\bar{k}}|)\Big)
\end{split}
\end{equation}
\begin{equation}
\begin{split}
W_2=&\frac{1}{12}\int_{\v k,\v{\bar{k}}}\Gamma(\v{q}_2,\v{q}_2- k,\v{q}_2+\v{\bar{k}}) \Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1-\v{\bar{k}}) D(\v{k})D(\v{\bar{k}}) \\
&
\times \Big[\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2-\v{k}|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{\bar{k}}|)+
\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{\bar{k}}|)\\
&
-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2-\v{\bar{k}}+\v{k}|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2|)\Big)\Big]
\end{split}
\end{equation}
\begin{equation}
\begin{split}
W_{3}=& - \frac{1}{12} \int_{\v k,\v{\bar{k}}}\Gamma(\v{q}_2,\v{q}_2-\v{\bar{k}},\v{q}_2+\v{k})\Big(\frac{1}{\v{q}_1^2} +\frac{(\v{q}_1+\v{k}) \cdot(\v{q}_1-\v{\bar{k}})}{(\v{q}_1+\v{k})^2(\v{q}_1-\v{\bar{k}})^2} +\Gamma(\v{q}_1,-\v{q}_1-\v{k},-\v{q}_1+\v{\bar{k}}) \Big)
\\
&
\times D(\v{k})D(\v{\bar{k}})
\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2-\v{\bar{k}}|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}|)+
\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{\bar{k}}|)\\&-\mathcal{I}_{\varepsilon_f}(|\v{k}+\v{\bar{k}}|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}-\v{\bar{k}}|)\Big)\\
&
-\frac{1}{8}\int_{k,\v{\bar{k}}}\Gamma(\v{q}_2,\v{q}_2-\v{\bar{k}},\v{q}_2+\v{k})\Gamma(\v{q}_1,-\v{q}_1-\v{k},-\v{q}_1+\v{\bar{k}})D(\v{k})D(\v{\bar{k}})\\&
\times \Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2-\v{\bar{k}}|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}|)
+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}|)
-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{\bar{k}}|)\\&-\mathcal{I}_{\varepsilon_f}(|\v{k}+ \v{\bar{k}}|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}-\v{\bar{k}}|)\Big)\,.
\end{split}
\end{equation}
We will use this equations for the numerical evaluations.
\section{Inclusive two gluon production: the MV model.}\label{sec5}
In this section we compute two gluon production by approximating the wave function of the (almost) real photon using the MV model. Note that although we use MV model, we do not discard terms which are suppressed by powers of color charge density, and which are customarily neglected in the CGC literature.
We start with $\Sigma^{\rm MV}_2$, which is trivial to compute,
\begin{equation}
\begin{split}
\Sigma^{\rm MV}_2(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )&=
4\int d^2\bm{x} \mu^2(\bm{x}) f^i(\bar{\v u}_1-\bm{x})f^i(\v u_1-{\bm{x}})f^j(\bar{\v u}_2-\bar{\v u}_1)f^j(\v u_2-\v u_1)
\\
&\left<\Tr \left[ U_{\v u_1}T^{a}[U^{\dagger}_{\v u_2}-U^{\dagger}_{\v u_1}] [U_{\bar{\v u}_2}-U_{\bar{\v u}_1}] T^{a}U^{\dagger}_{\bar{\v u}_1}\right] \right>_{T}
\end{split}
\end{equation}
This expression modulo the projectile wave function is very much the same as the one we obtained in the previous section, see Eq.~\eqref{Eq:N2}.
The contribution in $\Sigma_3^{\rm MV}$ involves two terms, both of which are of the same order in color charge density as $\Sigma^{\rm MV}_2$,
\begin{equation}
\begin{split}
\Sigma^{\rm MV}_3(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )&
= 2\int d^2\bm{x}\mu^2(\bm{x})f^j(\bar{\v u}_2-\bar{\v u}_1)f^i(\bar{\v u}_1-\bm{x})f^i(\v u_1-\bm{x})f^j(\v u_2-\bm{x})\\
&
\Tr\left[ [ U_{\bm{x}}+U_{\v u_1} ] T^a [U^{\dagger}_{\bm{x}}-U^{\dagger}_{\v u_2}] [U_{\bar{\v u}_2}-U_{\bar{\v u}_1}] T^{a} U^{\dagger}_{\bar{\v u}_1}]\right] \\
&
+2\int d^2\bm{x}\mu^2(\bm{x})f^j(\v u_2-\v u_1)f^i(\v u_1-\bm{x})f^i(\bar{\v u}_1-\bm{x})f^j(\bar{\v u}_2-\bm{x})\\
&
\Tr \left[ U_{\v u_1} T^{a} [U^{\dagger}_{\v u_2}-U^{\dagger}_{\v u_1}][U_{\bm{x}}-U_{\bar{\v u}_2}]T^a [U^{\dagger}_{\bm{x}}+U^{\dagger}_{\bar{\v u}_1}]\right]
\end{split}
\end{equation}
Finally, $\Sigma^{\rm MV}_4$ has two contributions, which can be organized by the power of $\mu^2$. We refer to $\mu^2$ contribution as quantum and $\mu^4$ as classical. We start with the former. Using Eq.~\eqref{Eq:rho4mv}
we obtain
\begin{equation}
\begin{split}
\Sigma_{4,Q}^{\rm MV}(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )
&
=\int d^2\bm{x}\mu^2(\bm{x}) f^i(\bar{\v u}_1-\bm{x})f^j(\bar{u}_2-\bm{x})f^i(\v u_1-\bm{x})f^j(u_2-\bm{x})\\
&
\Big(\Tr( [U^{\dagger}_{\v u_1}-U^{\dagger}_{\bm{x}}][U_{\bar{\v u}_1}-U_{\bm{x}}]T^a[U^{\dagger}_{\bm{x}}-U^{\dagger}_{\bar{\v u}_2}][U_{\bm{x}}-U_{\v u_2}]T^a)\\
&
-\frac13\Tr([U^{\dagger}_{\v u_1}+U^{\dagger}_{\bm{x}}][U_{\bar{\v u}_1}+U_{\bm{x}}]T^a)\Tr( [U^{\dagger}_{\bm{x}}-U^{\dagger}_{\v u_2}][U_{\bm{x}}-U_{\bar{\v u}_2}]T^a)\\
&
-\frac13\tr\Big([U^{\dagger}_{\v u_1}+U^{\dagger}_{\bm{x}}][U_{\bar{\v u}_1}+U_{\bm{x}}]T^a [U^{\dagger}_{\bm{x}}-U^{\dagger}_{\v u_2}][U_{\bm{x}}-U_{\bar{\v u}_2}]T^a \Big)
\end{split}
\end{equation}
The classical part is dominant at large color charge density in the projectile and reads
\begin{equation}
\begin{split}
\Sigma_{4,C}^{\rm MV}(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )&=4\int d^2\bm{x}d^2\bm{y}\int d^2\bar{\bm{x}}d^2\bar{\bm{y}}f^i(\bar{\v u}_1-\bm{x})f^j(\bar{\v u}_2-\bm{y})f^i(\v u_1-\bar{\bm{x}})f^j(\v u_2-\bar{\bm{y}})\\
&
\times \left< \left[[U^{\dagger}_{\v u_1}-U^{\dagger}_{\bar{\bm{x}}}][U_{\bar{\v u}_1}-U_{\bm{x}}]\right]_{e'e} \left[ [U^{\dagger}_{ \v u_2}-U^{\dagger}_{\bar{\bm{y}}}][U_{\bar{\v u}_2}-U_{\bm{y}}]\right]_{c'c} \right>_T \\&\times \Big(\delta^{c'e'}\delta^{ce}\mu^2(\bm{x})\mu^2(\bar{\bm{x}})\delta^{(2)}(\bm{x}-\bm{y})\delta^{(2)}(\bar{\bm{x}}-\bar{\bm{y}})\\
&
\quad+\delta^{c'e}\delta^{e'c}\mu^2(\bm{x})\mu^2(\bm{y})\delta^{(2)}(\bm{x}-\bar{\bm{y}})\delta^{(2)}(\bar{\bm{x}}-\bm{y})\\&\quad +\delta^{c'c}\delta^{e'e}\mu^2(\bm{x})\mu^2(\bm{y})\delta^{(2)}(\bm{x}-\bar{\bm{x}})\delta^{(2)}(\bm{y}-\bar{\bm{y}}) \Big)
\end{split}
\end{equation}
Here the first term in the classification of Ref.~\cite{Kovchegov:2012nd}
corresponds to the ``square'' diagrams involving the quadrupole, the second to ``cross'' diagrams. The last contribution represents emission where the correlations can only originate from the correlations in the target; this term also gives the leading contribution to the uncorrelated production.
\subsection{Momentum space and FDA}
As in the dipole model, we convert our expressions to momentum space and average over target using FDA.
We get
\begin{equation}
\begin{split}
\frac{dN_2}{d\v q^2_1d\v q^2_2 d\eta d\xi}&=
\frac{4 g^4}{(2\pi)^4}\int_{\v{k}_1,\v{k}_2,\v{k}_3,\v{k}_4}\mu^2(\v{k}_2+\v{k}_4-\v{k}_1-\v{k}_3)\left<\Tr \left[ U_{\v{k}_1}T^{a}U^{\dagger}_{\v{k}_2} U_{\v{k}_3} T^{a}U^{\dagger}_{\v{k}_4}\right] \right>_{T}\\
&L^i(\v{q}_2,\v{q}_2+\v{k}_2)L^j(\v{q}_2,\v{q}_2+\v{k}_3) \frac{(\v{q}_1+\v{q}_2+\v{k}_3-\v{k}_4)^i(\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_1)^j}{(\v{q}_1+\v{q}_2+\v{k}_3-\v{k}_4)^2(\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_1)^2}
\end{split}
\end{equation}
Using FDA, this simplifies into
\begin{equation}
\begin{split}
\frac{dN_2}{d\v q^2_1d\v q^2_2 d\eta d\xi}&=\frac{4N_cg^4}{(2\pi)^4}\int d^2 \v x\mu^2(\v x) \int_{\v{k}_1,\v{k}_2} D(\v{k}_1+\v{q}_1)D(-\v{k}_2+\v{q}_2)\\
&\Big\{ (N_c^2-1) \left(\frac{(\v{k}_2-\v{k}_1) \cdot L(\v{q}_2,\v{k}_2)}{(\v{k}_2-\v{k}_1)^2}\right)^2
\\ & \quad -\frac{(\v{k}_2-\v{k}_1)\cdot L(\v{q}_2,\v{k}_2)}{(\v{k}_2-\v{k}_1)^2}\frac{(\v{k}_2-\v{k}_1) \cdot L(\v{q}_2,\v{q}_2-\v{k}_2-\v{k}_1)}{(\v{k}_2-\v{k}_1)^2}\Big\}
\end{split}
\end{equation}
Similarly for $\Sigma_3$ contribution we obtain
\begin{equation}
\begin{split}
\frac{dN_3}{d\v q^2_1d\v q^2_2 d\eta d\xi}=& \frac{4 g^4}{(2\pi)^4}\int_{\v{k}_1,\v{k}_2,\v{k}_3,\v{k}_4} \mu^2\Big(\v{k}_4+\v{k}_2-\v{k}_1-\v{k}_3\Big) \Tr[U_{\v{k}_1}T^aU^{\dagger}_{\v{k}_2}U_{\v{k}_3}T^aU^{\dagger}_{\v{k}_4}]\\
&\Gamma(\v{q}_2,\v{q}_2+\v{k}_2,\v{q}_2+\v{k}_3)L^i(\v{q}_1,\v{k}_1-\v{q}_1)\frac{(\v{k}_4-\v{k}_3-\v{q}_1-\v{q}_2)^i}{(\v{k}_4-\v{k}_3-\v{q}_1-\v{q}_2)^2}
\end{split}
\end{equation}
where we assumed that $\mu^2(-\v{k}) = \mu^2(\v{k})$.
The FDA leads to
\begin{equation}
\begin{split}
\frac{dN_3}{d\v q^2_1d\v q^2_2 d\eta d\xi}
=&\frac{4N_cg^4}{(2\pi)^4}\int d^2 \v x\mu^2(\v x) \int_{\v{k}_1,\v{k}_2} D(\v{k}_1+\v{q}_1)D(-\v{k}_2+\v{q}_2)\\
&\Big\{(N_c^2-1) L^i(\v{q}_1,\v{k}_1)\frac{(\v{k}_1-\v{k}_2)^i}{(\v{k}_1-\v{k}_2)^2}\Gamma(\v{q}_2,\v{k}_2,\v{k}_2)\\& \quad -L^i(\v{q}_1,\v{k}_1)\frac{(\v{k}_1-\v{k}_2)^i}{(\v{k}_1-\v{k}_2)^2}\Gamma(\v{q}_2,\v{k}_2,\v{q}_2-\v{q}_1-\v{k}_1)\Big\}\,.
\end{split}
\end{equation}
Finally the $\Sigma_4$ contribution is given by two terms proportional to $\mu^2$ and $\mu^4$. Starting with the former, in FDA we obtain
\begin{equation}
\begin{split}
&\frac{dN_4^Q}{d\v q^2_1d \v q^2_2 d\eta d\xi}\\
=&\frac{g^4N_c}{(2\pi)^4}\int_{\v k,\v{\bar{k}}} D(-\v{k})D(-\v{\bar{k}})\int d^2\bm{x}\mu^2(\bm{x})\\&
\Big\{(N_c^2-1)\Big[\Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1+\v{k})-\frac13\Gamma(\v{q}_1,-\v{q}_1-\v{k},-\v{q}_1-\v{k}) \Big]\Gamma(\v{q}_2, \v{q}_2+ \v{\bar{k}},\v{q}_2+\v{\bar{k}})\\
&+\frac23\Gamma(\v{q}_1,-\v{q}_1-\v{k},-\v{q}_1-\v{\bar{k}})\Gamma(\v{q}_2, \v{q}_2+ \v{\bar{k}},\v{q}_2+\v k)\\
&+\Big[ \Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1+\v{\bar{k}}) -\frac13\Gamma(\v{q}_1,-\v{q}_1+\v{k},-\v{q}_1-\v{\bar{k}})\Big] \Gamma(\v{q}_2, \v{q}_2+ \v k,\v{q}_2+\v{\bar{k}})\Big\}
\end{split}
\end{equation}
The latter is given by
\begin{equation}
\begin{split}
\frac{dN_4^C}{d\v q^2_1d \v q^2_2 d\eta d\xi}
&=\frac{4g^4}{(2\pi)^4} \int_{\v k,\v{\bar{k}}}D(-\v{k})D(\v{\bar{k}})\Big\{\Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1+\v{k})
\Gamma(\v{q}_2,\v{q}_2-\v{\bar{k}},\v{q}_2-\v{\bar{k}})\\&
\left[|\mu^2(0)|^2(N_c^2-1)^2+(N_c^2-1)|\mu^2(\v{q}_1-\v{q}_2+\v{k}+\v{\bar{k}})|^2 \right]\\&
+(N_c^2-1)|\mu^2(\v{q}_1+\v{q}_2+\v{k}+\v{\bar{k}})|^2\Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1+\v{k})\Gamma(\v{q}_2,\v{q}_2+ \v{\bar{k}},\v{q}_2+\v{\bar{k}})\\
&+\Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1+\v{\bar{k}})\Gamma(\v{q}_2,\v{q}_2+ k,\v{q}_2+\v{\bar{k}})\Big( |\mu^2(\v{q}_1+\v{q}_2+\v{k}+\v{\bar{k}})|^2\\&
+(N_c^2-1)|\mu^2(\v{q}_1-\v{q}_2)|^2+|\mu^2(\v{k}-\v{\bar{k}})|^2\Big)\\
&+\Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1+\v{\bar{k}})\Gamma(\v{q}_2,\v{q}_2-\v{k},\v{q}_2-\v{\bar{k}})\Big((N_c^2-1)|\mu^2(\v{q}_1+\v{q}_2)|^2\\&
+|\mu^2(\v{q}_1-\v{q}_2+\v{k}+\v{\bar{k}})|^2+|\mu^2(\v{k}-\v{\bar{k}})|^2\Big)
\end{split}
\end{equation}
The physical origin of different terms can be easily identified:
\begin{itemize}
\item Uncorrelated production: the contribution proportional to $(N_c^2-1)^2|\mu^2(0)|^2$ describes uncorrelated two gluon production
\item Bose Enhancement in the projectile: this term is proportional to $(N_c^2-1)|\mu^2(\v{q}_1 \pm \v{q}_2 +\v{k}+\v{\bar{k}})|^2$. It is suppressed by the factor of $1/N_c^2$ compared to the uncorrelated production as the initial gluons have to have the same color. As in the dipole case, there is a correction to Bose Enhancement in the projectile which is suppressed by a further factor of $1/N_c^2$. It is proportional to $|\mu^2(\v{q}_1\pm \v{q}_2+\v{k}+\v{\bar{k}} )|^2$.
\item Bose Enhancement in the target: the terms proportional to $|\mu^2(\v{k}-\v{\bar{k}} )|^2$ represent Bose Enhancement in the target; they are suppressed by $1/N_c^4$ in the dense regime.
\item HBT: finally the terms $(N_c^2-1)|\mu^2(\v{q}_1\pm \v{q}_2)|^2$ represent HBT correlation. This term is suppressed by $1/N_c^2$ relative to the uncorrelated production as the color of the final state gluons should be the same.
\end{itemize}
\section{Numerical results}
\label{Sec:Num}
In this section we evaluate numerically inclusive two gluon production in both models for the projectile wave function.
We do not have ambitions to fit experimental data, and we therefore do not try to optimize parameters of the model. Nevertheless we want to be able to have at least a qualitative idea on the ballpark of the CGC predictions for relevant observables.
The first question we are faced with is that of subtraction of the background.
In experiment, low multiplicity events are used to subtract trivial contribution from high multiplicity data, which is then analyzed in terms of flow harmonics. In our calculation using the MV model for the projectile, this procedure naturally corresponds to subtracting the $\mu^2$ terms as they play a dominant role at low multiplicity and are negligible compared to $\mu^4$ terms at high multiplicity. We thus take as our ``signal'' only the $\mu^4$ contribution, which corresponds to the classical gluon production as defined in the previous section. In the dipole model for the projectile wave function, the distinction between low and high multiplicities is trickier to make since we do not have an explicit parameter $\mu^2$. Nevertheless motivated by the MV model, we will use the symmetric part of the matrix element $\Sigma_4$ as our high multiplicity ``signal''. For $N_c=3$, as we demonstrated in Sect.~\ref{Sect:2glDPM}, it has a compact form and is clearly not dominated by back-to-back production.
For numerical calculations we will use the following set of parameters: the saturation momentum of the target is chosen as $Q_s=2$ GeV, the infrared cut-off entering into the Poisson equation of the MV model is set to be equal to $\Lambda = 200$ MeV, additionally to insure that the logarithm of the argument of the logarithm entering into the adjoint dipole of Eq.~\eqref{Eq:KovnersHammer} does not change sign we use
\begin{align}
D(\v r) = \exp\left[-\frac{1}{4}Q_s^2r^2\ln(\frac{1}{\Lambda^2r^2}+e) \right]\,.
\end{align}
For the nearly real photon wave function, we use a small but non-zero value of $Q$ in order to regularize large distance physics. In our calculations we fix $\epsilon_f = 100$ MeV. This roughly corresponds to $Q\sim 200$ MeV.
To ensure that the projectile photon has a finite size,
in the projectile MV model we introduce the projectile size $R$ through the impact parameter dependent parameter $\mu^2(\v x)$:
\begin{align}
\mu^2(\v x) = {\cal N} \exp{-\frac{\v x^2}{R^2}}\,.
\end{align}
Since our signal is a homogeneous function of $\mu^2$, the constant $ {\cal N}$ does not play any role in the normalized quantities considered below and we do not need to specify it.
Here we do not explore the dependence on the shape of the projectile as it has little effect on the observable in this approach, except for unrealistically large ellipticities, see Ref.~\cite{Kovner:2018fxj}.
The code is publicly available in Ref.~\cite{UPCcode}.
In order to compute the two particle $v_2$ as a function of the transverse momentum, we define the $n$-th Fourier mode of the particle production cross section $V_n(p_\perp)$
\begin{align}
V_n(q_1) =
\int d \theta_1
\int_{0}^{p^{\rm max}_\perp} d^2 \v{q}_2
\exp(i n \Delta \theta )
\frac{d N}{d\v q^2_1 d \v q^2_2 d\eta d\xi}
\end{align}
where $\Delta \theta = \theta_2-\theta_1$ is the angle between $\v q_1$ and $\v q_2$.
Additionally, the integrated Fourier modes are defined according to
\begin{align}
V_n =
\int_0^{p^{\rm max}_\perp} d^2 \v{q}_1
\int_0^{p^{\rm max}_\perp} d^2 \v{q}_2
\exp(i n \Delta \theta )
\frac{d N}{d\v q^2_1 d \v q^2_2 d\eta d\xi}
\end{align}
Then we have
\begin{align}
\label{Eq:v22}
v^{(2)}_2(p_\perp) = \sqrt{\frac{V_2(p_\perp)}{V_0(p_\perp)}}
\end{align}
for the second Fourier harmonic of the two particle correlation function.
Experimental collaborations, in addition to $v^{(2)}_2(p_\perp)$ also extract $v_{2}$ whose definition is based on the assumption of the factorization of the hydrodynamic flow.
In hydrodynamics, a two-particle azimuthal harmonic $v^{(2)}_2$ is a product of single particle azimuthal anisotropies, $v_2$. This motivates considering the observable
\begin{align}
\label{Eq:v2}
v_{2}(p_\perp) = \frac{V_2(p_\perp)/V_0(p_\perp)}{\sqrt{V_2/V_0}}\,.
\end{align}
Experimentally the two observables are quite close to each other \cite{ATLAS:2021jhn}.
In our calculations, there is no reason at all to expect such hydrodynamics-like factorization. It is nevertheless useful to consider $v_{2}$ in order to get an idea of how much the two quantities differ from each other in our framework.
We plot these two quantities in Figs.~\ref{Fig:v22} and~\ref{Fig:v2}. To guide the eye we also added experimental data to compare it to $v_{2}$. Note that we did not plot the last experimental point which is located at rather large negative values of $v_2$, see ATLAS~\cite{ATLAS:2021jhn}.
The calculations are done for two upper limits of the integration with respect to the transverse momentum $p^{\rm max}_\perp = 2$ and 4 GeV. The former choice is the same as in the ATLAS measurements, while the later choice is motivated by comparison with the study in Ref.~\cite{Shi:2020djm}.
\begin{figure}
\centering
\includegraphics[width=0.49\linewidth]{Max2_v22.pdf}
\includegraphics[width=0.49\linewidth]{Max4_v22.pdf}
\caption{The second Fourier harmonic of the two particle correlation function, $v^{(2)}_2$, see Eq.~\eqref{Eq:v22}, as a function of the transverse momentum of the "trigger" gluon. The momentum of the "associated" gluon is integrated from $0$ to $p_\perp^{\rm max}$. }
\label{Fig:v22}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.49\linewidth]{Max2_v2.pdf}
\includegraphics[width=0.49\linewidth]{Max4_v2.pdf}
\caption{The elliptic flow $v_2$ extracted from two particle correlations function using the definition motivated by factorization, $v_2$, see Eq.~\eqref{Eq:v2}, as a function of the transverse momentum of the "trigger" gluon. The momentum of the associated gluon is integrated from $0$ to $p_\perp^{\rm max}$. To guide the eye we plotted ATLAS data from Ref.~\cite{ATLAS:2021jhn}. }
\label{Fig:v2}
\end{figure}
\section{Discussion}
\label{sec7}
We now discuss our results.
We stress again that our goal was not to fit the ATLAS data as we do not believe that our calculational framework is constrained enough for such an endeavor. First, we do not include hadronization in our calculation, and that may have a large effect on the correlated part of particle production. In addition the wave function of the real photon has to be modelled in some way, and this modelling leaves a lot of freedom. Instead we studied two ubiquitous models used in a variety of CGC calculations to understand the qualitative features of the effect.
Consequently we did not try to optimize the parameters of the MV model and the dipole model of the photon when calculating $v_2$.
Before we proceed, we want to put our study in the perspective by commenting on the existing literature and comment on the comparison of our results to those of Ref.~\cite{Shi:2020djm}. Direct comparison between the two is hard, since the calculational frameworks are very different, even though both calculations are based on the CGC approach. Our calculation is appropriate at non forward rapidities where particle production is dominated by gluons, whereas the authors of Ref.~\cite{Shi:2020djm} use the hybrid approach which restricts their calculation to the photon going direction. One could attempt to take the model used in Ref.~\cite{Shi:2020djm} for valence parton distribution in the photon and use it as yet another model for the photon. This is however not a well posed problem since no color correlations between valence partons have been taken into account in Ref.~\cite{Shi:2020djm} due to collinear approximation invoked there, while such color correlations are clearly important away from forward rapidity especially at relatively low transverse momenta.
There is however one qualitative observation that we can make with confidence. The $v_2$ is our calculation exhibits in all models the same qualitative trend: it rises at low transverse momentum to a maximum at $p_\perp$ around $ 1$ GeV, and drops thereafter, see Figs.~\ref{Fig:v22} and \ref{Fig:v2}. This is also a clear trend in the experimental data. On the other hand $v_2$ calculated in Ref.~\cite{Shi:2020djm} never reaches a maximum, but instead keeps growing at high momentum far beyond the point where experimentally it is observed to fall. Qualitative this difference in behavior is easy to understand. As we will explain below the origin of the turnover in our calculation is the narrowness of the gluon HBT peak. The calculation of \cite{Shi:2020djm} does not take the gluon HBT effect into account. Instead the correlations in \cite{Shi:2020djm} arise solely from the so called color domain structure in the target \cite{Kovner:2010xk}. The color domain effect is the more efficient, the higher the transverse momentum of produced particles. High $p_\perp$ particles are produced nearby in the coordinate space, and therefore are more probable to probe the same domain in the target. Thus the approach of \cite{Shi:2020djm} does not contain a mechanism that could lead to decreasing $v_2$ at high momentum.
\begin{figure}
\centering
\includegraphics[width=0.49\linewidth]{Factor_v2_mv_nb.pdf}
\includegraphics[width=0.49\linewidth]{Factor_v2_dip_nb.pdf}
\caption{The elliptic flow $v_2$ for three different kinematic ranges of the trigger particle. Here as in the previous figure, $Q_s=2$ GeV. The size of the projectile is set by $R=1$/GeV.}
\label{Fig:v21}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.49\linewidth]{Factor_v2_mv.pdf}
\includegraphics[width=0.49\linewidth]{Factor_v2_dip.pdf}
\caption{The same as Fig.~\ref{Fig:v21} but binned with the same bin size as the ATLAS analysis. }
\label{Fig:v23}
\end{figure}
Consider now our numerical results on Figs.~\ref{Fig:v22} and ~\ref{Fig:v2}. The first observation is that both the MV model and the dipole model produce $v_2$ of the same order as the experimental data. The shape of the momentum dependence in the two models is slightly different, with MV mode $v_2$ rather sharply peaked at low momentum.
One somewhat surprising feature of our results is that $v_2$ and $v_2^{(2)}$ turn out to be not very different. Naive expectation is that $v_2$ should drop very quickly once the momentum of the trigger particle is outside the integration interval of the momentum of the associate. The reasoning for this is that the two particle correlations in the CGC framework receive a very large contribution from the gluon HBT effect. The HBT peak is very narrow, in fact for a large projectile its width in momentum space is inversely proportional to the area. As a result for a large area projectile the correlation drops by several orders of magnitude once the momentum of the trigger is outside of the momentum bin of the associate~\cite{Altinoluk:2018ogz}. It is thus natural to expect $V_2/V_0\gg V_2(p_\perp)/V_0(p_\perp)$ and therefore $v_2\ll v_2^{(2)}(p_\perp)$ when $p_\perp>p_\perp^{\rm max}$.
Instead we observe that although $v_2$ does indeed drop somewhat faster for $p_\perp>p_\perp^{\rm max}$, the difference is not that spectacular especially for the dipole model photon. We believe this is due to the finite size of the projectile wave function, which leads to widening of the HBT peak and tempers the drop in the HBT correlation.
With these considerations in mind we plot the dependence of $v_2$ on the choice of the momentum bin of the associated particle. We choose the bins the same as in the ATLAS data~\cite{ATLAS:2021jhn}. As mentioned above, the hydrodynamic-like universality would suggest that this quantity does not depend on the bin size and location. Our results on the other hand are not expected to display such independence. In Fig.~\ref{Fig:v21} we plot $v_2$ as a function of the momentum of the trigger particle for three different bin choices of the associate. Indeed we observe a rather strong dependence on the bin choice consistent with the discussion above. Namely, $v_2$ is large when the momentum of the trigger particle $p_\perp$ is within the bin, and drops fairly quickly once $p_\perp$ is outside the bin. This behavior is common for the MV model and the dipole model of the photon. At some values of $p_\perp$ the difference between the values of $v_2$ is very large, with ratio between $v_2$ for different bins reaching about 7 or 8 at the maximum. This seems to be in stark contradiction with experimental data. However, one has to keep in mind that the experimental test of universality in Ref.~\cite{ATLAS:2021jhn} was not performed for a fixed momentum of a trigger particle, rather $p_\perp$ was also binned. Following the same procedure we plot on Fig.~\ref{Fig:v23} values of $v_2$ for four bins of $p_\perp$ chosen in the same way as in~Ref.\cite{ATLAS:2021jhn}. Interestingly, the binning has a very strong effect: since the curves in Fig.~\ref{Fig:v21} are all bell shaped on the scale comparable with the bin size, integrating over the $p_\perp$ bins leads to significant reduction of differences in the values of $v_2$. The ratio between the values of $v_2$ for different associate bins is now at most 3 or 4 rather than 7 or 8. This is still significantly larger than for the ATLAS data, where the maximal ratio is closer to 2, but the difference is now not as drastic. Interestingly, the systematics of momentum dependence of our results is quite similar to that in the data, with $v_2$ initially rising, and subsequently falling with $p_\perp$, and $v_2$ for some of the bins crossing over in the vicinity of $1$ GeV.
We stress again that our purpose in this exercise is not to force-fit the data. It is quite clear that fiddling with the model parameters would allow us to get much closer to experimental points, but it is unclear what it would teach us. Instead, our take home message is twofold. First, we conclude unambiguously that a CGC calculation does not yield results consistent with the hydrodynamic inspired universality of $v_2$, although the values of $v_2$ we obtain are roughly of the same magnitude as the experimental ones. Second, it seems to us that the test of the said universality conducted in Ref.~\cite{ATLAS:2021jhn} is inconclusive. The variation between the different bin values found in Ref.~\cite{ATLAS:2021jhn} albeit broadly consistent with universality, is also quite similar to our (not optimized) CGC results due to binning of the momentum of the trigger particle. A more detailed study of this point seems desirable.
\acknowledgments
We thank T.~Altinoluk, N.~Armesto, M.~Lublinsky, and A.~Milov for illuminating discussions.
A.K. is supported by the NSF Nuclear Theory grant 1913890. This material is based
upon work supported by the U.S. Department of Energy, Office of Science, Office of Nuclear
Physics through the Contract No. DE-SC0020081 (H.D. and V.S.).
\section{Introduction}
Deep inelastic scattering off a nuclear target at high energies constitutes the best environment to probe the small x tail of the hadron wave function and to confirm experimentally the existence of gluon saturation.
There is a trivial reason why one is interested in the smallest probe possible: to minimize the effect of final state interaction and make experimental data interpretable from the perspective of the hadron wave function. At high energy, the coherence length of the virtual photon is significantly larger than the proton radius. Thus the larger the nuclear target the denser the gluon field which the projectiles probes. While the Electron Ion Collider will be the best discovery machine for the saturation physics, the ultra-peripheral collisions~\cite{Bertulani:2005ru,Baltz:2007kq} at RHIC and the LHC may also be a very useful probe of the small x hadron wave function as a nearly real photon likely has a smaller partonic content than the proton, and consequently the influence of final state effects should be smaller as well.
In this paper we focus on correlations between produced particles in Ultraperipheral Collisions (UPC). Traditionally, hydrodynamics has been successful in dealing with correlations produced in heavy ion collisions. However, the discovery of such correlation in small systems brought to the fore the question of their origin. Following the high multiplicity pp measurement by CMS~\cite{CMS:2010ifv}, a set of measurements was done in the past decade, specifically, p-Pb at LHC~\cite{ALICE:2012eyl,ATLAS:2012cix,CMS:2012qk} and p-Au, d-Au, He-Au at RHIC~\cite{PHENIX:2016cfs,PHENIX:2013ktj,STAR:2014qsy}, see more details in review articles~\cite{Schenke:2021mxx,Nagle:2018nvi}. In addition similar correlations have been recently observed in UPC \cite{ATLAS:2021jhn}. Similar measurement was also performed in $e^+e^-$ collision \cite{Badea:2019vey}, and ep Deep-inelastic-scattering \cite{ZEUS:2019jya}, and no ridge correlation was found in this data.
Our focus in this paper is on the possibility that particle correlations originate from the quantum correlation in the initial state hadronic wave function.
From theoretical perspective, the wave-function of a nearly real photon is poorly understood due to uncontrollably large non-perturbative effects. Thus theoretical results describing scattering and particle production in UPC are unavoidably model dependent. In this paper, in order to at least provide a sense of the systematic uncertainty, we consider two models for the photon wave functions. The first one is based on a naive perturbative process of photon splitting into a virtual quark-antiquark pair. The second assumes that the real photon can be represented as, in general non-perturbative, hadronic state, e.g. $\rho$-meson~\cite{Sakurai:1960ju}. A $\rho$-meson wave-function evolved to asymptotically high energy/small x would be similar to the wave function of a nucleus or a proton and thus can be approximated by the McLerran-Venugopalan (MV) model. For the energies available at the LHC and especially at RHIC, this approximation can be expected to be wanting. Nonetheless, two diametrically opposite approximations may help to understand the influence and the importance of the associated systematics.
Within the Color Glass Condensate approach, the scattering and central rapidity particle production in photon-nucleus collisions pose yet another challenge (see Ref.~\cite{Zhao:2022ayk} for hydrodynamic approach to UPC).
At the moment, the two - particle production was only studied in the forward, photon going direction~\cite{Shi:2020djm}, which is not appropriate to describe the central rapidity production for two reasons. First, theoretically, forward production is a significantly different process than the production at central rapidity and as such it has qualitatively different systematics as a function of the control parameters. Second, the experimental data shows extremely strong rapidity dependence, see Ref.~\cite{ATLAS:2021jhn}. The lack of boost-invariance~\cite{Bozek:2015swa,Ke:2016jrd,Schenke:2016ksl,Shen:2020jwv,Wu:2021hkv} invalidates approximating of the mid-rapidity physics by the physics of the forward region. In particular only a small fraction of particles is produced in the forward region. Thus it is hard to imagine that particle correlations integrated over a sizable rapidity interval can be reasonably described in terms of these hadrons alone as in \cite{Shi:2020djm}. Thus addressing mid-rapidity production is absolutely crucial in order for the CGC based theoretical analysis to be taken seriously.
Here we undertake such a calculation, albeit within the constraints mentioned above.
As in the rest of the CGC based calculations, we consider inclusive two-gluon production from the projectile wave function in the eikonal approximation for the emission vertex. This approximation is justified if the gluons have sufficiently different rapidities. On the other hand when the rapidity difference becomes too large, $\sim \frac{1}{\alpha_s}$ this approximation has to be revised as one would be required to account for the evolution between the gluon emissions~\cite{Jalilian-Marian:2004vhw}. However experimentally one is not dealing with such large rapidity differences, and so our simple approximation should be adequate.
The calculation of particle production in the CGC framework necessitates averaging of products of Wilson lines over the color field ensemble of the target.
This in general is a rather complex endeavor since to study correlations we are forced to go beyond the large $N_c$ limit which then requires us to calculate correlators of a large number of Wilson lines. In order to make the calculations tractable and not overly numerically demanding, in this paper we use the so-called {\it factorized dipole approximation} (FDA) discussed in Ref.~\cite{Kovner:2017ssr,Kovner:2018vec} and successfully applied to study particle production and correlations in Refs.~\cite{Altinoluk:2018ogz,Kovner:2018fxj,Kovner:2018azs,Li:2021ntt,Altinoluk:2020psk}. This approximation is well suited for a dense nuclear target in and close to saturation regime. We will explain this in more detail in the following.
The paper is organized as follows.
We start from summarizing the machinery used to calculate particle production in dilute-dense scattering in Sec.~\ref{sec2}. In Sec.~\ref{sec3}, we calculate a general expression for double gluon production cross section. In Sec.~\ref{sec4} and Sec.~\ref{sec5}, we discuss projectile ensemble average for the dipole and MV models respectively. Sec.~\ref{Sec:Num} lays out the details of our numerical procedure. We conclude with the discussion of our results in Sec.~\ref{sec7}.
\section{Particle production in dilute-dense scattering}\label{sec2}
In this section we outline the most important ingredients of the formalism and define the observables.
The formalism of our choice is the wave function approach of Refs.~\cite{Kovner:2001vi,Baier:2005dv}. Alternatively one can use diagrammatic methods of Ref.~\cite{Kovchegov:2012mbw}. While the final expressions are the same, the representation we use is convenient for organizing calculations.
Consider a projectile at high energy before a collision with a hadronic/nuclear target. The projectile's wave function can be written in the general form
\begin{align}
| \Psi_{\rm in} \rangle =
\sum_{\{\v{x}, k^+, \alpha\}}
\psi (\{\v{x}, k^+, \alpha\})
\prod_i|{\v{x}_i, k^+_i, \alpha_i}\rangle
\end{align}
where $\v{x}_i$ is the $i$-th parton's transverse position, $k_i^+$ longitudinal momentum and $\alpha_i$ the color index. Subject to the physical process, the latter can belong to either (anti)fundamental
or adjoint representation describing a (anti)quark or a gluon in the incoming wave function.
At high energy, the separation of relevant timescales allows for significant simplifications of the analysis of the scattering process. The propagation time through the target in high energy collisions can be arbitrary small compared to effective inter-parton interaction time. One thus can neglect the interactions between the partons and approximate the scattering process by independent propagation of partons.
For a similar reason,
at high energy
the partons do not change their transverse positions during the scattering process. The high energy scattering in QCD is non-trivial due to the different eikonal phases acquired by different components of the wave function. This difference in phases is responsible for decoherence of the components of the wave function and leads to particle production in the final state. The wave function of the outgoing projectile right after scattering is
\begin{align}\label{out}
| \Psi_{\rm out} \rangle = S |\Psi_{\rm in}\rangle = \sum_{\{\v{x}, k^+, \alpha\}}
\psi (\{\v{x}, k^+, \alpha\})
\prod_i U_{\alpha_i \beta_i}( \v{x}_i) |\v{x}_i, k^+_i, \beta_i\rangle
\end{align}
where $U$ is the Wilson line defined in the representation of $SU(N_c)$ group appropriate to the color charge of a given parton:
\begin{align}
U_{\v{x}} = {\cal P} e^{ i \int d x^- T^a A^+_a(\v{x},x_-) }.
\end{align}
Here $A^+$ is the $+$ component of the gauge field of the target.
To calculate the expectation value of a gluon observable
in the final state, one has to allow for the propagation of the projectile state \eqref{out} to asymptotic time $t\rightarrow\infty$.
This evolution results in emission of additional gluons which dress the bare partons by their WW fields as well as in recombination of some of the outgoing soft gluons into the WW field of the outgoing fast partons.
It is straightforward to account for the evolution after scattering (or equivalently to change the basis from free to dressed partons) by introducing the coherent operator
\begin{align}\label{c}
C = {\cal P} e^{i \sqrt{2} \int d^2 x d\xi\,
\hat b_a^i(\xi, \v{x})
\left[
a^\dagger_{i,a} (\xi, \v{x})
+ a_{i,a} (\xi, \v{x})
\right]
}
\end{align}
where the WW field operator is defined through the total charge density operator $\hat{\rho}_{\rm P}^a(\xi,\v{y})$ integrated from the rapidity of the projectile to $\xi$
\begin{align}
\hat{b}^i_a(\xi,\v{x})=\frac{g}{2\pi}\int d^2y \frac{(\v{x}-\v{y})^i}{|\v{x}-\v{y}|^2} \hat{\rho}_{\rm P}^a(\xi,\v{y})+...
\end{align}
Here we have explicitly written out only the leading order in $\alpha_s$ expression. Higher order corrections can be included, but we will not take them into account in our calculations. The expectation value of a
gluon observable $O(a,a^\dagger)$ in the final state is then given by
$ \langle \Psi_{\rm out} | C \,O(a,a^\dagger) \, C^\dagger | \Psi_{\rm out} \rangle $.
For example the differential single inclusive gluon production is given by
\begin{align}
\frac{ dN } { dy d^2 \v q} = \langle \Psi_{\rm out} | C \,
a_{i,b}^\dagger(y, \v q)
a_{i,b}(y, \v q) \,
C^\dagger | \Psi_{\rm out} \rangle
\end{align}
This is nothing but the high energy counterpart of the Faddeev-Kulish construction of dressed states in theories with massless gauge bosons.
In this paper we will consider double gluon production, which is a straightforward generalization of the previous equation
\begin{align}
\frac{dN}{d\eta d^2\v{q}_1 d\xi d^2\v{q}_2} = \langle \Psi_{\rm out} | C \,
a_{i,b}^\dagger(\eta, \v{q}_1)
a_{i,b}^\dagger(\xi, \v{q}_2)
a_{i,b}(\eta, \v{q}_1) \,
a_{i,b}(\xi, \v{q}_2) \,
C^\dagger | \Psi_{\rm out} \rangle
\end{align}
To perform the actual calculations, we need of course a model for the wave function of the projectile. In general, in the high energy CGC approach,
this wave function has the form of a wave function for valence charges dressed by the WW field, that is \begin{align}
| \Psi_{\rm in} \rangle = C | v \rangle,
\end{align}
where $C$ is the coherent operator Eq.(\ref{c}) and $| v \rangle$ describes a state in the free Fock space of a small number of valence (large rapidity) partons. As mentioned in the introduction, we use two models for the valence wave function $|v \rangle $ - the dipole and MV model, which we describe in detail below.
Finally, the target is treated as an ensemble of classical gluon field configurations which have to be averaged over.
We now proceed with the detailed calculations within the framework just described.
\subsection{Modelling the Projectile and the Target}
As stated earlier, we explore two models for the wave function of a nearly on shell photon.
\subsection{Dilute dipole projectile}
For a nearly on-shell photon, the longitudinal polarization is suppressed by a square of the virtuality. Hence only the transverse polarization is of importance.
The leading order perturbative expression for the photon state is given by ~\cite {Beuf:2016wdz}
\begin{equation}
\begin{split}
|\gamma_{\lambda}^*\rangle \simeq &
\sum_{s_1,s_2}\int_0^{\infty} \frac{dz}{4(2\pi) z(1-z)} \int \frac{d^2 \v{k}_1 }{(2\pi)^2} \Psi^{T}_{\lambda}(z, \v{k}_1,s_1,s_2)b^{\dagger}_{\alpha,s_1}(\v{k}_1^+, k_{1})d^{\dagger}_{\alpha,s_2}(\v{k}_2^+, -\v{k}_{1})|0\rangle\\
=& \sum_{s_1,s_2}\int_0^{\infty} \frac{dz}{4(2\pi) z(1-z)} \int d^2 \bm{z}_1 d^2 \bm{z}_2 \Psi^{T}_{\lambda}(z, \v r,s_1,s_2)b^{\dagger}_{\alpha,s_1}(\v{k}_1^+, \v{z_1})d^{\dagger}_{\alpha,s_2}(\v{k}_2^+, \v{z_2})|0\rangle\\
\end{split}
\end{equation}
where $s_{1,2}$ are polarizations of quark and anti-quark, $\alpha$ is the color index in the fundamental representation, $\lambda=\pm 1$ is the photon polarization, $r=z_1-z_2$. The longitudinal momentum of the (anti)quark is ${k}_1^+ = z p^+$ (${k}_2^+ = (1-z) p^+$). The above expression is written in the reference frame where the photon has zero transverse momentum. The photon splitting functions in the momentum and coordinate space are
\begin{align}
\Psi^{T}_{\lambda}(z,\v k_{1},s_1,s_2)&=-2ee_f\delta_{s_1,-s_2}(2z-1+2\lambda s_1)\sqrt{z(1-z)}\frac{\v k_{1}\cdot \v{\epsilon_{\lambda}}}{\v k_{1}^2+\varepsilon_f^2}
\\
\Psi^{T}_{\lambda}(z, \v r ,s_1,s_2)&=-i\frac{2ee_f}{2\pi}\delta_{s_1,-s_2}(2z-1+2\lambda s_1)\sqrt{z(1-z)}\frac{\v r \cdot \v{\epsilon_{\lambda} }}{|\v r|}\varepsilon_fK_1(\varepsilon_f|\v r|),
\end{align}
where $\varepsilon_f^2 = z(1-z)Q^2$. In numerical calculations we use a small but nonzero value of $Q\sim 200$ MeV, see Sec.~\ref{Sec:Num} for details.
\subsubsection{Correlators of the charge density.}
For the calculation we will need to know correlators of up to four color charge density operators. Here we calculate those in the dipole model of the photon.
The charge density operator in terms of the quark and anti quark creation and annihilation operators is defined as
\begin{align}
\hat{\rho}^a(x^-,\v x)= b^{\dagger}_{\alpha \sigma}(x^-, \v{x}) t^a_{\alpha \beta} b_{\beta \sigma}(x^-,\v{x})-d^{\dagger}_{\alpha \sigma}(x^-, \v{x}) t^a_{\beta \alpha } d_{\beta \sigma}(x^-,\v{x})
\end{align}
For the purpose of a CGC calculation we need the correlators of the integrated quantity
\begin{equation}
\hat\rho^a(\v x)\equiv \int dx^-\hat{\rho}^a(x^-,\v x)
\end{equation}
We now evaluate the correlators of $\hat\rho$ in a dipole state
$|q\bar{q}\rangle = b^{\dagger}_{\alpha \sigma}(\v{z_1})d^{\dagger}_{\alpha \sigma}(\v{z_2})|0\rangle$,
where we have suppressed indices and arguments irrelevant for this computation. \\[10pt]
Using
\begin{equation}
\begin{split}
&\hat{\rho}^a(\v x) b^{\dagger}_{h}(\v {z_1})d^{\dagger}_{h}(\v{z_2})|0\rangle\\
=& \left[ t^a_{\alpha h}b^{\dagger}_{\alpha}(\v{z_1})d^{\dagger}_{h}(\v{z_2}) \delta^{(2)}(\v{x}-\v{z_1})
-t^a_{ h \alpha}b^{\dagger}_{h}(\v{z_1})d^{\dagger}_{\alpha}(\v{z_2})\delta^{(2)}(\v{x}-\v{z_2})\right]|0\rangle
\end{split}
\end{equation}
and
\begin{equation}
\begin{split}
& \langle 0|d_{l}(\v{z_2})b_{l}(\v{z_1})\hat{\rho}^a(\v{x})\\
=&\langle 0| \left[ t^a_{l \beta}d_{\beta}(\v{z_2})b_{l}(\v{z_1}) \delta^{(2)}(\v{x}-\v{z_1})-t^a_{ \beta l}d_{\beta}(\v{z_2})b_{l}(\v{z_1})\delta^{(2)}(\v{x}-\v{z_2})\right]
\end{split}
\end{equation}
we obtain
\begin{equation}
\langle q\bar{q} | \hat \rho^{a}(\v{x_1}) \hat\rho^{b}(\v{x_2}) | q\bar{q}\rangle
=\frac{\delta^{ab}}{2}
\prod_{i=1,2}
\left( \delta^{(2)}(\v{x_i}-\v{z_1}) - \delta^{(2)}(\v{x_i}-\v{z_2}) \right)\,.
\end{equation}
For higher correlators, to simplify expressions, we account explicitly for the symmetry of the photon splitting function
$\v{z_1} \leftrightarrow \v{z_2}$ (and $z \leftrightarrow 1 -z$)\footnote{That is we
consider $|q\bar{q}\rangle = \frac12 \left[ b^{\dagger}_{\alpha \sigma}(\v{z_1},z)d^{\dagger}_{\alpha \sigma}(\v{z_2},1-z) +
b^{\dagger}_{\alpha \sigma}(\v{z_2},1-z)d^{\dagger}_{\alpha \sigma}(\v{z_1},z)\right]
|0\rangle$
}
to obtain
\begin{equation}
\begin{split}
& \langle q\bar{q} | \hat\rho^{a}(\v{x_1}) \hat\rho^{b}(\v{x_2}) \hat\rho^{c}(\v{x_3}) | q\bar{q}\rangle \\
=& \frac{if_{abc}}{4} \left( \delta^{(2)}(\v{x_2}-\v{z_1}) + \delta^{(2)}(\v{x_2}-\v{z_2}) \right) \prod_{i=1,3} \left( \delta^{(2)}(\v{x_i}-\v{z_1}) - \delta^{(2)}(\v{x_i}-\v{z_2}) \right)
\end{split}
\end{equation}
and
\begin{equation}
\begin{split}
& \langle q\bar q| \hat\rho^{a}(\v{x_1}) \hat\rho^{b}(\v{x_2}) \hat\rho^{c}(\v{x_3}) \hat\rho^{d}(\v{x_4})|q\bar q\rangle
= \Tr(t^at^bt^ct^d)\prod_{i=1}^4\left[ \delta^{(2)}(\v{x_i}-\v{z_1}) - \delta^{(2)}(\v{x_i}-\v{z_2}) \right]\\
-&\frac{f_{cde}f_{abe}}{8} \left[\delta^{(2)}(\v{x_2}-\v{z_1})+\delta^{(2)}(\v{x_2}-\v{z_2})\right] \left[\delta^{(2)}(\v{x_3}-\v{z_1}) + \delta^{(2)}(\v{x_3}-\v{z_2})\right]
\\ &\quad \quad \quad \times
\prod_{i=1,4}\left[ \delta^{(2)}(\v{x_i}-\v{z_1}) - \delta^{(2)}(\v{x_i}-\v{z_2}) \right]
\end{split}
\end{equation}
where $\Tr(t^a t^b t^c t^d)=\frac{1}{4N_c}\delta^{ab}\delta^{cd}+\frac{1}{8}\left(d^{abe}+if^{abe}\right)\left(d^{cde}+if^{cde}\right)$.
In the following we will need a representation of this trace in terms its fully symmetric and the remaining parts:
$\Tr(t^a t^b t^c t^d) = \Tr(t^a t^b t^c t^d)_{\rm sym}
+ \Tr(t^a t^b t^c t^d)_{\rm ns}$
with
\begin{equation}
\begin{split}
\Tr(t^a t^b t^c t^d)_{\rm sym}&=\frac{1}{12}\Big[\frac{1}{N_c}(\delta^{ab}\delta^{cd}+\delta^{ac}\delta^{bd}+\delta^{ad}\delta^{bc})+\frac{1}{2}(d_{abe}d_{cde}+d_{ace}d_{bde}+d_{ade}d_{bce}) \Big]
\end{split}
\end{equation}
and
\begin{equation}
\begin{split}
\Tr(t^a t^b t^c t^d)_{\rm ns}&=
\frac{1}{12} (f_{ade} f_{bce} - f_{abe} f_{cde})
+ \frac{1}{8} i (d_{abe} f_{cde} + d_{cde} f_{abe} )\,.
\end{split}
\end{equation}
For $N_c=3$, one can use the following identity (the factor of two was missing in Refs.~\cite{haber2016useful}; the validity of our identity can be easily checked by summing with respect to $a=b$)
\begin{align}
d_{abe}d_{cde}+d_{ace}d_{bde}+d_{ade}d_{bce}=\frac{2}{3}(\delta^{ab}\delta^{cd}+\delta^{ac}\delta^{bd}+\delta^{ad}\delta^{bc})
\end{align}
to obtain
\begin{equation}
\label{Eq:TrSym}
\begin{split}
\Tr(t^a t^b t^c t^d)^{N_c=3}_{\rm sym}
&=\frac{1}{18} (\delta^{ab}\delta^{cd}+\delta^{ac}\delta^{bd}+\delta^{ad}\delta^{bc})\,.
\end{split}
\end{equation}
We stress again that this expression is only true for $N_c=3$.
\subsection{The MV model for the projectile}
The other model we use to describe the photon wave function is the McLerran-Venugopalan model. It
posits that the averaging over the ``classical'' color charge density of the projectile should be performed with Gaussian weight, so that all correlators are expressed in terms of Wick contractions of the basic "propagator"
\begin{align}
\label{Eq:rho2mv1}
%
\langle \rho_a(\v x) \rho_b(\v y) \rangle_{\rm MV}
= \mu^2(x) \delta^{(2)}(\v x- \v y) \delta_{ab}\,.
\end{align}
Note that \eqref{Eq:rho2mv1} goes beyond the conventional MV model in that it includes the dependence of the color charge density on the impact parameter $\v{b} = \frac{\v{x}+\v{y}}{2}$. In principle one can go even further and account for the color neutrality by substituting a suitable function instead of $\delta^{(2)}(\v x- \v y)$, see e.g. Ref.~\cite{McLerran:2016ivs}. We will not do this in the current paper.
One has to be aware that the ``classical'' charge density in \eqref{Eq:rho2mv1} is not identical to the color charge density operator $\hat \rho$. To calculate the correlator of color charge density operators, one has to perform the operator symmetrization first, and use the MV for the completely symmetric parts. This was shown in \cite{PhysRevD.72.074023} and explained from the point of view of quantum-classical correspondence in \cite{Li:2020bys}.
For bilinears, the symmetrization is straightforward:
\begin{align}
\hat \rho_a(\v x) \hat \rho_b(\v y)
&= \left\{ \hat \rho_a(\v x), \hat \rho_b(\v y) \right\} + \left[ \hat \rho_a(\v x), \hat \rho_b(\v y) \right] \notag \\
&= \rho_a(\v x) \rho_b(\v y) - \frac{1}{2} \delta^{(2)}(\v x- \v y) T^{c}_{ab}
\rho_c (\v x)
\end{align}
where we have replaced the fully symmetric combinations by the classical color density $\left\{ \hat \rho_a(\v x), \hat \rho_b(\v y) \right\} \to \rho_a(\v x) \rho_b(\v y)$ and took into account that the commutator $\left[ \hat \rho_a(\v x), \hat \rho_b(\v y) \right] = - \frac{1}{2} \delta^{(2)}(\v x-\v y) T^{c}_{ab} \rho_c (\v x) $ as follows from the definition of $\hat \rho$. We thus see that
\begin{align}
\label{Eq:rho2mv}
%
\langle \hat \rho_a(\v x) \hat \rho_b(\v y) \rangle_{\rm MV}
= \langle \rho_a(\v x) \rho_b(\v y) \rangle_{\rm MV}
= \mu^2(x) \delta^{(2)}(\v x- \v y) \delta_{ab}\,.
\end{align}
This procedure can be extended to any number of operators $\hat \rho$.
For three operators we obtain:
\begin{align}
\notag
&\hat \rho_a(\v x) \hat \rho_b(\v y) \hat
\rho_c(\v z)
= \rho_a(\v x) \rho_b(\v y)
\rho_c(\v z) \\ &-
\frac12 \big( \DDelta{\v y}{\v z} T^e_{ b c} \rho_a(\v x) \rho_e (\v y) +
\DDelta{\v x}{\v z} T^e_{ a c} \rho_b (\v y) \rho_e(\v x) +
\DDelta{\v x}{\v y} T^e_{a b} \rho_e(\v x) \rho_c(\v z)\big)\,,
\end{align}
with $T^{a}_{bc} = - i f_{abc}$.
Here the terms linear in $\hat \rho$ were omitted as they do not contribute to the expectation value over the color invariant ensemble~\footnote{Note that this expression also reproduces the result of the previous section. Indeed, taking into account the symmetry of the dipole wave function we need to account only for the term quadratic in $\rho$
\begin{align}
\langle \rho_a(\v x) \rho_b(\v y) \rangle
= \langle \{ \rho_a(\v x) , \rho_b(\v y) \} \rangle
= \frac{\delta_{ab}}{ 2 } (\delta(\v x-\v {z_1})-\delta(\v{x}-\v{z_2}))(\delta(\v{y}-\v{z_1})-\delta(\v{y}-\v{z_2})) \,.
\notag
\end{align}
Therefore
\begin{align}
\langle \rho_a(\v x) \rho_b(\v y) \rho_c(\v z) \rangle
= &-
\frac1{4} T^{a}_{bc}
\Big[ \DDelta{\v y}{\v z} (\delta(\v x-\v{z_1})-\delta(\v{x}-\v{z_2}))(\delta(\v{y}-\v{z_1})-\delta(\v{y}-\v{z_2}))
\notag \\ & -
\DDelta{\v x}{\v z}
(\delta(\v {x} -\v{z_1})-\delta(\v{x}-\v{z_2}))(\delta(\v{y}-\v{z_1})-\delta(\v{y}-\v{z_2}))
\notag \\ & +
\DDelta{\v x}{\v y} (\delta(\v x- \v{z_1})-\delta(\v x-\v{z_2}))(\delta(\v{z}-\v{z_1})-\delta(\v{z}-\v{z_2})) \Big] \notag \\
& = -
\frac1 {4} T^{a}_{bc}
(\delta(\v{x}-\v{z_1})-\delta(\v{x}-\v{z_2}))
(\delta(\v{z}-\v{z_1})-\delta(\v{z}-\v{z_2}))
(\delta(\v{y}-\v{z_1})+\delta(\v{y}-\v{z_2}))\,.
\notag
\end{align}
}.
Thus in the MV model :
\begin{align}
\label{Eq:rho3mv}
\langle
\hat \rho_a(\v x) \hat \rho_b(\v y) \hat
\rho_c(\v z)
\rangle_{\rm MV}
&=
- \frac12 \DDelta{\v x}{\v y} \DDelta {\v y} {\v z} T^a_{bc} \, \mu^2 (\v x)
\end{align}
Similarly for four operators, we obtain
\begin{align}
\notag
&\langle
\hat \rho_a(\v x) \hat \rho_b(\v y) \hat
\rho_c(\v z) \hat\rho_d(\v u)
\rangle_{\rm MV}
\\ & =
\mu^2(\v x) \mu^2(\v z) \delta^{ab}\delta^{cd} \DDelta{\v x}{\v y} \DDelta{\v z}{\v u}
+ \mu^2(\v x) \mu^2(\v y) \delta^{ac}\delta^{bd} \DDelta{\v x}{\v z} \DDelta{\v y}{\v u}
\notag \\ &
+ \mu^2(\v x) \mu^2(\v z) \delta^{ad}\delta^{bc} \DDelta{\v x}{\v u} \DDelta{\v z}{\v y}
\notag \\ & +
\frac{\mu^2(x)}{6} \DDelta{\v x}{\v y} \DDelta {\v x} {\v z} \DDelta {\v x} {\v u} \left(
f_{a d e} f_{b c e} - f_{a b e} f_{c d e}
\right)
\label{Eq:rho4mv}
\end{align}
The last term here is of lower order in $\mu^2$ and is thus subleading for a projectile with high charge density. We do not neglect it here, but as will become clear later, it does play a somewhat different role than the first two terms even for a dilute projectile.
\subsection{Small $x$ gluon component of the projectile wave function}
\label{Sec:Coherent}
At the leading order, the dressing of the color charge by gluons of rapidity $\eta$ is given by the coherent operator:
\begin{align}
C_{\eta}=\exp{i\sqrt{2} \int d^2 \v x \hat{b}^i_a(\v x)\left[ a^{i\dagger}_a(\eta, \v x)+a^i_a(\eta, \v x) \right]}
\end{align}
where the creation/annihilation operators are introduced as the decomposition of the transverse
gluon field
\begin{equation}
\begin{split}
A^i_a(x^+,\v x)=\int_{0}^{\infty}\frac{dk^+}{2k^+(2\pi)}\left(a^i_a(k^+,\v x)e^{-ik^-x^+} +a^{i\dagger}_a(k^+,\v x)e^{ik^-x^+} \right)\,.
\end{split}
\end{equation}
The rapidity variable is defined here as
\begin{align}
d\eta=\frac{dk^+}{k^+}
\end{align}
and the corresponding
\begin{align}
a^i_a(\eta, \v x)={\sqrt{\frac{k^+}{4\pi}}} a^i_a(k^+, \v x).
\end{align}
The commutation relations for the creation/annihilation operators are
\begin{align}
\left[a^i_a(\eta, \v x), a^{j\dagger}_b(\eta', \v y)\right]=\delta^{ij}\delta^{ab} \delta^2(\eta-\eta') \delta(\v x - \v y)\,.
\end{align}
The Weizs\"acker-Williams field at the leading order is proportional to the color charge density operator
\begin{align}
\hat{b}^i_a(\v x)=\frac{g}{2\pi}\int d^2\v y \frac{(\v x-\v y)^i}{(\v x- \v y)^2} \hat{\rho}^a(\v y)\, .
\end{align}
The distribution (wave function) of $\hat \rho$ can be modelled by the previously discussed dipole or MV models.
Our calculation requires the knowledge of the incoming wave function at the rapidities of both observed gluons: $\eta$ and $\xi<\eta$. As we alluded to in the introduction, we restrict our calculation to the case when the difference between the rapidites of the observed gluons $\eta-\xi$ is large enough to use the eikonal vertex, that is $\eta-\xi \ll 1$, but small enough to neglect corrections due to evolution, that is $\eta-\xi \ll 1/\alpha_s$.
The gluon with rapidity $\eta$ itself contributes to the source for emission of gluons at lower rapidity. This can be accounted for by an appropriate redefinition of the Weizs\"acker-Williams field at lower rapidity $\hat{b}^i_a(\v x) \to \hat{b}^i_a(\v x) + \delta \hat{b}^i_a(\v x)$.
Thus at rapidity $\xi$ we have
\begin{align}
C_{\xi}=\exp{i\sqrt{2} \int d^2\v x \left(\hat{b}^i_a(\v x)+\delta \hat{b}^i_a(\v x)\right)\left[ a^{i\dagger}_a(\xi, \v x)+a^i_a(\xi, \v x) \right]}
\end{align}
where the extra contribution due to the presence of the gluon in the dipole wave function is given by
\begin{equation}
\begin{split}
\delta \hat{b}^i_a(\v x)&=\frac{g}{2\pi} \int_{\xi}^{\eta} d\zeta \int d^2y \frac{(\v x-\v y)^i}{(\v x- \v y)^2} \hat{\rho}_g^a(\zeta, \v y)
\end{split}
\end{equation}
where
\begin{align}
\hat{\rho}^a_g(\zeta, \v x)= a^{i\dagger}_b(\zeta, \v x) T^a_{bc} a_c(\zeta, \v x)\,.
\end{align}
Thus to the required order, the valence state dressed by the gluon cloud is
\begin{equation} |\Psi_{in}\rangle=C_{\xi}C_{\eta} |0\rangle |v\rangle
\end{equation}
\subsection{Scattering off the dense target}
\label{Sec:Scattering}
To describe a large nucleus, we use the classic field approximation for the gluon field of the target. In the eikonal approximation we have,
\begin{align}
\label{Eq:S_g}
\hat{S}^{\dagger}a^i_{a}(\zeta, \v{\bar{u}})\hat{S}&=U^{aa'}_{\v{\bar{u}}}a^i_{a'}(\zeta, \v{\bar{u}})
\end{align}
for gluon scattering, where $U$ is the infinite Wilson line in the adjoint representation:
\begin{equation}
U_{\v{x}}={\cal P}\exp{ig\int_{-\infty}^\infty dx^{+}T^aA_a^-(x^+,\v{x})}\,.
\end{equation}
Similarly for quark and antiquark
\begin{equation}
\begin{split}
\hat{S}^{\dagger}b_{\alpha}(\v{x})\hat{S}&=V_{\alpha \beta}(\v{x})b_{\beta}(\v{x})\\
\hat{S}^{\dagger}d_{\alpha}(\v{x})\hat{S}&=V^{\dagger}_{\beta \alpha }(\v{x})d_{\beta}(\v{x})
\end{split}
\end{equation}
where the Wilson line in the fundamental representation is given by
\begin{equation}
V(\v{x})={\cal P}\exp{ig\int_{-\infty}^\infty dx^{+}t^aA_a^-(x^+,\v{x})}\,.
\end{equation}
The consequence of these equations and the definition of the color charge density is that
\begin{align}
\hat{S}^{\dagger}\hat{\rho}_a(\v{\bar{u}})\hat{S}&=U^{aa'}_{\v{\bar{u}}}\hat{\rho}_{a'}(\v{\bar{u}})
\end{align}
and similarly
\begin{equation}
\hat{S}^{\dagger}\hat{\rho}_a(\zeta,\v{ \bar{u}})\hat{S}=a^{i\dagger}_{b}(\zeta, \v{\bar{u}}) a_{c}(\zeta, \v{\bar{u}})\left[U^{\dagger}_{\v{\bar{u}}} T^a U_{\v{\bar{u}}} \right]_{bc}
= U^{aa'}_{\v{\bar{u}}}
\hat{\rho}_{a'}(\zeta,\v{ \bar{u}})
\,.
\end{equation}
A calculation of a physical observable requires averaging over the ensemble of target fields which determines various correlators of the Wilson lines.
For a large nucleus at small x, the ensemble of the color fields of the target is frequently taken as an MV model with a relatively large value of the saturation momentum. This procedure in practice works well for calculation of a correlator of a small number of Wilson lines, but becomes increasingly complicated and cumbersome for more complicated correlators.
As we will see, the observable we work with leads to the expectations values of four Wilson lines in the adjoint representation. In the limit of large number of colors, this can be reduced to a product of up to two dipoles and a quadrupole in the fundamental representation in the coordinate space.
In order to get particle production cross section, this combination has to be Fourier-transformed with respect to four two-dimensional transverse momenta. Although, there are analytical expressions for the dipole and quadrupole in the McLerran-Venugopalan model, the Fourier transforms render the problem numerically prohibitively complex, as fast Fourier transform is not feasible in this multidimensional space while Monte-Carlo integration is extremely ineffective due to the presence of the oscillatory phases.
The two particle correlations at large $N_c$
only arise at order $1/N_c^2$~\cite{Kovner:2010xk,Kovchegov:2012nd,Kovchegov:2013ewa}. Partly this can be attributed to the nature of the correlations arising either from gluon HBT or gluon Bose-Einstein correlations~\cite{Altinoluk:2018ogz}. Therefore the large $N_c$ limit does not capture the physics we want to study.
Going beyond the large $N_c$ limit provides additional level of complication, as the cross section also involves sixtupoles and octupoles for which explicit analytical expressions, if exist, are too complex for a numerical simulation.
Instead we will use another approach applicable for a well evolved, near black disc, target. The approximation in question is the so-called {\it factorized dipole approximation} (FDA) discussed in Ref.~\cite{Kovner:2017ssr,Kovner:2018vec} and successfully applied to study particle production and correlations in Refs.~\cite{Altinoluk:2018ogz,Kovner:2018fxj,Kovner:2018azs,Li:2021ntt,Altinoluk:2020psk}.
The idea behind this approximation is extremely intuitive. Consider an arbitrary combination of even number of Wilson lines
$\langle U_{x_1} U_{x_2} \ldots U_{x_{2n}} \rangle$ which is multiplied by a non-trivial function of coordinates and integrated with respect to all two-dimensional vectors $x_i$, $i=1, \ldots, 2n$ over some finite, but large compared to the inverse saturation scale of the target squared, transverse area $S_\perp$. If a single Wilson line were to have a nonzero expectation value the largest contribution to the integral would be of order $S^{2n}_\perp$ arising from the integration over the part of the phase space where the coordinates of all Wilson lines are far away from each other. However for a dense target $\langle U_{ab} \rangle = 0$, and so the largest power of the transverse area possible is $S^{n}_\perp$, with corrections of order $S^{n-1}_\perp$.
The leading order contribution includes terms which contain smallest color singlets in the projectile propagating through the target. On one hand, any non singlet state that in the transverse plane separated by more than $1/Q_s$ from other propagating partons must have a vanishing S-matrix on the dense target due to color neutralization on the scale of $1/Q_s$, see e.g. \cite{Iancu:2001md,Ferreiro:2002kv,Iancu:2002xk,Mueller:2002pi}. On the other hand, if the singlet state contains more than two partons, one looses a power of the area when integrating over the coordinates of the partons. Thus the leading contribution in the dense target regime is the one where only dipole contribution to the S-matrix should be accounted for.
For illustration lets consider a four Wilson line observable
\begin{align}
&\int \prod_{i=1}^{4} d^2 \v x_i f(\v x_1, \ldots , \v x_4)
\langle {\rm tr} [U_{\v x_1} U_{\v x_2} U_{\v x_3} U_{\v x_4} ]\rangle
\notag \\ &=
\int \prod_{i=1}^{4} d^2 \v x_i f(\v x_1, \ldots ,\v x_4)\Big[
\langle U^{ab}_{\v x_1} U^{bc}_{\v x_2} \rangle
\langle U^{cd}_{\v x_3} U^{da}_{\v x_4} \rangle
\notag \\ & +
\langle U^{da}_{\v x_4} U^{ab}_{\v x_1}\rangle
\langle U^{bc}_{\v x_2} U^{cd}_{\v x_3}\rangle
+
\langle U^{ab}_{\v x_1} U^{cd}_{\v x_3} \rangle
\langle U^{bc}_{\v x_2} U^{da}_{\v x_4} \rangle
\Big]
+ {\cal O}(S_\perp) \notag
\end{align}
where the function $f$ is assumed to have support for all $x$ within the area $S_\perp$. The correlator $\langle U^{ab}_{\v x} U^{cd}_{\v y} \rangle$ is non zero only for a color singlet state, in this case -- the dipole:
\begin{align}
\label{Eq:KovnersHammer}
\langle U^{ab}_{\v x} U^{cd}_{\v y} \rangle = \frac{1}{N_c^2-1} \delta_{ac} \delta_{bd} D(|\v x-\v y|)\,,
\end{align}
where $D(r) = \frac{1}{N_c^2-1} {\rm tr} [ U^{\dagger}_{\v r} U_{0} ]$.
The quality of this approximation was explicitly checked for a number of Wilson line combinations in Ref.~\cite{Li:2021ntt}. The key necessary (albeit not sufficient) condition for its applicability is $Q^2_s S_{{\rm proj}, \perp} \gg 1$, where $Q_s$ is the saturation momentum of the nucleus and $S_{{\rm proj}, \perp}$ is the effective transverse area of the projectile. For LHC energies
this condition is satisfied if we use the natural $ S_{{\rm proj}, \perp} = \frac{1}{Q_{\rm eff}^2}$ where $Q_{\rm eff}$ is the larger between the QCD non-perturbative scale $\Lambda_{\rm QCD}$ and the virtuality of the photon. For the nearly real photon we thus have $S_{{\rm proj}, \perp} \sim \Lambda_{\rm QCD}^{-2}$. One has to be aware that when large momentum particle production is considered the effective area of the projectile might be significantly smaller than $\Lambda_{\rm QCD}^{-2}$. Although this is a process-dependent statement, in general it means that at large momenta $k\gg Q_s$, the FDA is not uniformly applicable; it however still can be justified for certain momentum configurations.
For calculations in this paper we apply the FDA to evaluate the correlators of more than
two Wilson lines. As our calculations are mostly in momentum space, the basic correlator used in the FDA is the Fourier transform of Eq.~\eqref{Eq:KovnersHammer}:
\begin{align}
\langle U^{ab}_{\v{p}} U^{cd}_{\v{q}} \rangle_T=\frac{(2\pi)^2}{N_c^2-1}\delta_{ac}\delta_{bd}\delta^{(2)}( \v{p}+ \v{q})D(\v{p})
\end{align}
where $D(\v{p})$ is
\begin{align}
D(\v{p})&=\frac{1}{(N^2_c-1) S_{\perp}}\langle \Tr( U^{\dagger}_{\v{p}}U_{\v{p}})\rangle_T\,.
\end{align}
We will also assume spatial isotropy of the target field ensemble, so that $D( \v{p}) = D(|\v p|)$.
We will use the MV model for the dipole. That is in the coordinate space, we have
\begin{align}
D(\v r) = \exp\left[-\frac{1}{4}Q_s^2r^2\ln(\frac{1}{\Lambda^2r^2}) \right]\,
\end{align}
with $D(|\v p|) = \int d^2 r e^{i \v p\cdot \v r} D(\v r)$.
\section{Double inclusive gluon production - general expressions}\label{sec3}
The differential cross section for double inclusive gluon production with rapidities and the transverse momenta $(\eta, \v{q}_1)$ and $(\xi, \v{q}_2)$ is given by
\begin{equation}
\label{Eq:Double}
\begin{split}
\frac{d{\cal N}}{d\eta d \v{q}_1^2 d\xi d \v{q}_2^2}=&\frac{1}{(2\pi)^4} \int
d^2 \v u_1 d^2\v u_2
d^2 \bar{\v u}_1 d^2 \bar{\v u}_2
e^{-i\v{q}_1(\v{u}_1-\v{\bar{u}}_1)}e^{-i\v{q}_2(\v{u}_2-\v{\bar{u}}_2)}
\\ &\times
\langle \gamma^*|C^{\dagger}\hat{S}^{\dagger}C a_{i,a}^{\dagger}(\eta, \v{u}_1)a_{j,b}^{\dagger}(\xi, \v{u}_2)
a_{i,a}(\eta, \v{\bar{u}}_1)a_{j,b}(\xi, \v{\bar{u}}_2)C^{\dagger}\hat{S}C| \gamma^* \rangle
\end{split}
\end{equation}
The coherent operator $C$ is given by the product of $C = C_\xi C_\eta$, see Sec.~\ref{Sec:Coherent}. For double-inclusive gluon production, it is sufficient to expand the coherent operator to the leading order in the argument:
\begin{equation}
\begin{split}
C_{\eta}\simeq& 1+i\sqrt{2} \int d^2\v v_1 \hat{b}^i_{a}(\v{v}_1)\left[ a^{i\dagger}_a(\eta, \v{v}_1)+a^i_a(\eta, \v{v}_1) \right]\,,\\
C_{\xi}\simeq& 1+i\sqrt{2} \int d^2\v v_2 \left(\hat{b}^j_{b}(\v{v}_2)+\delta \hat{b}^j_b(\eta, \v{v}_2)\right)\left[ a^{j\dagger}_b(\xi, \v{v}_2)+a^j_b(\xi, \v{v}_2)\right] \,.
\end{split}
\end{equation}
Hence to the required order, the coherent operator can be approximated by $C_{\rm eff}$
\begin{equation}
\begin{split}
C_{\rm eff} =&1+i\sqrt{2} \int d^2\v v_1 \hat{b}^i_{a}(\v v_1)\left[ a^{i\dagger}_a(\eta, \v{v}_1)+a^i_a(\eta, \v{v}_1) \right]\\
&+i\sqrt{2} \int d^2\v v_2 \left(\hat{b}^j_{b}(\v{v}_2)+\delta \hat{b}^j_b(\eta, \v{v}_2)\right)\left[ a^{j\dagger}_b(\xi, \v{v}_2)+a^j_b(\xi, \v{v}_2)\right] \\
&-2 \int d^2\v v_1d^2\v v_2 \left(\hat{b}^j_{b}(\v{v}_2)+\delta \hat{b}^j_b(\eta, \v{v}_2)\right)\left[ a^{j\dagger}_b(\xi, \v{v}_2)+a^j_b(\xi, \v{v}_2)\right] \notag \\ &\quad \quad \quad \quad \quad \quad \times \hat{b}^i_{a}(\v{v}_1)\left[ a^{i\dagger}_a(\eta, \v{v}_1)+a^i_a(\eta, \v{v}_1) \right]\,.
\end{split}
\end{equation}
The matrix element in Eq.~\eqref{Eq:Double}, can be decomposed into three different contributions classified by the number of gluons emitted directly from the photon state $\mathcal{M}=\Sigma_2+\Sigma_3+\Sigma_4$, or equivalently, the powers of color charge density, schematically,
\begin{align}
b^2(b+\delta b)^2=b^4+2b^3 \delta b+ b^2(\delta b)^2\,.
\end{align}
Here every factor of $b$ corresponds to a gluon emitted directly from the photon, while a factor of $\delta b$ corresponds to a gluon emitted from another gluon with larger rapidity.
We will denote the corresponding contribution to the matrix element by $\Sigma_n$, where $n$ is the number of gluons emitted directly from the projectile. To elaborate on this: $\Sigma_2$ contributes to the probability of the process where a gluon with rapidity $\eta$ is emitted from the photon, and subsequently this gluon splits into a pair of gluons which are then observed in the final state. The splitting can occur either before of after the scattering on the target field. $\Sigma_4$ contributes to the probability of the process where both gluons are emitted directly from the photon state. Finally $\Sigma_3$ represents the interference of the previous two contributions.
We now turn to evaluating the production amplitude.
\begin{figure}[t!]
\centering
\includegraphics[width=0.4\textwidth]{Sigma2.pdf} \quad \quad
\includegraphics[width=0.4\textwidth]{Sigma4.pdf}
\caption{An illustration of two types of contribution to the two gluon production amplitude. Note that the number of actual diagrams is significantly larger, as the position of the shock wave can vary. For example, in total, there are six diagrams for the topology of the left diagram. }
\end{figure}
\subsection{Production amplitude}
Our goal is to evaluate the production amplitude in the coordinate space defined by
\begin{align}
{\cal A}(\bar{\v u}_1,\bar{\v u}_2) =
\hat S^\dagger a_{i,a}(\eta, \v{\bar{u}}_1)a_{j,b}(\xi, \v{\bar{u}}_2)C^{\dagger}\hat{S}C| \gamma^* \rangle\,.
\end{align}
We have introduced the factor $\hat S^\dagger$ for convenience, although it cancels in the calculation in the matrix element ${\cal M} = {\cal A}^\dagger {\cal A}$.
Using Eq.~\eqref{Eq:S_g}, we arrive at
\begin{align}
{\cal A}(\bar{\v u}_1,\bar{\v u}_2) =
U^{aa'}_{\v{\bar{u}}_1}U^{bb'}_{\v{\bar{u}}_2}
a_{ i,a}(\eta, \v{\bar{u}}_1)a_{j,b}(\xi, \v{\bar{u}}_2)
\hat S^\dagger C^{\dagger}\hat{S}C| \gamma^* \rangle \equiv {\cal A}_1+{\cal A}_2\,.
\end{align}
The operator $\hat S^\dagger C^{\dagger}\hat{S}C$ has terms proportional to $ b$ and to $b^2$, and these terms have been separated in the previous equation into the two amplitudes ${\cal A}_1$ and ${\cal A}_2$.
These amplitudes are directly related to the aforementioned $\Sigma_n$, $\Sigma_2 = {\cal A}_1{\cal A}^*_1$, $\Sigma_3 = {\cal A}_1{\cal A}^*_2+{\cal A}^*_1{\cal A}_2$ and $\Sigma_4 = {\cal A}_2{\cal A}^*_2$.
\subsubsection{Amplitude ${\cal A}_1$}
Expanding the coherent operators and collecting the terms of first order in $b$, we obtain
the production amplitude for two gluons with colors $a,b$, polarizations $i,j$ at positions $\v{\bar{u}}_1$ and $\v{\bar{u}}_2$:
\begin{equation}
\begin{split}
&{\cal A}_1(\bar{\v u}_1,\bar{\v u}_2)\\
&= 2\int d^2\v v_1d^2\v v_2
U^{aa'}_{\v{\bar{u}}_1} U^{bb'}_{\v{\bar{u}}_2}a^i_{a'}(\eta, \v{\bar{u}}_1)
a^j_{b'}(\xi, \v{\bar{u}}_2)
\left[ \hat S^{\dagger} \delta \hat{b}^k_c(\eta, \v{v}_2)a^{k\dagger}_c(\xi, \v{v}_2)\hat S\hat{b}^l_{d}(\v{v}_1)a^{l\dagger}_d(\eta, \v{v}_1) \right.\\
&\left.- \hat S^{\dagger} \hat{b}^l_{d}(\v{v}_1)a^{l\dagger}_d(\eta, \v{v}_1)\delta \hat{b}^k_c(\eta, \v{v}_2)a^{k\dagger}_c(\xi, \v{v}_2)\hat S- \delta \hat{b}^k_c(\eta, \v{v}_2)a^{k\dagger}_c(\xi, \v{v}_2)\hat{b}^l_{d}(\v{v}_1)a^{l\dagger}_d(\eta, \v{v}_1) \right]|\gamma^*\rangle
\end{split}\label{Eq:am1}
\end{equation}
The second term in this expression gives zero when the operator $\delta \hat{b}^k_c(\eta, \v{v}_2)$ acts on the gluon vacuum at rapidity $\eta$ (we always assume that $\xi < \eta$).
We thus have two non-trivial terms which can be readily evaluated. The first one gives
\begin{equation}
\begin{split}
&U^{aa'}_{\v{\bar{u}}_1}U^{bb'}_{\v{\bar{u}}_2}\int d^2\v v_1d^2\v v_2 a^i_{a'}(\eta, \v{\bar{u}}_1)a^j_{b'}(\xi, \v{\bar{u}}_2) \hat S^{\dagger} \delta \hat{b}^k_c(\eta, \v{v}_2)a^{k\dagger}_c(\xi, \v{v}_2)\hat S\hat{b}^l_{d}(\v{v}_1)a^{l\dagger}_d(\eta, \v{v}_1)|\gamma^*\rangle\\
&= \int d^2 \v x f^j({\bar{\v u}}_2-{\bar{\v u}}_1)f^i(\v{\bar{u}}_1-\v{x})\left[ T^b U_{\v{\bar{u}}_1} \right]_{af}\hat\rho_f(\v{x})|\gamma^*\rangle
\end{split}
\end{equation}
where
\begin{align}
f^i(\v{x}) = \frac{ g } { 2\pi} \frac{ \v x_i} {\v x^2} \,.
\end{align}
The last term in Eq.~\eqref{Eq:am1} simplifies into
\begin{equation}
\begin{split}
&U^{aa'}_{\v{\bar{u}}_1}U^{bb'}_{\v{\bar{u}}_2}\int d^2\v v_1d^2\v v_2 a^i_{a'}(\eta, \v{\bar{u}}_1)a^j_{b'}(\xi, \v{\bar{u}}_2) \delta \hat{b}^k_c(\eta, \v{v}_2)a^{k\dagger}_c(\xi, \v{v}_2)\hat{b}^l_{d}(\v{v}_1)a^{l\dagger}_d(\eta, \v{v}_1)|\gamma^*\rangle\\
=&\int d^2 \v x U^{aa'}_{\v{\bar{u}}_1}U^{bb'}_{\v{\bar{u}}_2}T^{a'}_{db'} f^j(\v{\bar{u}}_2-\v{\bar{u}}_1)f^i(\v{\bar{u}}_1-\v{x})\hat\rho_d(\v{x})|\gamma^*\rangle\,.
\end{split}
\end{equation}
Combining these two contributions we complete ${\cal A}_1$
\begin{equation}
\begin{split}
{\cal A}_1(\bar{\v u}_1,\bar{\v u}_2)=&2\int d^2 \v x f^j(\v{\bar{u}}_2-\v{\bar{u}}_1)f^i(\v{\bar{u}}_1-\v{x}) (U_{\v{\bar{u}}_2}-U_{\v{\bar{u}}_1})^{bb'}[U^{\dagger}_{\v{\bar{u}}_1}T^{a}U_{\v{\bar{u}}_1}]_{b'd} \hat\rho_d(\v{x})
\end{split}
\end{equation}
where to obtain the final expression we used the identity
\begin{align}
U^{aa'}_{\v{\bar{u}}_1}T^{a'}_{db'}= [U^{\dagger}_{\v{\bar{u}}_1}T^{a}U_{\v{\bar{u}}_1}]_{db'} \,.
\end{align}
\subsubsection{Amplitude ${\cal A}_2$}
Following similar calculations, we obtain:
\begin{equation}
\begin{split}
&{\cal A}_2(\bar{\v u}_1,\bar{\v u}_2)\\
&=2\int d^2\bm{x}d^2\bm{y}f^i(\bar{\v u}_1-\bm{x})f^j(\bar{\v u}_2-\bm{y})\left[U^{ae}_{\bm{x}}U^{bc}_{\bar{\v u}_2}\hat\rho_e(\bm{x})\hat\rho_{c}(\bm{y})+U^{bc}_{\bm{y}}U^{ae}_{\bar{\v u}_1}\hat\rho_{c}(\bm{y})\hat\rho_{e}(\bm{x})\right.\\
&\left.-U^{ae}_{\bm{x}}U^{bc}_{\bm{y}}\hat\rho_e(\bm{x})\hat\rho_c(\bm{y})-U^{ae}_{\bar{\v u}_1}U^{bc}_{\bar{\v u}_2}\hat{\rho}_{c}(\bm{y})\hat{\rho}_{e}(\bm{x})\right]
\\&= 2\int d^2\bm{x}d^2\bm{y}f^i(\bar{\v u}_1-\bm{x})f^j(\bar{\v u}_2-\bm{y})
\\ &\times \left[U^{ae}_{\bm{x}} \left( U^{bc}_{\bar{\v u}_2}
- U^{bc}_{\bm{y}}
\right)
\hat\rho_e(\bm{x})\hat\rho_{c}(\bm{y})+
U^{ae}_{\bar{\v u}_1}
\left(U^{bc}_{\bm{y}}-U^{bc}_{\bar{\v u}_2}\right)\hat\rho_{c}(\bm{y})\hat\rho_{e}(\bm{x})\right]\,.
\end{split}
\end{equation}
If we were to forget about the quantum nature of the color charge density operators (neglect their commutation relations) the later expression would be identical to the frequently used expression for two gluon production in the dilute-dense limit, see e.g. Appendix A of Ref.~\cite{Kovner:2012jm}.
\subsection{Matrix elements $\Sigma$}
We now compute the matrix elements.
The matrix element proportional to the second power of the projectile charge density is:
\begin{equation}
\begin{split}
&\Sigma_2(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )= {\cal A}_{1'}^*(\v u_1,\v u_2){\cal A}_1(\bar{\v u}_1,\bar{\v u}_2) \\
&=4\int d^2\bm{x}\int d^2\bar{\bm{x}}f^i(\bar{\v u}_1-\bm{x})f^i(\v u_1-\bar{\bm{x}})f^j(\bar{\v u}_2-\bar{\v u}_1)f^j(\v u_2-\v u_1)\left<\hat \rho_{d'}(\bar{\bm{x}})\hat \rho_{d}(\bm{x})\right>_{P}\\
&\left<\left[ [U^{\dagger}_{\v u_1}T^{a}U_{\ vu_1}][U^{\dagger}_{\v u_2}-U^{\dagger}_{\v u_1}] [U_{\bar{\v u}_2}-U_{\bar{\v u}_1}][U^{\dagger}_{\bar{\v u}_1}T^{a}U_{\bar{\v u}_1}]\right]_{d'd} \right>_{T}\\
&=4\int d^2\bm{x}\int d^2\bar{\bm{x}}f^i(\bar{\v u}_1-\bm{x})f^i(\v u_1-\bar{\bm{x}})f^j(\bar{\v u}_2-\bar{\v u}_1)f^j(\v u_2-\v u_1)\left<\hat \rho_{d'}(\bar{\bm{x}})\hat \rho_{d}(\bm{x})\right>_{P}\\
&\left<\Tr(U_{\v u_1}T^{e}[U^{\dagger}_{\v u_2}-U^{\dagger}_{\v u_1}][U_{\bar{\v u}_2}-U_{\bar{\v u}_1}]T^eU^{\dagger}_{\bar{\v u}_1})\right>_T
\end{split}
\end{equation}
The interference terms involving the cubic power of the projectile charge density:
\begin{equation}
\begin{split}
&\Sigma_3(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )= {\cal A}_{2'}^*(\v u_1,\v u_2){\cal A}_1(\bar{\v u}_1,\bar{\v u}_2)+{\cal A}_{1'}^*(\v u_1,\v u_2){\cal A}_{2}(\bar{\v u}_1,\bar{\v u}_2) \\
&=4 \int d^2\bar{\bm{x}}d^2\bar{\bm{y}} d^2\bm{x}f^j(\bar{\v u}_2-\bar{\v u}_1)f^i(\bar{\v u}_1-\bm{x})f^i(\v u_1-\bar{\bm{x}})f^j(\v u_2-\bar{\bm{y}})\\
& \times \left(\left<\hat{\rho}_{c'}(\bar{\bm{y}})\hat{\rho}_{e'}(\bar{\bm{x}}) \hat{\rho}_{d}(\bm{x})\right>_{P} \right.\\
& \times \left< U^{\dagger e'a}(\bar{\bm{x}}) \left[ [U^{\dagger}_{\v u_2}-U^{\dagger}_{\bar{\bm{y}}}] [U_{\bar{\v u}_2}-U_{\bar{\v u}_1}][U^{\dagger}_{\bar{\v u}_1}T^{a}U_{\bar{\v u}_1}]\right]_{c'd} \right>_T\\
&\left. + \left<\hat{\rho}_{e'}(\bar{\bm{x}})\hat{\rho}_{c'}(\bar{\bm{y}}) \hat{\rho}_{d}(\bm{x})\right>_{P} \right.\\
& \left. \times \left< U^{\dagger e'a}(\v u_1) \left[ [U^{\dagger}_{\bar{\bm{y}}}-U^{\dagger}_{\v u_2}] [U_{\bar{\v u}_2}-U_{\bar{\v u}_1}][U^{\dagger}_{\bar{\v u}_1}T^{a}U_{\bar{\v u}_1}]\right]_{c'd} \right>_T \right) \\
&+4 \int d^2\bar{\bm{x}}d^2\bm{y} d^2\bm{x}f^j(\v u_2-\v u_1)f^i(\v u_1-\bar{\bm{x}})f^i(\bar{\v u}_1-\bm{x})f^j(\bar{\v u}_2-\bm{y})\\
& \times \left(\left< \hat{\rho}_{d'}(\bar{\bm{x}})\hat{\rho}_{e}(\bm{x})\hat{\rho}_{c}(\bm{y})\right>_{P} \right.\\
& \times \left< \left[ [U^{\dagger}_{\v u_1}T^{a}U_{\v u_1}] [U^{\dagger}_{\v u_2}-U^{\dagger}_{\v u_1}][U_{\bar{\v u}_2}-U_{\bm{y}}]\right]_{d'c} U^{ ae}(\bm{x}) \right>_T\\
&\left. + \left< \hat{\rho}_{d'}(\bar{\bm{x}})\hat{\rho}_{c}(\bm{y})\hat{\rho}_{e}(\bm{x})\right>_{P} \right.\\
& \left. \times \left< \left[ [U^{\dagger}_{\v u_1}T^{a}U_{\v u_1}] [U^{\dagger}_{\v u_2}-U^{\dagger}_{\v u_1}][U_{\bm{y}}-U_{\bar{\v u}_2}]\right]_{d'c} U^{ ae}(\bar{\v u}_1) \right>_T \right)
\end{split}
\end{equation}
Finally the term with the highest power of the projectile density is
\begin{equation}
\begin{split}
&\quad\Sigma_4(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )\\&= {\cal A}_{2'}^*(\v u_1,\v u_2){\cal A}_2(\bar{\v u}_1,\bar{\v u}_2) \\
&=4\int d^2\bm{x}d^2\bm{y}\int d^2\bar{\bm{x}}d^2\bar{\bm{y}}f^i(\bar{\v u}_1-\bm{x})f^j(\bar{\v u}_2-\bm{y})f^i(\v u_1-\bar{\bm{x}})f^j(\v u_2-\bar{\bm{y}})\\
& \times \left(\left<\hat{\rho}_{c'}(\bar{\bm{y}})\hat{\rho}_{e'}(\bar{\bm{x}}) \hat{\rho}_{e}(\bm{x})\hat{\rho}_{c}(\bm{y})\right>_{P} \left< [U^{\dagger}_{\bar{\bm{x}}}U_{\bm{x}}]_{e'e} \left[ [U^{\dagger}_{\v u_2}-U^{\dagger}_{\bar{\bm{y}}}][U_{\bar{\v u}_2}-U_{\bm{y}}]\right]_{c'c} \right>_T \right.
\\
&\left.+\left<\hat{\rho}_{e'}(\bar{\bm{x}})\hat{\rho}_{c'}(\bar{\bm{y}}) \hat{\rho}_{e}(\bm{x})\hat{\rho}_{c}(\bm{y})\right>_{P} \left< [U^{\dagger}_{\v u_1}U_{\bm{x}}]_{e'e} \left[ [U^{\dagger}_{\bar{\bm{y}}}-U^{\dagger}_{\v u_2}][U_{\bar{\v u}_2}-U_{\bm{y}}]\right]_{c'c} \right>_T \right.
\\
&\left.+\left<\hat{\rho}_{e'}(\bar{\bm{x}})\hat{\rho}_{c'}(\bar{\bm{y}}) \hat{\rho}_{c}(\bm{y})\hat{\rho}_{e}(\bm{x})\right>_{P} \left< [U^{\dagger}_{\v u_1}U_{\bar{\v u}_1}]_{e'e} \left[ [U^{\dagger}_{\bar{\bm{y}}}-U^{\dagger}_{\v u_2}][U_{\bm{y}}-U_{\bar{\v u}_2}]\right]_{c'c} \right>_T \right.\\
&\left.+\left<\hat{\rho}_{c'}(\bar{\bm{y}})\hat{\rho}_{e'}(\bar{\bm{x}}) \hat{\rho}_{c}(\bm{y})\hat{\rho}_{e}(\bm{x})\right>_{P} \left< [U^{\dagger}_{\bar{\bm{x}}}U_{\bar{\v u}_1}]_{e'e} \left[ [U^{\dagger}_{\v u_2}-U^{\dagger}_{\bar{\bm{y}}}][U_{\bm{y}}-U_{\bar{\v u}_2}]\right]_{c'c} \right>_T \right)
\end{split}
\end{equation}
\section{Inclusive two gluon production: the dipole model}\label{sec4}
\label{Sect:2glDPM}
The two particle inclusive cross section in terms of the matrix element $\Sigma = \Sigma_2 + \Sigma_3 + \Sigma_4$ computed previously, reads
\begin{equation}
\begin{split}
\frac{dN}{d\eta d \v{q}_1^2 d\xi d \v{q}_2^2}&=\frac{1}{4} \sum_{s_1,s_2,\lambda}\int_{0}^{1}\frac{dz}{2\pi z(1-z)} \int_{\v z_1,\v z_2} \Psi^{T*}_{\lambda}(z,\v z_1-\v z_2,s_1)\Psi^{T}_{\lambda}(z,\v z_1-\v z_2,s_1)\\ & \times \frac{1}{(2\pi)^4} \int_{\v u_1, \bar{\v u}_1,\v u_2, \bar{\v u}_2} e^{-i \v{q}_1(\v u_1-\bar{\v u}_1)}e^{-i \v{q}_2(\v u_2-\bar{\v u}_2)}\Sigma(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )\,.
\end{split}
\end{equation}
Summing over the quark and (transverse) photon polarizations we obtain
\begin{align}
\sum_{s_1,s_2,\lambda}|\Psi^{T}_{\lambda}(z,\v z_1-\v z_2,s_1)|^2
=&\frac{32e^2e_f^2}{(2\pi)^2}(z^2+ {\bar z}^2)z \bar z \varepsilon^2_fK_1(\varepsilon_f|\v r|)^2
\end{align}
where we introduced $\bar z=1-z$ to simplify notation.
In principle $\varepsilon_f$ is a function of $z$; however for small $\varepsilon_f \propto Q$, the combination $\varepsilon_f^2 K_1^2(\varepsilon_f r)$ is approximately $\varepsilon_f$ and therefore $z$ independent. We thus set $z$ to $1/2$ in $\varepsilon_f^2 K_1^2(\varepsilon_f r)$ and integrate the rest of $z$ dependence explicitly:
\begin{align}
\int^1_0 dz \, (z^2+{\bar z}^2)=\frac{2}{3}\,.
\end{align}
Hence the cross section reads
\begin{equation}
\begin{split}
\frac{dN}{d\eta d \v{q}_1^2 d\xi d \v{q}_2^2}&=\frac{16e^2e_f^2}{3 (2\pi)^3}\int_{\bm{z}_1,\bm{z}_2}\varepsilon^2_fK_1(\varepsilon_f|\bm{z}_1-\bm{z}_2|)^2 \\
&\times \frac{1}{(2\pi)^4} \int_{u_1, \bar{u}_1,u_2, \bar{u}_2} e^{-i \v{q}_1(\v u_1-\bar{\v u}_1)}e^{-i \v{q}_2(\v u_2-\bar{\v u}_2)}\Sigma(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )
\end{split}
\end{equation}
where $\varepsilon_f = Q/2$.
Numerical calculations are easier performed in momentum space, therefore in the following we present all expressions in terms of momentum integrals.
It is convenient to split the production cross section into three terms corresponding to $\Sigma_2$, $\Sigma_3$ and $\Sigma_4$.
Using the correlators of the color charge density in the dipole, calculated in Section 2, we obtain for the contribution of $\Sigma_2$:
\begin{equation}
\label{Eq:N2}
\begin{split}
\frac{dN^{(2)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}&=\frac{16 e^2e_f^2}{3 \pi^4} \alpha_s^2\int_{P, \v{k} ,\v{\bar{k}} } \frac{\mathcal{I}(\varepsilon_f,P)}{P^2}\, \Gamma(\v{q}_2,\v k,\v{\bar{k}})
\\
&\Tr(U_{\v{q}_1+\v{k}-\v{P}}T^{e}U^{\dagger}_{\v{k}-\v{q}_2}U_{\v{\bar{k}}-\v{q}_2}T^eU^{\dagger}_{\v{q}_1+\v{\bar{k}} -\v{P}})\,.
\end{split}
\end{equation}
In Eq.~\eqref{Eq:N2} we introduced
\begin{align}
\mathcal{I}_{\varepsilon_f}(|\v P|) &=
\int_0^{\infty} dr r \epsilon_f^2 K^2_1(\epsilon_f |\v r|)(1-J_0(|\v r||\v P|))\notag \\ &=
-\frac{1}{2}+\frac{(\v P^2+2\epsilon_f^2)\ln(\frac{|\v P|}{2\epsilon_f}+\sqrt{1+\frac{\v P^2}{4 \epsilon_f^2}}) }{|\v P|\sqrt{\v P^2+4\epsilon_f^2}} ,
\\
\Gamma (\v{q}_2,\v{k},\v{\bar{k}}) &= \left(\frac{\v q^i_2}{\v{q}_2^2}-\frac{ \v k^i}{ \v k^2}\right) \left(\frac{\v q^i_2}{\v{q}_2^2}-\frac{\v{\bar{k}}^i}{\v{\bar{k}}^2}\right)\,.
\end{align}
Note that the ratio $\mathcal{I}(\varepsilon_f,P) / P^2$ as a function of $P^2$ approaches a finite constant in the IR($P/\varepsilon_f\to0$).
The contributions involving $\Sigma_3$ cannot be simplified into one homogeneous expression.
Instead we have
\begin{equation}
\frac{dN^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}=\frac{dN_1^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}+\frac{dN_2^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}+\frac{dN_3^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}+\frac{dN_4^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}
\end{equation}
We list all four terms separately.
\begin{equation}
\begin{split}
\frac{dN_1^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}
=&
- \frac{8 e^2e_f^2}{3 \pi^4} \alpha_s^2
\int_{\v{k}_1,\v{k}_2,\v{k}_3} \left\langle \Tr \left[T^d U^{\dagger}_{\v{k}_1}U_{\v{k}_2}T^dU^{\dagger}_{\v{k}_3}U_{\v{k}_1-\v{k}_2+\v{k}_3}\right]\right\rangle_T \\
&\times\frac{\v{q}_1 \cdot (\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_3)}{\v{q}_1^2(\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_3)^2}\Gamma(\v{q}_2,\v{q}_2+\v{k}_1,\v{q}_2+\v{k}_2)\\
&\times\Big[\mathcal{I}_{\varepsilon_f}( |\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_3|)+\mathcal{I}_{\varepsilon_f}( |\v{q}_2+\v{k}_1|)-\mathcal{I}_{\varepsilon_f}( |\v{q}_1-\v{k}_1+\v{k}_2-\v{k}_3|)\Big]\,.
\end{split}
\end{equation}
Note that this expression does not have poles at $\v{q}_2+\v{k}=0$ and $\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_3=0$ as the combination of $\mathcal{I}$ functions vanishes at these points.
Similarly for the remaining three contributions we have
\begin{equation}
\begin{split}
\frac{dN_2^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}&=-
\frac{8 e^2e_f^2}{3 \pi^4} \alpha_s^2
\int_{\v{k}_1,\v{k}_2,\v{k}_3}\left\langle \Tr \left[T^d U^{\dagger}_{\v{k}_1}U_{\v{k}_2}T^dU^{\dagger}_{\v{k}_3}U_{\v{k}_1-\v{k}_2+\v{k}_3} \right]\right\rangle_T \\
&\times\frac{(\v{q}_1-\v{k}_1+\v{k}_2-\v{k}_3) \cdot
(\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_3)}{(\v{q}_1-\v{k}_1+\v{k}_2-\v{k}_3)^2(\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_3)^2} \Gamma(\v{q}_2,\v{q}_2+\v{k}_1,\v{q}_2+\v{k}_2) \\
&\times \Big[\mathcal{I}_{\varepsilon_f}( |\v{q}_1-\v{k}_1+\v{k}_2-\v{k}_3|)+\mathcal{I}_{\varepsilon_f}( |\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_3|)-\mathcal{I}_{\varepsilon_f}( |\v{q}_2+\v{k}_1|)\Big]\,,
\end{split}
\end{equation}
\begin{equation}
\begin{split}
\frac{dN_3^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}&=
- \frac{8 e^2e_f^2}{3 \pi^4} \alpha_s^2
\int_{\v{k}_1,\v{k}_2,\v{k}_3}
\left \langle \Tr \left [T^{d} U^{\dagger}_{\v{k}_1}U_{\v{k}_2}T^{d}U^{\dagger}_{\v{k}_3}U_{\v{k}_1-\v{k}_2+\v{k}_3} \right] \right\rangle_T \\
&\times\frac{\v{q}_1\cdot (\v{q}_1+\v{q}_2-\v{k}_2+\v{k}_3)}{\v{q}_1^2(\v{q}_1+\v{q}_2-\v{k}_2+\v{k}_3)^2}\Gamma(\v{q}_2,\v{q}_2+\v{k}_3,\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3)\\
&\times \Big[\mathcal{I}_{\varepsilon_f}( |\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)+\mathcal{I}_{\varepsilon_f}( |\v{q}_1+\v{q}_2-\v{k}_2+\v{k}_3|)-\mathcal{I}_{\varepsilon_f}( |\v{q}_1-\v{k}_1|)\Big]\,,
\end{split}
\end{equation}
and, finally,
\begin{equation}
\begin{split}
\frac{dN_4^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}&=
- \frac{8 e^2e_f^2}{3 \pi^4} \alpha_s^2
\int_{\v{k}_1,\v{k}_2,\v{k}_3}\left\langle \Tr \left[T^d U^{\dagger}_{\v{k}_1}U_{\v{k}_2}T^{d}U^{\dagger}_{\v{k}_3}U_{\v{k}_1-\v{k}_2+\v{k}_3}\right]\right\rangle_T \\
&\times\frac{(\v{q}_1-\v{k}_1)\cdot (\v{q}_1+\v{q}_2-\v{k}_2+\v{k}_3)}{(\v{q}_1-\v{k}_1)^2(\v{q}_1+\v{q}_2-\v{k}_2+\v{k}_3)^2}\Gamma(\v{q}_2,\v{q}_2+\v{k}_3,\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3)\\
&\times \Big[\mathcal{I}_{\varepsilon_f}( |\v{q}_1-\v{k}_1|)+\mathcal{I}_{\varepsilon_f}( |\v{q}_1+\v{q}_2-\v{k}_2+\v{k}_3|)-\mathcal{I}_{\varepsilon_f}( |\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)\Big]\,.
\end{split}
\end{equation}
Switching to the contribution from $\Sigma_4$ involving four $\rho(x)$ correlators, we again obtain four terms that cannot be obviously combined:
\begin{equation}
\frac{dN^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}=\frac{dN_1^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}+\frac{dN_2^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}+\frac{dN_3^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}+\frac{dN_4^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}
\end{equation}
The explicit expressions are:
\begin{equation}
\begin{split}
&\frac{dN_1^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}\\&= \frac{32 e^2e_f^2}{3 \pi^4} \alpha_s^2 \int_{\v{k}_1,\v{k}_2,\v{k}_3}\frac{1}{\v{q}_1^2}\Gamma(\v{q}_2,\v{q}_2+\v{k}_3,\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3)
\\
&\times[U^{\dagger}_{\v{k}_1}U_{\v{k}_2}]_{e'e} \left[U^{\dagger}_{\v{k}_3}U_{\v{k}_1+\v{k}_3-\v{k}_2}\right]_{c'c}
\\
&\times\Big[\Tr(t^{c'}t^{e'}t^{e}t^{c})\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)
\\
+&\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}_1+\v{k}_3|)\Big)
\\
+&\frac{T_{ec}^aT_{c'e'}^a}{8}
\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)
\\
-&\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}_1+\v{k}_3|)\Big)
\Big]
\end{split}
\end{equation}
One in principle can substitute an explicit expression for the trace of four Gell-Mann matrices, in practice however, this does not help to simplify the expression.
\begin{equation}
\begin{split}
&\frac{dN_2^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}\\=&-\frac{32 e^2e_f^2}{3 \pi^4} \alpha_s^2 \int_{\v{k}_1,\v{k}_2,\v{k}_3}\frac{\v{q}_1 \cdot(\v{q}_1+\v{k}_1)}{\v{q}_1^2(\v{q}_1+\v{k}_1)^2}\Gamma(\v{q}_2,\v{q}_2+\v{k}_3,\v{q}_2+\v{k}_1+\v{k}_3-\v{k}_2)
\\
&\times[U^{\dagger}_{\v{k}_1}U_{\v{k}_2}]_{e'e} \left[U^{\dagger}_{\v{k}_3}U_{\v{k}_1+\v{k}_3-\v{k}_2}\right]_{c'c}
\\
&\times\Big[\Tr(t^{e'}t^{c'}t^{e}t^{c})\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)
\\
+&\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2 + \v{k}_1+\v{k}_3|)\Big)
\\
+&\frac{T_{ec}^aT_{e'c'}^a}{8}
\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)
\\
-&\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)+\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2 + \v{k}_1+\v{k}_3|)\Big)\Big]
\end{split}
\end{equation}
\begin{equation}
\begin{split}
&\frac{dN_3^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}\\=&
\frac{32 e^2e_f^2}{3 \pi^4} \alpha_s^2
\int_{\v{k}_1,\v{k}_2,\v{k}_3}\frac{(\v{q}_1+\v{k}_2) \cdot(\v{q}_1+\v{k}_1)}{(\v{q}_1+\v{k}_2)^2(\v{q}_1+\v{k}_1)^2}\Gamma(\v{q}_2,\v{q}_2+\v{k}_3,\v{q}_2+\v{k}_1+\v{k}_3-\v{k}_2)
\\
&\times[U^{\dagger}_{\v{k}_1}U_{\v{k}_2}]_{e'e} \left[U^{\dagger}_{\v{k}_3}U_{\v{k}_1+\v{k}_3-\v{k}_2}\right]_{c'c}
\\
&\times\Big[\Tr(t^{e'}t^{c'}t^{c}t^{e})\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)
\\
+&\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2 + \v{k}_1+\v{k}_3|)\Big)
\\
&\frac{T_{ce}^aT_{e'c'}^a}{8}\Big(-\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)
\\
&+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2 + \v{k}_1+\v{k}_3|)\Big)\Big]
\end{split}
\end{equation}
\begin{equation}
\begin{split}
&\frac{dN_4^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}\\=&-\frac{32 e^2e_f^2}{3 \pi^4} \alpha_s^2 \int_{\v{k}_1,\v{k}_2,\v{k}_3}\frac{(\v{q}_1+\v{k}_2) \cdot \v{q}_1}{(\v{q}_1+\v{k}_2)^2\v{q}_1^2}\Gamma(\v{q}_2,\v{q}_2+\v{k}_3,\v{q}_2+\v{k}_1+\v{k}_3-\v{k}_2)
\\
&\times[U^{\dagger}_{\v{k}_1}U_{\v{k}_2}]_{e'e} \left[U^{\dagger}_{\v{k}_3}U_{\v{k}_1+\v{k}_3-\v{k}_2}\right]_{c'c}
\\
&\times\Big[\Tr(t^{c'}t^{e'}t^{c}t^{e})\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)
\\
+&\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2 + \v{k}_1+\v{k}_3|)\Big)
\\
+&\frac{T_{ce}^aT_{c'e'}^a}{8}\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_2 + \v{k}_1-\v{k}_2 + \v{k}_3|)
\\
&+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)+\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}_1+\v{k}_3|)\Big)\Big]
\end{split}
\end{equation}
Note that the combinations of the functions $\mathcal{I}$ multiplying the trace of the four Gell-Mann matrices are identical, while those multiplying the product of two adjoint matrices are different in different terms due to distinct patterns of signs.
It is interesting to separate out the contribution which is analogous to calculations performed in the ``classical'' MV model. As we have discussed above, it corresponds to keeping only those terms in $\Sigma_4$ which are completely symmetric in the color charge density operators. Here this corresponds to symmetrizing the four generators in the trace with respect to the color indices.
For this symmetric part of the trace, see Eq.~\eqref{Eq:TrSym}, and $N_c=3$ we obtain
\begin{equation}
\begin{split}
\frac{dN_{\rm sym}^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}
=&\frac{16 e^2e_f^2}{27 \pi^4} \alpha_s^2 \int_{\v{k}_1,\v{k}_2,\v{k}_3}\Gamma(\v{q}_1,\v{q}_1+\v{k}_1,\v{q}_1+\v{k}_2)\Gamma(\v{q}_2,\v{q}_2+\v{k}_3,\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3)
\\
&\times\Big[\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_3|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_1|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3|)
\\
&\quad \quad +\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{k}_1-\v{k}_2|)
\\
&\quad \quad -
\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}_2-\v{k}_3|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}_1+\v{k}_3|)\Big)
\Big]\\
&\Big( \Tr[U^{\dagger}_{\v{k}_1}U_{\v{k}_2}]\Tr\left[U^{\dagger}_{\v{k}_3}U_{\v{k}_1+\v{k}_3-\v{k}_2}\right]
\\
&\quad +
\Tr\left[U^{\dagger}_{\v{k}_1}U_{\v{k}_2}U^{\dagger}_{\v{k}_3}U_{\v{k}_1+\v{k}_3-\v{k}_2}\right]
\\
&\quad +
\Tr\left[U^{\dagger}_{\v{k}_1}U_{\v{k}_2}U^{\dagger}_{-\v{k}_1+\v{k}_2-\v{k}_3}U_{-\v{k}_3}\right] \Big)\,.
\end{split}
\end{equation}
This expression has a few physically transparent properties. First, it vanishes if any of the four interactions with the target is set to zero, $U_{\v{k}} \to \delta^{(2)}(\v{k})$, due to $\Gamma(\v q,\v k,\v q) = \Gamma(\v q,\v q,\v k)=0$. Second, the color neutrality of the dipole suggests that a zero momentum gluon will not be able to resolve the dipole and thus lead to a trivial contribution. Gluons originating from the dipole have the momenta $\v{q}_1+\v{k}_1$, $\v{q}_1+\v{k}_2$, $\v{q}_2+\v{k}_3$, and $\v{q}_2+\v{k}_1-\v{k}_2+\v{k}_3$ (this can be read off the arguments of $\Gamma$ functions). Setting any of this to zero leads to vanishing contribution in the integral due to the combination of $\mathcal{I}$ functions.
Finally, one can trivially show that the cross section is symmetric under $\v{q}_2\to -\v{q}_2$.
\subsection{Averaging over the target in FDA}
We now need to average over the target field ensemble using the FDA.
We present here the detailed computation only for the term $\Sigma_2$. We have
\begin{equation}
\begin{split}
&\langle \Tr(U_{\v{q}_1+\v{k}-\v{P}}T^{e}U^{\dagger}_{\v{k}-\v{q}_2}U_{\v{\bar{k}}-\v{q}_2}T^eU^{\dagger}_{\v{q}_1+\v{\bar{k}} -\v{P}})\rangle_T\\
=&T^e_{bc}T^e_{fg} \langle U^{ab}_{\v{q}_1+\v{k}-\v{P}} U^{\dagger cd}_{\v{k}-\v{q}_2} \rangle_T \langle U^{df}_{\v{\bar{k}}-\v{q}_2} U^{\dagger ga}_{\v{q}_1+\v{\bar{k}} -\v{P}})\rangle_T\\
+&T^e_{bc}T^e_{fg} \langle U^{ab}_{\v{q}_1+\v{k}-\v{P}}U^{df}_{\v{\bar{k}}-\v{q}_2} \rangle_T \langle U^{\dagger cd}_{\v{k}-\v{q}_2} U^{\dagger ga}_{\v{q}_1+\v{\bar{k}} -\v{P}})\rangle_T\\
+&T^e_{bc}T^e_{fg} \langle U^{ab}_{\v{q}_1+\v{k}-\v{P}} U^{\dagger ga}_{\v{q}_1+\v{\bar{k}} -\v{P}}) \rangle_T \langle U^{\dagger cd}_{\v{k}-\v{q}_2} U^{df}_{\v{\bar{k}}-\v{q}_2} \rangle_T\\
\end{split}
\end{equation}
The first term vanishes since $T^e_{bc} \delta_{bc}=0$. The second term
\begin{equation}
\begin{split}
&T^e_{bc}T^e_{fg} \langle U^{ab}_{\v{q}_1+\v{k}-\v{P}}U^{df}_{\v{\bar{k}}-\v{q}_2} \rangle_T \langle U^{\dagger cd}_{\v{k}-\v{q}_2} U^{\dagger ga}_{\v{q}_1+\v{\bar{k}} -\v{P}})\rangle_T\\
&=T^e_{bc}T^e_{b c}\frac{(2\pi)^4}{N_c^2-1} \delta^{(2)}(0) \delta^{(2)}(\v{q}_1+\v{k}+\v{\bar{k}}-\v{P}-\v{q}_2)D(\v{q}_2-\v{\bar{k}})D(\v{q}_2- \v k)
\\
& =
- (2\pi)^2 N_c S_\perp \delta^{(2)}(\v{q}_1+\v{k}+\v{\bar{k}}-\v{P}-\v{q}_2)D(\v{q}_2-\v{\bar{k}})D(\v{q}_2- \v k) ,
\end{split}
\end{equation}
where we took into account that in the momentum space $\delta^{(2)}(0) = S_\perp/(2\pi)^2$.
The third term
\begin{equation}
\begin{split}
&T^e_{bc}T^e_{fg} \langle U^{ab}_{\v{q}_1+\v{k}-\v{P}} U^{\dagger ga}_{\v{q}_1+\v{\bar{k}} -\v{P}}) \rangle_T \langle U^{\dagger cd}_{\v{k}-\v{q}_2} U^{df}_{\v{\bar{k}}-\v{q}_2} \rangle_T\\
&= (2\pi)^2 \Tr (T^e T^e) S_\perp
\delta^{(2)}(\v{k}-\v{\bar{k}})D(\v{q}_1+\v{k}-\v{P})D(\v{q}_2-\v{k})\\
&= (2\pi)^2 N_c (N_c^2-1) S_\perp
\delta^{(2)}(\v{k}-\v{\bar{k}})D(\v{q}_1+\v{k}-\v{P})D(\v{q}_2-\v{k})\,.
\end{split}
\end{equation}
Hence the cross section for the contribution involving $\Sigma_2$ is
\begin{equation}
\begin{split}
\frac{dN^{(2)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}=&\frac{64e^2e_f^2 }{3\pi^2} N_c \alpha_s^2 S_{\perp} \int_{k,\v{\bar{k}} } \frac{\mathcal{I}(\varepsilon,|\v{q}_1+\v{q}_2+\v{k}+\v{\bar{k}}|)}{(\v{q}_1+\v{q}_2+\v{k}+\v{\bar{k}})^2}\\
&\Big((N_c^2-1)\Gamma(\v{q}_2,\v{q}_2+\v{k},\v{q}_2+\v{k})- \Gamma(\v{q}_2,\v{q}_2+ k,\v{q}_2+ \v{\bar{k}})\Big) D(\v{\bar{k}})D(\v{k})\,.
\end{split}
\end{equation}
Similarly we obtain
\begin{equation}
\begin{split}
\frac{dN^{(3)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}
&=\frac{32 e^2e_f^2}{3 \pi^2} N_c\alpha_s^2 S_{\perp} \int_{k,\v{\bar{k}}}
\Big[\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}-\v{\bar{k}}|) L(\v{q}_1, \v{\bar{k}}-\v{q}_1)\\
&+\Big(\mathcal{I}_{\varepsilon_f}(|\v{k}+\v{q}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{\bar{k}}-\v{q}_1|)\Big) L(\v{q}_1, \v{q}_1-\v{\bar{k}})\Big] \cdot \frac{
(\v{\bar{k}}-\v{k}-\v{q}_1-\v{q}_2)}{(\v{\bar{k}}-\v{k}_1-\v{q}_1-\v{q}_2)^2}
\\ &\times
\Big((N_c^2-1)\Gamma(\v{q}_2,\v{q}_2+\v{k},\v{q}_2+\v{k})-\Gamma(\v{q}_2,\v{q}_2+\v{k},\v{q}_2-\v{\bar{k}})\Big)
D(\v{k})D(\v{\bar{k}} )
\end{split}
\end{equation}
For the symmetric part of $\Sigma_4$ at $N_c=3$ and using FDA, we obtain
\begin{equation}
\begin{split}
\frac{dN^{(4)}_{\rm sym}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}
&
=\frac{16e^2e_f^2S_{\perp}}{27\pi^4} \alpha^2_s (N_c^2+1)
(\mathcal{W}_1+\mathcal{W}_2+\mathcal{W}_3)
\end{split}
\end{equation}
where $\mathcal{W}_i$ are the following integrals (although our equations are valid for $N_c=3$ only, we keep explicit variable $N_c$ in order to distinguish different contributions)
\begin{equation}
\label{Eq:W1}
\begin{split}
\mathcal{W}_1&=(N_c^2-1) \int_{\v k,\v{\bar{k}}}\Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1+\v{k})\Gamma(\v{q}_2,\v{q}_2+\v{\bar{k}},\v{q}_2+\v{\bar{k}}) D(\v{k})D(\v{\bar{k}})
\\
&\times\Big(2\mathcal{I}_{\varepsilon_f}(| \v{q}_2 +\v{\bar{k}}|)+2\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}|)
- \\& \quad -\mathcal{I}(\varepsilon_f,|\v{q}_2-\v{q}_1
+\v{\bar{k}}-\v{k}|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}+\v{\bar{k}}|)\Big) \,,
\end{split}
\end{equation}
\begin{equation}
\label{Eq:W2}
\begin{split}
\mathcal{W}_2&= \int_{\v k,\v{\bar{k}}}\Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1-\v{\bar{k}})\Gamma(\v{q}_2,\v{q}_2-\v{k},\v{q}_2+\v{\bar{k}}) D(\v{k})D(\v{\bar{k}})
\\
&
\times\Big({\mathcal{I}_{\varepsilon_f}(|\v{q}_2-\v{k}|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{\bar{k}}|)+
\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{\bar{k}}|)}\\&-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2|)
-\mathcal{I}_{\varepsilon_f}(|\v{k}+\v{\bar{k}}|)
-\mathcal{I}_{\varepsilon_f}(|\v{q}_2-\v{q}_1 + \v{\bar{k}}-\v{k}|)
\Big)\,,
\end{split}
\end{equation}
\begin{equation}
\label{Eq:W3}
\begin{split}
\mathcal{W}_3&= \int_{\v k,\v{\bar{k}}}\Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1+\v{\bar{k}})\Gamma(\v{q}_2,\v{q}_2+\v{\bar{k}},\v{q}_2+ k) D(\v{k})D(\v{\bar{k}})
\\
&
\times\Big({\mathcal{I}_{\varepsilon_f}(|\v{q}_2 + \v{\bar{k}}|)
+\mathcal{I}_{\varepsilon_f}(|\v{q}_1 +\v k|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1 + \v{\bar{k}}|)+
\mathcal{I}(\varepsilon_f,|\v{q}_2 +\v k|}\\
&-\mathcal{I}_{\varepsilon_f}(|\v{q}_2-\v{q}_1|)
-\mathcal{I}_{\varepsilon_f}(|\v{k}-\v{\bar{k}}|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2 + \v{k}+\v{\bar{k}}|)\Big)\,.
\end{split}
\end{equation}
In these contributions one can identify the underlying physical origin:
\begin{itemize}
\item Uncorrelated production:
the terms appearing in the second line of Eq.~\eqref{Eq:W1} and proportional to
$\mathcal{I}(\varepsilon_f,|\v{q}_1 +\v k|$ and $\mathcal{I}_{\varepsilon_f}(|\v{q}_2 +\v{\bar{k}}|)$
represent {uncorrelated production}. This is
manifested by factorization of the momenta $\v{q}_1$ and $\v{q}_2$ as well as the integral with respect to $\v k$ and $\v{\bar{k}}$.
\item Bose Enhancement in the projectile:
the last line of Eq.~\eqref{Eq:W1} is due to Bose Enhancement in the projectile wave function. The two gluons from the projectile scatter on the target independently.
This contribution is peaked at the momenta when the emitted gluons from the projectile have
either colinear or anti-colinear momenta, that is
when the arguments of the $\mathcal{I}$ function vanish. There is also an $N_c^2$-suppressed corrections to this contribution in the last terms of Eqs.~\eqref{Eq:W2} and ~\eqref{Eq:W3}.
\item Hanbury Brown and Twiss effect: the terms $\mathcal{I}(\varepsilon_f,|\v{q}_1 \pm \v{q}_2| )$ originate from the HBT effect.
\item Bose Enhancement in the target:
Gluons originating from the target and having
either colinear or anti-colinear momenta lead to the Bose Enhancement in target wave function. This is the origin of the rest of terms in the above expression. This is obvious for the contribution is proportional to $\mathcal{I}_{\varepsilon_f}(|\v k\pm\v{\bar{k}}|)$. The rest of the contributions are also due to Bose enhancement of the target: they correspond to the situation where the two gluons from the projectile scatter in the correlated manner from the target-by exchanging either the same or opposite transverse momenta. The somewhat unusual momentum combinations in factors of ${\cal I}$, like $\mathcal{I}_{\varepsilon_f}(|\v{\bar{k}}-\v{q}_1|)$ express the fact that gluons with momentum of order of inverse dipole size are emitted from the dipole in coherent fashion.
\end{itemize}
The remaining, non-symmetric part has kinematic factors that cannot be combined to Lipatov vertices.
We again write it as a sum of three contributions
\begin{equation}
\begin{split}
\frac{dN_{\rm ns}^{(4)}}{d\eta d\v{q}_1^2 d\xi d\v{q}_2^2}\approx
& \frac{32 e^2e_f^2}{3 \pi^4} \alpha_s^2 S_\perp N_c\Big( W_1+ W_2+ W_3\Big)
\end{split}
\end{equation}
with
\begin{equation}
\begin{split}
W_1
=&\frac{(N_c^2-1)}{12}\int_{\v k,\v{\bar{k}}}\Big[\Gamma(\v{q}_1,-(\v{q}_1+\v{k}),-(\v{q}_1+\v{k}))+2 \frac{(\v{q}_1+ \v k) \cdot \v{q}_1}{(\v{q}_1+ \v k)^2\v{q}_1^2} \Big]\Gamma(\v{q}_2,\v{q}_2+\v{\bar{k}},\v{q}_2+\v{\bar{k}})\\
&\times D(\v{k})D(\v{\bar{k}})
\Big(2\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{\bar{k}}|)+2\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}|)
-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}-\v{\bar{k}}|)\\&\quad\quad\quad\quad-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}+\v{\bar{k}}|)\Big)\\
&
+
\frac{(N_c^2-1)}{8}\int_{\v k,\v{\bar{k}}}\Gamma(\v{q}_2,\v{q}_2+\v{\bar{k}},\v{q}_2+\v{\bar{k}})\Gamma(\v{q}_1,-(\v{q}_1+\v{k}),-(\v{q}_1+\v{k}))D(\v{k})D(\v{\bar{k}})
\\
&\times \Big(2\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{\bar{k}}|)-2\mathcal{I}_{\varepsilon_f}(|\v{q}_1+ \v k|)
+\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2+\v{k}-\v{\bar{k}}|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}+ \v{\bar{k}}|)\Big)
\end{split}
\end{equation}
\begin{equation}
\begin{split}
W_2=&\frac{1}{12}\int_{\v k,\v{\bar{k}}}\Gamma(\v{q}_2,\v{q}_2- k,\v{q}_2+\v{\bar{k}}) \Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1-\v{\bar{k}}) D(\v{k})D(\v{\bar{k}}) \\
&
\times \Big[\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2-\v{k}|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{\bar{k}}|)+
\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{\bar{k}}|)\\
&
-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2-\v{\bar{k}}+\v{k}|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2|)\Big)\Big]
\end{split}
\end{equation}
\begin{equation}
\begin{split}
W_{3}=& - \frac{1}{12} \int_{\v k,\v{\bar{k}}}\Gamma(\v{q}_2,\v{q}_2-\v{\bar{k}},\v{q}_2+\v{k})\Big(\frac{1}{\v{q}_1^2} +\frac{(\v{q}_1+\v{k}) \cdot(\v{q}_1-\v{\bar{k}})}{(\v{q}_1+\v{k})^2(\v{q}_1-\v{\bar{k}})^2} +\Gamma(\v{q}_1,-\v{q}_1-\v{k},-\v{q}_1+\v{\bar{k}}) \Big)
\\
&
\times D(\v{k})D(\v{\bar{k}})
\Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2-\v{\bar{k}}|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}|)+
\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{\bar{k}}|)\\&-\mathcal{I}_{\varepsilon_f}(|\v{k}+\v{\bar{k}}|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}-\v{\bar{k}}|)\Big)\\
&
-\frac{1}{8}\int_{k,\v{\bar{k}}}\Gamma(\v{q}_2,\v{q}_2-\v{\bar{k}},\v{q}_2+\v{k})\Gamma(\v{q}_1,-\v{q}_1-\v{k},-\v{q}_1+\v{\bar{k}})D(\v{k})D(\v{\bar{k}})\\&
\times \Big(\mathcal{I}_{\varepsilon_f}(|\v{q}_2-\v{\bar{k}}|)-\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{k}|)
+\mathcal{I}_{\varepsilon_f}(|\v{q}_2+\v{k}|)
-\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{\bar{k}}|)\\&-\mathcal{I}_{\varepsilon_f}(|\v{k}+ \v{\bar{k}}|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1-\v{q}_2|)+\mathcal{I}_{\varepsilon_f}(|\v{q}_1+\v{q}_2+\v{k}-\v{\bar{k}}|)\Big)\,.
\end{split}
\end{equation}
We will use this equations for the numerical evaluations.
\section{Inclusive two gluon production: the MV model.}\label{sec5}
In this section we compute two gluon production by approximating the wave function of the (almost) real photon using the MV model. Note that although we use MV model, we do not discard terms which are suppressed by powers of color charge density, and which are customarily neglected in the CGC literature.
We start with $\Sigma^{\rm MV}_2$, which is trivial to compute,
\begin{equation}
\begin{split}
\Sigma^{\rm MV}_2(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )&=
4\int d^2\bm{x} \mu^2(\bm{x}) f^i(\bar{\v u}_1-\bm{x})f^i(\v u_1-{\bm{x}})f^j(\bar{\v u}_2-\bar{\v u}_1)f^j(\v u_2-\v u_1)
\\
&\left<\Tr \left[ U_{\v u_1}T^{a}[U^{\dagger}_{\v u_2}-U^{\dagger}_{\v u_1}] [U_{\bar{\v u}_2}-U_{\bar{\v u}_1}] T^{a}U^{\dagger}_{\bar{\v u}_1}\right] \right>_{T}
\end{split}
\end{equation}
This expression modulo the projectile wave function is very much the same as the one we obtained in the previous section, see Eq.~\eqref{Eq:N2}.
The contribution in $\Sigma_3^{\rm MV}$ involves two terms, both of which are of the same order in color charge density as $\Sigma^{\rm MV}_2$,
\begin{equation}
\begin{split}
\Sigma^{\rm MV}_3(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )&
= 2\int d^2\bm{x}\mu^2(\bm{x})f^j(\bar{\v u}_2-\bar{\v u}_1)f^i(\bar{\v u}_1-\bm{x})f^i(\v u_1-\bm{x})f^j(\v u_2-\bm{x})\\
&
\Tr\left[ [ U_{\bm{x}}+U_{\v u_1} ] T^a [U^{\dagger}_{\bm{x}}-U^{\dagger}_{\v u_2}] [U_{\bar{\v u}_2}-U_{\bar{\v u}_1}] T^{a} U^{\dagger}_{\bar{\v u}_1}]\right] \\
&
+2\int d^2\bm{x}\mu^2(\bm{x})f^j(\v u_2-\v u_1)f^i(\v u_1-\bm{x})f^i(\bar{\v u}_1-\bm{x})f^j(\bar{\v u}_2-\bm{x})\\
&
\Tr \left[ U_{\v u_1} T^{a} [U^{\dagger}_{\v u_2}-U^{\dagger}_{\v u_1}][U_{\bm{x}}-U_{\bar{\v u}_2}]T^a [U^{\dagger}_{\bm{x}}+U^{\dagger}_{\bar{\v u}_1}]\right]
\end{split}
\end{equation}
Finally, $\Sigma^{\rm MV}_4$ has two contributions, which can be organized by the power of $\mu^2$. We refer to $\mu^2$ contribution as quantum and $\mu^4$ as classical. We start with the former. Using Eq.~\eqref{Eq:rho4mv}
we obtain
\begin{equation}
\begin{split}
\Sigma_{4,Q}^{\rm MV}(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )
&
=\int d^2\bm{x}\mu^2(\bm{x}) f^i(\bar{\v u}_1-\bm{x})f^j(\bar{u}_2-\bm{x})f^i(\v u_1-\bm{x})f^j(u_2-\bm{x})\\
&
\Big(\Tr( [U^{\dagger}_{\v u_1}-U^{\dagger}_{\bm{x}}][U_{\bar{\v u}_1}-U_{\bm{x}}]T^a[U^{\dagger}_{\bm{x}}-U^{\dagger}_{\bar{\v u}_2}][U_{\bm{x}}-U_{\v u_2}]T^a)\\
&
-\frac13\Tr([U^{\dagger}_{\v u_1}+U^{\dagger}_{\bm{x}}][U_{\bar{\v u}_1}+U_{\bm{x}}]T^a)\Tr( [U^{\dagger}_{\bm{x}}-U^{\dagger}_{\v u_2}][U_{\bm{x}}-U_{\bar{\v u}_2}]T^a)\\
&
-\frac13\tr\Big([U^{\dagger}_{\v u_1}+U^{\dagger}_{\bm{x}}][U_{\bar{\v u}_1}+U_{\bm{x}}]T^a [U^{\dagger}_{\bm{x}}-U^{\dagger}_{\v u_2}][U_{\bm{x}}-U_{\bar{\v u}_2}]T^a \Big)
\end{split}
\end{equation}
The classical part is dominant at large color charge density in the projectile and reads
\begin{equation}
\begin{split}
\Sigma_{4,C}^{\rm MV}(\v u_1, \v u_2, \bar{\v u}_1,\bar{\v u}_2 )&=4\int d^2\bm{x}d^2\bm{y}\int d^2\bar{\bm{x}}d^2\bar{\bm{y}}f^i(\bar{\v u}_1-\bm{x})f^j(\bar{\v u}_2-\bm{y})f^i(\v u_1-\bar{\bm{x}})f^j(\v u_2-\bar{\bm{y}})\\
&
\times \left< \left[[U^{\dagger}_{\v u_1}-U^{\dagger}_{\bar{\bm{x}}}][U_{\bar{\v u}_1}-U_{\bm{x}}]\right]_{e'e} \left[ [U^{\dagger}_{ \v u_2}-U^{\dagger}_{\bar{\bm{y}}}][U_{\bar{\v u}_2}-U_{\bm{y}}]\right]_{c'c} \right>_T \\&\times \Big(\delta^{c'e'}\delta^{ce}\mu^2(\bm{x})\mu^2(\bar{\bm{x}})\delta^{(2)}(\bm{x}-\bm{y})\delta^{(2)}(\bar{\bm{x}}-\bar{\bm{y}})\\
&
\quad+\delta^{c'e}\delta^{e'c}\mu^2(\bm{x})\mu^2(\bm{y})\delta^{(2)}(\bm{x}-\bar{\bm{y}})\delta^{(2)}(\bar{\bm{x}}-\bm{y})\\&\quad +\delta^{c'c}\delta^{e'e}\mu^2(\bm{x})\mu^2(\bm{y})\delta^{(2)}(\bm{x}-\bar{\bm{x}})\delta^{(2)}(\bm{y}-\bar{\bm{y}}) \Big)
\end{split}
\end{equation}
Here the first term in the classification of Ref.~\cite{Kovchegov:2012nd}
corresponds to the ``square'' diagrams involving the quadrupole, the second to ``cross'' diagrams. The last contribution represents emission where the correlations can only originate from the correlations in the target; this term also gives the leading contribution to the uncorrelated production.
\subsection{Momentum space and FDA}
As in the dipole model, we convert our expressions to momentum space and average over target using FDA.
We get
\begin{equation}
\begin{split}
\frac{dN_2}{d\v q^2_1d\v q^2_2 d\eta d\xi}&=
\frac{4 g^4}{(2\pi)^4}\int_{\v{k}_1,\v{k}_2,\v{k}_3,\v{k}_4}\mu^2(\v{k}_2+\v{k}_4-\v{k}_1-\v{k}_3)\left<\Tr \left[ U_{\v{k}_1}T^{a}U^{\dagger}_{\v{k}_2} U_{\v{k}_3} T^{a}U^{\dagger}_{\v{k}_4}\right] \right>_{T}\\
&L^i(\v{q}_2,\v{q}_2+\v{k}_2)L^j(\v{q}_2,\v{q}_2+\v{k}_3) \frac{(\v{q}_1+\v{q}_2+\v{k}_3-\v{k}_4)^i(\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_1)^j}{(\v{q}_1+\v{q}_2+\v{k}_3-\v{k}_4)^2(\v{q}_1+\v{q}_2+\v{k}_2-\v{k}_1)^2}
\end{split}
\end{equation}
Using FDA, this simplifies into
\begin{equation}
\begin{split}
\frac{dN_2}{d\v q^2_1d\v q^2_2 d\eta d\xi}&=\frac{4N_cg^4}{(2\pi)^4}\int d^2 \v x\mu^2(\v x) \int_{\v{k}_1,\v{k}_2} D(\v{k}_1+\v{q}_1)D(-\v{k}_2+\v{q}_2)\\
&\Big\{ (N_c^2-1) \left(\frac{(\v{k}_2-\v{k}_1) \cdot L(\v{q}_2,\v{k}_2)}{(\v{k}_2-\v{k}_1)^2}\right)^2
\\ & \quad -\frac{(\v{k}_2-\v{k}_1)\cdot L(\v{q}_2,\v{k}_2)}{(\v{k}_2-\v{k}_1)^2}\frac{(\v{k}_2-\v{k}_1) \cdot L(\v{q}_2,\v{q}_2-\v{k}_2-\v{k}_1)}{(\v{k}_2-\v{k}_1)^2}\Big\}
\end{split}
\end{equation}
Similarly for $\Sigma_3$ contribution we obtain
\begin{equation}
\begin{split}
\frac{dN_3}{d\v q^2_1d\v q^2_2 d\eta d\xi}=& \frac{4 g^4}{(2\pi)^4}\int_{\v{k}_1,\v{k}_2,\v{k}_3,\v{k}_4} \mu^2\Big(\v{k}_4+\v{k}_2-\v{k}_1-\v{k}_3\Big) \Tr[U_{\v{k}_1}T^aU^{\dagger}_{\v{k}_2}U_{\v{k}_3}T^aU^{\dagger}_{\v{k}_4}]\\
&\Gamma(\v{q}_2,\v{q}_2+\v{k}_2,\v{q}_2+\v{k}_3)L^i(\v{q}_1,\v{k}_1-\v{q}_1)\frac{(\v{k}_4-\v{k}_3-\v{q}_1-\v{q}_2)^i}{(\v{k}_4-\v{k}_3-\v{q}_1-\v{q}_2)^2}
\end{split}
\end{equation}
where we assumed that $\mu^2(-\v{k}) = \mu^2(\v{k})$.
The FDA leads to
\begin{equation}
\begin{split}
\frac{dN_3}{d\v q^2_1d\v q^2_2 d\eta d\xi}
=&\frac{4N_cg^4}{(2\pi)^4}\int d^2 \v x\mu^2(\v x) \int_{\v{k}_1,\v{k}_2} D(\v{k}_1+\v{q}_1)D(-\v{k}_2+\v{q}_2)\\
&\Big\{(N_c^2-1) L^i(\v{q}_1,\v{k}_1)\frac{(\v{k}_1-\v{k}_2)^i}{(\v{k}_1-\v{k}_2)^2}\Gamma(\v{q}_2,\v{k}_2,\v{k}_2)\\& \quad -L^i(\v{q}_1,\v{k}_1)\frac{(\v{k}_1-\v{k}_2)^i}{(\v{k}_1-\v{k}_2)^2}\Gamma(\v{q}_2,\v{k}_2,\v{q}_2-\v{q}_1-\v{k}_1)\Big\}\,.
\end{split}
\end{equation}
Finally the $\Sigma_4$ contribution is given by two terms proportional to $\mu^2$ and $\mu^4$. Starting with the former, in FDA we obtain
\begin{equation}
\begin{split}
&\frac{dN_4^Q}{d\v q^2_1d \v q^2_2 d\eta d\xi}\\
=&\frac{g^4N_c}{(2\pi)^4}\int_{\v k,\v{\bar{k}}} D(-\v{k})D(-\v{\bar{k}})\int d^2\bm{x}\mu^2(\bm{x})\\&
\Big\{(N_c^2-1)\Big[\Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1+\v{k})-\frac13\Gamma(\v{q}_1,-\v{q}_1-\v{k},-\v{q}_1-\v{k}) \Big]\Gamma(\v{q}_2, \v{q}_2+ \v{\bar{k}},\v{q}_2+\v{\bar{k}})\\
&+\frac23\Gamma(\v{q}_1,-\v{q}_1-\v{k},-\v{q}_1-\v{\bar{k}})\Gamma(\v{q}_2, \v{q}_2+ \v{\bar{k}},\v{q}_2+\v k)\\
&+\Big[ \Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1+\v{\bar{k}}) -\frac13\Gamma(\v{q}_1,-\v{q}_1+\v{k},-\v{q}_1-\v{\bar{k}})\Big] \Gamma(\v{q}_2, \v{q}_2+ \v k,\v{q}_2+\v{\bar{k}})\Big\}
\end{split}
\end{equation}
The latter is given by
\begin{equation}
\begin{split}
\frac{dN_4^C}{d\v q^2_1d \v q^2_2 d\eta d\xi}
&=\frac{4g^4}{(2\pi)^4} \int_{\v k,\v{\bar{k}}}D(-\v{k})D(\v{\bar{k}})\Big\{\Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1+\v{k})
\Gamma(\v{q}_2,\v{q}_2-\v{\bar{k}},\v{q}_2-\v{\bar{k}})\\&
\left[|\mu^2(0)|^2(N_c^2-1)^2+(N_c^2-1)|\mu^2(\v{q}_1-\v{q}_2+\v{k}+\v{\bar{k}})|^2 \right]\\&
+(N_c^2-1)|\mu^2(\v{q}_1+\v{q}_2+\v{k}+\v{\bar{k}})|^2\Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1+\v{k})\Gamma(\v{q}_2,\v{q}_2+ \v{\bar{k}},\v{q}_2+\v{\bar{k}})\\
&+\Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1+\v{\bar{k}})\Gamma(\v{q}_2,\v{q}_2+ k,\v{q}_2+\v{\bar{k}})\Big( |\mu^2(\v{q}_1+\v{q}_2+\v{k}+\v{\bar{k}})|^2\\&
+(N_c^2-1)|\mu^2(\v{q}_1-\v{q}_2)|^2+|\mu^2(\v{k}-\v{\bar{k}})|^2\Big)\\
&+\Gamma(\v{q}_1,\v{q}_1+\v{k},\v{q}_1+\v{\bar{k}})\Gamma(\v{q}_2,\v{q}_2-\v{k},\v{q}_2-\v{\bar{k}})\Big((N_c^2-1)|\mu^2(\v{q}_1+\v{q}_2)|^2\\&
+|\mu^2(\v{q}_1-\v{q}_2+\v{k}+\v{\bar{k}})|^2+|\mu^2(\v{k}-\v{\bar{k}})|^2\Big)
\end{split}
\end{equation}
The physical origin of different terms can be easily identified:
\begin{itemize}
\item Uncorrelated production: the contribution proportional to $(N_c^2-1)^2|\mu^2(0)|^2$ describes uncorrelated two gluon production
\item Bose Enhancement in the projectile: this term is proportional to $(N_c^2-1)|\mu^2(\v{q}_1 \pm \v{q}_2 +\v{k}+\v{\bar{k}})|^2$. It is suppressed by the factor of $1/N_c^2$ compared to the uncorrelated production as the initial gluons have to have the same color. As in the dipole case, there is a correction to Bose Enhancement in the projectile which is suppressed by a further factor of $1/N_c^2$. It is proportional to $|\mu^2(\v{q}_1\pm \v{q}_2+\v{k}+\v{\bar{k}} )|^2$.
\item Bose Enhancement in the target: the terms proportional to $|\mu^2(\v{k}-\v{\bar{k}} )|^2$ represent Bose Enhancement in the target; they are suppressed by $1/N_c^4$ in the dense regime.
\item HBT: finally the terms $(N_c^2-1)|\mu^2(\v{q}_1\pm \v{q}_2)|^2$ represent HBT correlation. This term is suppressed by $1/N_c^2$ relative to the uncorrelated production as the color of the final state gluons should be the same.
\end{itemize}
\section{Numerical results}
\label{Sec:Num}
In this section we evaluate numerically inclusive two gluon production in both models for the projectile wave function.
We do not have ambitions to fit experimental data, and we therefore do not try to optimize parameters of the model. Nevertheless we want to be able to have at least a qualitative idea on the ballpark of the CGC predictions for relevant observables.
The first question we are faced with is that of subtraction of the background.
In experiment, low multiplicity events are used to subtract trivial contribution from high multiplicity data, which is then analyzed in terms of flow harmonics. In our calculation using the MV model for the projectile, this procedure naturally corresponds to subtracting the $\mu^2$ terms as they play a dominant role at low multiplicity and are negligible compared to $\mu^4$ terms at high multiplicity. We thus take as our ``signal'' only the $\mu^4$ contribution, which corresponds to the classical gluon production as defined in the previous section. In the dipole model for the projectile wave function, the distinction between low and high multiplicities is trickier to make since we do not have an explicit parameter $\mu^2$. Nevertheless motivated by the MV model, we will use the symmetric part of the matrix element $\Sigma_4$ as our high multiplicity ``signal''. For $N_c=3$, as we demonstrated in Sect.~\ref{Sect:2glDPM}, it has a compact form and is clearly not dominated by back-to-back production.
For numerical calculations we will use the following set of parameters: the saturation momentum of the target is chosen as $Q_s=2$ GeV, the infrared cut-off entering into the Poisson equation of the MV model is set to be equal to $\Lambda = 200$ MeV, additionally to insure that the logarithm of the argument of the logarithm entering into the adjoint dipole of Eq.~\eqref{Eq:KovnersHammer} does not change sign we use
\begin{align}
D(\v r) = \exp\left[-\frac{1}{4}Q_s^2r^2\ln(\frac{1}{\Lambda^2r^2}+e) \right]\,.
\end{align}
For the nearly real photon wave function, we use a small but non-zero value of $Q$ in order to regularize large distance physics. In our calculations we fix $\epsilon_f = 100$ MeV. This roughly corresponds to $Q\sim 200$ MeV.
To ensure that the projectile photon has a finite size,
in the projectile MV model we introduce the projectile size $R$ through the impact parameter dependent parameter $\mu^2(\v x)$:
\begin{align}
\mu^2(\v x) = {\cal N} \exp{-\frac{\v x^2}{R^2}}\,.
\end{align}
Since our signal is a homogeneous function of $\mu^2$, the constant $ {\cal N}$ does not play any role in the normalized quantities considered below and we do not need to specify it.
Here we do not explore the dependence on the shape of the projectile as it has little effect on the observable in this approach, except for unrealistically large ellipticities, see Ref.~\cite{Kovner:2018fxj}.
The code is publicly available in Ref.~\cite{UPCcode}.
In order to compute the two particle $v_2$ as a function of the transverse momentum, we define the $n$-th Fourier mode of the particle production cross section $V_n(p_\perp)$
\begin{align}
V_n(q_1) =
\int d \theta_1
\int_{0}^{p^{\rm max}_\perp} d^2 \v{q}_2
\exp(i n \Delta \theta )
\frac{d N}{d\v q^2_1 d \v q^2_2 d\eta d\xi}
\end{align}
where $\Delta \theta = \theta_2-\theta_1$ is the angle between $\v q_1$ and $\v q_2$.
Additionally, the integrated Fourier modes are defined according to
\begin{align}
V_n =
\int_0^{p^{\rm max}_\perp} d^2 \v{q}_1
\int_0^{p^{\rm max}_\perp} d^2 \v{q}_2
\exp(i n \Delta \theta )
\frac{d N}{d\v q^2_1 d \v q^2_2 d\eta d\xi}
\end{align}
Then we have
\begin{align}
\label{Eq:v22}
v^{(2)}_2(p_\perp) = \sqrt{\frac{V_2(p_\perp)}{V_0(p_\perp)}}
\end{align}
for the second Fourier harmonic of the two particle correlation function.
Experimental collaborations, in addition to $v^{(2)}_2(p_\perp)$ also extract $v_{2}$ whose definition is based on the assumption of the factorization of the hydrodynamic flow.
In hydrodynamics, a two-particle azimuthal harmonic $v^{(2)}_2$ is a product of single particle azimuthal anisotropies, $v_2$. This motivates considering the observable
\begin{align}
\label{Eq:v2}
v_{2}(p_\perp) = \frac{V_2(p_\perp)/V_0(p_\perp)}{\sqrt{V_2/V_0}}\,.
\end{align}
Experimentally the two observables are quite close to each other \cite{ATLAS:2021jhn}.
In our calculations, there is no reason at all to expect such hydrodynamics-like factorization. It is nevertheless useful to consider $v_{2}$ in order to get an idea of how much the two quantities differ from each other in our framework.
We plot these two quantities in Figs.~\ref{Fig:v22} and~\ref{Fig:v2}. To guide the eye we also added experimental data to compare it to $v_{2}$. Note that we did not plot the last experimental point which is located at rather large negative values of $v_2$, see ATLAS~\cite{ATLAS:2021jhn}.
The calculations are done for two upper limits of the integration with respect to the transverse momentum $p^{\rm max}_\perp = 2$ and 4 GeV. The former choice is the same as in the ATLAS measurements, while the later choice is motivated by comparison with the study in Ref.~\cite{Shi:2020djm}.
\begin{figure}
\centering
\includegraphics[width=0.49\linewidth]{Max2_v22.pdf}
\includegraphics[width=0.49\linewidth]{Max4_v22.pdf}
\caption{The second Fourier harmonic of the two particle correlation function, $v^{(2)}_2$, see Eq.~\eqref{Eq:v22}, as a function of the transverse momentum of the "trigger" gluon. The momentum of the "associated" gluon is integrated from $0$ to $p_\perp^{\rm max}$. }
\label{Fig:v22}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.49\linewidth]{Max2_v2.pdf}
\includegraphics[width=0.49\linewidth]{Max4_v2.pdf}
\caption{The elliptic flow $v_2$ extracted from two particle correlations function using the definition motivated by factorization, $v_2$, see Eq.~\eqref{Eq:v2}, as a function of the transverse momentum of the "trigger" gluon. The momentum of the associated gluon is integrated from $0$ to $p_\perp^{\rm max}$. To guide the eye we plotted ATLAS data from Ref.~\cite{ATLAS:2021jhn}. }
\label{Fig:v2}
\end{figure}
\section{Discussion}
\label{sec7}
We now discuss our results.
We stress again that our goal was not to fit the ATLAS data as we do not believe that our calculational framework is constrained enough for such an endeavor. First, we do not include hadronization in our calculation, and that may have a large effect on the correlated part of particle production. In addition the wave function of the real photon has to be modelled in some way, and this modelling leaves a lot of freedom. Instead we studied two ubiquitous models used in a variety of CGC calculations to understand the qualitative features of the effect.
Consequently we did not try to optimize the parameters of the MV model and the dipole model of the photon when calculating $v_2$.
Before we proceed, we want to put our study in the perspective by commenting on the existing literature and comment on the comparison of our results to those of Ref.~\cite{Shi:2020djm}. Direct comparison between the two is hard, since the calculational frameworks are very different, even though both calculations are based on the CGC approach. Our calculation is appropriate at non forward rapidities where particle production is dominated by gluons, whereas the authors of Ref.~\cite{Shi:2020djm} use the hybrid approach which restricts their calculation to the photon going direction. One could attempt to take the model used in Ref.~\cite{Shi:2020djm} for valence parton distribution in the photon and use it as yet another model for the photon. This is however not a well posed problem since no color correlations between valence partons have been taken into account in Ref.~\cite{Shi:2020djm} due to collinear approximation invoked there, while such color correlations are clearly important away from forward rapidity especially at relatively low transverse momenta.
There is however one qualitative observation that we can make with confidence. The $v_2$ is our calculation exhibits in all models the same qualitative trend: it rises at low transverse momentum to a maximum at $p_\perp$ around $ 1$ GeV, and drops thereafter, see Figs.~\ref{Fig:v22} and \ref{Fig:v2}. This is also a clear trend in the experimental data. On the other hand $v_2$ calculated in Ref.~\cite{Shi:2020djm} never reaches a maximum, but instead keeps growing at high momentum far beyond the point where experimentally it is observed to fall. Qualitative this difference in behavior is easy to understand. As we will explain below the origin of the turnover in our calculation is the narrowness of the gluon HBT peak. The calculation of \cite{Shi:2020djm} does not take the gluon HBT effect into account. Instead the correlations in \cite{Shi:2020djm} arise solely from the so called color domain structure in the target \cite{Kovner:2010xk}. The color domain effect is the more efficient, the higher the transverse momentum of produced particles. High $p_\perp$ particles are produced nearby in the coordinate space, and therefore are more probable to probe the same domain in the target. Thus the approach of \cite{Shi:2020djm} does not contain a mechanism that could lead to decreasing $v_2$ at high momentum.
\begin{figure}
\centering
\includegraphics[width=0.49\linewidth]{Factor_v2_mv_nb.pdf}
\includegraphics[width=0.49\linewidth]{Factor_v2_dip_nb.pdf}
\caption{The elliptic flow $v_2$ for three different kinematic ranges of the trigger particle. Here as in the previous figure, $Q_s=2$ GeV. The size of the projectile is set by $R=1$/GeV.}
\label{Fig:v21}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.49\linewidth]{Factor_v2_mv.pdf}
\includegraphics[width=0.49\linewidth]{Factor_v2_dip.pdf}
\caption{The same as Fig.~\ref{Fig:v21} but binned with the same bin size as the ATLAS analysis. }
\label{Fig:v23}
\end{figure}
Consider now our numerical results on Figs.~\ref{Fig:v22} and ~\ref{Fig:v2}. The first observation is that both the MV model and the dipole model produce $v_2$ of the same order as the experimental data. The shape of the momentum dependence in the two models is slightly different, with MV mode $v_2$ rather sharply peaked at low momentum.
One somewhat surprising feature of our results is that $v_2$ and $v_2^{(2)}$ turn out to be not very different. Naive expectation is that $v_2$ should drop very quickly once the momentum of the trigger particle is outside the integration interval of the momentum of the associate. The reasoning for this is that the two particle correlations in the CGC framework receive a very large contribution from the gluon HBT effect. The HBT peak is very narrow, in fact for a large projectile its width in momentum space is inversely proportional to the area. As a result for a large area projectile the correlation drops by several orders of magnitude once the momentum of the trigger is outside of the momentum bin of the associate~\cite{Altinoluk:2018ogz}. It is thus natural to expect $V_2/V_0\gg V_2(p_\perp)/V_0(p_\perp)$ and therefore $v_2\ll v_2^{(2)}(p_\perp)$ when $p_\perp>p_\perp^{\rm max}$.
Instead we observe that although $v_2$ does indeed drop somewhat faster for $p_\perp>p_\perp^{\rm max}$, the difference is not that spectacular especially for the dipole model photon. We believe this is due to the finite size of the projectile wave function, which leads to widening of the HBT peak and tempers the drop in the HBT correlation.
With these considerations in mind we plot the dependence of $v_2$ on the choice of the momentum bin of the associated particle. We choose the bins the same as in the ATLAS data~\cite{ATLAS:2021jhn}. As mentioned above, the hydrodynamic-like universality would suggest that this quantity does not depend on the bin size and location. Our results on the other hand are not expected to display such independence. In Fig.~\ref{Fig:v21} we plot $v_2$ as a function of the momentum of the trigger particle for three different bin choices of the associate. Indeed we observe a rather strong dependence on the bin choice consistent with the discussion above. Namely, $v_2$ is large when the momentum of the trigger particle $p_\perp$ is within the bin, and drops fairly quickly once $p_\perp$ is outside the bin. This behavior is common for the MV model and the dipole model of the photon. At some values of $p_\perp$ the difference between the values of $v_2$ is very large, with ratio between $v_2$ for different bins reaching about 7 or 8 at the maximum. This seems to be in stark contradiction with experimental data. However, one has to keep in mind that the experimental test of universality in Ref.~\cite{ATLAS:2021jhn} was not performed for a fixed momentum of a trigger particle, rather $p_\perp$ was also binned. Following the same procedure we plot on Fig.~\ref{Fig:v23} values of $v_2$ for four bins of $p_\perp$ chosen in the same way as in~Ref.\cite{ATLAS:2021jhn}. Interestingly, the binning has a very strong effect: since the curves in Fig.~\ref{Fig:v21} are all bell shaped on the scale comparable with the bin size, integrating over the $p_\perp$ bins leads to significant reduction of differences in the values of $v_2$. The ratio between the values of $v_2$ for different associate bins is now at most 3 or 4 rather than 7 or 8. This is still significantly larger than for the ATLAS data, where the maximal ratio is closer to 2, but the difference is now not as drastic. Interestingly, the systematics of momentum dependence of our results is quite similar to that in the data, with $v_2$ initially rising, and subsequently falling with $p_\perp$, and $v_2$ for some of the bins crossing over in the vicinity of $1$ GeV.
We stress again that our purpose in this exercise is not to force-fit the data. It is quite clear that fiddling with the model parameters would allow us to get much closer to experimental points, but it is unclear what it would teach us. Instead, our take home message is twofold. First, we conclude unambiguously that a CGC calculation does not yield results consistent with the hydrodynamic inspired universality of $v_2$, although the values of $v_2$ we obtain are roughly of the same magnitude as the experimental ones. Second, it seems to us that the test of the said universality conducted in Ref.~\cite{ATLAS:2021jhn} is inconclusive. The variation between the different bin values found in Ref.~\cite{ATLAS:2021jhn} albeit broadly consistent with universality, is also quite similar to our (not optimized) CGC results due to binning of the momentum of the trigger particle. A more detailed study of this point seems desirable.
\acknowledgments
We thank T.~Altinoluk, N.~Armesto, M.~Lublinsky, and A.~Milov for illuminating discussions.
A.K. is supported by the NSF Nuclear Theory grant 1913890. This material is based
upon work supported by the U.S. Department of Energy, Office of Science, Office of Nuclear
Physics through the Contract No. DE-SC0020081 (H.D. and V.S.).
|
\section{Introduction}\label{sec:intro}
Extracting clean speech from noisy speech signals is an important task for various real-world applications \cite{vincent2018audio}, e.g., speech recognition, assisted listening, and mobile communications. The current state-of-the-art methods in the domain of speech enhancement employ some variant of deep neural networks (DNNs), e.g., \cite{xu2014regression,lu2013speech, pascual17_interspeech, hu2020dccrn,hao2021fullsubnet}, which have outperformed the classical methods like spectral subtraction \cite{boll1979suppression}, Wiener filtering \cite{ lim1979enhancement} and minimum mean squared error estimation \mbox{(MMSE) \cite{ephraim1985speech, martin2005speech}}.
Although multi-microphone technologies have gathered a lot of attention over the past decades \cite{alma9910634259105973}, single channel noise reduction remains important as it is robust against variations in microphone locations and device acoustics. In other words, very little tuning is needed in the final stages of production. The main objective of noise suppression is to reduce the background noise level in noisy speech while preserving speech quality, speech intelligibility, and naturalness of background noise. The old reason to do this is to reduce listener fatigue: codecs are optimized for speech and especially in low bit-rate modes background noise tends to become unnatural and rather annoying. A newer reason is to enable voice conferences and other such communication use cases, where there are more than two lines open, e.g., in a conference call it is desirable to hear everyone's voice but no one's background. In such use cases the background noise can be strongly time-varying, e.g., family, pets, bells, and keyboards have been quite typical noise sources during the covid-19 era.
The real-world environments consist of both stationary as well as non-stationary noise types. The latter has been particularly challenging for traditional enhancement \mbox{methods \cite{cohen2001speech}}. Although DNNs have achieved good progress \cite{alexandre20, xu2014regression, goehring2019using} in tackling this, proper evaluation of these methods is still a significant challenge. Most published methods rely on objective metrics computed on simulated mixtures, improvements in which, may not translate to real-world benefits \cite{Gelderblom}.
However, for real-world applications such as mobile communication, the end goal is better measured by subjective metrics. Moreover, speech in simulated mixtures does not account for various real-world phenomena, e.g., the Lombard effect \cite{brumm2011evolution}. Due to these real-world constraints, it becomes imperative to conduct subjective evaluations of these models on real-world recordings. Currently, most studies have relied on subjective evaluation on simulated mixtures (e.g., \cite{reddy2020interspeech, reddy2021icassp, rao2021interspeech}), which is a good starting point for evaluation before going to real recordings. Finally, the most reliable results are obtained in conversational tests that require real-time implementation on hardware.
DNN models for real-world applications, e.g., mobile telephony, are deployed on edge devices that are constrained by limited computational capacity. The latency of the system should also be small enough to allow for real-time operation. The former can be expressed in terms memory footprint of the model, i.e., the number of parameters in the model, and, the number of operations often measured in terms of multiply-accumulate additions (MACs). Reducing the computational complexity of these models is a major motivation behind the current research in the field of knowledge distillation \cite{kim2021test} and model compression \cite{tan2021towards}. Additionally, these models need to process the audio in an online manner and output processed audio with minimum latency. For frequency-domain processing, the algorithmic latency is limited by the size of the synthesis window used for overlap-add processing.
In this paper, we present subjective listening tests for two low-latency real-time DNN variants: a unidirectional gated recurrent neural network (GRU) \cite{chung2014empirical} and a causal U-Net \cite{ronneberger2015u,choi2018phase}, and compare them with a traditional Wiener-filter baseline. The focus is on evaluating algorithms, and not their hardware implementations; therefore the tests are done using pre-recorded real-world material which is processed offline. Moreover, conversational tests are also beyond the scope of this work. The algorithmic latency of the models is 20 ms, which is suitable for communications. The networks are trained on simulated mixtures on data recorded using a device mockup. The listening tests are conducted with unsimulated recordings. The included noise types are stationary (e.g., car), slow-varying (e.g., traffic noise), and non-stationary (e.g., caf\'e) noises. The separated speech signals are evaluated by eleven expert listeners on the following four outcomes: \textit{speech intelligibility}, \textit{speech quality}, \textit{noise transparency}, and \textit{noise level with respect to speech}. Each outcome is rated as an opinion score on a scale of 1 to 5 and the final ratings are then obtained as an average, i.e., mean opinion score (MOS). The subjective evaluation framework used here is based on the ITU-T P.800 \cite{itu-tp800} and ITU-T P.835 \cite{itu-tp835} standards but refined to arrive at four highly perceptual and atomic evaluation attributes.
We show that DNNs are able to improve noise suppression for all noise conditions without considerable degradation in \textit{noise transparency} or \textit{speech quality}, and without any degradation in \textit{speech intelligibility}. Among the two DNN variants, U-Net performed slightly better but has higher computational complexity both in terms of memory footprint measured as number of model parameters, and, number of computations measured in terms of multiply-accumulate operations (MACs). We also report objective metrics for the two DNN variants on a simulated test set.
\section{Method }\label{sec:method}
The speech enhancement framework is depicted in \mbox{Fig. \ref{fig:dnn_block}}. An acoustic mixture is first subjected to short-time Fourier transform (STFT). Magnitude STFT, $|X(t, f)|$, is then decimated to a lower resolution for DNN processing. The DNN model is trained to predict a time-frequency (TF) mask which is then interpolated to restore the original STFT resolution. This TF mask, $\hat{G}(t, f)$, is then multiplied with mixture STFT to give TF spectra, $\hat{S}(t, f)$, of the enhanced speech signal. The loss function for DNN training is the mean-squared error between the estimated and ground truth speech TF representations. Thus, the mask prediction here is implicit in the sense that the network is trained to optimize directly for the speech TF output which is then subjected to inverse STFT and overlap-add processing to yield the time domain speech.
\subsection{Feature extraction}
The magnitude STFT spectra is computed on the principle of asymmetric windowing, i.e., the analysis window is longer than synthesis window to allow for better TF resolution, as was proposed in \cite{mauler2007low, wang_asym}. Since most of the spectral energy for human speech is concentrated in lower frequency bands, and the human ear is more sensitive to lower frequencies, the higher spectrum can be compressed. This also reduces the input/output size of the DNN and hence computational complexity. Roughly following the perceptually motivated Mel frequency scale, the first 54 of our 513 original frequency bins (FFT length of 1024) are used as such, while the rest of the spectrum is compressed to 12 bands of an increasing width by averaging the spectral bins within each band. The TF mask estimated by DNN is expanded to the original FFT resolution by piecewise constant interpolation.
\begin{figure}[!t]
\centering
\includegraphics[scale =0.4 ]{dnn_block_new}
\caption{\small DNN-based speech enhancement system. }
\label{fig:dnn_block}
\end{figure}
\begin{figure}[!b]
\centering
\includegraphics[scale=0.4]{unet_block_v2}
\caption{\small The U-Net architecture consisting of down-sampling blocks (1, 2), bottleneck block (3), and up-sampling blocks (4, 5). The sequence of operations in each block is also depicted.}
\label{fig:u-net}
\end{figure}
\subsection{DNN topologies}
Two types of DNN topologies are used: a unidirectional gated recurrent unit (GRU) network consisting of recurrent layers followed by a feedforward layer, and a U-Net architecture \cite{ronneberger2015u,choi2018phase} consisting of causal convolutional layers. The motivation behind this choice is to test two widely used architectures: recurrent layers \cite{hao2021fullsubnet, xu14, goehring2019using}, and convolutional layers \cite{hu2020dccrn, bulut, chandna2017monoaural, naithani2017low}, for the real-world speech enhancement task. The U-Net architecture, depicted in Fig. \ref{fig:u-net}, is a sequence of down-sampling and up-sampling blocks with a bottleneck block in between. The network consists of two down-sampling and two up-sampling blocks. Each of these blocks consists of two convolutional layers. There is skip connection between down-sampling and up-sampling blocks as depicted. The down-sampling is achieved by max-pooling and up-sampling is achieved through transposed convolutions. The bottleneck block does not have any down/up sampling operation. The sequence of operations is also depicted. The order of operation in an up-sampling block is as follows: the input is first up-sampled and then concatenated with the skip connection input along the channel dimension. This is then fed to convolutional layers in the block. The network is thus symmetrical. The Fig. \ref{fig:u-net} also shows the output dimension at each stage of processing for an input of shape $(1, \, N,\, 66)$, where $N$ is the number of STFT frames in the input. If $C$ is the number of filters in the convolutional layers in the first down-sampling block, the number of channels at the output of each block is given by,
$1 \textrm{(input)}\rightarrow C\rightarrow 2C\rightarrow 4C\rightarrow 2C\rightarrow C\rightarrow 1 \textrm{(output)},
$
\noindent where the arrows show the processing chain in Fig. \ref{fig:u-net}. The kernel size of $5 \times 5$ in time and frequency axes is used. The output block is a pointwise $1 \times1 $ convolution block. The convolutional layers are implemented using depth-wise separable convolutions \cite{chollet2017xception} to reduce the computational complexity of the network.
\subsection{Baseline}
The classical Wiener filter,
$$
G(t, f) = \frac{\xi(t, f)}{\xi(t, f)+1},\ \ \ \xi(t, f) = \frac{|S(t, f)|^2}{|N(t, f)|^2}
$$
is selected as the ground truth mask that DNNs estimate from input mixture spectrum, $X(t, f) = S(t, f)+N(t, f)$, where $X(t, f)$, $S(t, f)$ and $N(t, f)$
refer to STFT of noisy speech, speech and noise, respectively, and $\xi(t, f)$ is the \mbox{a priori} SNR.
Thus, it is fair to compare outputs of DNNs also to a traditional noise suppressor where noise is estimated separately.
Noise is estimated via
$$
|\hat{N}(t, f)|^2 = \alpha |\hat{N}(t-1, f)|^2 + \left(1- \alpha\right) |X(t, f)|^2,
$$
and \mbox{a priori SNR} $\xi(t, f)$ is estimated using "decision directed" approach of \cite{EphraimMalah} as
$$
\hat{\xi}(t, f) = \alpha G(t-1, f)^2\gamma(t, f) + (1-\alpha) \max(\gamma(t, f) - 1, 0),
$$
where \mbox{a posteriori} SNR, $\gamma(t, f) = \frac{ |X(t, f)|^2}{|N(t, f)|^2}$.
The parameter $\alpha$ controls update speeds of the estimates. In our version we have set it separately for each frequency band
and time instant via trial and error using estimates of pitch, other features typical to speech, the spectral distance between noisy speech and noise estimate
and minimum statistics estimate \cite{Martin94spectralsubtraction} \cite{MartinR2001Npsd}.
In practical applications \mbox{a priori} SNR is limited below and the limit dictates maximum noise attenuation, typically the limit is set corresponding to a maximum attenuation of 12 dB.
\begin{table*}[!ht]
\centering
\caption {\small The subjective evaluation framework to assign opinion scores.}
\label{tab:eval}
\vfill
\begin{tabular}{@{}cllll@{}}
\toprule
Score & \textit{Speech intelligibility} & \textit{Speech quality} & \textit{Noise transparency} & \textit{Noise level w.r.t speech} \\ \midrule
5 & Fully intelligible & Excellent & Natural & \begin{tabular}[c]{@{}l@{}}Noise level is low compared to speech \\ but not fully muted.\end{tabular} \\
4 & Slightly raised effort needed to understand & Good & Natural with minor issues & \begin{tabular}[c]{@{}l@{}}Noise level is moderately low, not \\ fully muted\end{tabular} \\
3 & Clearly raised effort needed to understand & Fair & \begin{tabular}[c]{@{}l@{}}Mostly natural with some annoying \\ features or fully muted\end{tabular} & Noise level is moderate or fully muted \\
2 & Difficult to understand & Poor & \begin{tabular}[c]{@{}l@{}}More unnatural than natural, e.g., \\ considerable discontinuties\end{tabular} & Noise level is moderately high \\
1 & Impossible to understand & Bad & Unnatural & Noise level is high w.r.t speech \\ \bottomrule
\end{tabular}
\end{table*}
\section{Evaluation} \label{sec:evaluation}
\subsection{Dataset}
For the training data, speech recordings were captured in an anechoic chamber, quiet listening rooms and quiet offices. The recordings consist of short sentences spoken in English or in Finnish. In the anechoic chamber, the speakers were instructed to keep voice levels steady while speech levels were more natural in the recordings that took place in an office or a listening room. There are 29 speakers, most of them native Finnish, aged 25--60, in the dataset, consisting of both male and female voice types. The noise recordings correspond to various real-world stationary and non-stationary environments like a train station, orchestra tuning session, wind-tunnel, vacuum cleaner noise, etc. The length of each recording is approximately 40 seconds and they were all recorded with four similar headset mockup devices. This dataset was augmented using CSTR-VCTK corpus for speech and TUT acoustic scenes 2017 dataset for noise. The former consists of speech read from a newspaper by 110 English speakers. Around 20 minutes of speech is available for each speaker. TUT Acoustic Scenes dataset consists of recordings from 15 acoustic scenes, e.g., bus, cafeteria, car, metro station, etc., and around 52 min of audio is available for each acoustic scene. To compensate for channel mismatch, we first convolved the speech signals with an impulse response derived from the mockup recordings. The impulse responses were calculated to estimate impulse response between the user's mouth and the microphone of the device and since there were many users, one impulse response was randomly selected for each speech signal. In addition, the speech signal levels were scaled to values randomly sampled from the histogram of speech signal levels derived from the mockup recordings.
The real-world recordings for subjective evaluation were captured using two different kinds of mockups and one product prototype in Tampere, Finland, and Shanghai, China, in order to ensure that the devised algorithms work properly with different kinds of devices in the same device category. The mockups were an in-house design with microphones and loudspeaker, but without a processing unit, i.e., they can be used for recording and playback in the early stages of algorithm development. Fifteen speakers were involved. The recorded samples for the listening test were about 10 seconds long. In total 168 such audio samples, i.e., 28 minutes of audio was included. The speakers, aged 25--50, native Finnish and Chinese, spoke English, Finnish and Chinese. The content of the speech was scripted conversations. Both female and male speakers were included. The signal-to-noise ratio (SNR) was not predefined but users talked naturally and had phone conversations in natural environments, e.g., in Tampere railway station, in a car on a highway from Tampere to Helsinki, or during lunch time in a canteen. The estimated average SNR ratio of the recordings estimated by averaging signal powers during speech absence and presence is 7 dB. However we must emphasize that this is estimated with noise level changing due to changes in background (e.g., a car passing by) and also the speaker raising and lowering her/his voice. Thus, the estimation is not fully comparable with the SNRs used in generating simulated mixtures. Following noisy scenarios are included: stationary (e.g., car), slowly varying (e.g., train station, street, etc.), non-stationary noises (e.g., metro area, orchestra noise, etc.) including babble (caf\'e) noise.
\subsection{Experiments}
The DNNs were trained on simulated mixtures while the evaluation was done on real-world unsimulated mixtures. For training, the dataset was split into training and validation folds consisting of 67 \% and 33 \% of the available data. Chunks of speech and noise recordings were randomly sampled and summed. The speech/noise signals were first downsampled to 32 kHz. Prior to the summation, a random signal-to-noise ratio (SNR) was chosen in the range [-5, 5] dB. The noise signal was scaled with respect to the speech thereby preserving the speech levels. The mixture signal was then high-pass filtered using a second-order Butterworth filter with a cut-off frequency at 150 Hz in order to suppress low-frequency components. This was done as some types of noises in the dataset were found to be dominated by low-frequency components that cannot be played back with relatively small loudspeakers typical in telephony. The asymmetric analysis window of effective length 22.5 ms (window design \cite{wang_asym}) and 50 \% frame shift was used. The algorithmic latency of the system, determined by length of the synthesis window, is 20 ms. The maximum noise attenuation was limited to 15 dB when processing for the listening tests, i.e., 3 dB more attenuation is allowed than in the baseline method.
A unidirectional GRU network with 1 layer and 128 GRU units followed by a feedforward layer was used. The U-Net configuration, depicted in Fig. \ref{fig:u-net} with $C$ as 16, was used. These network topologies were chosen from among several network candidates through grid hyper-parameter search based on acceptable objective performance on a simulated test set and desired computational complexity. The computational complexity for these model architectures in terms of the number of network parameters and multiply-accumulate operations (MACs) per second of input audio during a single inference pass, is reported in Table \ref{tab:obj_metrics}. For training, the Adam optimizer \cite{kingma2014adam} with default learning rate parameters was used as the optimizer. Early stopping was used to stop the training when the validation error stops decreasing for 20 consecutive epochs. The Librosa \cite{mcfee2015librosa} library was used for audio processing and the PyTorch \cite{paszke2017automatic} framework was used for training the models.
\subsection{Subjective evaluation}\label{sec:eval_listening}
The subjective evaluation was arranged as a blind listening test and 22 experts with a strong background both in critical listening and audio signal processing, with an age range of 26--60 years and known to have normal hearing within their age cohort. Eleven persons participated without the organizer knowing their identity within the invitation list. Each participant anonymously picked a randomized set of listening samples and conducted the evaluation using their high-quality PC sound interface and headphones.
The evaluation framework consists of computing mean opinion scores (MOS) on four characteristics of the speech processed via DNN/baseline noise reduction: \textit{speech intelligibility}, \textit{speech quality}, \textit{noise transparency}, and \textit{noise level w.r.t speech}. Here, \textit{speech intelligibility} refers to listening effort, according to a rating scale similar to the listening-effort scale of ITU-800 but slightly modified based on our experience to make judging between the scores even easier. \textit{Speech quality} refers to how pleasant and natural the speech sounds according to the respective scale of ITU-T P.800, regardless of how easy it is to understand the speech. For the assessment of the background noise, we have expanded from the one-dimensional background rating scale of ITU-T P.835 to acquire a more diverse understanding of the impact of speech enhancement processing on noise. Our first noise evaluation attribute is \textit{noise transparency} which refers to the similarity of the residual noise in the enhanced speech to the listener's expectation of the original constituent noise. Our definition of transparency is loosely based on \mbox{ITU-R BS.2399-0.} \cite{itu-rbs2399}. The transparency is high when the residual noise is well recognizable and sounds natural, though optimally quieter after noise suppression. \textit{Noise level w.r.t. speech}, on the other hand, indicates the experienced loudness of residual noise in comparison to the experienced loudness of speech. Comparing the result of this attribute for the processed speech to that for the input signal indicates the experienced suppression of noise by the speech enhancement system. The evaluation framework is described in detail in Table \ref{tab:eval}.
Note that we intentionally did not include an overall quality attribute in the subjective evaluation. Rather, we selected the four attributes described above that are highly perceptual --- and thus objective --- as well as atomic in nature. In our experience, this increases the reliability and reproducibility of the evaluation results. Also, we found it more meaningful to assess the listening effort rather than "hard" word intelligibility. Word intelligibility tends to decrease only in rather harsh noise conditions, whereas listening effort begins to increase already in less severe conditions, thus the latter is more broadly applicable to evaluation. A limitation of this approach is that none of these attributes can be directly compared to an instrumental overall quality or word intelligibility prediction metric. The chosen subjective attributes and instrumental metrics rather complement each other.
Although the algorithms were trained with 32 kHz data, the real-world recordings were captured with a sampling rate of 16 kHz. These were then upsampled to 32 kHz for DNN processing. The motivation behind this is a pragmatic one, i.e., using the same model for wideband and super-wideband applications instead of training separate models for each. It was verified that there is not much deviation between the objective metrics for DNNs trained/tested with 16 kHz data and DNNs trained/tested on 32 kHz data (upsampled from 16 kHz as mentioned above). The same was verified in informal listening tests. The objective metrics considered here are source to distortion ratio (SDR)\cite{vincent2006performance}, short-time objective intelligibility (STOI) \cite{taal2010short}, and perceptual evaluation of speech quality (PESQ) \cite{rix2001perceptual}.
\subsection{Results and discussion}
\begin{figure}[!t]
\centering
\includegraphics[scale=0.165, center]{Figure_1.eps}
\caption{\small MOS values for all conditions, babble, and quiet cases separately. The scores are depicted with 80 \% confidence interval.}
\label{fig:agg_noise}
\end{figure}
Fig. \ref{fig:agg_noise} shows the average mean opinion scores averaged over all conditions, babble noise, and quiet cases. We used 80 \% confidence interval in the analysis as we have found it sufficient when comparing the performance of solutions targeted to commercial applications. It is evident that DNNs suppressed noise in all conditions and the 3 dB difference in maximum attenuation explains some, but not all, improvement in speech vs. noise level, namely the difference is higher than average in the babble case. The U-Net preserved \textit{speech intelligibility} slightly better than the GRU, and both DNN variants were better than the baseline. The same was observed in informal listening. Otherwise, the two DNN variants performed similarly with no major statistical difference between them.
We see that, as expected, the use of DNNs was most beneficial in the babble noise case where the noise signal resembles the target signal most. Moreover, we observed that the more stationary the background noise is, the more the listener appreciated the more mature Wiener filtering based baseline solution. Thus, stationary noise cases need more attention, e.g., post-processing or simply more training data. Additionally, on average, DNN processing either preserved \textit{noise transparency} or did not heavily degrade it. The quiet condition in Fig \ref{fig:agg_noise} refers to recording scenarios where practically no background noise or only a low level of background noise was present (quiet offices, listening room, etc.). The DNNs and the baseline method performed equally well in the quiet case.
We also report the objective metrics for the initial and DNN processed conditions corresponding to a simulated test set as well as a comparison of the computational complexity of the solutions in Table \ref{tab:obj_metrics}. A notable observation here is that there is a divergence between the objective scores on simulated data, i.e., STOI and PESQ, and the subjective scores on real-world data. Although it is not fair to compare the two, it should also be noted that the two modes of evaluation can not be directly compared but rather complement each other as was described in Section \ref{sec:eval_listening}.
Since the baseline method involves a significant amount of smoothing in time and frequency when estimating noise and gain, we expected to have superior performance for DNNs in noise cases where there are sudden changes. In case of stationary or slowly varying noise types no significant improvement is expected. The DNNs at hand also share some limitations of traditional baseline noise suppressor, i.e., time frequency masking corresponds to filtering in the frequency domain with a time-varying zero-phase filter, i.e. it does not modify the phase of the signal and cyclic convolution produces artifacts that are handled in overlap and add processing. Thus, the maximum noise attenuation provided by suppressors is limited so that residual noise masks artifacts. At best noise suppressor preserves the speech and attenuates noise.
\vspace{2mm}
\begin{table}[!t]
\centering
\caption {\small The objective metrics on a simulated test set and computational complexity of the DNNs.}
\label{tab:obj_metrics}
\vfill
\begin{tabular}{@{}lccccc@{}}
\toprule
Model & SDR & STOI & PESQ & Param. & \begin{tabular}[c]{@{}l@{}}MACs \\(per sec) \end{tabular} \\ \midrule
Initial & 0.2 & 0.65 & 2.1 & - & - \\
GRU & 9.7 & 0.67 & 2.5 & 83 K & 8.5 M \\
U-Net & 9.6 & 0.68 & 2.6 & 24 K & 47 M \\ \bottomrule
\end{tabular}
\end{table}
\section{ Conclusions} \label{sec:conclusion}
In this paper, we evaluated two DNN based speech enhancement systems and reported subjective listening results in the form of average mean opinion scores corresponding to four attributes of separated speech: \textit{speech quality}, \textit{speech intelligibility}, \textit{noise transparency}, and \textit{noise level w.r.t. speech}. We showed that DNNs improve noise suppression in all noisy scenarios without a negative impact on speech intelligibility and without strong degradation of speech quality or noise transparency. Notably, the DNNs are capable of improving speech quality and noise transparency for babble and other non-stationary scenarios, for which traditional, non-DNN based methods are sub-optimal. Out of these selected methods, U-Net seemed to preserve speech intelligibility slightly better than GRU at the expense of higher computational complexity and memory consumption. With these DNNs at least 3 dB more noise attenuation is possible without degrading other quality attributes. For improving other quality attributes besides noise suppression, more sophisticated methods should be considered. Such methods could, e.g., employ multiple microphones or incorporate phase processing.
\section{Acknowledgements}
The authors would like to thank the CSC-IT Centre of Science Ltd., Finland, for providing computational resources used for the experiments reported in this paper.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
In this paper we study a median-of-means algorithm
for multidimensional randomized quasi-Monte Carlo (RQMC) sampling
over $[0,1]^s$ for $s\ge1$.
The problem in RQMC is to estimate
$\mu =\int_{[0,1]^s}f(\boldsymbol{x})\,\mathrm{d}\boldsymbol{x}$.
The familiar Monte Carlo estimate is the
mean $\hat\mu$ of $f(\boldsymbol{x}_i)$ for $n$ independent
$\boldsymbol{x}_i\sim\mathbb{U}[0,1]^s$, with a root
mean squared error (RMSE) of $O(n^{-1/2})$
when $f$ has finite variance.
A quasi-Monte Carlo (QMC) estimate \cite{nied92}
replaces those
$n$ points by deterministic points
strategically chosen to more
uniformly sample the unit cube \cite{dick:pill:2010}.
The resulting absolute error is $O(n^{-1+\epsilon})$
for any $\epsilon>0$ when $f$ has finite total
variation in the sense of Hardy and Krause.
Randomizing those points \cite{rtms}
in such a way that they remain digital nets
provides independent unbiased estimates of $\mu$
allowing one to estimate accuracy statistically.
For smooth enough $f$, the randomization also
improves the RMSE to $O(n^{-3/2+\epsilon})$
\cite{smoovar}.
The usual way to combine independent replicates of
randomized digital nets is to simply take the
average of the replicate estimates. The method
we study here is to instead take the median
estimate from $2r-1$ independent replicates
when using the random linear
scramble from \cite{mato:1998}.
In \cite{superpolyone} we studied the case $s=1$.
The median-of-means proposal in \cite{superpolyone}
uses a $(0,m,1)$-net in base $2$ randomized
with a random linear scramble of Matousek \cite{mato:1998}
and a digital shift.
For $f$ analytic on $[0,1]$ with integral $\mu$
estimated by an infinite precision RQMC
estimator denoted by $\hat\mu_\infty$ we saw
that the median of the randomization distribution of
$\hat\mu_\infty-\mu$ converges to $0$
as $O(n^{-c\log_2(n)})$ for any $c<3\log(2)/\pi^2\approx 0.21$.
That same rate could be attained by the sample median
of $2r-1$ independently replicated RQMC estimates so
long as $r=\Omega(m)$ by
which we mean $m=O(r)$ as both $r$ and $m$
go to infinity.
That paper also considered integrands whose
$\alpha$ derivative satisfied a $\lambda$-H\"older
condition and found an error of
$O(n^{-\alpha-\lambda+\epsilon})$ for that case.
The significance of this result is that we can
attain a better rate than the customary
mean of replicated
estimates and that rate
can adapt to an unknown smoothness
level of the integrand without the user
having to know the smoothness level. Indeed when
many integrals are computed from the same inputs
we might know that they have different smoothness
levels.
The previous paper was limited to $s=1$, where
there are many other good ways to integrate
a smooth function over $[0,1]$, as in \cite{davrab}.
That paper did however include a numerical
result for
the OTL circuit function on $[0,1]^6$
from \cite{surj:bing:2013}.
There the standard deviation of a median
of means estimator was superior to that of the usual
mean-of-means at practically relevant sample sizes.
In the present paper we consider analytic functions
$f:[0,1]^s\to\mathbb{R}$. We find that the median
value of $\hat\mu_\infty-\mu$ is now
$O(n^{-c\log_2(n)/s})$ for any $c<3\log(2)/\pi^2.$
In other words, there is still superlinear
convergence but with a dimension effect.
An outline of this paper is as follows.
Section~\ref{sec:background} introduces some
notation as well as the integration problem
and scrambling algorithms.
Section~\ref{sec:decomp} decomposes the
RQMC error into a sum over nonzero vectors
of $s$ nonnegative integers. It is a sum
of a randomly selected set of randomly
signed Walsh coefficients.
That section introduces some notation that
we need to describe the complexity of the
Walsh basis functions and then presents
an upper bound on Walsh coefficients from Yoshiki \cite{yosh:2017}.
Section~\ref{sec:asymptotic}
gives asymptotic properties of the median
of means estimator. It bounds
the probability that a Walsh coefficient
contributes to the error and it shows
that the probability of an integration error
above $2^{-\lambda m^2/s+O(m\log(m))}$
is $O(1/\sqrt{m})$ when scrambling
a $(t,m,s)$-net in base $2$.
It also shows superpolynomial convergence
for some finite precision estimates
where the number of bits in the sample
values grows faster than a certain
multiple of $m^2/s$ and the median
of $\Omega(m^2)$ independent copies is used.
Section~\ref{sec:finitesample}
looks at finite sample performance of the method
and gives conditions where a median-of-means
can outperform a mean-of-means for large $s$
and feasible $m$, despite the dimension
effect. This may happen when the
integrand is dominated by contributions from
a small set of important variables.
Section~\ref{sec:discussion} has a
discussion of the results focusing on
two remaining challenges: adaptation
to unknown smoothness, and quantifying uncertainty.
The median-of-means setting makes use of
techniques from analytic combinatorics that have
previously seen very little use in quasi-Monte
Carlo. That literature has quite different
methods and notational conventions, and the
results we derive with it are in an Appendix.
We close the introduction with some bibliographic
remarks on median-of-means. It is a longstanding
method in theoretical computer science. See
\cite{jerr:vali:vazi:1986} and \cite{lecu:lera:2020}
for some old and new uses, respectively.
Several uses in information based complexity are
discussed in \cite{kunsch2019solvable}.
Uses in quasi-Monte Carlo include \cite{superpolyone}
metioned above as well as
\cite{hofstadler2022consistency} for some laws
of large numbers, \cite{goda:lecu:2022:tr}
for some smoothness adaptive lattice rules and
\cite{gobe:lera:meti:2022:tr} for robust RQMC
estimates.
\section{Notation and background}\label{sec:background}
We use $\mathbb{N}$ for the natural numbers,
$\mathbb{N}_0=\mathbb{N}\cup\{0\}$
and $\mathbb{Z}_n=\{0,1,\dots,n-1\}$ for integers $n\ge2$.
For $K\subset\mathbb{N}$ we use $|K|$ for its cardinality.
We use $\mathcal{N}=\{K \subset\mathbb{N}\mid |K|<\infty\}$.
For $K\in\mathcal{N}$ we use $\lceil K\rceil$ to
denote the largest element of $K$ with
$\lceil\emptyset\rceil=0$ by convention.
When $x\in\mathbb{R}$ we use
$\lceil x\rceil$ for the smallest integer greater than
or equal to $x$. The context will make it clear
whether the argument to $\lceil \cdot\rceil$ is
a real number or a set of natural numbers.
We let $\boldsymbol{0}$ be a vector of $m$ zeros
and we set $\mathbb{N}^s_*=\mathbb{N}_0^s\setminus\{\boldsymbol{0}\}$.
We abuse notation slightly by letting $\boldsymbol{0}$ be
either a row or a column vector as needed.
For $s\in \mathbb{N}$ and $f\in L^2[0,1]^s$ we study approximation
of
$$
\mu=\int_{[0,1]^s}f(\boldsymbol{x})\,\mathrm{d}\boldsymbol{x}
\quad
\text{by}
\quad
\hat\mu=\frac1n\sum_{i=0}^{n-1}f(\boldsymbol{x}_i)
$$
for $n\ge1$ and $\boldsymbol{x}_i\in[0,1]^d$.
We use $1{:}s=\{1,2,\dots,s\}$ for the set
of input indices to $f$.
When $v\subseteq1{:}s$ we use
$-v$ for $1{:}s\setminus v$.
We use a van der Corput style mapping
between natural numbers and bit vectors and
points in $[0,1)$ as follows.
For $i\in\mathbb{Z}_{2^m}$ we let $\vec{i} = \vec{i}[m] = (i_1,i_2,\dots,i_m)^\mathsf{T}\in\{0,1\}^m$
where $i = \sum_{\ell=1}^mi_\ell2^{\ell-1}$.
For $a=\sum_{\ell=1}^m a_\ell2^{-\ell}\in[0,1)$ we let
$\vec{a} = \vec{a}[E] = (a_1,a_2,\dots,a_E)^\mathsf{T}$.
Here $E$ is the precision of $\vec{a}$
and we typically have $E\ge m$ in our use cases.
For $a$ with two binary expansions we choose the one ending
in infinitely many 0s.
For each $\vec{a}$ there is a unique
$a\in[0,1)$. When $E<\infty$, we can
have $\vec{a}=\vec{a\,}'$ for $a\ne a'$.
For an integer base $b\ge2$ and vectors $\boldsymbol{k},\boldsymbol{c}\in\mathbb{N}_0^s$ with $0\le c_j<b^{k_j}$,
an elementary interval in base $b$ is a Cartesian
product of the form
$$
\prod_{j=1}^s\Bigl[
\frac{c_j}{b^{k_j}},\frac{c_j+1}{b^{k_j}}\Bigr).
$$
For integers $m\ge t\ge0$, the points $\boldsymbol{x}_0,\dots,\boldsymbol{x}_{b^m-1}\in[0,1)^s$
form a $(t,m,s)$-net in base $b$ if every
elementary interval with $\sum_{j=1}^sk_j=m-t$
contains precisely $b^t$ of those points.
Here, $t$ is the quality parameter of the net
with smaller values being better. It is not
always possible to get $t=0$ for a given
choice of $b$ and $m$ and $s$.
The infinite sequence $\boldsymbol{x}_i\in[0,1)^s$ for $i\in\mathbb{N}_0$ forms a $(t,s)$-sequence in base
$b$ if for all integers $m\ge t$ and $r\ge0$,
the points $\boldsymbol{x}_{rb^m},\dots,\boldsymbol{x}_{(r+1)b^m-1}$
form a $(t,m,s)$-net in base $b$.
In this paper we consider $b=2$. This includes
the most widely used nets of Sobol' \cite{sobo:1967}
as well
as those of Niederreiter and Xing \cite{niedxing96} that have
some of the best available $t$ values.
Base $2$ digital nets of $n=2^m$ points are formed by setting
\begin{align}\label{eq:defveca}
\vec{a}_{ij} = C_j\vec{i}\ \ \mathsf{mod}\ 2
\end{align}
for $0\le i<2^m$ and $j=1,\dots,s$
for carefully chosen generator matrices
$C_j =C_j[E]\in\{0,1\}^{E\times m}$
where $E\ge m$ is a precision.
Our theoretical
analysis emphasizes $E=\infty$.
The attained value of $t$ is a property
of the chosen generator matrices.
We always assume that $C_j$ has full rank
over $\mathbb{Z}_2$.
The points $\boldsymbol{a}_i\in[0,1)^s$ have components
$a_{ij}$ determined by $\vec{a}_{ij}$ from equation~\eqref{eq:defveca}. That is, we give
expressions for $\vec{a}_{ij}$ with the
understanding that $a_{ij}\in[0,1)
=\sum_{\ell=1}^E2^{-\ell}a_{ij\ell}$
when $\vec{a}_{ij} = (a_{ij1},a_{ij2},\dots,a_{ijE})$.
For a base $2$ digital $(t,s)$-sequence one uses generator
matrices with infinitely many rows and columns.
For $m\ge t$, the first $n=2^m$ points of such a sequence
are a $(t,m,s)$-net in base $2$.
When we consider a digital sequence we suppose
that for each finite $m$ we are working with
$C_j$ equal to the upper left $E\times m$ submatrix
of the infinite generator matrix. Note that any entries
in $\vec{i}$ after the $m$'th are zero for $i\in\mathbb{Z}_{2^m}$
so columns of $C_j$ after the $m$'th do not affect $\hat\mu_n$.
Given points $\boldsymbol{a}_i = (a_{i1},\dots,a_{is})\in[0,1)^s$
of a digital net, we
define linearly scrambled points as follows.
For precision $E\ge m$ we choose random matrices $M_j\in\{0,1\}^{E\times m}$
and random vectors $D_j \in\{0,1\}^{E}$
and take
\begin{equation}\label{eqn:xequalMCiplusD}
\vec{x}_{ij}=\vec{x}_{ij}[E]=\vec{a}_{ij}+\vec{D}_j=M_jC_j\vec{i}+\vec{D}_j
\ \ \mathsf{mod}\ 2
\end{equation}
for $0\le i<2^m$ and $1\le j\le s$ to define $\boldsymbol{x}_i\in[0,1]^s$.
From here on, arithmetic operations on bit vectors are
taken modulo two unless otherwise indicated.
Our estimate of $\mu$ is now
$$
\hat\mu = \hat\mu_E = \frac1n\sum_{i=0}^{n-1}f(\boldsymbol{x}_i).
$$
For $E'<E$ we define $\hat\mu_{E'}$ as above
keeping only the first $E'$ rows of $M_j$ and
the first $E'$ entries in $D_j.$ Our reduced
precision estimate $\hat\mu_{E'}$ uses
$\vec{x}_{ij}[E']=M_j(1{:}E',:)C_j\vec{i}+\vec{D}_j(1{:}E')$.
\begin{lemma}\label{lem:precisionerror}
Let $f:[0,1]^s\to\mathbb{R}$ have modulus of continuity $\omega_f$.
Let $M_j$ and $D_j$
for $j=1,\dots,s$ be defined with
infinite precision. Then
$$|\hat\mu_\infty-\hat\mu_E|\leq \omega_f\Bigl(\frac{\sqrt{s}}{2^E}\Bigr).$$
\end{lemma}
\begin{proof} Let $\boldsymbol{x}_i[E]$ be $\boldsymbol{x}_i$ under scrambling with
precision $E$ and $\boldsymbol{x}_i[\infty]$ be $\boldsymbol{x}_i$ under scrambling in
the infinite precision limit. By Lemma 1 of \cite{superpolyone}, each coordinate of $\boldsymbol{x}_i[E]$ differs from $\boldsymbol{x}_i[\infty]$ by at most $2^{-E}$. Therefore
$\Vert\boldsymbol{x}_i[E]-\boldsymbol{x}_i[\infty]\Vert_2\leq \sqrt{s}2^{-E}$ and so
$$|\mu_\infty-\mu_E|\leq \frac{1}{n}\sum_{i=0}^{n-1} |f(\boldsymbol{x}_i[E])-f(\boldsymbol{x}_i[\infty])|\leq \omega_f\Bigl(\frac{\sqrt{s}}{2^E}\Bigr).
\qedhere$$
\end{proof}
We will use $\omega_f(\sqrt{s})$ as
shorthand for
$\sup_{\boldsymbol{x}\in [0,1]^s}f(\boldsymbol{x})-\inf_{\boldsymbol{x}\in[0,1]^s}f(\boldsymbol{x})$.
We focus on the random linear scrambling of \cite{mato:1998}.
The matrix $M_j\in\{0,1\}^{E\times s}$
is lower triangular with ones on the diagonal
and independent $\mathbb{U}\{0,1\}$ entries
below the diagonal. The digital shift
has independent $\mathbb{U}\{0,1\}$ elements.
That is
$$M_{j,\ell\ell'}
=\begin{cases}
0, & 1\le \ell <\ell'\le m\\
1, & 1\le \ell =\ell'\le m\\
\mathbb{U}\{0,1\}, &\text{else,}
\end{cases}
$$
and $D_{j,\ell}=\mathbb{U}\{0,1\}$
for $\ell=1,\dots,E$.
We sketch this setting for $m=3$ and $E=4$
as follows:
\begin{align}\label{eq:thesketch}
M_j =
\begin{pmatrix}
1 &\\
u & 1 \\
u & u & 1\\
u & u & u
\end{pmatrix}
\quad\text{and}\quad D_j = \begin{pmatrix}
u\\
u\\
u\\
u
\end{pmatrix},
\end{align}
with $u$ representing random elements.
All of the uniform random variables in
$M_1,\dots,M_s$ and $D_1,\dots,D_s$ are
independent.
\section{Error decomposition}\label{sec:decomp}
In order to analyze the convergence rate of median-of-means, we first derive an error decomposition formula for $\hat{\mu}_{\infty}-\mu$ using Walsh functions.
For $k\in \mathbb{N}_0$ and $x\in [0,1)$, we define
\begin{align}\label{eq:defwalk}
\mathrm{wal}_k(x)=(-1)^{\vec{k{}}^\mathsf{T} \vec{x}}.
\end{align}
Because $k$ is a finite integer, only finitely many
entries in $\vec{k}$ are nonzero and so the inner product
in~\eqref{eq:defwalk} is a finite
sum. For the multivariate generalization, the $\boldsymbol{k}$'th dyadic Walsh
function $\mathrm{wal}_{\bsk}(\boldsymbol{x})$ for $\boldsymbol{k}\in\mathbb{N}_0^s$ is defined to be
\begin{equation}\label{eqn:Walshfunction}
\mathrm{wal}_{\bsk}(\boldsymbol{x})=\prod_{j=1}^s \mathrm{wal}_{k_j}(x_j)=(-1)^{\sum_{j=1}^s\vec{k{}}_j^{\mathsf{T}} \vec{x}_j}.
\end{equation}
It is known that $\{\mathrm{wal}_{\bsk}(\boldsymbol{x})\mid \boldsymbol{k}\in \mathbb{N}_0^s\}$ form a
complete orthonormal basis of $L^2[0,1)^s$
\cite{dick:pill:2010}. Therefore for $f\in L^2[0,1)^s$
\begin{align}\label{eqn:Walshdecomposition}
f(\boldsymbol{x}) &=\sum_{\boldsymbol{k}\in \mathbb{N}_0^s}\hat{f}(\boldsymbol{k}) \mathrm{wal}_{\bsk}(\boldsymbol{x}),
\quad\text{where}\\
\hat{f}(\boldsymbol{k})&=\int_{[0,1)^s}f(\boldsymbol{x})\mathrm{wal}_{\bsk}(\boldsymbol{x})\,\mathrm{d}\boldsymbol{x}.
\label{eqn:Walshcoefficient}
\end{align}
Equation~\eqref{eqn:Walshdecomposition} holds in a mean square sense.
\begin{theorem}\label{thm:decomp}
Let $f\in L^2[0,1)^s$ and let $\boldsymbol{x}_i$ be defined
by~\eqref{eqn:xequalMCiplusD} for $0\le i<2^m$.
Then
\begin{equation}\label{eqn:errordecomposition}
\hat{\mu}_{\infty}-\mu=\sum_{\boldsymbol{k}\in \mathbb{N}_*^s}\boldsymbol{1}\Biggl\{\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\Biggr\}\hat{f}(\boldsymbol{k}) (-1)^{\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} \vec{D}_{j}}.
\end{equation}
\end{theorem}
\begin{proof} From equation~\eqref{eqn:Walshcoefficient} we see that $\mu=\hat{f}(\boldsymbol{0})$. So by equations~\eqref{eqn:Walshfunction} and \eqref{eqn:Walshdecomposition},
$$\hat{\mu}_{\infty}-\mu=\sum_{\boldsymbol{k}\in \mathbb{N}_*^s}\hat{f}(\boldsymbol{k})\frac{1}{n}\sum_{i=0}^{n-1} \mathrm{wal}_{\bsk}(\boldsymbol{x}_i)=\sum_{\boldsymbol{k}\in \mathbb{N}_*^s}\hat{f}(\boldsymbol{k})\frac{1}{n}\sum_{i=0}^{n-1} (-1)^{\sum_{j=1}^s\vec{k{}}_{j}^\mathsf{T} \vec{x}_{ij}}.$$
From equation~\eqref{eqn:xequalMCiplusD}, we have
\begin{align*}
\frac{1}{n}\sum_{i=0}^{n-1} (-1)^{\sum_{j=1}^s\vec{k{}}_{j}^\mathsf{T} \vec{x}_{ij}}&=\frac{1}{n}\sum_{i=0}^{n-1} (-1)^{\sum_{j=1}^s\vec{k{}}_{j}^\mathsf{T} (M_j C_j\vec{i}+\vec{D}_j)}\\
&=(-1)^{\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} \vec{D}_{j}} \frac{1}{n}\sum_{i=0}^{n-1} (-1)^{\sum_{j=1}^s\vec{k{}}_{j}^\mathsf{T} M_j C_j\vec{i}}\\
&=(-1)^{\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} \vec{D}_{j}}\boldsymbol{1}\Bigl\{\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\Bigl\}
\end{align*}
so the conclusion follows.
\end{proof}
We need to quantify several
properties of Walsh function indices
$k$ and $\boldsymbol{k}$.
Let $k\in\mathbb{N}_0$ have binary expansion $k=\sum_{\ell=1}^\infty b_\ell 2^{\ell-1}$
for bits $b_\ell\in\{0,1\}$.
First we let
\begin{align}\vec{k} &:= \vec{k}[\infty] = (b_1,b_2,\dots)^\mathsf{T},\quad\text{and}\\
\kappa & := \{ \ell\in\mathbb{N} \mid b_\ell=1\}.
\end{align}
We will study $\mathrm{wal}_k$ using the cardinality of $\kappa$, the
sum of its elements, and its last (largest) element.
For $k\ge1$, these are
$$
|\kappa|,\quad
\Vert\kappa\Vert_1=\sum_{\ell\in\kappa}\ell,
\quad\text{and}\quad
\lceil \kappa\rceil = \max_{\ell\in\kappa}\ell,
$$
respectively.
For $k=0$ we set
$\kappa=\emptyset$ and then
$\Vert\kappa\Vert_0=\Vert\kappa\Vert_1=\lceil\kappa\rceil=0$,
the last one by convention.
In the $s$ dimensional setting we need to vectorize
these quantities.
For $\boldsymbol{k}=(k_1,\dots,k_s)\in\mathbb{N}^s$, we define
the corresponding vectors $\vec{k}_1,\dots,\vec{k}_s$
and sets $\kappa_1,\dots,\kappa_s$ componentwise.
We need to keep track of those indices in
$\boldsymbol{k}$ for which $k_j>0$. We denote the
supports of $\boldsymbol{k}$ and $\boldsymbol{\kappa}$ as
$\boldsymbol{s}(\boldsymbol{k})=\{j\in 1{:}s\mid k_j>0\}$
and $\boldsymbol{s}(\boldsymbol{\kappa})
=\{j\in 1{:}s\mid \kappa_j\ne\emptyset\}$
respectively. Clearly $\boldsymbol{s}(\boldsymbol{k})=\boldsymbol{s}(\boldsymbol{\kappa})$.
Given $\boldsymbol{k}\in\mathbb{N}_0^s$ we now define
the corresponding
bit matrix $\vec{\boldsymbol{k}}=(\vec{k}_1,\dots,\vec{k}_s)$
along with
$\boldsymbol{\kappa}=(\kappa_1,\dots,\kappa_s)\in\mathcal{N}^s$,
a list of finite sets of natural numbers.
We need some componentwise quantities
for $\boldsymbol{\kappa}$ and some aggregate quantities.
The componentwise quantities are
$$
\lceil\boldsymbol{\kappa}\rceil
= (\lceil\kappa_1\rceil,\dots,
\lceil\kappa_s\rceil)\in\mathbb{N}_0^s
\quad\text{and}\quad
|\boldsymbol{\kappa}| = (|\kappa_1|,\dots,|\kappa_s|)
\in\mathbb{N}_0^s.
$$
The first two aggregate quantities are
$$
\Vert\boldsymbol{\kappa}\Vert_1 = \sum_{j=1}^s\Vert\kappa_j\Vert_1
\quad\text{and}\quad
\Vert\boldsymbol{\kappa}\Vert_0=\sum_{j=1}^s|\kappa_j|.
$$
Note that $\Vert\boldsymbol{\kappa}\Vert_0$ is the
number of one bits in $\vec{\boldsymbol{k}}$.
We also need the sum of largest indices
$$
\Vert\lceil\boldsymbol{\kappa}\rceil\Vert_1 = \sum_{j=1}^s\lceil\kappa_j\rceil.
$$
These quantities satisfy
$$
\Vert\boldsymbol{\kappa}\Vert_0 \le \Vert \lfloor\boldsymbol{\kappa}\rfloor\Vert_1 \le \Vert \boldsymbol{\kappa}\Vert_1.
$$
Theorem 2 of \cite{yosh:2017}
provides the following crucial bound on $|\hat{f}(\boldsymbol{k})|$.
\begin{lemma}\label{lem:Walshcoefficientbound}
Let $f\in C^{\infty}[0,1)^s$. Then
\begin{align*}|\hat{f}(\boldsymbol{k})|&\leq 2^{-\Vert\boldsymbol{\kappa}\Vert_1-\Vert\boldsymbol{\kappa}\Vert_0}\sup_{\boldsymbol{x}_{\boldsymbol{s}(\boldsymbol{k})}\in[0,1)^{|\boldsymbol{s}(\boldsymbol{k})|}}\Big|\int_{[0,1)^{s-|\boldsymbol{s}(\boldsymbol{k})|}}f^{|\boldsymbol{\kappa}|}(\boldsymbol{x})\,\mathrm{d}\boldsymbol{x}_{-\boldsymbol{s}(\boldsymbol{k})}\Big|
\end{align*}
where
\begin{align*}
f^{|\boldsymbol{\kappa}|}&=f^{(|\kappa_1|,\dots,|\kappa_s|)}=\frac{\partial^{\Vert\boldsymbol{\kappa}\Vert_0}f}{\partial x_1^{|\kappa_1|}\cdots\partial x_s^{|\kappa_s|}}.
\end{align*}
\end{lemma}
Yoshiki's Theorem 2 uses a norm defined in
his Theorem 1 for smoothness $\alpha\ge2$.
Our setting has $\alpha=\infty$.
We take his $p=\infty$.
Our $\Vert\boldsymbol{\kappa}\Vert_0+\Vert\boldsymbol{\kappa}\Vert_1$
is his $\mu'_\alpha(\boldsymbol{k}_v)$.
\section{Asymptotic convergence rate}\label{sec:asymptotic}
In this section we derive the super-polynomial convergence rate of median-of-means. Many parts of the analysis will be refined in the next section to derive a tighter finite sample bound.
As a first step, we want to know the probability that $\sum_{j=1}^s
\vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}$ when $M_j$ is generated by random
linear scrambling. Recall that we have assumed that each $C_j$ is nonsingular.
We let $C_j(1{:}q,:)$ denote the first $q\ge0$ rows
of $C_j$ and then for $q_1,\dots,q_s\in\mathbb{N}$ we write
$$C^{(q_1,\dots,q_s)}=\begin{bmatrix}
C_1(1{:}q_1,:)\\
\vdots\\
C_s(1{:}q_s,:)
\end{bmatrix}
\in\{0,1\}^{(\sum_{j=1}^sq_j)\times m}$$
with the convention that when $q_j=0$, $C_j(1{:}q_j,:)$ is an empty matrix.
If $(q_1,\dots,q_s)=\boldsymbol{0}$, we define $C^{(q_1,\dots,q_s)}$ to be a
$0\times m$ matrix and it has rank 0. We will use $\mathrm{Row}(C)$ to denote the row
space of matrix $C$ in $\{0,1\}^m$.
For $\boldsymbol{v}\subseteq1{:}s$
we let $\boldsymbol{1}\{\boldsymbol{v}\}\in\{0,1\}^s$
be the vector with $v_j=1$ for $j\in v$ and $v_j=0$ for $j\not\in v$.
A very important quantity that recurs in our
analysis is the matrix
$C^{\lceil\boldsymbol{\kappa}\rceil}$.
For every $j$ with $k_j>0$, this
matrix has all the rows of $C_j$
that will be relevant to $\mathrm{wal}_{\bsk}(\boldsymbol{x}_i)$, namely $C_j(1{:}\lceil\kappa_j\rceil,:)$. If we remove the last relevant
row of each $C_j$ we obtain
$C^{\lceil\boldsymbol{\kappa}\rceil-\boldsymbol{1}\{\boldsymbol{s}(\boldsymbol{\kappa})\}}$.
\begin{lemma} \label{lem:gaincoef}
If $\max_{1\le j\le s}\lceil\kappa_j\rceil>m$, then
$$\Pr\bigg(\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\bigg)=2^{-m}.$$
If $\max_{1\le j\le s}\lceil\kappa_j\rceil\leq m$ and $\sum_{j\in
\boldsymbol{s}(\boldsymbol{k})}C_j(\lceil\kappa_j\rceil,:)\in \mathrm{Row}(C^{\lceil\boldsymbol{\kappa}\rceil-\boldsymbol{1}\{\boldsymbol{s}(\boldsymbol{k})\}})$, then
$$\Pr\bigg(\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\bigg)=2^{-\mathrm{rank}(C^{\lceil\boldsymbol{\kappa}\rceil-\boldsymbol{1}\{\boldsymbol{s}(\boldsymbol{k})\}})}.$$
Otherwise
$$\Pr\bigg(\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\bigg)=0.$$
\end{lemma}
\begin{proof}
Because of the upper triangular form for $M_j$
(recall the sketch in equation~\eqref{eq:thesketch}),
we see that $\vec{k{}}_j^\mathsf{T} M_j$ has the same distribution as $M_j(\lceil\kappa_j\rceil,:)$.
Because $C_j$ is nonsingular, if $\lceil k_{j^*}\rceil> m$ for any $j^*\in 1{:}s$, then $\vec{k{}}_{j^*}^\mathsf{T} M_{j^*} C_{j^*}$ is uniformly distributed on the set of $2^m$ possible binary vectors so that
\begin{align*}
Pr\Big(\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\Big)
&=\Pr\Big(\vec{k{}}_{j^*}^\mathsf{T} M_{j^*} C_{j^*}+\sum_{j\in 1{:}s, j\neq j^*} \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\Big)\\
&=\Pr\Big(\vec{k{}}_{j^*}^\mathsf{T} M_{j^*} C_{j^*}=\sum_{j\in 1{:}s, j\neq j^*} \vec{k{}}_j^\mathsf{T} M_j C_j\Big)=2^{-m}
\end{align*}
establishing the first claim.
Now assume that all $\lceil\kappa_j\rceil\leq m$. Then
$$\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j
=\sum_{j\in \boldsymbol{s}(\boldsymbol{k})}C_j(\lceil\kappa_j\rceil,:)+\sum_{j\in \boldsymbol{s}(\boldsymbol{k})}
\Bigl(\vec{k{}}_j^\mathsf{T} M_jC_j-C_j(\lceil\kappa_j\rceil,:)\Bigr).$$
Observe that $\vec{k{}}_j^\mathsf{T} M_jC_j-C_j(\lceil\kappa_j\rceil,:)$ is uniformly distributed on the linear span of first $\lceil\kappa_j\rceil-1$ rows of $C_j$. Hence the second sum on the right is uniformly distributed on $\mathrm{Row}(C^{\lceil\boldsymbol{\kappa}\rceil-\boldsymbol{1}\{\boldsymbol{s}(\boldsymbol{k})\}})$. If $\sum_{j\in \boldsymbol{s}(\boldsymbol{k})}C_j(\lceil\kappa_j\rceil,:)\in \mathrm{Row}(C^{\lceil\boldsymbol{\kappa}\rceil-\boldsymbol{1}\{\boldsymbol{s}(\boldsymbol{k})\}})$, then
$$\Pr\Big(\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\Big)=\frac{1}{|\mathrm{Row}(C^{\lceil\boldsymbol{\kappa}\rceil-\boldsymbol{1}\{\boldsymbol{s}(\boldsymbol{k})\}})|}=2^{-\mathrm{rank}(C^{\lceil\boldsymbol{\kappa}\rceil-\boldsymbol{1}\{\boldsymbol{s}(\boldsymbol{k})\}})}$$
establishing the second claim.
If $\sum_{j\in \boldsymbol{s}(\boldsymbol{k})}C_j(\lceil\kappa_j\rceil,:)\notin \mathrm{Row}(C^{\lceil\boldsymbol{\kappa}\rceil-\boldsymbol{1}\{\boldsymbol{s}(\boldsymbol{k})\}})$, then the above probability is clearly 0,
establishing the final claim.
\end{proof}
\begin{corollary}\label{cor:tmsbound}
If $C_1,\dots,C_s$ generate a digital $(t,m,s)$-digital in base $2$, then
$$\Pr\Big(\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\Big)\leq 2^{-m+t+s}.$$
\end{corollary}
\begin{proof}
We only need to verify that
$\mathrm{rank}(C^{\lceil\boldsymbol{\kappa}\rceil-\boldsymbol{1}\{\boldsymbol{s}(\boldsymbol{k})\}})\geq m-t-s$ when
$\max_{1\le j\le s} \lceil\kappa_j\rceil\leq m$ and $\sum_{j\in \boldsymbol{s}(\boldsymbol{k})}C_j(\lceil\kappa_j\rceil,:)\in \mathrm{Row}(C^{\lceil\boldsymbol{\kappa}\rceil-\boldsymbol{1}\{\boldsymbol{s}(\boldsymbol{k})\}})$. Notice that in this case $C^{\lceil\boldsymbol{\kappa}\rceil}$ is rank-deficient. By the definition of $(t,m,s)$-digital net, a rank-deficient $C^{\lceil\boldsymbol{\kappa}\rceil}$ must contains $m-t$ linearly independent rows, so $\mathrm{rank}(C^{\lceil\boldsymbol{\kappa}\rceil})\geq m-t$. Hence
$$\mathrm{rank}(C^{\lceil\boldsymbol{\kappa}\rceil-\boldsymbol{1}\{\boldsymbol{s}(\boldsymbol{k})\}})\geq \mathrm{rank}(C^{\lceil\boldsymbol{\kappa}\rceil})-|\boldsymbol{s}(\boldsymbol{k})|\geq m-t-s$$
which proves the conclusion.
\end{proof}
\begin{corollary}\label{cor:concentration}
Let $\lambda=3\log(2)^2/\pi^2\approx 0.146$.
For $j=1,\dots,s$ let $C_j=C_j(m)$
be the first $m$ columns of the
generator matrices of
a digital $(t,s)$-net in base $2$.
Then
$$\Pr\bigg(\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0} \text{ for some } \boldsymbol{k}\ne\boldsymbol{0} \text{ with }
\Vert \boldsymbol{k}\Vert_1\leq
\frac{\lambda m^2}{s}\bigg)=O\Bigl(\frac{1}{\sqrt{m}}\Bigr)$$
as $m\to \infty$.
\end{corollary}
\begin{proof}
From Corollary~\ref{cor:growthrate} in the appendix, we know that
$$ \bigl|\{\boldsymbol{k}\in \mathbb{N}_*^s\mid \Vert\boldsymbol{\kappa}\Vert_1\leq {\lambda m^2}/{s}\}\bigr|=\Theta\Bigl(\frac{2^m}{\sqrt{m}}\Bigr).$$
So from the union bound on
the result of Corollary~\ref{cor:tmsbound}
\begin{align*}
&\Pr\bigg(\,\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}
\text{ for some } \boldsymbol{k}
\in\mathbb{N}_*^s \text{ with } \Vert\boldsymbol{\kappa}\Vert_1\leq \frac{\lambda m^2}{s}\bigg)\\
&\leq 2^{-m+t+s}\bigl|\{\boldsymbol{k}\in \mathbb{N}_*^s\mid \Vert\boldsymbol{\kappa}\Vert_1\leq {\lambda m^2}/{s}\}\bigr|\\
&=O\Bigl(\frac{1}{\sqrt{m}}\Bigr). \qedhere
\end{align*}
\end{proof}
Now we are ready to prove the main theorem that shows $|\hat{\mu}_{\infty}-\mu|=2^{-\lambda m^2/s+O(m\log m)}$ with high probability.
We note that for $f$ to be analytic over $[0,1]^s$ means
that it equals its infinite order Taylor expansion
on some open set containing $[0,1]^s$.
\begin{theorem} \label{thm:superconvergence}
Let $f$ be analytic over $[0,1]^s$.
Let $\boldsymbol{x}_i$ be from a $(t,s)$-sequence in
base $2$ with a random linear scramble
plus digital shift.
Then there exist constants $B_1$ and $B_2$ such that for all $m\geq 2$
$$\Pr\Bigl(|\hat{\mu}_{\infty}-\mu|\geq 2^{-\lambda m^2/s+B_1m\log m}\Bigr)\leq \frac{B_2}{\sqrt{m}}.$$
\end{theorem}
\begin{proof}
Because $[0,1]^s$ is compact, we can find $\epsilon>0$ such that for all $\boldsymbol{t}\in [0,1]^s$, the Taylor expansion of $f$ centered at $\boldsymbol{t}$
$$\sum_{n_1=0}^\infty\dots\sum_{n_s=0}^\infty \frac{\partial^{n_1+\dots+n_s}f}{\partial x_1^{n_1}\dots\partial x_s^{n_s}}(\boldsymbol{t})\prod_{j=1}^s \frac{(x_j-t_j)^{n_j}}{n_j!}$$
converges absolutely in an edge-length-$2\epsilon$ box centered at $\boldsymbol{t}$.
It follows that
$$\bigg|\frac{\partial^{n_1+\dots+n_s}f}{\partial x_1^{n_1}\dots\partial x_s^{n_s}}(\boldsymbol{t})\bigg|\prod_{j=1}^s \frac{\epsilon^{n_j}}{n_j!}\to 0$$
as $n_1+\dots+n_s\to \infty$.
There must then be a constant $A$ such that
$$\Big|\frac{\partial^{n_1+\dots+n_s}f}{\partial x_1^{n_1}\dots\partial x_s^{n_s}}(\boldsymbol{t})\Big|\leq \frac{A n!}{\epsilon^n}$$
holds for all $\boldsymbol{t}\in[0,1]^s$
where $n=n_1+\dots+n_s$. Lemma~\ref{lem:Walshcoefficientbound}
then implies that
\begin{align*}|\hat{f}(\boldsymbol{k})|&\leq 2^{-\Vert\boldsymbol{\kappa}\Vert_1-\Vert\boldsymbol{\kappa}\Vert_0}\sup_{\boldsymbol{t}\in [0,1]^s}\Big|\frac{\partial^{\Vert\boldsymbol{\kappa}\Vert_0}f}{\partial x_1^{|\kappa_1|}\cdots\partial x_s^{|\kappa_s|}}(\boldsymbol{t})\Big|\\
&\leq A 2^{-\Vert\boldsymbol{\kappa}\Vert_1} \Bigl(\frac{1}{2\epsilon}\Bigr)^{\Vert\boldsymbol{\kappa}\Vert_0} \Vert\boldsymbol{\kappa}\Vert_0!.
\end{align*}
Because $\epsilon$ can be chosen arbitrarily small, we assume without loss of generality that $2\epsilon<1$.
Let $\mathcal{E}$ be the event that
no $\boldsymbol{k}\in\mathbb{N}^s_*$ with $\Vert\boldsymbol{\kappa}\Vert_1\leq \lambda m^2/s$ has
$\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}$.
Corollary~\ref{cor:concentration} shows that $\Pr(\mathcal{E}) = 1-O(1/\sqrt{m})$ and
we take $B_2$ to be the implied constant
in that expression.
Conditionally on $\mathcal{E}$, equation~\eqref{eqn:errordecomposition} becomes
\begin{align*}
|\hat{\mu}_{\infty}-\mu|&=\Biggl|
\sum_{\boldsymbol{k}\in \mathbb{N}_*^s:\Vert\boldsymbol{\kappa}\Vert_1>\lambda m^2/s}
\boldsymbol{1}\biggl\{\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\biggr\}
\hat{f}(\boldsymbol{k}) (-1)^{\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} \vec{D}_{j}}\Biggr| \nonumber\\
&\leq \sum_{\boldsymbol{k}\in \mathbb{N}_*^s:\Vert\boldsymbol{\kappa}\Vert_1>\lambda m^2/s}|\hat{f}(\boldsymbol{k})| \nonumber\\
&\leq A\times\sum_{\boldsymbol{k}\in \mathbb{N}_*^s:\Vert\boldsymbol{\kappa}\Vert_1>\lambda m^2/s} 2^{-\Vert\boldsymbol{\kappa}\Vert_1}
\Vert\boldsymbol{\kappa}\Vert_0!/(2\epsilon)^{\Vert\boldsymbol{\kappa}\Vert_0} .
\end{align*}
Now for $k=\sum_{\ell=1}^\infty b_\ell 2^{\ell-1}=\sum_{\ell\in\kappa}2^{\ell-1}$
\begin{align*}
\Vert\kappa\Vert_1&=\sum_{\ell\in\kappa} \ell \geq\sum_{\ell=1}^{|\kappa|}\ell
\ge\frac{|\kappa|^2}{2},
\end{align*}
with equality holding for $\kappa=\emptyset$. Then
\begin{align*}
\Vert\boldsymbol{\kappa}\Vert_1&=\sum_{j=1}^s \Vert\kappa_j\Vert_1
\ge
\sum_{j=1}^s\frac{|\kappa_j|^2}{2}\geq
\frac{1}{2s}\biggl(\sum_{j=1}^s|\kappa_j|\biggr)^2=\frac{1}{2s}\Vert\boldsymbol{\kappa}\Vert_0^2,
\end{align*}
yielding $\Vert\boldsymbol{\kappa}\Vert_0\leq \sqrt{2s \Vert\boldsymbol{\kappa}\Vert_1}$. Hence
\begin{align*}
&\sum_{\boldsymbol{k}\in \mathbb{N}_*^s:\Vert\boldsymbol{\kappa}\Vert_1>\lambda m^2/s}A 2^{-\Vert\boldsymbol{\kappa}\Vert_1} \Vert\boldsymbol{\kappa}\Vert_0!/(2\epsilon)^{\Vert\boldsymbol{\kappa}\Vert_0} \\
&\leq \sum_{N=\lceil \lambda m^2/s\rceil }^\infty A 2^{-N}\Bigl(\frac{1}{2\epsilon}\Bigr)^{\sqrt{2s N}}\Gamma(\sqrt{2sN}+1)|\{\boldsymbol{k}\in \mathbb{N}_*^s\mid \Vert\boldsymbol{\kappa}\Vert_1=N\}|
\end{align*}
where $\Gamma(\cdot)$ is the Gamma function
and we have also used $2\epsilon<1$.
By Theorem~\ref{thm:finiteNbound} in the appendix,
$$|\{\boldsymbol{k}\in \mathbb{N}_*^s\mid \Vert\boldsymbol{\kappa}\Vert_1=N\}|\leq \frac{B}{\sqrt{N}}\exp\Bigl(\pi\sqrt{\frac{sN}{3}}\Bigr)$$
holds for $B=\pi\sqrt{s}/(2\sqrt{3})$. Hence
$$\Bigl(\frac{1}{2\epsilon}\Bigr)^{\sqrt{2s N}}\Gamma(\sqrt{2sN}+1)|\{\boldsymbol{k}\in \mathbb{N}_*^s\mid \Vert\boldsymbol{\kappa}\Vert_1=N\}|\leq 2^{D\sqrt{N}\log (N)}$$
for some constant $D$. Because $\sqrt{N+1}\log (N+1)-\sqrt{N}\log (N)$ converges to $0$ as $N\to \infty$, we can find $N_\rho$ for any $\rho>1$ such that $2^{D\sqrt{N}\log (N)}<\rho^{N}$ for $N>N_\rho$. Let us choose $\rho=3/2$ for simplicity. Then when $\lambda m^2/s>N_{3/2}$,
\begin{align*}
\sum_{N=\lceil \lambda m^2/s\rceil }^\infty 2^{-N}2^{D\sqrt{N}\log (N)}&\leq 2^{-\lceil \lambda m^2/s\rceil}2^{D\sqrt{\lceil \lambda m^2/s\rceil}\log (\lceil \lambda m^2/s\rceil)}\sum_{N=0}^\infty \Bigl(\frac{3}{4}\Bigr)^{N}\\
&\leq 2^{-\lambda m^2/s+B_1 m\log(m)}
\end{align*}
for some constant $B_1$. The conclusion follows once we increase $B_1$ sufficiently to cover all $m\geq 2$ cases.
\end{proof}
\begin{corollary}\label{cor:medianMSE}
Under the same condition as Theorem~\ref{thm:superconvergence}, if $E\geq \lambda m^2/s$ and $r=\Omega(m^2)$, then the sample median $\hat{\mu}^{(r)}_{E}$ of $2r-1$ independently generated values of $\hat{\mu}_E$ satisfies
$$\mathbb{E}(|\hat{\mu}^{(r)}_{E}-\mu|^2)\leq 4^{-\lambda m^2/s+O(m\log(m))}.$$
\end{corollary}
\begin{proof} By Lemma~\ref{lem:precisionerror}, with probability at least $1-{B_2}/{\sqrt{m}}$,
$$|\hat{\mu}_E-\mu|\leq 2^{-\lambda m^2/s+B_1m\log m}+\frac{\sqrt{s}}{2^E}
\sup_{\boldsymbol{x}\in [0,1]^s}\Vert\nabla f(\boldsymbol{x})\Vert_2 $$
where $B_1$ and $B_2$ come from Theorem~\ref{thm:superconvergence} and $\nabla f$ is the gradient of $f$.
In order for the sample median of $2r-1$ copies of $\hat{\mu}_E$ to violate the
above bound, there must be at least $r$ copies violating the bound. Because
there are ${2r-1\choose r}$ subsets of size $r$, the union bound implies
that the probability of $r$ such violations is at most
$${2r-1\choose r}\Bigl(\frac{B_2}{\sqrt{m}}\Bigr)^{r}=\prod_{j=2}^r \frac{(2j-1)(2j-2)}{j(j-1)} \Bigl(\frac{B_2}{\sqrt{m}}\Bigr)^{r}< \Bigl(\frac{4B_2}{\sqrt{m}}\Bigr)^{r}.$$
When the above described event happens, $|\hat{\mu}^{(r)}_{E}-\mu|$ is still
bounded by $\sup_{\boldsymbol{x}\in[0,1]^s}|f(\boldsymbol{x})|$. Hence
\begin{align*}
\mathbb{E}(|\hat{\mu}^{(r)}_{E}-\mu|^2)&\leq
\Bigl(2^{-\lambda m^2/s+B_1m\log m}+O\Bigl(\frac{1}{2^E}\Bigr)\Bigr)^2+O\Bigl(\Bigl(\frac{4B_2}{\sqrt{m}}\Bigr)^{r}\Bigr)\\
&=4^{-\lambda m^2/s+O(m\log(m))}
\end{align*}
under our assumptions on $E$ and $r$.
\end{proof}
\section{Finite sample analysis}\label{sec:finitesample}
Although the asymptotic convergence rate of median-of-means is
super-poly\-nomial, the bound in Corollary~\ref{cor:medianMSE} is of
limited use when $\lambda m^2/s$ is only moderately large or even
smaller than $m$. In this section, we derive results that
better describe the finite sample behavior of median-of-means. In
particular, we want to study under what conditions median-of-means can
outperform the usual RQMC estimator (mean-of-means) in terms of mean
squared error. For simplicity, we assume that the precision $E$ is high enough that the difference between $\hat{\mu}_\infty$ and $\hat{\mu}_E$ is negligible in comparison to their root mean squared error.
First let us work out the variance of $\hat{\mu}_\infty$.
\begin{lemma}\label{lem:uncorrelated}
For $\boldsymbol{k}\in \mathbb{N}^s_*$,
let $S(\boldsymbol{k}) = (-1)^{\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} \vec{D}_{j}}$. Then
$$\Pr(S(\boldsymbol{k})=1)=\Pr(S(\boldsymbol{k})=-1)=1/2.$$
For distinct $\boldsymbol{k},\boldsymbol{k}'\in\mathbb{N}^s_*$, $S(\boldsymbol{k})$ and $S(\boldsymbol{k}')$
are independent.
\end{lemma}
\begin{proof}
The proof is similar to Lemma 4 of \cite{superpolyone} and is omitted here.
\end{proof}
\begin{theorem} $\mathbb{E}(\hat{\mu}_\infty)=\mu$ and
\begin{equation*}
\mathrm{Var}(\hat{\mu}_\infty)=\sum_{\boldsymbol{k}\in \mathbb{N}_*^s}
\Pr\Biggl(\,\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\Biggr)\hat{f}(\boldsymbol{k})^2.
\end{equation*}
\end{theorem}
\begin{proof}
Let $\boldsymbol{M}=(M_1,M_2,\dots,M_s)$.
By equation~\eqref{eqn:errordecomposition} and
Lemma~\ref{lem:uncorrelated},
\begin{equation}\label{eqn:conditionallyunbiased}
\mathbb{E}(\hat{\mu}_\infty-\mu\!\mid\! \boldsymbol{M})=\sum_{\boldsymbol{k}\in \mathbb{N}_*^s}\boldsymbol{1}\Bigl\{\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\Bigl\}\hat{f}(\boldsymbol{k}) \mathbb{E}(S(\boldsymbol{k}))=0
\end{equation}
and
\begin{align}\label{eqn:conditionalvariance}
\mathrm{Var}(\hat{\mu}_\infty-\mu\!\mid\! \boldsymbol{M})&=\sum_{\boldsymbol{k}\in \mathbb{N}_*^s}\boldsymbol{1}\Bigl\{\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\Bigl\}\hat{f}(\boldsymbol{k})^2 \mathrm{Var}(S(\boldsymbol{k}))\notag\\
&=\sum_{\boldsymbol{k}\in \mathbb{N}_*^s}\boldsymbol{1}\Bigl\{\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\Bigl\}\hat{f}(\boldsymbol{k})^2.
\end{align}
The conclusion follows by taking expectations with respect to $M$.
\end{proof}
To illustrate when the median-of-means can outperform the mean-of-means,
suppose we can find a rare event $\mathcal{A}$ such that
$\mathbb{E}(\hat{\mu}_\infty\!\mid\!\mathcal{A}^c)=\mu$ and
$\mathrm{Var}(\hat{\mu}_\infty\!\mid\!\mathcal{A}^c)\ll\mathrm{Var}(\hat{\mu}_\infty)$.
Then if we generate independent copies of $\hat{\mu}_\infty$ and look
at the histogram, we should see a cluster with bandwidth
comparable to $\sqrt{\mathrm{Var}(\hat{\mu}_\infty\!\mid\! \mathcal{A}^c)}$
around $\mu$. Those
$\hat{\mu}_\infty$ for which $\mathcal{A}$ happens could
well be far into the tails away from $\mu$.
In such a setting, the sample median is robust with
respect to the event $\mathcal{A}$
and has mean square error close to $\mathrm{Var}(\hat{\mu}_\infty\!\mid\! \mathcal{A}^c)$.
We make this intuition precise with the following lemma.
\begin{lemma}
Let $\mathcal{A}$ be an event with
$\Pr(\mathcal{A})\leq \delta$ and
$\mathbb{E}(\hat{\mu}_\infty\!\mid\! \mathcal{A}^c)=\mu$. Then the sample median
$\hat{\mu}^{(r)}_{\infty}$ of $2r-1$ independently generated values of
$\hat{\mu}_\infty$ using a digital $(t,m,s)$-net satisfies
\begin{align}\label{eq:msebound}
&\mathbb{E}((\hat{\mu}^{(r)}_{\infty}-\mu)^2)\leq {\Pr(\mathcal{A}^c)}
\mathrm{Var}(\hat{\mu}_\infty\!\mid\! \mathcal{A}^c){\delta^{-1}}+(8\delta)^r \Delta^2_n
\end{align}
where
\begin{align*}\Delta_n=\min\biggl(
\omega_f(\sqrt{s}),
\frac{V_{\mathrm{HK}}(f)}{2^{m-t}}\sum_{i=0}^{s-1}{m-t \choose i}\biggr),
\end{align*}
$\omega_f$ gives the modulus of continuity for $f$,
and $V_{\mathrm{HK}}(f)$ is the total variation of $f$ in the sense of Hardy and Krause.
\end{lemma}
\begin{proof}
Conditionally on $\mathcal{A}^c$, we can apply Markov's inequality to get
$$\Pr\Bigl(|\hat{\mu}_\infty-\mu|^2\geq\frac{\Pr(\mathcal{A}^c)}{\delta}\mathrm{Var}(\hat{\mu}_\infty\!\mid\! \mathcal{A}^c)\!\bigm|\! \mathcal{A}^c\Bigr)\leq \frac{\delta}{\Pr(\mathcal{A}^c)}.$$
Hence
$$\Pr\Bigl(|\hat{\mu}_\infty-\mu|^2\geq\frac{\Pr(\mathcal{A}^c)}{\delta}\mathrm{Var}(\hat{\mu}_\infty\!\mid\! \mathcal{A}^c)\Bigr)\leq \frac{\delta}{\Pr(\mathcal{A}^c)} \Pr(\mathcal{A}^c)+\Pr(\mathcal{A})\leq 2\delta.$$
The rest of the proof is similar to that of Corollary~\ref{cor:medianMSE}. In particular,
$$\Pr\Bigl(|\hat{\mu}^{(r)}_{\infty}-\mu|^2\geq \frac{\Pr(\mathcal{A}^c)}{\delta}\mathrm{Var}(\hat{\mu}_\infty\!\mid\! \mathcal{A}^c)\Bigr)\leq {2r-1\choose r}(2\delta)^r\leq (8\delta)^r. $$
When the `bad event' $\mathcal{A}$ happens, we can bound the error in two ways:
first it is clear that both $\hat{\mu}^{(r)}_{\infty}$ and $\mu$ are between
$\inf_{\boldsymbol{x}\in [0,1]^s}f(\boldsymbol{x})$ and $\sup_{\boldsymbol{x}\in [0,1]^s}f(\boldsymbol{x})$, so their
difference is no larger than
$\omega_f(\sqrt{s})=\sup_{\boldsymbol{x}\in [0,1]^s}f(\boldsymbol{x})-\inf_{\boldsymbol{x}\in [0,1]^s}f(\boldsymbol{x})$. Second, if $f$ has
finite Hardy–Krause variation, we can apply the Koksma–Hlawka inequality \cite{hick:2014} to
conclude that $|\hat{\mu}_\infty-\mu|\leq V_{\mathrm{HK}}(f)D^*_n(\boldsymbol{x}_0,\dots,\boldsymbol{x}_{n-1})$ where
$D^*_n(\cdot)$ denotes the star discrepancy.
Because this is true for all $\hat{\mu}_\infty$, it is also true for
$\hat{\mu}^{(r)}_{\infty}$. Because $\boldsymbol{x}_0,\dots,\boldsymbol{x}_{n-1}$ is a $(t,m,s)$-net
regardless of the scrambling, we can apply the bound
$$D^*_n(\boldsymbol{x}_0,\dots,\boldsymbol{x}_{n-1})\leq \frac{1}{2^{m-t}}\sum_{i=0}^{s-1}{m-t \choose i}$$
from Corollary 5.3 of \cite{dick:pill:2010}. By combining the two bounds,
we derive $|\hat{\mu}^{(r)}_{\infty}-\mu|\leq \Delta_n$ and hence the bound
on $\mathbb{E}(\hat{\mu}^{(r)}_{\infty}-\mu)^2$.
\end{proof}
The $(8\delta)^r \Delta^2_n$ term in the bound~\eqref{eq:msebound} is exponentially small in $r$ if $\delta<1/8$.
As shown in Section 3 of \cite{superpolyone}, $\mathrm{Var}(\hat{\mu}_\infty)$ is in general $\Omega(n^{-3})$ for smooth functions $f$. Hence we only need $r\geq C_*m$ for some $C_*>0$ to make $(8\delta)^r D^2_n\ll \mathrm{Var}(\hat{\mu}_\infty)$. With the same computational effort, the mean-of-means has variance equal to $\mathrm{Var}(\hat{\mu}_\infty)/(2r-1)$. So heuristically, median-of-means can significantly outperform mean-of-means in terms of MSE if
there exists an event $\mathcal{A}$ such that $\Pr(\mathcal{A})\leq \delta<1/8$ and $\mathrm{Var}(\hat{\mu}_\infty\!\mid\! \mathcal{A}^c)\ll \mathrm{Var}(\hat{\mu}_\infty)/m$.
Motivated by Corollary~\ref{cor:concentration}, one way to choose $\mathcal{A}$ is to specify a set of frequencies $K\subseteq \mathbb{N}^s_*$ and let $\mathcal{A}=\{\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0} \text{ for some } \boldsymbol{k}\in K\}$. We know that $\mathbb{E}(\hat{\mu}_\infty\!\mid\! \mathcal{A}^c)=\mu$ because equation~\eqref{eqn:conditionallyunbiased} shows that $\hat{\mu}_\infty$ is unbiased conditionally on $\boldsymbol{M}=(M_1,\dots,M_s)$ and $\mathcal{A}$ belongs to the $\sigma$-algebra generated by $\boldsymbol{M}$. Moreover, as long as $\sum_{\boldsymbol{k}\in K}\Pr\bigl(\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\bigl)\leq \delta$, we know by the union bound that $\Pr(\mathcal{A})\leq \delta$ as well. According to equation~\eqref{eqn:conditionalvariance},
\begin{align*}
\Pr(\mathcal{A}^c)\mathrm{Var}(\hat{\mu}_\infty\!\mid\! \mathcal{A}^c)&=\Pr(\mathcal{A}^c)\mathbb{E}(\mathrm{Var}(\hat{\mu}_\infty-\mu\!\mid\! \boldsymbol{M})\!\mid\! \mathcal{A}^c)\\
&=\sum_{\boldsymbol{k}\in \mathbb{N}_*^s}\Pr(\mathcal{A}^c)\Pr\biggl(\,\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\!\mid\! \mathcal{A}^c\biggr)\hat{f}(\boldsymbol{k})^2\\
&\leq \sum_{\boldsymbol{k}\in \mathbb{N}_*^s\setminus K}\Pr\biggl(\,\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\biggr)\hat{f}(\boldsymbol{k})^2.
\end{align*}
So in principle, if one knows all $\hat{f}(\boldsymbol{k})^2$ and $\Pr\bigl(\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\bigl)$, then one can find a good candidate $\mathcal{A}$ by solving the following combinatorial optimization problem:
\begin{equation*}
\begin{aligned}
\max_{K\subseteq \mathbb{N}^s_*} \quad & \sum_{\boldsymbol{k}\in K}\Pr\biggl(\,\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\biggr)\hat{f}(\boldsymbol{k})^2\\
\textrm{s.t.} \quad & \sum_{\boldsymbol{k}\in K}\Pr\biggl(\,\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\biggr)\leq \delta.\\
\end{aligned}
\end{equation*}
In particular, if $\mathrm{Var}(\hat{\mu}_\infty)$ is dominated by a few $\boldsymbol{k}$ with large $\hat{f}(\boldsymbol{k})^2$, then we should see a significant variance reduction after we condition on the $\mathcal{A}^c$ specified by the above optimization problem.
To make the problem more tractable, we examine one case where our function $f$ is effectively low-dimensional. Suppose there are a few components $x_j$ that contribute most of the variability to $f$. More precisely, let
$$f_1(\boldsymbol{x}) = \mathbb{E}( f(\boldsymbol{x})\!\mid\! x_j, j\in u)-\mu \quad\text{and}\quad f_2(\boldsymbol{x}) = f(\boldsymbol{x})-f_1(\boldsymbol{x})-\mu.$$
Then $\mu$, $f_{1}$, and $f_{2}$ are orthogonal in the
$L^2[0,1)^s$ inner product, so that $\sigma^2(f)=\sigma^2(f_1)+\sigma^2(f_{2})$. We assume that $\sigma^2(f_1)\gg \sigma^2(f_{2})$, and then $f_1$ captures most of the variance of $f$.
Given such a function, it is natural to choose $K=\{\boldsymbol{k}\in\mathbb{N}_*^s\mid\boldsymbol{s}(\boldsymbol{k})\subseteq u\}$ because their associated $f(\boldsymbol{k})^2$ are relatively large. Then,
Corollary~\ref{cor:tmsbound} can be strengthened in the following way:
\begin{lemma}\label{lem:finitesampletmsbound}
For non-empty $u\subseteq 1{:}s$, define $\mathbb{N}^{u}\subset\mathbb{N}^s$ to be the set of
$\boldsymbol{k}\in\mathbb{N}_*^s$ with $\boldsymbol{s}(\boldsymbol{k})=u$.
Further define
\begin{align*}
t^*_u &=
m+1-\min_{\boldsymbol{k}\in\mathbb{N}^{u}}\bigl\{ \Vert\lceil\boldsymbol{\kappa}\rceil\Vert_1
\mid
C^{\lceil\boldsymbol{\kappa}\rceil}\ \mathrm{not\ of\ full\ rank\ }
\bigr\}.
\end{align*}
If $u=\emptyset$, we conventionally define $t^*_u=0$. Then
$$\Pr\bigg(\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\bigg)\leq 2^{-m+t^*_{\boldsymbol{s}(\boldsymbol{k})}+|\boldsymbol{s}(\boldsymbol{k})|}.$$
\end{lemma}
\begin{proof}
The proof is basically the same as Corollary~\ref{cor:tmsbound}. By the definition of $t^*_u$, a rank-deficient $C^{\lceil\boldsymbol{\kappa}\rceil}$ must contains $m-t^*_{\boldsymbol{s}(\boldsymbol{k})}$ linearly independent rows, so
$$\mathrm{rank}(C^{\lceil\boldsymbol{\kappa}\rceil-\boldsymbol{1}\{\boldsymbol{s}(\boldsymbol{k})\}})\geq \mathrm{rank}(C^{\lceil\boldsymbol{\kappa}\rceil})-|\boldsymbol{s}(\boldsymbol{k})|\geq m-t^*_{\boldsymbol{s}(\boldsymbol{k})}-|\boldsymbol{s}(\boldsymbol{k})|$$
which proves the conclusion.
\end{proof}
\begin{remark} To compare $t^*_u$ with $t$, consider for instance a Sobol' sequence constructed by the $s$ lowest order irreducible polynomials. As shown in Section 4.5 of \cite{sobol67}
the order of irreducible polynomials grows roughly like $\log(s)$ and $t$ is consequently $O(s\log(s))$. The supremum of $t^*_u$ on the other hand, grows no faster than $|u|\log(s)$, which is potentially much smaller than $t$.
\end{remark}
Now we can prove the finite sample version of Corollary~\ref{cor:concentration}.
\begin{theorem}\label{thm:finitesampleconcentration}
For non-empty $u\subseteq1{:}d$, let $t_u=\max_{v\subseteq u}t^*_v$. For $\delta>0$, let $N^*_m$ be the largest integer $N$ satisfying
\begin{align*}
&2^{t^*_u+|u|}{m + |u| \choose |u|-1}N\exp\Bigl(\pi\sqrt{\frac{|u|N}{3}}\Bigr)\leq \delta 2^m,
\intertext{and}
&\sqrt{3(|u|-1)N}\leq \frac{\pi}{2}(m-t_u).
\end{align*}
Then
$$\Pr\bigg(\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0} \text{ for some } \boldsymbol{k} \ne\boldsymbol{0} \text{ with } \boldsymbol{s}(\boldsymbol{k})\subseteq u, \ \Vert\boldsymbol{\kappa}\Vert_1\leq N^*_m+m-t_u\bigg)\leq \delta.$$
\end{theorem}
\begin{proof}
It is shown in Section 5 of \cite{pan:owen:2021:tr} that $t_v^*+|v|\leq t_u^*+|u|$ if $v\subseteq u$. So for $\boldsymbol{k}\ne\boldsymbol{0}$ with $\boldsymbol{s}(\boldsymbol{k})\subseteq u$,
\begin{equation}\label{eqn:usubsetbound}
\Pr\biggl(\,\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\biggr)\leq 2^{-m+t^*_{\boldsymbol{s}(\boldsymbol{k})}+|\boldsymbol{s}(\boldsymbol{k})|}\leq 2^{-m+t^*_u+|u|}.
\end{equation}
Now by the definition of $t^*_u$, $C^{\lceil\boldsymbol{\kappa}\rceil}$ has full rank if $\Vert\lceil\boldsymbol{\kappa}\rceil\Vert_1\leq m-t^*_{\boldsymbol{s}(\boldsymbol{k})}$, so $\Pr\big(\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\big)=0$.
Let
$$
\mathbb{N}_0^u = \{ \boldsymbol{k}\in\mathbb{N}_*^s \mid \boldsymbol{s}(\boldsymbol{k})\subseteq u\}
$$
By choosing $s=|u|$ and $R=m-t_u$ in Corollary~\ref{cor:finiteNbound2} from the appendix, we further get
\begin{multline*}
|\{\boldsymbol{k}\in\mathbb{N}_0^u\mid \Vert\boldsymbol{\kappa}\Vert_1\leq N^*_m+m-t_u, \Vert\lceil\boldsymbol{\kappa}\rceil\Vert_1> m-t_u\}|\\
<{m+|u|\choose |u|-1}N^*_m\exp\Bigl(\pi\sqrt{\frac{|u|N^*_m}{3}}\Bigr).
\end{multline*}
After taking a union bound over
all $\boldsymbol{k}$ in the above set, we finally get
\begin{align*}
&\Pr\Big(\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0} \text{ for some } \boldsymbol{k}\ne\boldsymbol{0} \text{ with } \boldsymbol{s}(\boldsymbol{k})\subseteq u, \ \Vert\boldsymbol{\kappa}\Vert_1\leq N^*_m+m-t_u\Big)\\
&\leq {m+|u|\choose |u|-1}N^*_m\exp\Bigl(\pi\sqrt{\frac{|u|N^*_m}{3}}\Bigr)2^{-m+t^*_u+|u|}\leq \delta.\qedhere
\end{align*}
\end{proof}
To interpret this result, let us consider the setting of Theorem~\ref{thm:superconvergence}. For simplicity, we will replace $f$ by $f_1(\boldsymbol{x})=\mathbb{E}(f(\boldsymbol{x})\!\mid\! x_j,j\in u)-\mu$ and pretend that the problem is $|u|$-dimensional, which is a useful approximation under our assumption on $f$. In view of Lemma~\ref{lem:Walshcoefficientbound} and equation~\eqref{eqn:conditionalvariance}, one can argue that $\mathrm{Var}(\hat{\mu}_\infty\!\mid\! \boldsymbol{M})$ is proportional to $4^{-\Vert\boldsymbol{\kappa}\Vert_1}$ for the $\boldsymbol{k}$ with the smallest $\Vert\boldsymbol{\kappa}\Vert_1$ among those satisfying $\sum_{j\in u} \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}$. This is certainly true in the asymptotic sense, as we have shown in the proof of Theorem~\ref{thm:superconvergence} that $\Vert\boldsymbol{\kappa}\Vert_0\leq \sqrt{2|u|\Vert\boldsymbol{\kappa}\Vert_1}$ and the supremum norm of partial derivatives grows no faster than $\Vert\boldsymbol{\kappa}\Vert_0!$. (More precisely, Lemma~\ref{lem:Walshcoefficientbound} only provides an upper bound on $\hat{f}(\boldsymbol{k})^2$, but section 3 of \cite{superpolyone}
shows the factor $4^{-\Vert\boldsymbol{\kappa}\Vert_1}$ is in general necessary.)
By the definition of $t_u$, there exists a set of $\boldsymbol{k}\in\mathbb{N}_*^s$
such that $C^{\lceil\boldsymbol{\kappa}\rceil}$ is rank-deficient and $\Vert\lceil\boldsymbol{\kappa}\rceil\Vert_1=m-t_u+1$. It is also true that $C^{\lceil\boldsymbol{\kappa}\rceil-\boldsymbol{1}\{\boldsymbol{s}(\boldsymbol{k})\}}$ has full rank, because otherwise $t_u$ would be even larger. Hence $\mathrm{rank} (C^{\lceil\boldsymbol{\kappa}\rceil-\boldsymbol{1}\{\boldsymbol{s}(\boldsymbol{k})\}})=\Vert\lceil\boldsymbol{\kappa}\rceil\Vert_1-|\boldsymbol{s}(\boldsymbol{k})|\leq m-t_u $ and $\Pr\bigl(\sum_{j\in u} \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}\bigr)\geq 2^{-m+t_u}$
from the second case in Lemma~\ref{lem:gaincoef}.
On the other hand, if we condition on the
event $\mathcal{A}$ specified by Theorem~\ref{thm:superconvergence}, the
smallest $\Vert\boldsymbol{\kappa}\Vert_1$ for which $\sum_{j\in u} \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0}$ is
possible is $N^*_m+m-t_u+1$ and the corresponding probability is no more
than $2^{-m+t^*_u+|u|}$ according to equation~\eqref{eqn:usubsetbound}. So
roughly speaking, $\mathrm{Var}(\hat{\mu}_\infty\!\mid\! \mathcal{A}^c)$ is a factor of $4^{-N^*_m}$
smaller than $\mathrm{Var}(\hat{\mu}_\infty)$. In view of our previous criterion
$\mathrm{Var}(\hat{\mu}_\infty\!\mid\! \mathcal{A}^c)$ needs to be much smaller than
$\mathrm{Var}(\hat{\mu}_\infty)/m$, we see that with a proper choice on the number
of replicates, median-of-means can significantly outperform mean-of-means
when $N^*_m\gg \log m$.
\begin{remark}
One can easily generalize the above discussion to cases where $f$ can be approximated by multiple low-dimensional functions. For instance, suppose $f$ has effective dimension $d$ in the superposition sense \cite{cafmowen}, namely $f\approx \sum_{u\subseteq 1{:}s, |u|\leq d}f_u$ where $f_u$ is the ANOVA term corresponding to subset $u$. We can define $t_d=\max_{u:|u|= d}t_u$ and $T_d=\max_{u:|u|= d}t^*_u+|u|$. By applying the above theorem to each of the ${s\choose d}$ size-$d$ subsets of $1{:}s$, we get
$$\Pr\biggl(\,\sum_{j=1}^s \vec{k{}}_j^\mathsf{T} M_j C_j=\boldsymbol{0} \text{ for some } \boldsymbol{k}\ne\boldsymbol{0} \text{ with } |\boldsymbol{s}(\boldsymbol{k})|\leq d, \ \Vert\boldsymbol{\kappa}\Vert_1\leq N^*_m+m-t_d\biggr)\leq \delta$$
where $N^*_m$ is the largest integer $N$ satisfying
\begin{align*}
&2^{T_d}{s\choose d}{m +d \choose d-1}N\exp\Bigl(\pi\sqrt{\frac{dN}{3}}\Bigr)\leq \delta 2^m
\quad\text{and}\quad
\sqrt{3(d-1)N}\leq \frac{\pi}{2}(m-t_d).
\end{align*}
Again when $N^*_m\gg \log m$, we expect to see median-of-means outperform mean-of-means.
\end{remark}
\section{Discussion}\label{sec:discussion}
We have shown that a median-of-means strategy
based on scrambled $(t,m,s)$-nets in base $2$
can attain superpolynomial accuracy for integration
of analytic functions on $[0,1]^s$. The main
nets we have in mind are those that arise as
the first $2^m$ points of a Sobol' sequence.
The superpolynomial rate comes with a dimension
effect that has lesser impact when the integrand
is dominated by low dimensional ANOVA components.
We have not shown that the method adapts to
lesser levels of smoothness of the integrand.
That is known to hold for $s=1$ from \cite{superpolyone}. It therefore also holds
for additive functions on $[0,1]^s$ with a rate
given by the worst smoothness of any of the
summands. We do not know the extent of adaptation
for more general functions.
It remains to quantify the uncertainty in
the median-of-means estimate using
the sample data. For the mean-of-means we can
get an unbiased estimate of the variance
of the combined estimate.
There is a central limit theorem (CLT) by Loh \cite{loh:2003} for scrambled nets as $n\to\infty$
but it only applies to nested uniform
scrambling from \cite{rtms} and is only proved
for $t=0$.
There is recent work by Nakayama and Tuffin \cite{naka:tuff:2021}
that describes CLTs for the mean-of-means
over scrambled nets as the number of replicates increases.
For the median-of-means, things are more
complicated. We can use nonparametric
statistical methods to get a confidence
interval for the median of $\hat\mu_{\infty,r}$
over all scrambles, but that is not the
same quantity as $\mu=\mathbb{E}(\hat\mu_\infty)$
and it generally depends on $m$.
There are confidence intervals for the
median-of-means (see e.g., \cite{devr:etal:2016}) but in our setting those would have width proportional to
${\mathrm{Var}(\hat\mu_{\infty,r})^{1/2}}$.
That standard deviation does not decrease at a super-polynomial rate and so the confidence intervals would not
reflect the increased precision that comes from
using the median-of-means.
The median-of-means works so well for random
linear scrambling because that estimate is usually
very accurate apart from outliers that raise its
variance. The presence of outliers implies that
the convergence to the Gaussian distribution
will be slow for the mean-of-means with the
random linear scrambles we study here.
This upper bound on the error has the same rate that we would
get in applying a one dimensional rule with error
$O(n^{-c\log_2(n)})$ in an $s$-fold product.
However, an $s$-fold product rule allows no nontrivial
sample sizes below $2^s$ which may be far too large to use
and still ineffective.
It is also not clear whether there would need to be $(2r-1)^s$-fold
computation in a product rule whose factors involve
medians of means.
Digital nets exist for sample sizes $2^m$ for $m\ge0$ so
we can get this rate along a practically usable sequence
of sample sizes and benefit from a good convergence
rate on the low dimensional ANOVA or other components.
The situation is similar to that in \cite{he2016extensible}
where the optimal rate under Lipschitz continuity is attained
by a grid but also by sampling along a Hilbert space-filling
curve.
\bibliographystyle{plain}
|
\section{Introduction}\label{sec:introduction}
Deep Neural Network (DNN) has paved the way for the application of artificial intelligence algorithms in many fields, including image recognition and natural language processing. The success of deep learning is due to the advancement in DNN algorithms~\cite{deeplearning-neuralnet}, programming interfaces and platforms~\cite{pytorch, tensorflow, theano}, and optimized hardware for deep learning, such as the general-purpose GPU and Tensor Processing Unit (TPU)~\cite{tpu}.
Among many DNN algorithms, Convolutional Neural Network (CNN)~\cite{lenet, vgg, alexnet} is widely used for various applications, such as object detection for a self-driving cars and image recognition. The complex internal structure of CNN implementations requires massive computing power and parallelism, particularly during the training step. Depending on the model complexity, training a CNN model can take a few weeks or months~\cite{cnn-detail, alibaba-dl-train-workload}, and building a cost-efficient and fast processing environment is crucial to increase the productivity of CNN algorithm developers. In addition, GPU devices can provide massive parallelism, but building a GPU cluster can be prohibitively expensive. Users can easily build a CNN model without purchasing GPU devices by using cloud resources. Training jobs may occur on occasion, and using GPU instances only when needed can result in cost savings.
The pace of cloud computing innovation is very fast, with new types of cloud services and instances being released on a regular basis by public vendors. For example, at the recent re:Invent, one of the largest annual cloud computing events, Amazon Web Services (AWS) introduced various instance types from new hardware vendors with enhanced network and storage capacity~\cite{reinvnet-2021}. It is difficult for deep learning algorithm developers to keep up with updates from public cloud vendors and understand the impact of the updates in order to apply them to their deep learning training pipeline. As a result, it is ideal if cloud service vendors provide an optimized deep learning environment natively so that algorithm developers can focus on core application development.
\begin{figure*}[t]
\centering
\includegraphics[width=0.8\textwidth]{figures/general-cnn-vgg16.pdf}
\caption{A general architecture of a CNN model with an example of VGG16}
\label{fig:cnn_architecture}
\end{figure*}
The Platform-as-a-Service (PaaS) cloud service model promoted as reducing development platform management overhead, allowing programmers to focus on critical software development tasks. Current deep learning pipeline platform services provided by public cloud vendors, such as AWS SageMaker and Google Cloud Datalab, are far from optimal because users must manually select cloud instance types and the number of instances to train a DNN model. Furthermore, an optimal DNN environment can vary across model architectures, dataset sizes, and model configurations, making deep learning platform management more difficult.
To understand the performance characteristics of the various CNN algorithm training latency on different public cloud GPU instances, we measured the training time on AWS EC2 instances that are equipped with GPU devices (Table~\ref{tab:aws_gpu_spec}), and the result is presented in Figure~\ref{fig:motivation}. Different CNN model architectures could result in a five-fold performance difference between the best and worst-performing GPU instance types. The training latencies with varying batch and input image pixel sizes do not change linearly with respect to the configured size, making prediction difficult. Despite such stark performance differences when training a CNN model on GPU cloud instances, algorithm developers rarely understand such performance characteristics and miss opportunities to optimize training environments. In addition, imposing training environment operation burdens on DNN algorithm developers might be too much overhead for them, as they are not generally system experts. Thus, the cloud service vendor should prepare an optimal training environment that reflects algorithm developers’ implementation characteristics.
Habitat~\cite{habitat}, Paleo~\cite{paleo}, NeuralPower~\cite{neuralpower}, and MLPredict~\cite{mlpredict-bigdata} proposed algorithms to predict the training time for various CNN algorithms on different GPU devices. The systems use the internal architecture of a CNN implementation and GPU device characteristics as input features. We call such approaches \emph{white-box} methods. Though they envisioned the feasibility of predicting the training time of CNN implementations on arbitrary GPU devices, the \emph{white-box} approach can be challenging to apply in an environment where a deep learning algorithm developer and a DNN development platform provider, such as a public cloud vendor, are not identical. In public cloud services, to provide an optimal DNN training environment using a \emph{white-box} approach, a service provider should know details of the model architecture. However, it is unlikely that algorithm developers would be willing to share the source code, which is generally confidential and is a private asset of an organization. Furthermore, they have the critical drawback of inflexibility for a new CNN architecture. Except Habitat~\cite{habitat}, which requires a profiling before prediction, the proposed algorithms cannot predict the training time of a CNN implementation with a new layer that is not involved in the prediction model training step.
To overcome the shortcomings of the previous works, we propose PROFET, which aims to predict the training time for arbitrary CNN implementations without revealing the internal model architecture. Hiding the model architecture in a training latency prediction model makes the proposed system appropriate for a public cloud system where algorithm developers and development platform maintainers are different. To meet the goal, PROFET uses abstracted profiling information from CNN training as prediction model input and proposes novel heuristics of median-ensemble modeling and feature clustering to enhance the prediction accuracy of a CNN implementation with unique operation names. In addition, PROFET can predict the CNN model training time on diverse GPU devices, batch sizes, and image pixel sizes to support diverse CNN training scenarios on the cloud. A thorough evaluation of PROFET reveals that it outperforms the state-of-the-art algorithms, Habitat~\cite{habitat}, Paleo~\cite{paleo}, and MLPredict~\cite{mlpredict-bigdata}, improving prediction accuracy by $35\%$, $38\%$, and $84\%$, respectively.
In summary, the major contributions of this paper are as follows.
\begin{itemize}
\item{Envisioning the importance of \emph{black-box} performance estimation of the CNN on a public cloud}
\item{A novel feature clustering algorithm based on the CNN operation name similarity}
\item{Unique modeling heuristics to predict training latency across distinct GPU devices}
\item{Demonstrating a training latency prediction system for distinct GPU, mini-batch, and input image pixel sizes}
\end{itemize}
\section{Training CNN on Cloud}
Training a model can take weeks or months, depending on the internal architecture of the CNN implementation~\cite{dnn-workload-sensetime, dnn-workload-microsoft, alibaba-dl-train-workload}. Various configurations influence CNN model training time, and it is critical to understand the performance diversity of CNN training to build an optimal training environment.
\subsection{Overview of CNN}\label{sec:cnn-overview}
CNN algorithms are widely used to analyze visual representations from given input dataset which generally includes images or videos. The most common form of a CNN model is constructed with input and output layers with a series of hidden layers between them. Each hidden layer contains a diverse set of operations, such as convolution, activation and pooling. The operations extract information from a previous layer and abstract the information by applying a designated operation to pass it to the next layer. In the early days of CNN implementation, small models were proposed, such as LeNet5~\cite{lenet}, which contains about $60,000$ parameters. Recently, more diverse and complex models from the perspective of depths, layers, activation function configurations are proposed which improve the model accuracy drastically. They include AlexNet~\cite{alexnet}, which contains about $600,000,000$ parameters, VGG~\cite{vgg}, ResNet~\cite{resnet}, and Inception~\cite{inception}. As the architecture of CNN gets more complex, an attempt to find optimal CNN architecture is proposed and getting relevant in the field~\cite{neural-architecture-search}.
Figure~\ref{fig:cnn_architecture} shows an architecture of VGG16 model~\cite{vgg} which classifies an input image of $224 \times 224$ pixels with three channels. In the model, multiple layers with different operations appear repeatedly. The operation contains a convolution with a window size of $3 \times 3$ followed by ReLU activation function~\cite{relu} and max-pooling with a size of $2 \times 2$. Multiple input images are trained at once with a unit of configured mini-batch size, and model weights are updated per mini-batch through back-propagation. The mini-batch size can impact the model accuracy and training latency.
\subsection{GPU Instances on Cloud}\label{sec:aws-gpu-instances}
As a CNN model training task requires extensive computing capacity, various accelerator hardwares, such as GPU, Field Programmable Gate Array (FPGA), and Application-Specific Integrated Circuit (ASIC), are widely used to expedite the model training. Among them, GPU is the most popular accelerator device because of its superb performance and well-established Software Development Kit (SDK). Due to its high cost of hardware and cluster management overhead, using GPU devices through public cloud services becomes norm, and most public cloud vendors provide the service. Similar to other cloud instance types, there exist various instance types which equip GPU devices. Table~\ref{tab:aws_gpu_spec} shows GPU-based instance types provided by AWS. In each instance family, we show the smallest size one. As shown in the table, different instance types have distinct GPU models which are manufactured by NVIDIA. The CNN training performance is expected to be mainly dependent on the processing capacity of GPU cores which can be expressed with the number of GPU cores, clock speed, and the number of Floating Point Operations Per Second (FLOPS). In our experiments and modeling, we could observe that not only the processing capacity of GPU but also the CPU capacity, GPU memory, and the bandwidth between CPU and GPU are important factors to determine the training performance. Other than the distinct processing capacity of each instance type, the price also differs significantly, and it becomes very challenging for CNN algorithm developers to decide optimal GPU cloud instance types for their model development.
\begin{table}
\def1.3{1.05}
\centering
\begin{tabular}{ |c||c|c|c|c| }
\hline
\multicolumn{5}{|c|}{AWS GPU Instance Specification} \\
\hline
Instance Family & G3s & G4(dn) & P2 & P3\\
\hline
\hline
GPU Model & M60 & T4 & K80 & V100\\
GPU Core & 2048 & 2560 & 2496 & 5120 \\
GPU Clock(MHz) & 1178 & 1590 & 875 & 1380\\
TFLOPS(FP32) & 4.825 & 8.141 & 4.113 & 14.13\\
\hline
Released Year & 2017 & 2019 & 2016 & 2017 \\
Price(\$/hr) & 0.75 & 0.526 & 0.9 & 3.06\\
\hline
\end{tabular}
\caption{Specification of different GPU Instances on AWS}
\label{tab:aws_gpu_spec}
\end{table}
\subsection{Understanding CNN Training Performance}
The diversity of CNN model architecture, configurations, and cloud instances with GPU devices can result in significant differences in model training time. To represent latency and cost difference under various settings, we conducted thorough experiments whose highlights are shown in Figure~\ref{fig:motivation}.
\begin{figure*}[t]
\centering
\subfloat[Different model]{
\includegraphics[width=0.33\textwidth]{figures/motiv-model-difference.pdf}
\label{fig:motivation_model}
}
\subfloat[Different input pixel size]{
\includegraphics[width=0.33\textwidth]{figures/motiv-dataset-difference.pdf}
\label{fig:motivation_dataset}
}
\subfloat[Different mini-batch size]{
\includegraphics[width=0.33\textwidth]{figures/motiv-batchsize-difference.pdf}
\label{fig:motivation_batchsize}
}
\caption{Performance variation of CNN model training using GPU cloud instances}
\label{fig:motivation}
\end{figure*}
Figure~\ref{fig:motivation_model} shows training time difference of distinct models on various cloud instances. We measured mini-batch training latency of LeNet5~\cite{lenet} and AlexNet~\cite{alexnet} setting the input dataset size as $32\times 32$ pixels with three channel colors and batch size as 16. The latencies are shown in the primary vertical axis through bars. In the secondary vertical axis, we present the relative cost to complete given training workload whose values are represented using star marks. For each workload, training latencies of different instance types are presented after normalization to the least value in each workload. A horizontal line at $1.0$ indicates the best performing case. For each workload, the cloud instances are shown in the order of \emph{g3s}, \emph{g4dn}, \emph{p2}, and \emph{p3} marked with right upper diagonal, left upper diagonal, horizontal, and vertical line patterns, respectively. For LeNet5, the \emph{g4dn} is the fastest, while the \emph{p3} is the fastest for AlexNet training. Comparing the best and worst performing instance types, the latency of LeNet5 is less than two times (\emph{g4dn} and \emph{p2}), but that of AlexNet is close to ten times (\emph{p3} and \emph{p2}). Regarding the cost, we can observe that \emph{g4dn} is the most cost efficient for both workloads, but choosing \emph{p3} workload can be more beneficial for AlexNet because its training latency is about one-third with only twice more cost.
In a CNN model training, an algorithm developer has to decide the input image pixel size because raw images are generally too large to train a model efficiently. Thus, it needs to be down-scaled for both train and inference steps. For example, photos taken from an ordinary smart phone or digital camera contains 10,000,000s (10 Megas) of pixels, while most widely used CNN models require input images are in the range of 1,000s - 100,000s of pixels. In such a scenario, a CNN algorithm developer should understand trade-offs from the input image sizes~\cite{cnn-datasize} and needs to decide an optimal model input pixel size considering application characteristics. To present the impact of different image sizes, Figure~\ref{fig:motivation_dataset} shows the latency and the corresponding cost of training a ResNet50 model when input image sizes are $32 \times 32$ pixels and $128 \times 128$ pixels with a fixed batch size of 16. Regardless of the input pixel size, \emph{p3} instance has the shortest latency, but higher price of \emph{p3} instance results in poorer cost efficiency than the \emph{g4dn} type. The latency difference of \emph{p3} and \emph{g4dn} is marginal (less than 10\%) when the input image is small, but it becomes noticeable (more than 100\%) with a larger image size. We can also observe that the newer instance types (\emph{p3} and \emph{g4dn}) shows better latency and cost-effectiveness than the older instance types (\emph{g3s} and \emph{p2}) regarding the input image sizes. It implies that timely upgrading CNN training environment is necessary for optimal operations.
In CNN training, a mini-batch size is an important hyper-parameter which impacts both model accuracy and training time. A batch size represents the number of images to be trained at once, and a batch latency gets longer as the batch size increases. Intuitively, the relationships between batch size change and latency should be proportional. For example, increasing batch size from $32$ to $64$ will result in twice the batch latency in a perfectly linear relation. To quantitatively present the impact from the batch size change, we conducted experiments varying batch sizes to 16, 32, 64, 128, and 256 after fixing the model, input image pixel size, and cloud instances. Figure~\ref{fig:motivation_batchsize} presents the ratio of batch train latency with respect to the different batch sizes. For an arbitrary set of model, input image pixel, and instance type, we set the latency of batch size of $16$ as $1.0$ and calculated latency of other batch sizes as the ratio to that of $16$. For each instance type, we marked the ratio value of minimum, 25\%, median, 75\%, and the maximum value. We could discover that the relationship between batch size and latency is not linear and the pattern is barely identified. For example, the latency of MobileNetV2, $32 \times 32$, \emph{p3}, increasing batch size from $16$ to $256$, which is $16$ times larger batch size, results in only $1.45$ times increase in the training latency. In another case of VGG13, $128 \times 128$, \emph{g4dn}, a batch latency of 256 is $13.5$ times more than that of batch size 16. Furthermore, \emph{p3} instances show a distinguished pattern of lower latency increase than other instance types as the batch size increases. This result implies that we cannot simply use a linear regressor to predict performance of CNN training time with respect to distinct batch sizes and need elaborate model for accurate prediction.
In summary, Figure~\ref{fig:motivation} presents drastic CNN training latency variations on diverse cloud instances with GPU devices under different model architectures, input image pixels, and batch sizes. Considering the continuous release of new cloud services, resources, and pricing mechanisms~\cite{deepspotcloud, spot-price-prediction, ri-marketplace}, it is challenging for CNN algorithm developers to follow new updates and timely apply them to their training environment. Furthermore, it is very cumbersome to try every instance type that might need custom device library setup to check how a CNN model performs. Therefore, a guidance of estimated performance of custom CNN algorithm implementations on diverse cloud instance types is mandatory to help developers to build an optimal training environment and focus on core application development.
\begin{figure*}[t]
\centering
\includegraphics[width=0.95\textwidth]{figures/overall-architecture.pdf}
\caption{High level description of the proposed system and working scenario.}
\label{fig:overall-architecture}
\end{figure*}
\section{PROFET : Modeling Training Time}\label{section:profet-overall}
The goal of PROFET is to predict the training time for arbitrary CNN implementations on various GPU instances provided by a public cloud service vendor with minimal exposure to implementation details. Minimally exposing the implementation details is important, especially in a public cloud environment. Setting up a CNN development environment using GPUs requires considerable system operation effort, which can be quite challenging for an algorithm developer. Thus, the responsibility of maintaining and operating an optimal development environment should be imposed on the cloud service vendor, which agrees with the recent cloud computing evolution direction~\cite{cloud-evolution} represented by the serverless computing~\cite{serverless-backward-forward, function-bench}. To allow public cloud vendors to provide an optimal training environment for arbitrary CNN implementations, CNN model characteristics should be provided while hiding the detailed internal architecture, and PROFET achieves this goal using abstracted operation information.
To build a prediction model, a set of feature vectors which we denote as $\vec{X}$ is generated from offline experiments. $\vec{X}$ is composed of $N$ workloads. Each workload return a feature vector with a dimension of $D$. Thus, the dimension of $\vec{X}$ is $N \times D$. We denote each workload scenario as $\vec{x}_i, i=1:N$. To note $j$-th feature of workload $i$, where $i=1:N, j=1:D$, we use $x_{ij}$.
To generate diverse CNN workloads, we variate CNN training scenarios with respect to the instance types ($G$), model architectures ($M$), batch sizes ($B$), and input image pixel sizes ($P$). For $G$, we assume GPU cloud instances provided by AWS, $G \owns \{g3s, g4dn, p2, p3\}$. For models, $M$, we used well-known CNN models in literature, $M \owns$ \{\emph{AlexNet}, \emph{LeNet5}, \emph{InceptionV3}, \emph{InceptionResNetV2}, \emph{MobileNetV2}, \emph{MNIST\_CNN}, \emph{CIFAR10\_CNN}, \emph{ResNetSmall}, \emph{ResNet18}, \emph{ResNet34}, \emph{ResNet50}, \emph{VGG11}, \emph{VGG13}, \emph{VGG16}, \emph{VGG19}\}. We used five batch sizes, $B \owns \{16, 32, 64, 128, 256\}$, and five input image pixel sizes, $P \owns \{32 \times 32, 64\times 64, 128 \times 128, 224 \times 224, 256 \times 256 \}$. To generate images with different pixel sizes, we used the Numpy library. All workload scenarios can be generated by conducting Cartesian product in the four dimensions ($G \times M \times B \times P = \{(g, m, b, p) : g \in G, m \in M, b \in B, p \in P\}$). All cases of $G \times M \times B \times P$ cannot be completed due to hardware or model constraints. Filtering out inexecutable cases, we finalize $1228$ cases which becomes the cardinality of input dataset, which is $N$. Each workload $\vec{x_i}$ is a vector with a length $D$ that is returned by the underlying profiler. Among the operations provided by TensorFlow Profiler~\cite{tf-profiler}, we generate $65$ aggregated high-level operations from $1228$ executable workloads. Thus, the dimension of $\vec{X}$ is $1228 \times 65$.
Figure~\ref{fig:overall-architecture} shows an end-to-end working scenario of PROFET. A client who wants to estimate training time of custom CNN implementation executes the code on an anchor instance, $g_a$, of arbitrary choice with profiling enabled (Chapter \ref{ch:feature-profiling}) and submits the profiled output to the prediction model. The profiled result does not contain the internal model architecture but high level categorical latency of various operations. Using the profiling result, PROFET predicts the training time on various cases, such as on different instance types, batch sizes, and input image pixel sizes. The prediction model of PROFET is composed of profiling CNN characteristics to generate train dataset (Chapter~\ref{ch:feature-profiling}) which is responsible for presenting diverse workload characteristics in an abstracted way. The operation clustering module (Chapter~\ref{ch:feature-name-clustering}) is responsible for grouping similar operations to improve prediction accuracy of a model which uses infrequently-used operation names. The latency prediction module creates a model to predict training latency in different instance types (Chapter~\ref{ch:cross-instance-modeling}) and distinct batch sizes and input image pixel sizes (Chapter~\ref{ch:batch-pixel-predict}). To improve prediction accuracy, PROFET proposes novel prediction models, median-ensemble and scale-based second order polynomial, based on its forecasting scenarios.
\subsection{Extracting CNN Characteristics on Cloud}\label{ch:feature-profiling}
A PaaS cloud model frees CNN algorithm developers from operating cloud infrastructures and DNN development platforms because a cloud service provider is responsible for maintaining them. To provide efficient deep learning platforms, cloud vendors should understand distinct workload characteristics submitted by clients. However, in a PaaS environment, most CNN algorithm developers are not willing to share source code or internal architectures of a developed model. Thus, previous works which reference internal model architectures to predict DNN training latency~\cite{paleo, neuralpower, mlpredict-bigdata} cannot be applied directly on a PaaS environment. To predict training latency of arbitrary CNN implementations on a public PaaS model, PROFET should be able to characterize a workload while minimally disclosing the internal model architecture.
\begin{figure}[t]
\centering
\includegraphics[width=0.45\textwidth]{figures/profiling-feature.pdf}
\caption{An example of profiling data generated from AlextNet model training using TensorFlow profiler}
\label{fig:tf-profiler-example}
\end{figure}
Other than referencing source code to characterize CNN workloads, PROFET proposes to use performance metrics provided by profilers of DNN programming platforms, such as TensorFlow~\cite{tensorflow}, MXNet~\cite{mxnet}, Torch~\cite{pytorch}, and Theano~\cite{theano}. Though information from metrics of each platform differs slightly, they provide response times for core DNN-specific operations in common.
To better understand performance metrics provided by a DNN development platform, we show an example of a CNN model profiling outcome generated by TensorFlow Profiler~\cite{tf-profiler} in Figure~\ref{fig:tf-profiler-example}. A profiling output contains Operation, Operation detail, and Latency fields. The operation field indicates a method name used in a source code which is specified by a development platform. The operation details field contains rich information of a CNN model which includes method name and its layer location, input and output tensor sizes, memory usage, and many more. The information contains internal architecture of custom implementation, and a model can be assembled using the information. Of the profiling outcome, PROFET proposes to use the operation field and the corresponding aggregated time field as features to represent CNN workloads and make a model to predict training time of arbitrary CNN implementations. With the high level information about operations, Hafeez et. al. proved that profiled outcome can well represent characteristics of arbitrary CNN models~\cite{aws-cnn-analysis}. Furthermore, hiding operation details can relieve CNN algorithm developers when they share the workload characters with public cloud PaaS vendors which can offer an optimal development environment for any kinds of CNN workloads. In summary, PROFET adopts a \emph{black-box} approach by using high-level expression of CNN workloads without using internal model architectures which is contrary to the \emph{white-box} approach adopted in previous work~\cite{paleo, mlpredict-bigdata, neuralpower}, and such black-box design fits very well with the public cloud computing environment where the service providers and consumers differ.
Using the operation names and aggregated latencies as features constructs input vectors, $\vec{X}$, and we define output vectors, $\vec{Y}$, as corresponding batch latencies. Formally speaking, for an arbitrary CNN workload profiling features, $x_i$, $y_i$ is a scalar value which represents a measured batch latency of workload $i$ where $i=1:N$. To note a value of specific feature, $j$, of model $i$, we use $x_{ij}$, where $i=1:N, j=1:D$, which is a scalar value. Creating feature vectors, $\vec{X}$, from a deep learning platform incurs non-negligible extra overhead that can impact batch latency, $\vec{Y}$. In our off-line experiments, about $20\% - 30\%$ larger batch latency is measured when we enabled profiling. To remove the impact from the profiling overhead, we conducted two sets of experiments with a workload, $i$, one with profiling enabled, another without profiling enabled. In the experiments with profiling, we gather $\vec{X}$. To get an accurate value of $\vec{Y}$, we measured the batch latency without enabling profiling. This procedure of separately generating $\vec{X}$ and $\vec{Y}$ is identical when a new CNN workload scenario is predicted with PROFET; an user enables profiling to get a feature vector to get predicted batch training latency without profiling enabled.
\subsection{Feature Engineering by Name Clustering}\label{ch:feature-name-clustering}
Through the profiling step, an offline train execution of model, $i$, generates a feature vector $x_i$. Due to the nature of distinct model architecture of different implementations, a set of features manifested from each model can differ even if some features are very similar. For instance, activation functions \emph{ReLU}~\cite{relu} and \emph{ReLU6} are very similar except the maximum value imposed by \emph{ReLU6}. In modeling VGG16, a \emph{ReLU} activation function is used. Meanwhile, a \emph{ReLU6} activation function is used in the MobileNetV2~\cite{mobilenetv2} modeling. In the profiling step, \emph{ReLU} and \emph{ReLU6} are treated as different features, and the similarity cannot be modeled. This situation can lower prediction accuracy of PROFET especially when a model generates a feature vector which includes unique features that are not observed during training. To embed operation similarity in the PROFET modeling, we propose to measure the quantitative similarity of operation names and use the similarity value as a metric to cluster analogous features. The clustered features are aggregated to a single feature by summing the profiled values.
\subsubsection{Measurement of Operation Name Similarity}
The operation names from profiling result are representative to characterize distinct operations. Thus, we propose to measure the string similarity of operation names to judge whether an arbitrary combination of operations is appropriate to be clustered. In the string similarity measurement, we use the Levenshtein~\cite{levenshtein} algorithm. The algorithm measures how many steps of modification, such as a character insertion, deletion, and replacement, must be repeated to make two comparing words identical. The lower Levenshtein distance implies more similarity. For example, to make two operations names of \emph{ReLU} and \emph{ReLU6} the same, only the \emph{6} character must be inserted into the \emph{ReLU}, and the distance is one. In another case, the distance of two different operations of \emph{ReLU} and \emph{Conv2D} is six because all four characters of \emph{ReLU} must be replaced, and additional two characters of \emph{2D} must be inserted to make two words the same. Calculating the Levenshtein distance among all pairs of $D$ features results in $D \times D$ distance matrix which is symmetric.
\subsubsection{Feature Name Clustering by Distance}
Using the Levenshtein distance matrix, PROFET needs to group similar operations for aggregation. To achieve the goal, we apply hierarchical clustering~\cite{hierarchical-clustering}. The clustering algorithm generates a tree structure based on distances among nodes. In the PROFET implementation, each operation name becomes a leaf node, and operations with a similar name lies in a close location. Using a generated tree, a dendrogram~\cite{dendrogram} is generated to express hierarchical relationships among nodes. A height in a dendrogram of any two or more objects that are joined together implies the similarity of the joined nodes, and we can quantitatively judge the similarity based on the height; the lower height means more similarity. Deciding distance between any two nodes is simple by referencing the $D \times D$ dimensional matrix. It becomes more complex when there are more than two nodes that exist under a same parent node. To calculate a height of a node with multiple nodes in an already established cluster, various linkage heuristics exist, such as average, single, complete, and Ward's~\cite{linkage}. Among them, PROFET applies the average linkage method based on empirical analysis. The average method calculates pair-wise distances of a new candidate node to all nodes which are already in a cluster and averages the calculated distances to use as a height metric to build a dendrogram. For example, assuming three operation names of \emph{MaxPoolGrad, AvgPoolGrad}, and \emph{ArgMax}, the Levenshtein distance of \emph{MaxPoolGrad} and \emph{AvgPoolGrad} is three which is the smallest of the combinations. Assuming the two operations are clustered, the distances to the new candidate, \emph{ArgMax}, are ten and eight for \emph{MaxPoolGrad} and \emph{AvgPoolGrad}, respectively. Adopting the average linkage heuristic results in the distance between the new candidate operation and the cluster as nine, which is the average.
\subsubsection{Clustered Feature Aggregation}
After building a hierarchical dendrogram tree, it is important to generate clusters based on text similarity. In creating a cluster, we should determine the maximum height which decides cluster boundaries. The higher maximum height will result in lower number of clusters and higher number of operations in a cluster. In other words, the lower maximum height will result in more number of clusters with a lower number of operations in a cluster. In our thorough empirical analysis, setting the maximum height as six results in the best prediction accuracy. By setting the maximum height as six, all the operations whose height are less than six in a dendrogram tree are clustered in a single group. For each group of operations, a single feature is generated by aggregating the values of each operation. For aggregation, we use the sum operator. Some representative aggregated operations are [\emph{FusedBatchNormV3, FusedBatchNormGradV3}], [\emph{AssignSubVariableOp, AssignAddVariableOp}], [\emph{Softmax, ArgMax}], [\emph{MaxPoolGrad, AvgPoolGrad}], [\emph{DepthwiseConv2dNativeBackpropInput, DepthwiseConv2dNativeBackpropFilter}], [\emph{BiasAddGrad, BiasAdd}], [\emph{MatMul, MaxPool, AvgPool}], [\emph{Relu6Grad, RsqrtGrad, ReluGrad}]. In most cases, the clustered result is reasonable, and similar operations with a slightly different name are clustered together. However, in some cases, irrelevant operations with similar names are also clustered together, such as \emph{MatMul} and \emph{MaxPool}. Though they are clustered in a same group, the impact from the features after clustering does not change because we linearly sum the value in the aggregation.
\begin{figure}[t]
\centering
\includegraphics[width=0.45\textwidth]{figures/feature-clustering.pdf}
\caption{A procedure of feature clustering based on the operation name similarity}
\label{fig:feature-clustering}
\end{figure}
Figure~\ref{fig:feature-clustering} explains the overall procedure of feature clustering. In the \emph{Phase 1}, a Levenshtein distance matrix is generated using the features, and similar operation names are clustered in the \emph{Phase 2} using a dendrogram tree. In the \emph{Phase 3}, operations in a same cluster is aggregated to a single feature by summing the value. The feature aggregation by name similarity greatly enhances PROFET's prediction accuracy especially when a model uses a unique operation that are not generally used. For example, aggregating \emph{ReLU} and \emph{ReLU6} operation benefits a model which uses \emph{ReLU6} operation because it is not a widely used operation and might not be used during training.
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{figures/anchor-prediction-system.pdf}
\caption{Predicting the training latency on a target instance type based on profiled feature from an anchor instance}
\label{fig:anchor_prediction}
\end{figure*}
\subsection{Modeling CNN Performance on Cloud}\label{ch:modeling}
PROFET predicts the training latency of arbitrary CNN implementations ($M$) on different GPU-based instance types ($G$), batch sizes ($B$), and input image pixel sizes ($P$). To formally express input datasets, $x_i$, in a finer-grained way, we use the lowercase letter of each category as a subscript, $x_{mgbp} = \{(m, g, b, p) : m \in M, g \in G, b \in B, p \in P\}$. For example, to specify input datasets of an arbitrary model ($m$), batch size ($b$), and input pixel size ($p$) for all the possible GPU instances in $G$, we use $x_{mbp}$ which excludes $g$ from the subscript, and the cardinality of $x_{mbp}$ equals the size of $\arrowvert G \arrowvert$, which means four instance types with GPU devices used in this proposed work. The prediction model of PROFET is composed of two phases, cross-instance performance prediction and batch/input image pixel size prediction.
\subsubsection{Cross-Instance Performance Modeling}\label{ch:cross-instance-modeling}
The goal of cross-instance performance modeling phase is to predict the training latency of an arbitrary workload of $x_{mbp}$ on various instance types. This phase requires a profiled feature set of a target workload on an instance type which we call as an anchor instance, $g_a$. Using the profiling outcome from $g_a$, PROFET predicts the training latency on a target instance type, $g_t$ where $g_a \neq g_t$ and $g_a,g_t \in G$, of workload $x_{mbp}$. Let us define a training latency prediction model from an anchor instance, $g_a$, to a target instance, $g_t$, as $f_{g_a \rightarrow g_t}$. The training dataset of a model $f_{g_a \rightarrow g_t}$ is defined as $\mathcal{D}_{g_a \rightarrow g_t}$. To build a prediction model of training latency on a target instance by using an observed profiled features of an anchor instance, we define the training dataset as follows.
\begin{equation*}
\mathcal{D}_{g_a \rightarrow g_t} \triangleq \{(x_{mbp|G=g_a},y_{mbp|G=g_t}) | mbp \in (M \cup B \cup P)\}
\end{equation*}
For workloads from a combination of various model architectures, batch sizes, and input image pixel sizes ($mbp$), an input feature $x$ is returned from a deep learning training platform profiler after executing on an anchor instance, $g_a$. The output of the model, $y$, is a scalar value which represents a batch latency for the same workload, $mbp$, which is executed from a target instance, $g_t$.
Using anchor-target instance experiment datasets, $\mathcal{D}_{g_a \rightarrow g_t}$, PROFET builds a prediction model using an ensemble algorithm~\cite{bagging} with a median operator~\cite{median-ensemble}. An ensemble modeling in machine learning combines multiple models that are trained individually to have better prediction accuracy than the original models. The bootstrap aggregating (bagging)~\cite{bagging} trains multiple models using a same set of train dataset and allocate weights for each model to calculate the final prediction. Adjusting model weights equally results in the average ensembling. Lang et. al.~\cite{median-ensemble} proposed median-based bagging algorithm which adopts the median predicted values among multiple candidate models. The authors claimed that using the median value helps to improve the model accuracy by removing noises in real-world signal processing applications. To build a median bagging ensemble model, PROFET builds three independent prediction models which use a deep neural-net (DNN) architecture, random forest~\cite{random-forest}, and linear regressor. We build a DNN model of $128 \times 64 \times 32 \times 16 \times 1$ dense layer architecture with Rectified Linear Unit (ReLU) activation~\cite{relu} in each layer. We set the model to minimize the combined loss of MAPE and RMSE while using Adam optimizer~\cite{adam-optimizer} with a learning rate of $0.001$. To build models with random forest and linear regression algorithm, we use Python's Scikit-Learn library~\cite{scikit-learn} with the default hyper-parameters provided by the library.
Figure~\ref{fig:anchor_prediction} explains the overall procedure of cross-instance batch latency prediction. The upper layer of the figure expresses train dataset generation phase. In the step, all the workloads are executed on all the GPU instance types in $G$, and profiling features with the corresponding batch latencies are recorded. To use the experimental result as a model training input, we match the profiled feature ($x_{mbp}$) of an anchor instance ($x_g$) and batch latencies ($y_{mbp}$) of target instances ($x_t$). In the figure, the anchor instance type is \emph{g3s}, and the target instance types are \emph{g4, p2, p3}. It shows three training input cases; for a same profiled feature from \emph{g3}, three batch latencies from \emph{g4, p2, p3} are matched as outputs. In the prediction model building step, separate models are trained per each anchor and target instance type combination using distinct training dataset of $\mathcal{D}_{g_a \rightarrow g_t}$. Using the unique dataset per anchor and target instance types, three models of linear regressor, random forest, DNN are trained for median-ensemble modeling. In the figure, three separate ensemble models are built, $f_{g3s \rightarrow g4}, f_{g3s \rightarrow p2}$, and $f_{g3s \rightarrow p3}$.
\subsubsection{Modeling Latency of Dataset and Batch Size Change}\label{ch:batch-pixel-predict}
In the first phase of latency prediction, PROFET predicts the expected training time of an arbitrary model, $x_{mbp}$, on any target instances based on the profiling outcome from an anchor instance. In the second phase of the prediction, using the predicted latency on any target instance types, PROFET predicts the training latency on distinct batch and input image pixel sizes on all the target instance types.
Formally speaking, given the predicted training time of an arbitrary workload of $mbp$ on an instance $g$ which is $y_{mbpg}$, PROFET provides expected training latency on any batch and input image pixel size combinations, $bp \in B \times P$. Modeling training time for different instance types requires expressing non-linear relationships of hardware specifications across distinct instance types. Different from the instance types, the impact from changing batch and input image pixel sizes can be expressed using the ratio of numerical value differences.
While building a training time prediction model for different batch and input image pixel sizes, PROFET first defines the minimum and maximum value of batch and input image pixel sizes. Given the defined values, PROFET applies a min-max scaler for training latency and builds a second-order polynomial regression model~\cite{polynomial-regression} to express the relationships between normalized train latency and batch, input image pixel sizes. The regressor model is expressed as $T_N(b) = \alpha_2 b^2 + \alpha_1 b + \alpha_0$ for an arbitrary batch size $b$. $T_N(b)$ is the normalized latency of a batch size $b$, and the regressor model decides $\alpha_2, \alpha_1,$ and $\alpha_0$ which minimizes the prediction error for $T_N(b)$. Please note that a separate regressor model is built for batch size and input image pixel size per each instance type, and $b$ can be replaced with $p$ to express the prediction model for different input image pixel sizes.
Using the second-order polynomial regressor model, given a batch ($b$) or input image pixel ($p$) size, PROFET predicts a normalized time ($T_N$) in the range of $0.0 - 1.0$, and the value needs to be denormalized to reflect the real training latency range. To achieve the goal, PROFET requires two original denormalized training latencies with a minimum and maximum configuration which are defined as $T_O(min)$ and $T_O(max)$, respectively. Assuming that the normalized predicted value from a second-order polynomial regressor for a batch size $T_N(b)$, the denormalized original latency of a target configuration, $T_O(b)$, is calculated as Equation~\ref{eq:range-latency-predict}.
\begin{equation}\label{eq:range-latency-predict}
T_O(b) = T_N(b) \times (T_O(max) - T_O(min)) + T_O(min)
\end{equation}
Figure~\ref{fig:batchsize_prediction} explains the overall procedure of predicting training latency with varying batch sizes. In the example, the minimum batch size is $16$, and the maximum batch size is $256$. On upper side of the figure, we show the first step of predicting a normalized latency of an arbitrary batch size, $b$, which is $T_N(b)$. In building a regressor model, it uses the normalized latency from all the models and image pixel sizes as a training input. The input datasets are grouped by an instance type which results in a separate model per each instance type. After predicting a normalized latency of target batch size, in the bottom of the figure, it denormalizes $T_N(b)$ to $T_O(b)$ by using a minimum ($T_O(min)$) and maximum ($T_O(max)$) latency of the target model and image pixel size. Prediction of latency for a distinct image pixel size is similar to a process shown in Figure~\ref{fig:batchsize_prediction}.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figures/minmax-polynomial-system.pdf}
\caption{Prediction of training latency with respect to the diverse batch sizes}
\label{fig:batchsize_prediction}
\end{figure}
\subsubsection{Predicting Latency of New GPU Model}\label{sec:new-gpu-support}
A new type of GPU is released quite often, and it is important to anticipate training latency on such new hardware which was not available during the prediction modeling. PROFET does not use hardware specifications as prediction model features as previous work did~\cite{paleo, mlpredict-bigdata, habitat}, and it cannot predict training latency for a new GPU device which was not available at a modeling phase. However, this does not become a significant issue when the prediction service is provided by a cloud vendor. In the cloud, service providers decide when to make a new instance type available to customers, and they have enough time to prepare a latency prediction model for the new hardware. As cloud computing becomes more prevalent for diverse applications, cloud vendors should become more responsible for dealing with complex system operations. From this context, preparation of prediction model by a cloud service vendor for a new hardware type is rational.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figures/demo-implementation.pdf}
\caption{The implementation of PROFET using serverless computing services}
\label{fig:demo-implementation}
\end{figure}
\section{Implementation : PROFET Service}
We make all the materials of PROFET publicly accessible\footnote{PROFET is publicly available at \url{http://profet.ddps.cloud}}. The source code include extracting profiled features from TensorFlow profiler on AWS EC2 and an ensemble model trainer. We also built an publicly available end-to-end demonstration system adopting a serverless architecture~\cite{serverless-backward-forward}. Figure~\ref{fig:demo-implementation} shows the architecture of implemented demo system. A static HTML file is served through Amazon S3, an object storage service that provides a publicly accessible HTTP end-point URL for a stored file object. A user's prediction request is served in AWS Lambda, which provides a fully-managed source code execution environment. AWS Lambda and other similar services relieve users of the burden of maintaining function run-time and allow developers to concentrate on core application development. However, it has diverse limitations in the run-time configuration, and the limited local storage size was the most significant issue to implement PROFET's prediction model. To solve the issue, we adopted a custom-built container image for a function run-time with all the necessary packages~\cite{lambda-container-support}. The custom container image is stored in Amazon Elastic Container Registry (ECR) and can be accessed from AWS Lambda. We used Amazon API Gateway, which provides an HTTP end-point with the RESTful API, to trigger PROFET's prediction model served in AWS Lambda.
\begin{figure*}[t]
\centering
\subfloat[g3s.xlarge]{
\includegraphics[width=0.23\textwidth]{figures/anchor-eval-g3s.pdf}
\label{fig:anchor-eval-g3s}
}
\subfloat[g4.xlarge]{
\includegraphics[width=0.23\textwidth]{figures/anchor-eval-g4.pdf}
\label{fig:anchor-eval-g4}
}
\subfloat[p2.xlarge]{
\includegraphics[width=0.23\textwidth]{figures/anchor-eval-p2.pdf}
\label{fig:anchor-eval-p2}
}
\subfloat[p3.xlarge]{
\includegraphics[width=0.23\textwidth]{figures/anchor-eval-p3.pdf}
\label{fig:anchor-eval-p3}
}
\caption{True (x-axis) and predicted (y-axis) latencies for different anchor instances}
\label{fig:anchor-evaluation}
\end{figure*}
\section{Evaluation}
We thoroughly evaluate PROFET to present its practicality and superb prediction performance. To evaluate the model accuracy, we mainly use Mean Absolute Percentage Error (MAPE) which is calculated as the ratio of the absolute error value to the true value. We also use Root Mean Squared Error (RMSE) and coefficient of determination which is denoted as $R^2$ where necessary. The $R^2$ metric represents quantitative measurement of how the predicted outcome resembles the true value. The value is range between $0.0$ and $1.0$, where the higher value means better prediction quality. $R^2$ metric is mostly used to evaluate regressor models' prediction quality.
The experiments are conducted on AWS GPU instances, CNN models, training batch sizes, and input image pixel sizes, which are defined in Section~\ref{section:profet-overall}, as $G, M, B,$ and $P$, respectively. In running CNN training on AWS, we used Deep Learning AMI (Ubuntu 18.04) Version $35.0$ which contains NVidia GPU driver library version of 450.80.02 and CUDA version of 10.1. As the CNN development platform, we use TensorFlow 2.3.0.
\subsection{Cross-Instance Latency Predictor}
Figure~\ref{fig:anchor-evaluation} presents the prediction accuracy of cross-instance performance modeling presented in Chapter~\ref{ch:cross-instance-modeling}. Each sub-figure represent different anchor instance; Figures~\ref{fig:anchor-eval-g3s},~\ref{fig:anchor-eval-g4},~\ref{fig:anchor-eval-p2},~\ref{fig:anchor-eval-p3} represents the anchor instances of \emph{g3s.xlarge, g4dn.xalrge, p2.xlarge}, and \emph{p3.2xlarge}, respectively. In the figure, the horizontal axis shows the true training latency, and the vertical axis shows the predicted latency by PROFET. The real and predicted values are expressed using a scatter plot, and the value close to an equation $y=x$ means the accurate prediction. As shown in the figure, most predicted values are very close to the real CNN training time which present superb prediction accuracy of PROFET.
To quantitatively evaluate the prediction accuracy of the median ensemble approach of PROFET for different instance types, we compare the MAPE, RMSE, and $R^2$ values with other approaches in Figure~\ref{fig:performance-model-efficiency}. The primary vertical axis represents the MAPE whose values are shown in solid gray bars. The secondary vertical axis shows the RMSE whose values are shown in the bars with the right upper diagonal pattern. Both MAPE and RMSE, lower values are better, and lower bar means better prediction accuracy. The star markers presents $R^2$ where the higher value means more precise prediction performance. In the horizontal axis, we show distinct prediction models of a linear regressor (\emph{Linear}), a tree-based non-linear regressor (\emph{RandomForest}), \emph{DNN}, and the proposed algorithm (\emph{PROFET}). In the \emph{Linear} approach, different from others, we use batch latency measured from the anchor instance as input data. The linear prediction model is an order-1 regressor which can be expressed as $\alpha x + \beta$, where $\alpha$ means coefficient of batch latency, and $\beta$ means a bias value of the model. Details of \emph{RandomForest} and \emph{DNN} implementations are presented in Section~\ref{ch:cross-instance-modeling}. Using the three different models, the median ensemble approach of PROFET takes the three predicted values and determines the median value as the predicted latency. As shown in the figure, PROFET shows the best prediction accuracy for all three metrics, where the MAPE is 11.4159\%, RMSE is 66.2311, and the $R^2$ is $0.9749$. Comparing to the MAPE of \emph{DNN}, the best among single models, the PROFET shows $3.9$\% improvement. Comparing the RMSE of the \emph{Linear}, the best among single models, the PROFET improves by $6.2\%$. Using three distinct models with different complexity in an ensemble manner compensates weakness of each model, and it greatly improves the generality of the prediction model. For instance, the MAPE and RMSE of \emph{Linear} model show drastic difference. Careful investigation reveals that it shows poor prediction accuracy for small models which can make MAPE worse even for small error values. We count how many times different algorithms are selected in the median ensemble modeling. Under our practical evaluation scenarios, \emph{Linear} is adopted $25.8\%$, \emph{RandomForest} is selected $32.8\%$, and the \emph{DNN} is adopted $41.43\%$.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figures/performance-model-efficiency.pdf}
\caption{Higher prediction accuracy of the proposed median ensemble approach of PROFET}
\label{fig:performance-model-efficiency}
\end{figure}
\subsection{Batch/Input-Pixel Size Predictor Accuracy}
We conducted experiments to evaluate the prediction accuracy of training time for different batch and input image pixel sizes after setting the minimum and maximum batch sizes as $16$ and $256$, respectively. For the input image pixel sizes, we set the minimum and maximum as $32\times 32$ and $256\times 256$, respectively. Figure~\ref{fig:batch-size-prediction} presents the prediction accuracy with respect to different batch sizes except minimum and maximum which were used in the training. The result of the input image pixel size shows a similar pattern, and we omit it.
There are two ways to get the training latencies of minimum and maximum configuration; one using a truly measured value (\emph{True}), and another using a predicted value from PROFET (\emph{Predict}). Figure~\ref{fig:minmax-scaler-true} shows the accuracy of a predictor based on the measured latency (\emph{True}). Meanwhile, Figure~\ref{fig:minmax-scaler-pred} shows the accuracy of a predictor which uses the latencies predicted by a cross-instance performance model of PROFET (\emph{Predict}). The vertical axis shows the MAPE, and the horizontal axis shows different batch sizes for prediction. As shown in the figure, the MAPE of predicting for different batch sizes is around $5\%$ when using a \emph{True} latency. When using a \emph{Predict} latency, the average MAPE is about $11\%$. In summary, to get the higher prediction accuracy ($5\%$ in \emph{True}), one has to run two training tasks, with minimum and maximum batch sizes, on a target GPU instance. If a user wants to avoid additional execution of training tasks on a target instance type, one can use the cross-instance latency prediction model of PROFET to get minimum and maximum batch size latency in a target instance type. Though such additional prediction hurts the overall prediction accuracy, it still shows $11\%$ of MAPE which is very accurate comparing to other state-of-the-art work.
\begin{figure}[t]
\centering
\subfloat[Prediction from true values]{%
\includegraphics[width=0.25\textwidth]{figures/minmax-scaler-true.pdf}%
\label{fig:minmax-scaler-true}%
}
\subfloat[Prediction from predicted values]{%
\includegraphics[width=0.25\textwidth]{figures/minmax-scaler-anchorpred.pdf}%
\label{fig:minmax-scaler-pred}%
}
\caption{Training latency prediction when batch size changes}
\label{fig:batch-size-prediction}
\end{figure}
PROFET uses the second-order polynomial regressor with a min-max scaler when predicting training latencies with different batch and input image pixel sizes. To check the impact of the model's complexity, we compare the prediction accuracy of the order-1 polynomial regressor which is presented in Figure~\ref{fig:polynomial-order-compare}. In the figure, we show the MAPE in the primary vertical axis, RMSE in the secondary vertical axis, and the $R^2$ using a star marker. The lower MAPE and RMSE are better, and the higher $R^2$ is better. The horizontal axis shows instance types. In the left part of the figure, the prediction accuracy of the order-1 polynomial regressor is presented, and the right part shows the values of the order-2 polynomial regressor. The order-2 polynomial regressor outperforms which indicates a complex relationship between batch/input image pixel sizes and training latency, and simply referring to the change ratio can result in poor prediction accuracy.
PROFET separates the training latency prediction models of cross-instance and batch/input image pixel size change. Each model makes use of unique input features that are specifically designed to predict the target latency. One can build a single model that predicts the training latency of a CNN implementation for different GPU cloud instance types, batch sizes, and input image pixel sizes without separating two prediction models. To validate the model architecture choice of PROFET which separates two distinct phases, we compare the prediction accuracy of PROFET with a single model which takes a target instance type, batch size, and input image pixel size for prediction, and Table~\ref{table:joint-seperate-compare} compares the prediction accuracy of \emph{Joint Modeling} which uses a single model and \emph{Separate Modeling} which is applied in PROFET. In the \emph{Joint Modeling}, in addition to features proposed in PROFET, an instance type, batch size, and input image pixels sizes are passed as input features expressed in a one-hot-encoding format. In each method, we build a prediction modeling using Random Forest and DNN algorithms. As shown in Table~\ref{table:joint-seperate-compare}, separating the prediction models shows significant accuracy improvement comparing to creating a single model for both Random Forest and DNN. We tried many variants of \emph{Joint Modeling} to improve the prediction accuracy of a single model. Failing to generate a single model with reasonable prediction accuracy implies that training latency differs for different cloud instance types and batch/input image pixel sizes, and this result validates the prediction model's design choice of PROFET, which generates two separate models.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figures/scaler-order-compare.pdf}
\caption{Prediction accuracy of batch size changes with respect to different polynomial equation orders}
\label{fig:polynomial-order-compare}
\end{figure}
\begin{table}
\def1.3{1.3}
\centering
\begin{tabular}{|c||c|c|c|c|}
\hline
& \multicolumn{2}{c|}{Joint Modeling} & \multicolumn{2}{c|}{Separate Modeling (PROFET)} \\ \hline
& RandomForest & DNN & RandomForest & DNN \\ \hline
MAPE &126.0314 &90.4146 &16.8430 &11.8899 \\
$R^2$ &0.5990 &-0.0765 &0.9515 &0.9615 \\
RMSE &247.6350 &405.7258 &91.9789 &81.9791 \\
\hline
\end{tabular}
\caption{The prediction accuracy of Joint Modeling and Separate Modeling (PROFET)}
\label{table:joint-seperate-compare}
\end{table}
\begin{figure}[t]
\centering
\subfloat[Models containing unique features]{
\includegraphics[width=0.45\textwidth]{figures/feature-clustering-improve.pdf}
\label{fig:feature-clustering-improve}
} \\
\subfloat[Models containing common features]{
\includegraphics[width=0.45\textwidth]{figures/feature-clustering-sustain.pdf}
\label{fig:feature-clustering-sustain}
}
\caption{Prediction accuracy improvement from feature clustering}
\label{fig:feature-clustering-experiment}
\end{figure}
\subsection{Efficiency of Profiled Feature Clustering}
PROFET proposed a feature engineering heuristic by clustering profiled operation names for aggregation. To present the efficiency of operation clustering as a feature engineering heuristic, Figure~\ref{fig:feature-clustering-experiment} shows the MAPE values in the primary vertical axis of PROFET. The gray solid bars present MAPE value when PROFET disables feature clustering, and the upper right diagonal bars shows the prediction accuracy when we enable the feature clustering module. Figure~\ref{fig:feature-clustering-improve} shows the MAPE of models which contain unique operations that might not exist during the prediction model training phase. For each model, we also add the MAPE improvement by enabling the feature clustering. The MAPE improves the most with a model of \emph{InceptionV3} which is $29.9\%$. We can observe that using the feature clustering improves MAPE at least $8.3\%$.
To compare training latency prediction accuracy of models which do not contain unique features, we present MAPE changes of \emph{ResNet} and \emph{VGG} models. Both model categories have many variants by repeatedly locating similar layers. Thus, the models in a same category are likely to use a similar set of operations, and they might not get much benefit from the feature clustering. As shown in Figure~\ref{fig:feature-clustering-sustain}, enabling feature clustering does not result in drastic prediction accuracy improvement. However, one thing we should note is that the clustering features and aggregation do not result in negative impact for the prediction accuracy. Thus, we can conclude that the similar operation clustering heuristic results in a noticeable prediction accuracy improvement for models with unique operations without sacrificing prediction accuracy of models with general operations.
\begin{table}
\def1.3{1.3}
\centering
\begin{tabular}{|p{1.2cm}||p{2.88cm}|p{2.88cm}|}
\hline
& \makecell{PALEO} & \makecell{PROFET} \\ \hline
MAPE & \makecell{10.1072} & \makecell{6.22435} \\
$R^2$ & \makecell{0.99399} & \makecell{0.99615} \\
RMSE & \makecell{32.3637} & \makecell{19.3185} \\ \hline
\end{tabular}
\caption{The prediction accuracy of PALEO and PROFET}
\label{table:paleo-profet-compare}
\end{table}
\begin{table}
\def1.3{1.3}
\centering
\begin{tabular}{ |c||c|c|c|c| }
\hline
& \multicolumn{2}{c|}{MAPE (\%)} & \multicolumn{2}{c|}{RMSE} \\
\hline
Batchsize & MLPredict & PROFET & MLPredict & PROFET \\
\hline
16 &15.68 &3.74 &90.82 &17.8 \\
32 &17.89 &3.35 &135.63 &24.61 \\
64 &24.27 &3.96 &408.81 &61.43 \\
128 &115.39 &6.16 &2254.41 &205.46 \\
\hline
\end{tabular}
\caption{The prediction accuracy of VGG16 model for diverse GPU instances of MLPredict and PROFET}
\label{tab:mlpred-profet-compare}
\end{table}
\subsection{Comparison to the State-of-the-Art}
Most contemporary CNN model training latency prediction algorithms adopt the \emph{white-box} approach while referencing the internal model architecture and hardware configurations as features. We qualitatively argue that the \emph{black-box} approach of the PROFET is more appropriate in a public cloud environment. To quantitatively present the prediction accuracy of PROFET with the most recent related work, we compare its performance with Paleo~\cite{paleo}, MLPredict~\cite{mlpredict-bigdata}, and Habitat~\cite{habitat}.
First, reproducing the experiment result from Paleo requires building a CNN model training environment which is identical to the one Paleo used. We discovered that it is impractical to build an identical environment due to the incompatible TensorFlow and GPU driver and library version. Applying the Paleo's prediction algorithm on a contemporary CNN development environment does not result in as accurate result as one presented in the original paper, and we decided to compare the result presented in the original publication~\cite{paleo}. Table~\ref{table:paleo-profet-compare} compare the prediction accuracy of PROFET and Paleo. Among experiment results conducted by PROFET, we compare CNN models which are common to Paleo, which are AlexNet~\cite{alexnet} and VGG-16~\cite{vgg}. In PROFET, we used the true and predicted values of the common models using the cross-instance prediction modeling algorithm. As shown in the table, PROFET outperforms Paleo for all three metrics. For MAPE, PROFET is $38\%$ better and $40\%$ better for RMSE. Analyzing Paleo's experiment result reveals that the theoretical modeling of Paleo cannot represent the real operation characteristics of a CNN development platform, such as TensorFlow. Different from Paleo, PROFET uses a profiled outcome as features which represent the practical execution characteristics.
After Paleo~\cite{paleo} is proposed, MLPredict~\cite{mlpredict-bigdata} presented an algorithm to predict training time of DNN models across different GPU devices for different batch sizes, and it presented significant accuracy improvement. To compare the prediction accuracy of PROFET with MLPredict, we implemented the algorithm of MLPredict following details in the paper~\cite{mlpredict-bigdata} while communicating with the authors of the paper. Table~\ref{tab:mlpred-profet-compare} shows the MAPE and RMSE of MLPredict and PROFET with different batch sizes. In the MLPredict publication, the prediction result of VGG16 model is presented the most and showed the best result, and we compare the prediction accuracy of VGG16.
\begin{table}[t]
\def1.3{1.3}
\centering
\begin{tabularx}{0.45\textwidth}{|c||X|X|}
\hline
& \makecell{Habitat} & \makecell{PROFET} \\ \hline
T4 $\rightarrow$ V100 & \makecell{12.16} & \makecell{7.04} \\
V100 $\rightarrow$ T4 & \makecell{7.99} & \makecell{5.59} \\ \hline
\end{tabularx}
\caption{The prediction accuracy (MAPE) of Habitat and PROFET with different combination of anchor and target GPUs}
\label{table:habitat-profet-compare}
\end{table}
The prediction accuracy of PROFET outperforms MLPredict for both MAPE and RMSE. We could observe that the accuracy of MLPredict is poorer than the result presented in the original paper~\cite{mlpredict-bigdata}. We contacted the authors to discuss the poor prediction accuracy and concluded this result is reasonable. In the original MLPredict paper, the authors mainly predicted training latency of small batch sizes, in the range of $1$ to $16$. The error rate of MLPredict increases as the batch size becomes larger, and we could conclude the MLPredict algorithm is rather optimized for small batch sizes. However, such small batch sizes are impractical in real-world model training because they increase training time. Recent GPUs available device memory can accommodate larger batch sizes for many well-known DNN implementations, and we believe it is more important to accurately predict the training time of larger batch sizes. In summary, PROFET improves the RMSE by $84.3\%$ comparing to MLPredict.
Last, we compare PROFET with the most recent related work, Habitat~\cite{habitat}, which uses detailed profiling results to build a model and predict the training latency across different GPUs. Habitat does not support prediction for varying batch sizes, and we measured the prediction accuracy with fixed batch sizes of $16, 32$, and $64$ both for PROFET and Habitat. Habitat's system implementation is open-sourced, and we could reproduce the experimental results in the publication. After completing the experiments, we contacted the authors and confirmed that the results were reasonable. Table~\ref{table:habitat-profet-compare} presents the prediction accuracy of PROFET and Habitat. For both systems, we use two GPUs of Nvidia T4 and V100. The row with T4 $\rightarrow$ V100 indicates that the anchor instance is T4, and the target instance to predict the training latency is V100. For different anchors and target GPUs, we predict the training latency of \emph{Resnet50, InceptionV3}, and \emph{VGG16} models and present the average MAPE. We select the GPU device and CNN models that are common to the PROFET and Habitat experiments. PROFET and Habitat both have decent accuracy. Overall, the average MAPE of PROFET is $35\%$ lower than that of Habitat.
In summary, PROFET presents the best prediction accuracy among contemporary related work by lowering the MAPE by $35\%$ (Habitat), $38\%$ (Paleo), and $84\%$ (MLPredict). We claim that the higher prediction accuracy of PROFET while using less detailed information than the \emph{white-box} approach is due to the creative definition of input features from the anchor instance and target latency from the instance type that PROFET needs to predict. Feature clustering also helps to improve the prediction model's generality.
\subsection{Supporting New GPU Devices}
As new GPU devices are released, PROFET should be able to provide the training latency prediction service for the new devices. To mimic such a case, we assume that a new GPU instance is released. One of the most recent GPU instance release by AWS was G5 instances which equip Nvidia A10 GPU. This represents a case when an user wants to try a new GPU instance type. To assume another interesting scenario of using a different cloud service provider, we select IBM AC1 instance type which equips NVidia P100 GPU which is not provided by AWS. This represents a case when an user wants to try another GPU device which is not supported by the currently using cloud vendor. Table~\ref{table:support-new-gpus} shows the MAPE of the prediction. The columns in $2-5$ represent the anchor instances. The rows in $2-3$ represent the target instances. We can discover that it shows consistent result with other anchor-target instance combinations with low prediction error. If cloud vendors provide this type of latency prediction service, users can easily decide whether migrating to a new GPU or another cloud vendor is appropriate. Please note that the NVidia A10 supported by AWS G5 instance is built with the Ampere architecture~\cite{nvidia-ampere}, the newest one at the time of writing, and it is not adopted in the other anchor GPUs. Despite the architecture difference, PROFET can still provide superb prediction accuracy.
\begin{table}[t]
\def1.3{1.3}
\centering
\begin{tabularx}{0.485\textwidth}{|c||c|c|c|c|}
\hline
\multirow{2}{*}{Target Instances}& \multicolumn{4}{c|}{Anchor Instances} \\ \cline{2-5}
& \makecell{M60(G3s)} & \makecell{T4(G4dn)} & \makecell{K80(P2)} & \makecell{V100(P3)} \\ \hline
A10 (AWS G5) & \makecell{11.1} & \makecell{11.8} & \makecell{13.52} & \makecell{8.23} \\
P100 (IBM AC1) & \makecell{9.73} & \makecell{9.11} & \makecell{9.84} & \makecell{7.31} \\ \hline
\end{tabularx}
\caption{The prediction accuracy (MAPE) of existing anchor GPUs to new target GPUs}
\label{table:support-new-gpus}
\end{table}
\section{Related Work}
\textbf{Performance Modeling on Various Hardware}: Paleo~\cite{paleo} is a DNN training time estimation algorithm. To accurately predict the training time, it separates the latency for computation and communication and uses the internal model architecture, FLOPS of an underlying GPU device, data size, and device bandwidths. NeuralPower~\cite{neuralpower} predicts training time and power usage in DNN workloads for resource-constrained hardware environments. It uses a polynomial regressor with an internal model architecture as a feature. MLPredict~\cite{mlpredict-bigdata} suggests an algorithm to predict various DNN algorithms by using internal model architectures and GPU device characteristics as features. Zancato et al.~\cite{pred-without-training} proposes a performance model that predicts the number of steps required for convergence of pre-trained model fine-tuning. It uses a stochastic differential equation to reduce prediction overhead, but it does not consider various models and GPU types. Habitat~\cite{habitat} uses a profiling result as PROFET does. However, Habitat uses detailed profiling output which might be inappropriate in a public cloud environment. The authors claim that the proposed algorithm can be executed by algorithm developers, but imposing a lot of system operation overheads to algorithm developers does not align with the current cloud computing evolution direction. The aforementioned works refer to the internal architecture of target models (\emph{white-box}), and algorithm developers may be hesitant to share the architecture. This is an important issue particularly in cloud. PROFET's \emph{black-box} modeling is better suited to cloud.
AI Gauge~\cite{ai-gauge} estimates cost and execution time for CNN model training by using a decision tree in the prediction. Despite of its accurate prediction result, it utilizes vendor specific information provided by IBM cloud which hurts the model's general applicability. DTT/B~\cite{evolving-dnn} estimates training time with features extracted from DNN models by using a genetic algorithm. The algorithm supports batch size of only 32 which limits the algorithm's applicability. Pei et al.~\cite{iter-pred-multigpu} have developed a performance model to analyze and predict data parallel training of CNN with multiple GPU devices, and it relies on details of model architecture to generate features. Srifty~\cite{srift} also proposes a performance prediction model for distributed cloud systems with multiple GPUs. The study combines profiling and model-based approach to estimate distributed training latency accurately by modeling communication overhead. The authors claim that the computation and communication time among multiple GPU devices can be modeled independently, and PROFET can utilize the communication performance estimation modeling to support parallel CNN processing.
Few work presented the performance modeling of DNN or machine learning jobs on HPC environments with heterogeneous resources either training or inference~\cite{ml-performance-modeling-hpc, ml-model-serve-heterogeneous, ai-turbo, ribbon, energy-efficient-training-gpu-fpga, marble-multigpu-dnn-scheduler-hpc}. On a heterogeneous HPC environment, the current prediction model of PROFET might not provide as accurate result as in the cloud GPU environment. Referencing the work by Han et. al.~\cite{dnn-hpc-performance-evaluation} which thoroughly evaluated the DNN performance on HPC environment can be helpful when extending the prediction model of PROFET to a HPC environment.
\textbf{Using Profiler for DNN Performance Analysis}: Various performance metrics generated during training phase can be helpful to understand workload characteristics and to estimate the overall execution time. To access detailed execution statistics, we can use a profiler provided by a hardware vendor, such as NVIDIA nsight~\cite{nsight}, or DNN development platform, such as TensorFlow Profiler~\cite{tf-profiler}. Many previous work focused on analyzing profiled outcomes to better understand DNN workload characteristics. Training Benchmark for DNN (TBD)~\cite{tbd} analyzed profiling result of various DNN models, hardware, and DNN development frameworks, and presented processing throughput, utilization, and memory consumption. In addition, the authors developed a profiler which can be utilized in various DNN development platform which can be a beneficial feature for PROFET to allow profiling. Yeung et al.~\cite{gpu-util-cloud} suggested a performance model which leverages GPU profiling outcome to measure memory usage and analyze the structure of DNN models to predict GPU memory utilization for arbitrary DNN model training tasks.
Ceer~\cite{aws-cnn-analysis} uses DNN development framework profiler to generate features to represent CNN model characteristics. Using the profiled outcome, the work classifies operations into heavy and light ones to build tree-based performance estimation model to predict training time. Ceer is similar to PROFET from the perspective of predicting CNN training latency by using profiled result. However, Ceer does not support predicting training latency across different GPU devices which is a vital feature on cloud. Liu et al.~\cite{dnn-mobile-prof} demonstrates performance analysis on CPU, memory, GPU, and power utilization through profiling result from both GPU hardware profiler and DNN development framework. PerfNetV2~\cite{perfnet-v2} uses a DNN profiler to collect detailed DNN operation data for inference and training to build performance model for execution time estimation. However, it supports only a pre-defined set of DNN operations which can limit its applicability. Meanwhile, PROFET does not limit supported operations, and it is applicable to arbitrary new DNN models.
\textbf{DNN System with Performance Modeling}: Accurate estimation of DNN training time can be beneficial especially when building a DNN training cluster. The following work showed the improved cluster utilization and efficiency when using a DNN training time estimation models. Optimus~\cite{optimus} and Cynthia~\cite{cynthia} proposed a system that maximizes DNN training cluster utilization and decreases training time. It recommends the optimal placement of parameter servers and workers on a distributed DNN cluster based on training time estimation to achieve the goal. Horus~\cite{horus} proposed a DNN job scheduling strategy to minimize interference among multiple training jobs on a single GPU device. Once-For-All(OFA)~\cite{ofa} is aimed to find DNN models that are executed efficiently on limited resources, such as mobile devices. Correct estimation of training time is critical in the process of finding an optimal DNN architecture in order to provide an applicable recommendation. RubberBand~\cite{rubberband} proposed a performance prediction model which anticipates job completion time and provisioning cost for DNN models. Using the performance model, the authors proposed an elastic hyper-parameter tuning system. Chaudhary et. al.~\cite{dl-gpu-cluster-sharing} presented a GPU scheduling algorithm for deep learning which enhances fairness and utilization. If the DNN training time on various GPU devices is fairly predictable when using a PROFET-proposed algorithm, we are confident that scheduling efficiency will improve dramatically.
\textbf{Performance Estimation on Cloud Environment}: As cloud computing becomes more popular for big-data analysis tasks, recommendation algorithms for optimal cloud environment of various machine learning algorithms are proposed in the literature. To estimate the training time of big-data machine learning jobs, Ernest~\cite{ernest} applies an NNLS equation to model interactions among different hardware configurations. It also suggests an offline experiment design algorithm on cloud and downsampling data heuristic to decrease cloud usage cost. Similar to Ernest, CherryPick~\cite{cherrypick} proposed an algorithm to provide an optimal big-data analysis environment. It applied the Bayesian optimization algorithm~\cite{bayesian-optimization} to efficiently search the space of cloud configuration. PARIS~\cite{paris} suggests a performance modeling algorithm considering the trade-off of accuracy and cost while using hybrid offline and online analyzer for arbitrary workloads. In the offline modeling step, it applies scale-based downsampling. While running an analysis job using the sampled dataset, it collects hardware usage metrics and applied Random Forest~\cite{random-forest} to forecast expected running time. MPEC~\cite{mpec-tcc} focused on recommending an optimal cloud environment to execute a dense matrix multiplication task which is a core kernel operation for many machine learning jobs. To accomplish this, the authors proposed a number of features to represent matrix multiplication scenarios as well as various cloud instances. They also proposed a cloud-based mechanism for generating optimal experiment scenarios.
The ultimate goal of the works mentioned above is similar to PROFET, which provides an optimal cloud environment. However, the computational environment for executing DNN and general machine learning algorithms is quite different. The scale-out approach is preferred for big-data processing because it focuses on processing throughput and the majority of computation occurs on the CPU. Meanwhile, a scale-up approach is more preferred to train a DNN model with specialized hardware, such as GPU. Popular development platforms are also different. For big-data processing, Map and Reduce primitives~\cite{mapreduce} are widely adopted, and Hadoop~\cite{hadoop} and Spark~\cite{spark} are popular choices. Because of the stark difference in workload characteristics, a performance estimator model should be designed separately for DNN tasks.
\section{Discussions}
PROFET can predict the training time of any arbitrary CNN implementation for a wide range of GPU devices, mini-batch sizes, and input image pixel sizes. Despite PROFET's extensive prediction capability, there are still topics to be addressed in order to provide an optimal CNN model training environment on the PaaS cloud, where CNN algorithm developers and development platform maintainers are not the same.
\textbf{Modeling train latency on different deep learning frameworks}: During PROFET evaluation, we suffered from replaying CNN development and training platforms. Most of previous related work listed the version of development platforms for their evaluations, such as TensorFlow, Keras, GPU driver, and CUDA SDK. However, we discovered that replaying the previous work algorithms on the most recent development platform did not reproduce the prediction accuracy shown in the original papers, so we had to create a development environment using specific older versions. We anticipate a similar situation for PROFET. As new SDK versions are released, the DNN training performance is more likely to be improved, but there is not yet an algorithm or a system that can model the training latency as algorithm developers change underlying SDK library versions. Though it can be challenging, we believe that modeling deep learning training performance with respect to a different SDK version can be useful, particularly for deep learning algorithm developers who need to decide whether an SDK version upgrade is worth the effort.
\textbf{Training latency prediction when using multiple GPU devices}: A large portion of DNN training tasks is conducted on a single GPU device. According to Microsoft Philly trace~\cite{dnn-workload-character-microsoft}, $86\%$ of DNN tasks are conducted on a single GPU. Hu et. al.~\cite{dnn-workload-character-sensetime} analyzed the DNN training logs of SenseTime and revealed that about $70\%$ of tasks are executed with a single GPU. Though the majority of DNN training tasks are executed on a single GPU, using multiple GPUs in parallel can be an option to expedite a training task. During a parallel CNN model training, majority of time is spent on calculating weight parameters in a GPU device and communicating weight parameters among GPU devices. PROFET recommends a method for predicting training latency in each GPU device. To expand the proposed algorithm for parallel model training, we can apply a communication overhead estimation model proposed in the literature. Hafeez et. al.~\cite{aws-cnn-analysis} thoroughly investigated the training latency of arbitrary CNN algorithms while changing the number of GPUs during training. The authors discovered that as more GPUs are added for CNN training, the performance gain ratio becomes more static, regardless of GPU instance type. As we presented in this paper, prediction of training time across distinct GPU instance types requires complex models which is more challenging, and we are sure that applying the static multiplier suggested by Hafeez et. al.~\cite{aws-cnn-analysis} to PROFET would result in accurate training latency estimation using multiple GPU devices.
\textbf{Training latency prediction for non-CNN models} : The current PROFET supports training latency prediction of CNN models. Different types of DNN models, such as language processing models of Transformer~\cite{transformer} and BERT~\cite{bert}, use different operations, and the latency prediction models built with CNN profiling result might not show as good result. The proposed feature clustering algorithm in PROFET can help to enhance prediction accuracy to some extent, but the precise performance should be measured fairly. We are currently working on extending PROFET to support general DNN models.
\section{Conclusion}
Training a CNN model with a GPU device has become the norm because it requires significant computing power, and public cloud service vendors offer a variety of GPU devices elastically. Due to the dynamically changing performance of arbitrary CNN implementations on various GPU devices, it is difficult for an algorithm developer to create an optimal training environment. This paper presented PROFET, which can predict the training latency of arbitrary CNN implementations with diverse configurations on various GPU devices, to aid in the development of an efficient CNN training environment. Without revealing implementation details of CNN implementation, PROFET can predict training latency on multiple distinct GPU devices. In addition, PROFET can predict training latency when an algorithm developer changes mini-batch sizes and input image pixel sizes. Thorough experimental result reveals PROFET outperforms contemporary related work. Other than the quantitative superiority of PROFET, predicting training latency without revealing implementation detail makes PROFET suitable in a cloud where an algorithm developer and resource providers differ.
\section*{Acknowledgments}
This work is supported by the National Research Foundation of Korea (NRF) Grant funded by the Korean Government (MSIP) (No. NRF-2015R1A5A7037615, and NRF-2020R1A2C1102544), the ICT R\&D program of IITP (2017-0-00396), and the SW Star Lab (RS-2022-00144309).
\bibliographystyle{unsrt}
|
\section{Introduction}
In \cite{bh}, we introduced the theory of equivariant Seiberg--Witten--Floer cohomology and established its basic properties. In this paper we further develop this theory, with a particular emphasis on Brieskorn homology spheres. Applications include a new proof of the Milnor conjecture and obstructions to extending group actions over a bounding $4$-manifold.
For pairwise coprime positive integers $a_1 , \dots , a_r > 1$, the Brieskorn manifold $Y = \Sigma(a_1 , \dots , a_r)$ is an integral homology Seifert $3$-manifold. The Seifert structure defines a circle action on $Y$. Restricting the circle action to finite subgroups, we obtain an action of the cyclic group $\mathbb{Z}_p$ on $Y$ for each integer $p > 1$. We obtain our main results by considering the equivariant Seiberg--Witten--Floer cohomology of $Y$ with respect to such $\mathbb{Z}_p$-actions.
\subsection{Knot concordance invariants and the Milnor conjecture}
Let $p$ be a prime number. One way of producing $\mathbb{Z}_p$-actions on rational homology $3$-spheres is to take $Y = \Sigma_p(K)$, the cyclic $p$-fold cover of $S^3$ branched over a knot $K$. From the equivariant Seiberg--Witten--Floer cohomology of $Y$ one may extract invariants of the knot $K$. In \cite{bar}, this construction was used to obtain a series of knot concordance invariants $\theta^{(p)}(K)$. These invariants are lower bounds for the slice genus $g_4(K)$, that is, $g_4(K) \ge \theta^{(p)}(K)$ for all primes $p$. More generally, the invariants $\theta^{(p)}$ can be used to bound the genus of surfaces bounding $K$ in negative definite $4$-manifolds with $S^3$ boundary.
We are interested in the case that $K$ is a torus knot. If $K = T_{a,b}$ is an $(a,b)$ torus knot and $c$ is a prime not dividing $ab$, then $\Sigma_c(T_{a,b})$ is the Brieskorn homology sphere $\Sigma(a,b,c)$ and the $\mathbb{Z}_c$-action arising from the branched covering construction coincides with the restriction to $\mathbb{Z}_c$ of the Seifert circle action. By studying the equivariant Seiberg--Witten--Floer homology of $\Sigma(a,b,c)$, we deduce the following:
\begin{theorem}\label{thm:theta1}
Let $a,b > 1$ be coprime integers and let $c$ be a prime not dividing $ab$. Then $\theta^{(c)}(T_{a,b}) = \dfrac{1}{2}(a-1)(b-1)$.
\end{theorem}
Since $\theta^{(c)}$ is a lower bound for the slice genus, we obtain the Milnor conjecture as an immediate corollary:
\begin{corollary}
Let $a,b > 1$ be coprime. Then $g_4(T_{a,b}) = \dfrac{1}{2}(a-1)(b-1)$.
\end{corollary}
\begin{proof}
The Milnor fibre of the singularity $x^a = y^b$ has genus $(a-1)(b-1)/2$ \cite{mil}, hence $g_4(T_{a,b}) \le (a-1)(b-1)/2$. On the other hand, if we let $c$ be any prime not dividing $ab$, then Theorem \ref{thm:theta1} gives $g_4(T_{a,b}) \ge (a-1)(b-1)/2$.
\end{proof}
The original proof of the Milnor conjecture due to Kronheimer and Mrowka uses gauge theory and adjunction inequalities \cite{krmr}. The result was proven again by Ozsv\'ath and Szab\'o using the $\tau$-invariant of Knot Floer homology \cite{os1} and by Rasmussen using the $s$-invariant of Khovanov homology \cite{ras}. Although our proof uses gauge theory, it does not use adjunction inequalities but rather is based on Floer theoretic methods. Thus our proof has more in common with Ozsv\'ath--Szab\'o and Rasmussen than with Kronheimer--Mrowka. It is interesting to note that our proof, like those of Ozsv\'ath--Szab\'o and Rasmussen, is based on finding a knot concordance invariant which bounds the slice genus and equals $(a-1)(b-1)/2$ for the torus knot $T_{a,b}$.
\subsection{Equivariant delta invariants of Brieskorn homology spheres}
Our next result concerns the equivariant delta invariants of Brieskorn homology spheres. The equivariant delta invariants, introduced in \cite{bh}, are a certain equivariant generalisation of the Ozsv\'ath--Szab\'o $d$-invariant and are equivariant homology cobordism invariants. Given a rational homology $3$-sphere $Y$, an action of $\mathbb{Z}_p$ on $Y$ by orientation preserving diffeomorphisms and a $\mathbb{Z}_p$-invariant spin$^c$-structure $\mathfrak{s}$, we obtain a sequence of invariants $\delta_j^{(p)}(Y , \mathfrak{s}) \in \mathbb{Q}$ indexed by a non-negative integer $j$. We call $\delta_j^{(p)}(Y , \mathfrak{s})$ the equivariant delta invariants of $(Y , \mathfrak{s})$. When $Y$ is an integral homology $3$-sphere, it has a unique spin$^c$-structure which is automatically $\mathbb{Z}_p$-invariant. In this case we may write the invariants as $\delta^{(p)}_j(Y)$. The most important property of these invariants is that they satisfy an equivariant version of the Froyshov inequality \cite{bh}. In particular, this implies that they are invariant under equivariant homology cobordism. Consequently, the $\delta_j^{(p)}$ define obstructions to extending the $\mathbb{Z}_p$-action over a contractible $4$-manifold bounding $Y$:
\begin{proposition}[Proposition 7.6, \cite{bh}]
Let $Y$ be an integral homology $3$-sphere on which $\mathbb{Z}_p$ acts by orientation preserving diffeomorphisms. Suppose that $Y$ is bounded by a contractible compact smooth $4$-manifold $W$. If the $\mathbb{Z}_p$-action extends smoothly over $W$ then $\delta_j^{(p)}(Y) = \delta_j^{(p)}(-Y) = 0$ for all $j \ge 0$.
\end{proposition}
The sequence of invariants $\{ \delta_j^{(p)}(Y) \}_{j \ge 0}$ is decreasing and eventually constant. We set $\delta_\infty^{(p)}(Y) = \lim_{j \to \infty} \delta_j^{(p)}(Y)$.
Let $Y = \Sigma(a_1, \dots , a_r)$ be a Brieskorn homology sphere and let $p$ be any prime. We assume $a_1, \dots , a_r > 1$ and $r \ge 3$ so that $Y \neq S^3$. In \textsection \ref{sec:deltabri} we prove the following results (see Proposition \ref{prop:deltaprop}):
\begin{itemize}
\item[(1)]{$\delta_j^{(p)}(Y) = \delta_\infty^{(p)}(Y)$ for all $j \ge 0$.}
\item[(2)]{$\delta(Y) \le \delta_\infty^{(p)}(Y) \le -\lambda(Y)$.}
\item[(3)]{$\lambda(Y) \le \delta_j^{(p)}(-Y) \le -\delta(Y)$ for all $j \ge 0$.}
\end{itemize}
Here $\lambda(Y)$ is the Casson invariant of $Y$ and $\delta(Y) = d(Y)/2$ is half the Ozsv\'ath--Szab\'o $d$-invariant. In particular, if $Y$ is a Brieskorn homology sphere which bounds a contractible $4$-manifold, then $\delta(Y) = 0$, $\delta_j^{(p)}(Y) = \delta_\infty^{(p)}(Y)$ and
\[
\delta_\infty^{(p)}(-Y) \le \delta_j^{(p)}(-Y) \le 0
\]
for all $j \ge 0$. Thus $\delta_j^{(p)}(\pm Y) = 0$ for all $j$ if and only if $\delta_\infty^{(p)}( \pm Y) = 0$. This justifies restricting attention to the invariants $\delta_\infty^{(p)}( \pm Y)$.
We first consider the case of free $\mathbb{Z}_p$-actions. Given a prime $p$, the restriction of the circle action on $Y = \Sigma(a_1 , \dots , a_r)$ to $\mathbb{Z}_p$ acts freely if and only if $p$ does not divide $a_1 \cdots a_r$. Hence the action is free for all but finitely many primes. In fact any free action of a finite group on $Y$ is conjugate to a finite subgroup of the Seifert circle action \cite[Proposition 4.3]{lusj}. In the free case we have:
\begin{theorem}\label{thm:free0}
Let $Y = \Sigma(a_1, \dots , a_r)$ be a Brieskorn homology sphere and let $p$ be a prime not dividing $a_1 \cdots a_r$. Set $Y_0 = Y/\mathbb{Z}_p$. Then for any spin$^c$-structure $\mathfrak{s}_0$ on $Y_0$, we have
\[
\delta^{(p)}_\infty(Y) - \delta(Y) = {\rm rk}( HF_{red}^+(Y) ) - {\rm rk}( HF_{red}^+(Y_0 , \mathfrak{s}_0 ) ).
\]
\end{theorem}
Furthermore, we have:
\begin{theorem}\label{thm:free1}
We have that ${\rm rk}( HF_{red}^+(Y) ) > {\rm rk}( HF_{red}^+( Y_0 , \mathfrak{s}_0 ))$ except in the following cases:
\begin{itemize}
\item[(1)]{$Y = \Sigma(2,3,5)$ and $p$ is any prime.}
\item[(2)]{$Y = \Sigma(2,3,11)$ and $p=5$.}
\end{itemize}
In case (1) we have ${\rm rk}( HF_{red}^+(Y) ) = {\rm rk}( HF_{red}^+( Y_0 , \mathfrak{s}_0 )) = 0$ and in case (2) we have ${\rm rk}( HF_{red}^+(Y) ) = {\rm rk}( HF_{red}^+( Y_0 , \mathfrak{s}_0 )) = 1$.
\end{theorem}
Combining these two results gives:
\begin{corollary}\label{cor:dinffree0}
Let $Y = \Sigma(a_1 , a_2 , \dots , a_r)$ be a Brieskorn homology sphere and let $p$ be a prime not dividing $a_1 \cdots a_r$. Then $\delta_\infty^{(p)}(Y) > \delta(Y)$ except in the following cases:
\begin{itemize}
\item[(1)]{$Y = \Sigma(2,3,5)$ and $p$ is any prime.}
\item[(2)]{$Y = \Sigma(2,3,11)$ and $p=5$.}
\end{itemize}
In both cases we have $\delta_\infty^{(p)}(Y) = \delta(Y) = 1$.
\end{corollary}
\begin{corollary}
Let $Y = \Sigma(a_1 , a_2 , \dots , a_r)$ be a Brieskorn homology sphere and let $m>1$ be an integer not dividing $a_1 \cdots a_r$. Suppose that $W$ is a contractible compact smooth $4$-manifold bounding $Y$. Then the $\mathbb{Z}_m$-action on $Y$ does not extend smoothly to $W$.
\end{corollary}
\begin{proof}
Let $p$ be a prime divisor of $m$. It suffices to show that the subgroup $\mathbb{Z}_p \subseteq \mathbb{Z}_m$ does not extend over $W$. Since $Y$ is bounded by a contractible $4$-manifold we have $\delta(Y) = 0$. Then Corollary \ref{cor:dinffree0} implies that $\delta_\infty^{(p)}(Y) > 0$, unless $Y = \Sigma(2,3,5)$ or $Y = \Sigma(2,3,11)$ and $p=5$. However these cases do not bound contractible $4$-manifolds as they have $\delta(Y) = 1$. So $\delta_\infty^{(p)}(Y) > 0$ which implies that the $\mathbb{Z}_p$-action does not extend smoothly to $W$.
\end{proof}
The $r=3$ case of the above result was proven by Anvari--Hambleton \cite{anha1}.
We can also show that for all sufficiently large primes, $\delta_\infty^{(p)}(Y)$ equals the Casson invariant $\lambda(Y)$.
\begin{theorem}\label{thm:free2}
Let $Y = \Sigma(a_1 , a_2 , \dots , a_r)$ be a Brieskorn homology sphere and let $p$ be a prime not dividing $a_1 \cdots a_r$. Suppose that $p > N$ where
\[
N = a_1 \cdots a_r \left( (r-2) - \sum_{i=1}^r \frac{1}{a_i} \right).
\]
Then $\delta_\infty^{(p)}(Y) = \lambda(Y)$.
\end{theorem}
\begin{proof}
This follows from Theorem \ref{thm:free0} and Proposition \ref{prop:deltap}.
\end{proof}
We now consider the case of branched coverings. Let $Y = \Sigma(a_1, \dots , a_r)$ be a Brieskorn homology sphere and let $p$ be a prime dividing $a_1 \cdots a_r$. Without loss of generality we may assume that $p$ divides $a_1$. Then the quotient space $Y_0 = Y/\mathbb{Z}_p$ is the Brieskorn homology sphere $\Sigma(a_1/p , a_2 , \dots , a_r)$ and $Y \to Y_0$ is a cyclic branched covering. Our main result is the following:
\begin{theorem}\label{thm:deltabranch}
We have that
\[
\delta(-Y) - \delta^{(p)}_\infty(-Y) \ge {\rm rk}( HF_{red}^+(Y)) - p \, {\rm rk}( HF_{red}^+(Y_0) ).
\]
\end{theorem}
\begin{remark}
Karakurt--Lidman have shown that ${\rm rk}( HF_{red}^+(Y)) \ge p \, {\rm rk}( HF_{red}^+(Y_0) )$. Thus the right hand side of the inequality in Theorem \ref{thm:deltabranch} is non-negative. Moreover, if $\delta_\infty^{(p)}(-Y) = \delta(-Y)$ then we must have an equality: ${\rm rk}( HF_{red}^+(Y)) = p \, {\rm rk}( HF_{red}^+(Y_0) )$.
\end{remark}
\begin{theorem}
Let $Y = \Sigma(a_1 , a_2 , \dots , a_r)$ be a Brieskorn homology sphere and let $p$ be a prime dividing $a_1 \cdots a_r$. Suppose that $W$ is a contractible compact smooth $4$-manifold bounding $Y$. If ${\rm rk}( HF_{red}^+(Y)) > p \, {\rm rk}( HF_{red}^+(Y_0) )$, then the $\mathbb{Z}_p$-action on $Y$ does not extend smoothly to $W$.
\end{theorem}
The $r=3$ case of this result was proved by Anvari--Hambleton \cite{anha2} without requiring the assumption that ${\rm rk}( HF_{red}^+(Y)) > p \, {\rm rk}( HF_{red}^+(Y_0) )$.
We expect that the condition ${\rm rk}( HF_{red}^+(Y)) = p \, {\rm rk}( HF_{red}^+(Y_0) )$ is rarely satisfied, however there are some cases where it does hold. One family of examples is given by $Y = \Sigma(2,3,30n+5)$ and $p=5$, in which case $Y_0 = \Sigma(2,3,6n+1)$ and ${\rm rk}( HF_{red}^+(Y)) = 5 \, {\rm rk}( HF_{red}^+(Y_0) ) = 5n$. All of these examples have $\delta(Y) = 1$, so such a $Y$ can not bound a contractible $4$-manifold. We suspect that there are no examples where ${\rm rk}( HF_{red}^+(Y)) = p \, {\rm rk}( HF_{red}^+(Y_0) )$ and $Y$ bounds a contractible $4$-manifold.
Suppose that $Y_1, \dots , Y_m$ are Brieskorn homology spheres and $p$ is a prime such that for each $i$, the $\mathbb{Z}_p$-action on $Y_i$ is not free. Then we can form an equivariant connected sum $Y = Y_1 \# \cdots \# Y_m$ by attaching the summands to each other along fixed points. From \cite[Proposition 3.1]{bar} we have that $\delta_{j_1+ \cdots + j_m}^{(p)}(-Y) \le \sum_{k=1}^m \delta_{j_k}^{(p)}(-Y_k)$. Taking $j_1, \dots , j_m$ sufficiently large, we obtain
\[
\delta_{\infty}^{(p)}(-Y) \le \sum_{k=1}^m \delta_{\infty}^{(p)}(-Y_k).
\]
Then Theorem \ref{thm:deltabranch} implies that
\[
\delta(-Y) - \delta_{\infty}^{(p)}(-Y) \ge \sum_{k=1}^m \left( {\rm rk}( HF_{red}^+(Y_k)) - p \, {\rm rk}( HF_{red}^+(Y_k/\mathbb{Z}_p) ) \right),
\]
which gives us the following result:
\begin{corollary}
Let $Y_1, \dots , Y_m$ be Brieskorn homology spheres and let $p$ be a prime such that for each $i$, the $\mathbb{Z}_p$-action on $Y_i$ is not free. Let $Y = Y_1 \# \cdots \# Y_m$ be the equivariant connected sum. Suppose that $W$ is a contractible compact smooth $4$-manifold bounding $Y$. If ${\rm rk}( HF_{red}^+(Y_i)) > p \, {\rm rk}( HF_{red}^+(Y_i/\mathbb{Z}_p) )$ for some $i$, then the $\mathbb{Z}_p$-action on $Y$ does not extend smoothly to $W$.
\end{corollary}
\subsection{Structure of the paper}
The paper is structured as follows. In \textsection \ref{sec:esw} we recall the basic results on equivariant Seiberg--Witten--Floer cohomology from \cite{bh} and the associated knot concordance invariants. In \textsection \ref{sec:fdelta} we examine in great detail the spectral sequence relating equivariant and non-equivariant Floer cohomology and use this to deduce more refined information about the equivariant delta invariants. In \textsection \ref{sec:tab}, we study the Floer homology of Brieskorn spheres $\Sigma(a,b,c)$ and use this to compute the invariants $\theta^{(c)}(T_{a,b})$ leading to the proof of Theorem \ref{thm:theta1}. In \textsection \ref{sec:branched}, we consider the case where $\mathbb{Z}_p$ acts non-freely on $\Sigma(a_1, \dots , a_r)$ and prove Theorem \ref{thm:deltabranch}. Finally in \textsection \ref{sec:free}, we consider the case where $\mathbb{Z}_p$ acts freely on $\Sigma(a_1, \dots , a_r)$ and prove Theorems \ref{thm:free0}, \ref{thm:free1} and \ref{thm:free2}.
\section{Equivariant Seiberg--Witten--Floer cohomology and knot concordance invariants}\label{sec:esw}
\subsection{Seiberg--Witten--Floer cohomology}\label{sec:swf}
Let $Y$ be a rational homology $3$-sphere and $\mathfrak{s}$ a spin$^c$-structure. For such a pair $(Y,\mathfrak{s})$, Manolescu constructed an $S^1$-equivariant stable homotopy type whose equivariant (co)homology groups are isomorphic to the Heegaard Floer or Monopole Floer (co)homology groups of $(Y,\mathfrak{s})$ \cite{man}. We denote the $S^1$-equivariant reduced cohomology groups with coefficients in $\mathbb{F}$ by $HSW^*(Y , \mathfrak{s} ; \mathbb{F})$ and refer to them as the {\em Seiberg--Witten--Floer cohomology} of $(Y , \mathfrak{s})$. If the coefficient group is understood then we will write $HSW^*(Y , \mathfrak{s})$. If $Y$ is an integral homology $3$-sphere, then it has a unique spin$^c$-structure and in this case we simply write $HSW^*(Y)$. We have that $HSW^*(Y , \mathfrak{s})$ is a graded module over the ring $H^*_{S^1} = H^*_{S^1}( pt ; \mathbb{F})$. Note that $H^*_{S^1} \cong \mathbb{F}[U]$, where $deg(U) = 2$.
There exists a chain of isomorphisms relating Seiberg--Witten--Floer homology to monopole Floer homology \cite{lima2} and to Heegaard Floer homology \cite{klt1,klt2,klt3,klt4,klt5,cgh1,cgh2,cgh3,tau}. In particular we have isomorphisms
\[
HSW^*(Y , \mathfrak{s}) \cong HF_+^*(Y , \mathfrak{s})
\]
where $HF_+^*(Y , \mathfrak{s})$ denotes the plus version of Heegaard Floer cohomology with coefficients in $\mathbb{F}$. Unless stated otherwise, we will take our coefficient group $\mathbb{F}$ to be a field. Then the universal coefficient theorem implies that the Heegaard Floer cohomology $HF_+^*(Y , \mathfrak{s})$ is isomorphic to the Heegaard Floer homology $HF^+_*(Y , \mathfrak{s})$, except that the action of $\mathbb{F}[U]$ on $HF^+_*(Y,\mathfrak{s})$ gets replaced by its dual, so $deg(U) = 2$ whereas in Heegaard Floer homology one has $deg(U)=-2$. We will frequently identify $HSW^*(Y , \mathfrak{s})$ with $HF^+_*(Y , \mathfrak{s})$, equipped with the dual $\mathbb{F}[U]$-module structure.
Let $d(Y , \mathfrak{s})$ denote the Ozsv\'ath--Szab\'o $d$-invariant. Due to the isomorphism $HSW^*(Y , \mathfrak{s}) \cong HF^+_*(Y , \mathfrak{s})$, we have that $d(Y , \mathfrak{s})$ is the minimal degree $i$ for which there exists an $x \in HSW^i(Y , \mathfrak{s})$ with $U^k x \neq 0$ for all $k \ge 0$. For notational convenience we define $\delta(Y , \mathfrak{s}) = d(Y,\mathfrak{s})/2$.
We define $HSW_{red}^*(Y , \mathfrak{s}) = \{ x \in HSW^*(Y , \mathfrak{s}) \; | \; U^k x = 0 \text{ for some } k \ge 0\}$. More generally, given an $\mathbb{F}[U]$-module $M$, we write $M_{red}$ for the submodule of elements $x \in M$ such that $U^k x = 0$ for some $k \ge 0$.
Recall that for any $(Y , \mathfrak{s})$, there is an isomorphism of $\mathbb{F}[U]$-modules
\[
HF^+(Y , \mathfrak{s}) \cong \mathbb{F}[U]_{d(Y,\mathfrak{s})} \oplus HF^+_{red}(Y , \mathfrak{s}),
\]
where for any $\mathbb{F}[U]$-module $M^*$ and any $d \in \mathbb{Q}$, we define $M^*_d$ by $(M^i_d) = M^{i-d}$. It follows that we similarly have an isomorphism
\[
HSW^*(Y , \mathfrak{s}) \cong \mathbb{F}[U]_{d(Y , \mathfrak{s})} \oplus HSW_{red}(Y , \mathfrak{s}).
\]
\subsection{Equivariant Seiberg--Witten--Floer cohomology}\label{sec:eswfc}
Let $Y$ be a rational homology $3$-sphere and suppose that $\tau \colon Y \to Y$ an orientation preserving diffeomorphism of order $p$, where $p$ is prime. This gives an action of the finite group $G = \mathbb{Z}_p$ on $Y$ generated by $\tau$. Let $\mathfrak{s}$ be a spin$^c$-structure preserved by $\tau$. In \cite{bh}, the authors constructed the equivariant Seiberg--Witten--Floer cohomology groups $HSW^*_G(Y , \mathfrak{s})$. Except where stated otherwise, we take Floer cohomology with respect to the coefficient field $\mathbb{F} = \mathbb{Z}_p$. Then $HSW^*_G(Y , \mathfrak{s})$ is a module over the ring $H^*_{S^1 \times G} = H^*_{S^1 \times G}( pt ; \mathbb{F})$. If $p=2$, then $H^*_{S^1 \times G} \cong \mathbb{F}[U,Q]$, where $deg(U) = 2$, $deg(Q) = 1$. If $p$ is odd, then $H^*_{S^1 \times G} \cong \mathbb{F}[U,R,S]/(R^2)$, where $deg(U) = 2$, $deg(R)=1$, $deg(S)=2$. As in the non-equivariant case, the grading on $HSW^*_G(Y , \mathfrak{s})$ can in general take rational values. However, if $Y$ is an integral homology sphere then the grading is integer-valued.
The localisation theorem in equivariant cohomology implies that the localisation $U^{-1} HSW^*_G(Y , \mathfrak{s})$ is a free $U^{-1}H^*_{S^1 \times G}$-module of rank $1$. Letting $\mu$ denote a generator of $U^{-1} HSW^*_G(Y , \mathfrak{s})$, we have an isomorphism of the form
\[
\iota \colon U^{-1}HSW^*_G(Y , \mathfrak{s}) \to \mathbb{F}[Q , U , U^{-1}]\mu
\]
for $p=2$ and
\[
\iota \colon U^{-1}HSW^*_G(Y , \mathfrak{s}) \to \frac{\mathbb{F}[R , S , U , U^{-1}]}{(R^2)} \mu
\]
for $p$ odd. Following \cite[\textsection 3]{bh} we define a sequence of equivariant $\delta$-invariants as follows. The cases $p=2$ and $p \neq 2$ need to be treated separately. First suppose $p=2$. For each $j \ge 0$, we define $\delta^{(p)}_{j}(Y , \mathfrak{s} , \tau)$ to be $i/2 - j/2$, where $i$ is the least degree for which there exists an element $x \in HSW^i_G(Y , \mathfrak{s})$ and a $k \in \mathbb{Z}$ such that
\[
\iota x = Q^{j} U^k \tau \; ({\rm mod} \; Q^{j+1}).
\]
If $p \neq 2$, then for each $j \ge 0$, we define $\delta_{j}^{(p)}(Y , \mathfrak{s} , \tau)$ to be $i/2 - j$, where $i$ is the least degree for which there exists an element $x \in HSW^i_G(Y , \mathfrak{s})$ and a $k \in \mathbb{Z}$ such that
\[
\iota x = S^{j} U^k \tau \; ({\rm mod} \; S^{j+1}, RS^j).
\]
When the diffeomorphism $\tau$ is understood we will omit it from the notation and simply write the delta invariants as $\delta_j^{(p)}(Y , \mathfrak{s})$.
Various properties of the $\delta$-invariants are shown in \cite{bh}. In particular, we have:
\begin{itemize}
\item[(1)]{$\delta^{(p)}_{0}(Y,\mathfrak{s}) \ge \delta(Y,\mathfrak{s})$, where $\delta(Y,\mathfrak{s}) = d(Y,\mathfrak{s})/2$ and $d(Y,\mathfrak{s})$ is the Ozsv\'ath--Szab\'o $d$-invariant.}
\item[(2)]{$\delta_{j+1}^{(p)}(Y,\mathfrak{s}) \le \delta_{j}^{(p)}(Y , \mathfrak{s})$ for all $j \ge 0$.}
\item[(3)]{The sequence $\{ \delta_{j}^{(p)}(Y,\mathfrak{s}) \}_{j \ge 0}$ is eventually constant.}
\end{itemize}
Using property (3), we may define two additional invariants of $(Y , \mathfrak{s} , \tau)$ as follows. We define $\delta_\infty^{(p)}(Y , \mathfrak{s},\tau) = \lim_{j \to \infty} \delta_j^{(p)}(Y , \mathfrak{s},\tau)$ and we define $j^{(p)}(Y , \mathfrak{s},\tau)$ to be the smallest $j$ such that $\delta_j^{(p)}(Y , \mathfrak{s},\tau) = \delta_\infty^{(p)}(Y , \mathfrak{s},\tau)$. If $\tau$ is understood we will simply write $\delta_\infty^{(p)}(Y , \mathfrak{s})$ and $j^{(p)}(Y,\mathfrak{s})$.
\subsection{Knot concordance invariants}\label{sec:kci}
Given a knot $K \subset S^3$ and a prime number $p$, we let $Y = \Sigma_p(K)$ denote the degree $p$ cyclic cover of $S^3$ branched over $K$. Then $Y$ is a rational homology $3$-sphere \cite[Corollary 3.2]{liv} and it comes equipped with a natural $\mathbb{Z}_p$-action. Let $\pi \colon Y \to S^3$ denote the covering map. From \cite[Corollary 2.2]{jab}, any spin$^c$-structure on $Y \setminus \pi^{-1}(K)$ uniquely extends to $Y$. Then since $H^2( S^3 \setminus K ; \mathbb{Z}) = 0$, there is a unique spin$^c$-structure on $S^3 \setminus K$. The pullback of this spin$^c$-structure under $\pi$ extends uniquely to a spin$^c$-structure on $Y$. Following \cite{jab}, we denote this spin$^c$-structure by $\mathfrak{s}_0 = \mathfrak{s}_0(K , p)$. Uniqueness of the extension implies that $\mathfrak{s}_0$ is $\mathbb{Z}_p$-invariant. Thus for any prime $p$ and any $j \ge 0$, we obtain a knot invariant $\delta^{(p)}_j(K)$ by setting
\[
\delta_j^{(p)}(K) = 4 \delta_j^{(p)}( \Sigma_p( K ) , \mathfrak{s}_0 ).
\]
The invariants $\delta^{(p)}_j(K)$ are knot concordance invariants and they satisfy a number of properties \cite{bh}, \cite{bar}. In particular, for $p=2$ we have:
\begin{itemize}
\item[(1)]{$\delta^{(2)}_0(K) \ge \delta^{(2)}(K)$, where $\delta^{(2)}(K)$ is the Manolescu--Owens invariant \cite{mo}.}
\item[(2)]{$\delta^{(2)}_{j+1}(K) \le \delta^{(2)}_j(K)$ for all $j \ge 0$.}
\item[(3)]{$\delta^{(2)}_j(K) \ge -\sigma(K)/2$ for all $j \ge 0$ and $\delta^{(2)}_j(K) = -\sigma(K)/2$ for $j \ge g_4(K)-\sigma(K)/2$.}
\end{itemize}
Here $g_4(K)$ is the slice genus of $K$ and $\sigma(K)$ is the signature. Similarly for $p \neq 2$, we have:
\begin{itemize}
\item[(1)]{$\delta^{(p)}_0(K) \ge \delta^{(p)}(K)$, where $\delta^{(p)}(K)$ is the Jabuka invariant \cite{jab}.}
\item[(2)]{$\delta^{(p)}_{j+1}(K) \le \delta^{(p)}_j(K)$ for all $j \ge 0$.}
\item[(3)]{$\delta^{(p)}_j(K) \ge -\sigma^{(p)}(K)/2$ for all $j \ge 0$ and $\delta^{(p)}_j(K) = -\sigma^{(p)}(K)/2$ for $2j \ge (p-1)g_4(K)-\sigma^{(p)}(K)/2$.}
\end{itemize}
Here $\sigma^{(p)}(K)$ is defined as
\[
\sigma^{(p)}(K) = \sum_{j=1}^{p-1} \sigma_{K}( e^{2\pi i j/p}),
\]
where $\sigma_K(\omega)$ is the Levine--Tristram signature of $K$. Define $j^{(p)}(K)$ to be the smallest $j$ such that $\delta^{(p)}_j(K)$ attains its minimum. Thus
\[
j^{(p)}(K) = j^{(p)}( \Sigma_p(K) , \mathfrak{s}_0).
\]
From property (3), we see that the minimum value of $\delta^{(2)}_j(K)$ is precisely $-\sigma(K)/2$ and that
\[
j^{(2)}(K) \le g_4(K) - \sigma(K)/2.
\]
This can be re-arranged into a lower bound for the slice genus:
\[
g_4(K) \ge j^{(2)}(K) + \frac{\sigma(K)}{2}.
\]
Similarly for $p \neq 2$, the minimum value of $\delta^{(p)}_j(K)$ is $-\sigma^{(p)}(K)/2$ and that
\[
2 j^{(p)}(K) \le (p-1)g_4(K) - \frac{\sigma^{(p)}(K)}{2}.
\]
This can be re-arranged to
\[
g_4(K) \ge \frac{2 j^{(p)}(K)}{(p-1)} + \frac{\sigma^{(p)}(K)}{2(p-1)}.
\]
Replacing $K$ by its mirror $-K$, we obtain the following slice genus bounds:
\[
g_4(K) \ge j^{(2)}(-K) - \frac{\sigma(K)}{2}
\]
for $p=2$, and
\[
g_4(K) \ge \frac{2 j^{(p)}(-K)}{(p-1)} - \frac{\sigma^{(p)}(K)}{2(p-1)}
\]
for $p \neq 2$. Following \cite{bar}, we define knot concordance invariants $\theta^{(p)}(K)$ for all primes $p$ by setting
\[
\theta^{(2)}(K) = \max\left\{ 0 , j^{(2)}(-K) - \frac{\sigma(K)}{2} \right\}
\]
for $p=2$, and
\[
\theta^{(p)}(K) = \max\left\{ 0 , \frac{2 j^{(p)}(-K)}{(p-1)} - \frac{\sigma^{(p)}(K)}{2(p-1)} \right\}
\]
for $p \neq 2$. Then we have the slice genus bounds $g_4(K) \ge \theta^{(p)}(K)$ for all $p$.
\section{Further properties of the delta invariants}\label{sec:fdelta}
In this section we will use the spectral sequence of \cite{bh} relating equivariant and non-equivariant Seiberg--Witten--Floer cohomologies in order to deduce more precise information on the $\delta$-invariants. In particular, we will be able to determine the value of $j^{(p)}(Y)$ under certain assumptions on the Floer homology of $Y$.
\subsection{Spectral sequence}\label{sec:ss}
Let $Y$ be a rational homology $3$-sphere and $\tau \colon Y \to Y$ an orientation preserving diffeomorphism of prime order $p$. Let $\mathfrak{s}$ be a spin$^c$-structure preserved by $\tau$. Let the coefficient group be $\mathbb{F} = \mathbb{Z}_p$ and let $G = \langle \tau \rangle = \mathbb{Z}_p$.
Recall that $H^*_G$ is isomorphic to $\mathbb{F}[Q]$, where $deg(Q) = 1$ if $p=2$ and is isomorphic to $\mathbb{F}[R,S]/(R^2)$ where $deg(R) = 1$, $deg(S) = 2$ if $p$ is odd. Thus in either case we have $H^{ev}_G \cong \mathbb{F}[S]$, where in the $p=2$ case we set $S = Q^2$.
Let $\{ E_r^{p,q} , d_r \}$ denote the spectral sequence for the $G$-equivariant Seiberg--Witten--Floer cohomology $HSW^*_{G}(Y , \mathfrak{s})$ \cite[Theorem 3.2]{bh}. In detail, this means there is a filtration $\{ \mathcal{F}_j^* \}_{j \ge 0}$ on $HSW^*_G(Y , \mathfrak{s})$ such that $E_\infty$ is isomorphic to the associated graded group
\[
E_\infty \cong Gr( HSW^*_{G}(Y , \mathfrak{s}) ) = \bigoplus_{j \ge 0} \mathcal{F}^*_j/\mathcal{F}^*_{j+1}
\]
as $H^*_{S^1 \times G}$-modules. In terms of bigrading this means that $E^{p,q}_r \cong \mathcal{F}_p^{p+q}/\mathcal{F}_{p+1}^{p+q}$. Furthermore, we have that
\[
E_2^{p,q} = H^p( \mathbb{Z}_p ; HSW^q(Y , \mathfrak{s})).
\]
\subsection{Behaviour of the delta invariants}\label{sec:beh}
Let $(Y , \mathfrak{s} , \tau)$ be as in Section \ref{sec:ss}. We will examine the spectral sequence $\{ E_r^{p,q} \}$ to deduce some properties of the delta invariants $\{ \delta_j^{(p)}(Y,\mathfrak{s}) \}$.
To simplify notation, we will set $H^i = HSW^i(Y , \mathfrak{s})$ and $\widehat{H}^i = HSW^i_G(Y , \mathfrak{s})$. Let $d = d(Y , \mathfrak{s})$ and $\delta = d/2$. We will make the assumption that $H^i = 0$ unless $i = d \; ({\rm mod} \; 2)$. This assumption is satisfied if $Y$ is a Seifert homology sphere oriented so that $-Y$ is the boundary of a negative definite plumbing.
\begin{lemma}\label{lem:even}
If $p=2$ and if $HSW^*(Y , \mathfrak{s})$ is concentrated in degrees equal to $d(Y,\mathfrak{s})$ mod $2$, then for all $j \ge 0$ we have $\delta_{2j+1}^{(2)}(Y , \mathfrak{s}) = \delta_{2j}^{(2)}(Y ,\mathfrak{s})$.
\end{lemma}
\begin{proof}
Adapting the proof of \cite[Lemma 5.7]{bh} to the $p=2$ case, one can show that $Q \colon E_\infty^{p,q} \to E_\infty^{p+1,q}$ is surjective for all $p,q$. This implies that $Q \colon \mathcal{F}_j \to \mathcal{F}_{j+1}$ is surjective. Now let $x \in HSW_{\mathbb{Z}_2}^i( Y , \mathfrak{s})$ satisfy $\iota x = Q^{2j+1} U^k \mu \; ({\rm mod} \; Q^{2j+2})$ for some $k$ and assume that $x$ has the minimal possible degree, so $$\delta_{2j+1}^{(2)}(Y , \mathfrak{s}) = deg(x)/2 - (2j+1)/2 = i/2 - j - 1/2.$$ Note that since $deg(Q) = 1$, $deg(U) = 2$, we have that $deg(x) = 1 + d(Y,\mathfrak{s}) \; ({\rm mod} \; 2)$. This means that the image of $x$ in $HSW_{\mathbb{Z}_2}^i( Y , \mathfrak{s})/\mathcal{F}_1 = E_\infty^{0,q}$ is zero, because $E_\infty^{0,q} \subseteq E_2^{0,q} = H^0( \mathbb{Z}_2 ; HSW^q(Y , \mathfrak{s}))$ is concentrated in degrees equal to $d(Y , \mathfrak{s})$ mod $2$. Hence $x \in \mathcal{F}_1$. This implies that $x = Qy$ for some $y \in HSW^{i-1}_{\mathbb{Z}_2}(Y , \mathfrak{s})$. Then since $x =Qy$, we have
\[
\iota x = Q \iota y = Q^{2j+1} U^k \mu \; ({\rm mod} \; Q^{2j+2}).
\]
Moreover, $Q$ is injective on $U^{-1} HSW_{\mathbb{Z}_2}^*(Y , \mathfrak{s}) \cong \mathbb{F}[Q , U , U^{-1}]$, so we deduce that
\[
\iota y = Q^{2j} U^k \mu \; ({\rm mod} \; Q^{2j+1}).
\]
Therefore, from the definition of $\delta_{2j}^{(2)}(Y , \mathfrak{s})$, it follows that
\[
\delta_{2j}^{(2)}(Y , \mathfrak{s}) \le deg(y)/2 - j = deg(x)/2 - 1/2 - j = \delta_{2j+1}^{(2)}(Y , \mathfrak{s}).
\]
But we always have $\delta_{2j+1}^{(2)}(Y , \mathfrak{s}) \le \delta_{2j}^{(2)}(Y , \mathfrak{s})$, so we must have an equality $\delta_{2j}^{(2)}(Y , \mathfrak{s}) = \delta_{2j+1}^{(2)}(Y , \mathfrak{s})$.
\end{proof}
\begin{remark}\label{rem:j2}
By Lemma \ref{lem:even}, if $p=2$ and $HSW^*(Y , \mathfrak{s})$ is concentrated in degrees equal to $d(Y,\mathfrak{s})$ mod $2$, then $j^{(2)}(Y , \mathfrak{s})$ is even.
\end{remark}
In what follows, we find it more convenient to consider not the whole equivariant Seiberg--Witten--Floer cohomology but only the part concentrated in degrees equal to $d(Y , \mathfrak{s})$ mod $2$. For this reason we consider even counterparts of the relevant ingredients, such as the spectral sequence and filtration. One advantage is that it allows us to treat the $p=2$ and $p \neq 2$ cases simultaneously.
Let $H^{ev}_{S^1 \times G}$ denote the subring of $H^*_{S^1 \times G}$ given by elements of even degree. Then $H^{ev}_{S^1 \times G} \cong \mathbb{F}[U,S]$. Let $\widehat{H}^{ev}$ denote the $H^{ev}_{S^1 \times G}$-submodule of $\widehat{H}^*$ given by elements of degree equal to $d$ mod $2$. The filtration $\{ \mathcal{F}_j^* \}$ defines a filtration $\{ \mathcal{F}_j^{ev} \}$ on $\widehat{H}^{ev}$, where $\mathcal{F}_j^{ev} = \mathcal{F}_{2j} \cap \widehat{H}^{ev}$. The associated graded $H^{ev}_{S^1 \times G}$-module of this filtration is isomorphic to
\[
E_\infty^{ev} = \bigoplus_{p,q} E_\infty^{2p,q}.
\]
By \cite[Lemma 5.7]{bh}, the map $S \colon E_\infty^{2p,q} \to E_\infty^{2p+2 , q}$ is surjective for all $p,q$. It follows that $S \colon \mathcal{F}^{ev}_j \to \mathcal{F}^{ev}_{j+1}$ is surjective for each $j \ge 0$. In particular, since $\mathcal{F}^{ev}_0 = \widehat{H}^{ev}$, we see that $\mathcal{F}^{ev}_j = S^j \widehat{H}^{ev}$.
Recall the localisation isomorphism
\[
\iota \colon U^{-1} \widehat{H}^* \cong H^*_G[U , U^{-1}] \mu
\]
for some $\mu$. Restricting to $\widehat{H}^{ev}$, we get a localisation isomorphism
\[
\iota \colon U^{-1} \widehat{H}^{ev} \cong H^{ev}_G[U,U^{-1}]\mu \cong \mathbb{F}[S,U,U^{-1}]\mu.
\]
Let $\delta_j$ denote $\delta_j^{(p)}(Y,\mathfrak{s})$ if $p$ is odd and $\delta_{2j}^{(2)}(Y , \mathfrak{s})$ if $p=2$. Since $S = Q^2$ in the $p=2$ case, it follows that regardless of whether $p$ is even or odd, we have $\delta_j = a/2 - j$, where $a$ is the least degree such that there exists an element $x \in \widehat{H}^{a}$ with
\[
\iota x = S^j U^k \mu \; ({\rm mod} \; S^{j+1})
\]
for some $k \in \mathbb{Z}$.
Observe that $E_2^{0,q} = H^0( \mathbb{Z}_p ; H^q )$ is a submodule of $H^q$. Then since $E_{r+1}^{0,q}$ is a submodule of $E_{r}^{0,q}$ for each $r \ge 2$, we see that $E_\infty^{0,*} = \mathcal{F}^{ev}_0/\mathcal{F}^{ev}_1$ can be identified with an $\mathbb{F}[U]$-submodule of $H^*$. We will denote this submodule by $J^* \subset H^*$. Now since $H^* \cong \mathbb{F}[U] \lambda \oplus H^*_{red}$, where $\lambda$ has degree $d$ and $H^*_{red} = \{ x \in H^* \; | \; U^m x = 0 \text{ for some } m \ge 0\}$. It follows that $J^* \cong \mathbb{F}[U]\theta \oplus J^*_{red}$, where $\theta$ has degree at least $d$ and $J^*_{red} \subseteq H^*_{red}$. Since $U^m \theta \neq 0$ for all $m \ge 0$, we see that the image of $\theta$ under the localisation map $\iota$ is non-zero. More precisely,
\[
\iota \theta = U^{\alpha} \mu \; ({\rm mod} \; S )
\]
for some $\alpha \in \mathbb{Z}$. On the other hand, any $x \in J^*$ has $U^m x = 0 \; ({\rm mod} \; \mathcal{F}_1^{ev})$ for some $m \ge 0$. This means that $U^m x = Sy$ for some $y$ and thus $\iota x$ is a multiple of $S$. It follows that $deg(\theta) = \delta_0$.
By \cite[Proposition 3.14]{bh}, we then have that $\delta_j = a/2 - j$, where $a$ is the least degree such that there exists an element $x \in \widehat{H}^a$ such that
\[
U^m x = S^j U^k \theta \; ({\rm mod} \; \mathcal{F}_{j+1}^{ev})
\]
for some $m, k \ge 0$.
Consider again the sequence $\delta_0 \ge \delta_1 \ge \delta_2 \ge \cdots$. This sequence is eventually constant, hence there exists a finite set of indices $0 < j_1 < j_2 < \cdots < j_r$ such that for each $j > 0$, $\delta_j < \delta_{j-1}$ if and only if $j = j_i$ for some $i$. We let $n_i = \delta_{j_{i-1}} - \delta_{j_i} > 0$ for $1 < i \le r$ and $n_1 = \delta_{j_1} - \delta_0$ for $i=1$. Thus $\delta_{j_i} = \delta_0 - (n_1 + \cdots + n_i )$.
\begin{lemma}\label{lem:indep}
There exists $x_1, \dots , x_r \in J^*_{red}$ such that:
\begin{itemize}
\item[(1)]{The set $\{ U^a x_i \; | \; 1 \le i \le r, 0 \le a \le n_i -1 \}$ is linearly independent over $\mathbb{F}$. In particular, $U^a x_i \neq 0$ for $a < n_i$.}
\item[(2)]{We have $\delta_{j_i} = deg(x_i)/2 - j_i$.}
\item[(3)]{Each $x_i$ has a lift to an element $y_i \in \widehat{H}^i$ satisfying
\[
U^{m_i} y_i = S^{j_i} U^{k_i} \theta \; ({\rm mod} \; \mathcal{F}_{j_i+1}^{ev})
\]
for some $m_i, k_i \ge 0$.}
\end{itemize}
\end{lemma}
\begin{proof}
By the definition of $\delta_{j_i}$, there exists $y_i \in \widehat{H}^*$ such that $\delta_{j_i} = deg(y_i)/2 - j_i$ and $U^{m_i} y_i = S^{j_i} U^{k_i} \theta \; ({\rm mod} \; \mathcal{F}_{j_i+1}^{ev})$ for some $m_i,k_i$. Let $x_i \in H^*$ be the image of $y_i$ under the natural map $\widehat{H}^{ev} \to \widehat{H}^{ev}/\mathcal{F}_1^{ev} \cong J^*$. We claim that $x_i \in J^*_{red}$. If not, then $x_i = c U^k \theta + w$ for some $c \in \mathbb{F}^*$, $k \ge 0$ and $w \in J^*_{red}$. Then $U^m x_i = c U^{k+m} \theta$ for some $m, k \ge 0$ and hence $U^m y_i = c U^{k+m} \theta \; ({\rm mod} \; \mathcal{F}^{ev}_1)$. This contradicts $U^{m_i} y_i = S^{j_i} U^{k_i} \theta \; ({\rm mod} \; \mathcal{F}_{j_i+1}^{ev})$ as $S^{j_i} U^{k_i} \theta = 0 \; ({\rm mod} \; \mathcal{F}^{ev}_1)$ (since $j_i > 0$). So $x_i \in J^*_{red}$. It is now evident that (2) and (3) are satisfied. We claim that (1) also holds, that is, the set $\{ U^a x_i \; | \; 1 \le i \le r, 0 \le a \le n_i -1 \}$ is linearly independent over $\mathbb{F}$.
Suppose we have a non-trivial linear relation amongst elements of $\mathcal{B} = \{ U^a x_i \; | \; 1 \le i \le r, 0 \le a \le n_i -1 \}$ of degree $s$. So
\[
\sum_i c_i U^{\beta_i} x_i = 0
\]
for some $c_i \in \mathbb{F}$, not all zero and $\beta_i = (s - deg(x_i) )/2$. Furthermore, the sum is restricted to those $i$ such that $0 \le \beta_i \le n_i - 1$ (so that $U^{\beta_i} x_i$ belongs to $\mathcal{B}$). Lifting this to $\widehat{H}^{ev}$, we have $\sum_i c_i U^{\beta_i} y_i \in \mathcal{F}^{ev}_1$ and hence
\begin{equation}\label{equ:w}
\sum_i c_i U^{\beta_i} y_i = Sw
\end{equation}
for some $w \in \widehat{H}^{s-2}$. Let $u$ be the smallest value of $i$ such that $c_u \neq 0$. Multiplying by a sufficiently large power of $U$ and taking the result modulo $\mathcal{F}^{ev}_{j_u+1}$, we see that
\[
S U^m w = c_i S^{j_u} U^{b} \theta \; ({\rm mod} \; \mathcal{F}^{ev}_{j_u+1})
\]
for some $m,b \ge 0$. Hence $S \iota (U^m w) = S \iota (c_i S^{j_u-1} U^b \theta) \; ({\rm mod} \; S^{j_u+1})$, where $\iota$ is the localisation map. Since $S$ is injective on $\mathbb{F}[S,U,U^{-1}]\mu$, we also have $\iota w = c_i S^{j_u-1} U^{b-m} \iota \theta \; ({\rm mod} \; S^{j_u})$. Now we recall that $\iota \theta = U^\alpha \mu \; ({\rm mod} \; S)$ for some $\alpha$, hence $\iota (c_i^{-1} w ) = S^{j_u-1} U^{b-m+\alpha} \mu \; ({\rm mod} \; S^{j_u})$. By the definition of $\delta_{j_u-1}$, we must therefore have $deg(w)/2 \ge \delta_{j_u-1} + (j_u-1)$. But $deg(w) = s-2$, so $s \ge \delta_{j_u-1} + j_u$. On the other hand, equating degrees in Equation (\ref{equ:w}) gives $s = \beta_u + deg(x_u)/2 = \beta_u + \delta_{j_u} + j_u$. This gives
\[
\beta_u = s - \delta_{j_u} - j_u \ge \delta_{j_u-1} - \delta_{j_u} = n_u.
\]
But this contradicts $\beta_u \le n_u-1$. So no such linear relation can exist.
\end{proof}
\begin{corollary}\label{cor:ineq}
We have that $\delta_0 - \delta_j \le dim_{\mathbb{F}}( J_{red}^*) \le dim_{\mathbb{F}}( HF_{red}^*(Y , \mathfrak{s}) )$ for all $j \ge 0$.
\end{corollary}
\begin{proof}
Since $\delta_j = \delta_{j_r}$ for all $j \ge j_r$, we have that $\delta_0 - \delta_j \le \delta_0 - \delta_{j_r}$ for all $j$. But the subspace $span_{\mathbb{F}}\{ U^a x_i \; | \; 1 \le i \le r, 0 \le a \le n_i -1 \} \subseteq J^*$ of Lemma \ref{lem:indep} has dimension $n_1 + \cdots + n_r = \delta_0 - \delta_{j_r}$.
\end{proof}
Let $j'(Y,\mathfrak{s})$ denote the smallest $j$ such that $\delta_j$ attains its minimum. Thus $j'(Y,\mathfrak{s}) = j^{(p)}(Y,\mathfrak{s})$ if $p$ is odd and $j'(Y,\mathfrak{s}) = j^{(2)}(Y,\mathfrak{s})/2$ if $p=2$, by Remark \ref{rem:j2}.
\begin{proposition}\label{prop:jinv}
Let $Y$ be a rational homology $3$-sphere, $\tau \colon Y \to Y$ an orientation preserving diffeomorphism of order $p$, where $p$ is prime. Let $\mathfrak{s}$ be a spin$^c$-structure preserved by $\tau$. Suppose that the following conditions hold:
\begin{itemize}
\item[(1)]{$HF^+(Y , \mathfrak{s})$ is non-zero only in degrees $i = d(Y , \mathfrak{s}) \; ({\rm mod} \; 2)$.}
\item[(2)]{$\delta_0 - \delta_j = dim_{\mathbb{F}}( HF_{red}^*(Y , \mathfrak{s}) )$ for some $j \ge 0$.}
\item[(3)]{$HF_{red}^*(Y , \mathfrak{s}) \neq 0$.}
\item[(4)]{Let $\ell^+(Y,\mathfrak{s})$ denote the highest non-zero degree in $HF^+_{red}(Y,\mathfrak{s})$. Any non-zero element in the image of $U \colon HF_{red}^+(Y , \mathfrak{s}) \to HF_{red}^+(Y , \mathfrak{s})$ has degree strictly less than $\ell^+(Y,\mathfrak{s})$.}
\end{itemize}
Then
\[
j'(Y,\mathfrak{s}) = \ell^+(Y,\mathfrak{s})/2 - \delta_0(Y,\mathfrak{s}) + dim_{\mathbb{F}}( HF_{red}^*(Y , \mathfrak{s}) ).
\]
Moreover, if $Y$ is an integral homology sphere and $\delta(Y) \in 2 \mathbb{Z}$, then
\[
j'(Y) = \ell^+(Y)/2 - \delta_0(Y) + \delta(Y) + \lambda(Y),
\]
where $\lambda(Y)$ is the Casson invariant of $Y$.
\end{proposition}
\begin{proof}
Let $x_1, \dots , x_r \in J_{red}^*$ be as in Lemma \ref{lem:indep}. It follows that $j'(Y,\mathfrak{s}) = j_r$. Choose $j$ such that $\delta_0 - \delta_j = dim_{\mathbb{F}}( HF_{red}^*(Y , \mathfrak{s}) )$. From Corollary \ref{cor:ineq}, we have
\[
dim_{\mathbb{F}}( HF_{red}^*(Y , \mathfrak{s}) ) = \delta_0 - \delta_j \le \delta_0 - \delta_{j_r} = n_1 + \cdots + n_r \le dim_{\mathbb{F}}( J_{red}^*) \le dim_{\mathbb{F}}( HF_{red}^*(Y , \mathfrak{s}) ).
\]
It follows that we must have equalities throughout. So $\delta_0 - \delta_{j_r} = n_1 + \cdots + n_r = dim_{\mathbb{F}}( HF_{red}^*(Y , \mathfrak{s}) )$. So $\{ U^a x_i \; | \; 1 \le i \le r, 0 \le a \le n_i -1 \}$ is a basis for $HF_{red}^*(Y,\mathfrak{s})$. Set $\alpha_i = deg(x_i)/2$. Lemma \ref{lem:indep} (2) implies that $\delta_{j_i} = \alpha_i - j_i$ for $1 \le i \le r$. Setting $\alpha_0 = \delta_0$ and $j_0 = 0$, we also have $\delta_{j_i} = \alpha_i - j_i$ for $i=0$. Further, we have that $\delta_{j_i} - \delta_{j_{i-1}} = n_i$ for $1 \le i \le r$. Hence we obtain $(\alpha_i + n_i - 1) - \alpha_{i-1} = j_i - j_{i-1} - 1$. But $j_i > j_{i-1}$, so $j_i - j_{i-1} - 1 \ge 0$, giving
\begin{equation}\label{equ:alphain}
(\alpha_i + n_i - 1) \ge \alpha_{i-1}.
\end{equation}
Now we observe that $2(\alpha_i + n_i - 1)$ is the degree of $U^{n_i-1} x_i$ and $2\alpha_{i-1}$ is the degree of $\alpha_{i-1}$. Consider the space $A = HF^{\ell^+(Y,\mathfrak{s})}_{red}(Y , \mathfrak{s})$ of top degree elements in $HF^+_{red}(Y , \mathfrak{s})$. Let $a = dim_{\mathbb{F}}(A) > 0$. By assumption (4), no non-zero element of $A$ is in the image of $U$. Hence there must exist indices $k_1 < k_2 < \dots < k_a$ such that $x_{k_1}, \dots , x_{k_a}$ is a basis for $A$. In fact, $K = \{k_1, \dots , k_a\}$ is precisely the set of indices $i$ such that $2\alpha_i = \ell^+(Y , \mathfrak{s})$. Suppose $i \in K$ and $i < r$. Then (\ref{equ:alphain}) gives $2(\alpha_{i+1} + n_{i+1} - 1) \ge 2\alpha_{i} = \ell^+(Y,\mathfrak{s})$. But $2(\alpha_{i+1} + n_{i+1} - 1)$ is the degree of $U^{n_{i+1}-1} x_{i+1}$ and $2 \ell^+(Y , \mathfrak{s})$ is the highest degree in $HF^+_{red}(Y,\mathfrak{s})$. So we must have $2(\alpha_{i+1} + n_{i+1} - 1) = \ell^+(Y,\mathfrak{s})$. But from assumption (4), this can only happen if $n_{i+1} = 1$, hence $2\alpha_{i+1} = \ell^+(Y , \mathfrak{s})$ and so $i+1 \in K$. So if $i \in K$ and $i < r$, then $i+1 \in K$. It follows that $K = \{ r-a+1 , r-a +2 , \dots , r\}$. In particular, $r \in K$ and $\alpha_r = \ell^+(Y,\mathfrak{s})/2$. Therefore
\[
j'(Y , \mathfrak{s}) = j_r = \alpha_r - \delta_{j_r} = \ell^+(Y , \mathfrak{s})/2 - \delta_0(Y , \mathfrak{s}) + dim_{\mathbb{F}}( HF_{red}^*(Y , \mathfrak{s}) ),
\]
where the last equality holds since $\delta_0 - \delta_{j_r} = dim_{\mathbb{F}}( HF_{red}^*(Y , \mathfrak{s}) )$.
Lastly, suppose that $Y$ is an integral homology sphere and $\delta(Y) \in 2 \mathbb{Z}$. From assumption (1), it follows that $HF_{red}^+(Y)$ is non-zero only in even degrees. Then from \cite[Theorem 1.3]{os}, it follows that $ dim_{\mathbb{F}}( HF_{red}^*(Y)) = \delta(Y) + \lambda(Y)$, where $\lambda(Y)$ is the Casson invariant. Hence we have that $j'(Y) = \ell^+(Y)/2 - \delta_0(Y) + \delta(Y) + \lambda(Y)$.
\end{proof}
\subsection{Delta invariants of Brieskorn spheres}\label{sec:deltabri}
Given pairwise coprime integers $a_1 , \dots , a_r$ with $r \ge 3$ define the Brieskorn homology sphere $\Sigma(a_1 , \dots , a_r)$ to be the link of the singularity at the origin of the variey $\{ (z_1 , \dots , z_r ) \in \mathbb{C}^n \; | \; b_{i1} z_1^{a_1} + b_{i2}z_2^{a_2} + \cdots + b_{ir}z_r^{a_r} = 0, \; 1 \le i \le r-2 \}$, where $( b_{ij} )$ is a sufficiently generic $(r-2) \times r$ matrix. Then $\Sigma(a_1 , a_2 , \dots , a_r)$ is the Seifert homology sphere $M(e_0 , (a_1,b_1) , \dots , (a_r , b_r) )$, where $e_0 , b_1, \dots , b_r$ are uniquely determined by the conditions that $0 < b_j < a_j$ and
\begin{equation}\label{equ:integral}
e_0 + \sum_{j=1}^r \frac{b_j}{a_j} = -\frac{1}{a_1 a_2 \cdots a_r}.
\end{equation}
Note that $e_0 < 0 $ because $b_j > 0$ for all $j$.
Let $Y = \Sigma(a_1 , \dots , a_r)$ be a Brieskorn homology sphere. The Seifert structure of $Y$ gives a circle action. For any prime $p$, the restriction of the circle action to $\mathbb{Z}_p \subset S^1$ defines an action of the finite cyclic group $G = \mathbb{Z}_p$. We are interested in studying the equivariant Seiberg--Witten--Floer cohomology groups of $Y$ and $-Y$ with respect to this action. Since $Y$ has a unique spin$^c$-structure we will omit it from the notation and write $HSW^*_G(Y)$ and $HSW^*_G(-Y)$. Similarly the $\delta$-invariants will be denoted $\delta_j^{(p)}(Y)$ and $\delta_j^{(p)}(-Y)$.
To understand the equivariant Seiberg--Witten--Floer cohomology groups of $Y$ and $-Y$, we will use the spectral sequence relating it to the corresponding non-equivariant groups. We have
\begin{align*}
HSW^*(Y) \cong HF^+(Y) & \cong \mathbb{F}[U]_{d(Y)} \oplus HF_{red}^+(Y) \\
HSW^*(-Y) \cong HF^+(-Y) & \cong \mathbb{F}[U]_{-d(Y)} \oplus HF_{red}^+(-Y)
\end{align*}
From \cite{os2}, we have that $HF^+(-Y)$ is concentrated in even degrees. It follows that $d(Y)$ is even and $HF_{red}^+(Y)$ is concentrated in {\em odd} degrees. Furthermore $HF^+(-Y)$ can be computed from the graded roots algorithm \cite{nem}. This algorithm implies that $HF_{red}^+(-Y)$ is concentrated in degrees at least $-d(Y)$. Dually this implies that $HF_{red}^+(Y)$ is concentrated in degrees at most $d(Y)-1$.
\begin{proposition}\label{prop:deltaprop}
Let $Y = \Sigma(a_1, \dots , a_r)$ be a Brieskorn homology sphere. We have:
\begin{itemize}
\item[(1)]{${\rm rk}( HF_{red}^+(Y) ) = -\delta(Y) - \lambda(Y)$ where $\lambda(Y)$ is the Casson invariant of $Y$.}
\item[(2)]{$\delta_j^{(p)}(Y) = \delta_\infty^{(p)}(Y)$ for all $j \ge 0$.}
\item[(3)]{$\delta(Y) \le \delta_\infty^{(p)}(Y) \le -\lambda(Y)$ and $\lambda(Y) \le \delta_j^{(p)}(-Y) \le -\delta(Y)$ for all $j \ge 0$.}
\end{itemize}
\end{proposition}
\begin{proof}
From \cite[Theorem 1.3]{os} we have that $\chi( HF_{red}^+(Y) ) = \delta(Y) + \lambda(Y)$. But $HF_{red}^+(Y)$ is concentrated in odd degrees, which gives (1).
We will prove (2) and (3) in the case that $p$ is odd. The case $p=2$ is similar and omitted for brevity. Recall that for odd primes we have $H^*_{\mathbb{Z}_p} \cong \mathbb{F}[R,S]/(R^2)$ with $deg(R)=1$, $deg(S) = 2$.
As in Section \ref{sec:ss}, let $\{ E_r^{p,q}(Y) , d_r \}$ denote the spectral sequence for the equivariant Seiberg--Witten--Floer cohomology $HSW^*_{\mathbb{Z}_p}(Y)$. We have a filtration $\{ \mathcal{F}_j^*(Y) \}_{j \ge 0}$ on $HSW^*_{\mathbb{Z}_p}(Y)$ such that $E_\infty(Y)$ is isomorphic to the associated graded group
\[
E_\infty(Y) \cong Gr( HSW^*_{\mathbb{Z}_p}(Y)) = \bigoplus_{j \ge 0} \mathcal{F}_j(Y)/\mathcal{F}_{j+1}(Y)
\]
as $H^*_{S^1 \times \mathbb{Z}_p}$-modules. Furthermore, we have that
\[
E_2^{p,q}(Y) = H^p( \mathbb{Z}_p ; HF^{+}_q(Y)).
\]
Now since the $\mathbb{Z}_p$-action on $Y$ is contained in a circle action, the generator of the $\mathbb{Z}_p$-action is smoothly isotopic to the identity and acts trivially on $HF^+(Y)$. Hence we further have
\[
E_2^{p,q}(Y) \cong HF^{+}_q(Y) \otimes_{\mathbb{F}} H^p_{\mathbb{Z}_p}.
\]
Choose an element $\theta \in HF^+_{d(Y)}(Y)$ such that $U^k \theta \neq 0$ for all $k \ge 0$. Then $HF^+(Y) \cong \mathbb{F}[U]\theta \oplus HF^+_{red}(Y)$ and hence
\begin{align*}
E_2(Y) &\cong H^*_{\mathbb{Z}_p}[U]\theta \oplus \left( HF_{red}^+(Y) \otimes_{\mathbb{F}} H^*_{\mathbb{Z}_p} \right) \\
& \cong \frac{\mathbb{F}[U,R,S]}{(R^2)} \theta \oplus \frac{HF_{red}^+(Y)[R,S]}{(R^2)}.
\end{align*}
We have a similar spectral sequence $E_r^{p,q}(-Y)$ and filtration $\{ \mathcal{F}_j(-Y) \}$ on $HSW^*_{\mathbb{Z}_p}(-Y)$ such that $E_\infty^{*,*}(-Y)$ is isomorphic to the associated graded $H^*_{S^1 \times \mathbb{Z}_p}$-module of the filtration. We have
\[
E_2(-Y) \cong \frac{\mathbb{F}[U,R,S]}{(R^2)} \omega \oplus \frac{HF_{red}^+(-Y)[R,S]}{(R^2)}
\]
for some $\omega$ of bi-degree $(0 , -d(Y))$. Now recall that $HF_{red}^+(-Y)$ is concentrated in degrees at least $-d(Y)$. Hence $d_r(\omega) = 0$ for all $r \ge 2$. It immediately follows that $\delta_0^{(p)}(-Y) = -\delta(Y)$ and hence $\delta_j^{(p)}(-Y) \le -\delta(Y)$ for all $j \ge 0$.
Define $E_r(Y)_{red}$ to be the subgroup of $E_r(Y)$ consisting of elements $x$ such that $U^k x = 0$ for some $k \ge 0$. So for $r = 2$ we have $E_2(Y)_{red} \cong HF_{red}^+(Y)[R,S]/(R^2)$ and
\[
E_2(Y) \cong \frac{\mathbb{F}[U,R,S]}{(R^2)} \theta \oplus E_2(Y)_{red}.
\]
As explained in \cite[\textsection 5]{bh}, the image of the differential $d_r$ is contained in $E_r(Y)_{red}$. It follows that for each $r \ge 2$, $E_r(Y)$ is of the form
\[
E_r(Y) \cong \frac{\mathbb{F}[U,R,S]}{(R^2)} U^{m_r}\theta \oplus E_r(Y)_{red}
\]
for some increasing sequence $0 = m_0 \le m_1 \le \dots $. It is further shown in \cite[\textsection 5]{bh} that the sequence $m_r$ must be eventually constant. Hence we have
\[
E_\infty(Y) \cong \frac{\mathbb{F}[U,R,S]}{(R^2)} U^{m}\theta \oplus E_\infty(Y)_{red}
\]
where $m = \lim_{r \to \infty} m_r$.
Since we work over a field $\mathbb{F}$, it is possible to (non-canonically) split the filtration $\{ \mathcal{F}_j(Y) \}$ giving an isomorphism of $\mathbb{F}$-vector spaces
\[
HSW^*_{\mathbb{Z}_p}(Y) \cong E_\infty^{*,*}(Y).
\]
Under this isomorphism $\mathcal{F}_j(Y)$ corresponds to the subspace of $E_\infty(Y)$ spanned by homogeneous elements of bi-degree $(p,q)$ where $p \ge j$. We fix a choice of such an isomorphism $\varphi \colon HSW^*_{\mathbb{Z}_p}(Y) \to E_\infty^{*,*}(Y)$ and henceforth identify $HSW^*_{\mathbb{Z}_p}(Y)$ with $E_\infty^{*,*}(Y)$. This isomorphism will typically not be an isomorphism of $H^*_{S^1 \times \mathbb{Z}_p}$-modules. Nevertheless we can use the isomorphim $\varphi$ to induce a new $H^*_{S^1 \times \mathbb{Z}_p}$-module action on $E_\infty^{*,*}(Y)$ corresponding to the one on $HSW^*_{\mathbb{Z}_p}(Y)$. To be precise, if $c \in H^*_{S^1 \times \mathbb{Z}_p}$, then we define $\widehat{c} \colon E_\infty \to E_\infty$ by $\widehat{c} \, x = \varphi( c \varphi^{-1}(x) )$. Consider in particular $\widehat{U}$. The action of $\widehat{U}$ will typically not respect the bi-grading, but we can decompose it into homogeneous components as
\[
\widehat{U} = U_{(0,2)} + U_{(1,1)} + U_{(2,0)} + U_{(3,-1)} + \cdots
\]
where $U_{j , 2-j} \colon E_\infty^{p,q}(Y) \to E_\infty^{p+j , 2-j}(Y)$. Note that there are only terms of bi-degree $(j,2-j)$ for $j \ge 0$ as $\widehat{U}$ respects the filtration $\{ \mathcal{F}_j(Y) \}$. Note also that $U_{(0,2)} = U$ because taking the associated graded module of the filtration recovers the original $H^*_{S^1 \times \mathbb{Z}_p}$-module structure on $E_\infty(Y)$.
Recall that $\delta_{j}^{(p)}(Y)$ is given by $i/2 - j$ where $i$ is the least degree such that there exists an $x \in HSW^i_{\mathbb{Z}_p}(Y)$ with $U^k x = S^j U^l \theta \; ({\rm mod} \; \mathcal{F}_{2j+1}(Y) )$ for some $k,l \ge 0$.
Recall that $E_\infty(Y) \cong \mathbb{F}[U,R,S]/(R^2) U^{m}\theta \, \oplus \, (E_\infty(Y))_{red}$. Since $\theta$ has bi-degree $(0, d(Y) )$, each homogeneous element in $ \mathbb{F}[U,R,S]/(R^2) U^{m}\theta$ has bi-degree $(p,q)$ where $q \ge d(Y) + 2m$. On the other hand, since $E_\infty(Y)_{red}$ is a subquotient of $HF_{red}^+(Y)[R,S]/(R^2)$ and $HF_{red}^+(Y)$ is concentrated in degrees at most $d(Y)-1$, we see that each homogeneous element $x$ in $E_\infty(Y)_{red}$ has bi-degree $(p,q)$, where $q \le d(Y)-1$. Then $U x = U_{(0,2)}x + U_{(1,1)}x + \cdots $ is a sum of homogeneous terms $U_{(j,2-j)}x$ of bidegree $(p+j , q + 2-j )$. Consider first $U_{(0,2)}x$. Since $U_{(0,2)} = U$, we see that $U_{(0,2)}x \in E_\infty(Y)_{red}$. Each of the remaining terms $U_{(j,2-j)} x$ for $j \ge 1$ has bi-degree of the form $(p',q') = (p+j , q+2-j)$, hence $q' = q+2-j \le q+1 \le d(Y)$. Thus if $m > 0$, then $U_{(j,2-j)}x$ must belong again to $E_\infty(Y)_{red}$. This implies that $E_\infty(Y)_{red}$ is an $\mathbb{F}[\widehat{U}]$-submodule of $E_\infty(Y)$. It follows easily from this that for any fixed $j \ge 0$, there exists a $k \ge 0$ such that $U^k E_\infty(Y)_{red} \subseteq \mathcal{F}_{2j+1}(Y)$. This further implies that $x = S^j U^m \theta$ is a minimal degree element such that $U^k x = S^j U^l \theta \; ({\rm mod} \; \mathcal{F}_{2j+1}(Y) )$ for some $k,l \ge 0$, hence $\delta_j^{(p)}(Y) = m + d(Y)/2$ for all $j \ge 0$. So $\delta_j^{(p)}(Y) = \delta_\infty^{(p)}(Y) = m + \delta(Y) > \delta(Y)$. From \cite[Proposition 5.10]{bh}, we also have that $m \le {\rm rk}( HF_{red}^+(Y) ) = -\delta(Y) - \lambda(Y)$, so $\delta_\infty^{(p)}(Y) = m + \delta(Y) \le -\lambda(Y)$.
If $m=0$, then we still have that $\widehat{U} E_\infty(Y)_{red} \subseteq E_\infty(Y)_{red}$, for if this were not the case then we would have some $x \in E_{\infty}(Y)_{red}^{a,d(Y)-1}$ such that $U_{(1,1)}x \notin E_{\infty}(Y)_{red}$. Further, since $d(Y)-1$ is the highest degree in $HF_{red}^+(Y)$, it follows that $U_{(0,2)}x = 0$ and that $U_{(1,1)}x = c \theta$ for some non-zero element $c \in H_{\mathbb{Z}_p}^{a+1}$. Then since $U_{(j,2-j)}x \in \mathcal{F}_{a+2}(Y)$ for $j \ge 2$, it follows that
\[
\widehat{U}x = c \theta \; ({\rm mod} \; \mathcal{F}_{a+2}(Y) ).
\]
Hence $\delta_{\mathbb{Z}_p , c }(Y) < \delta(Y)$. On the other hand, we have already shown that $\delta_{\mathbb{Z}_p , 1} = \delta_0^{(p)}(-Y) = -\delta(Y)$. Using this and \cite[Theorem 4.4]{bh}, we have
\[
0 \le \delta_{\mathbb{Z}_p , c}(Y) + \delta_{\mathbb{Z}_p , 1}(-Y) < \delta(Y) - \delta(Y) = 0,
\]
a contradiction. So even in the $m=0$ case we still have that $\widehat{U} E_{\infty}(Y)_{red} \subseteq E_{\infty}(Y)_{red}$ and so we again conclude that $\delta_j^{(p)}(Y) = \delta_\infty^{(p)}(Y)$ for all $j \ge 0$ and that $\delta_\infty^{(p)}(Y) \le -\lambda(Y)$. This proves (2).
We have proven everything except the inequality $\delta_j^{(p)}(-Y) \ge \lambda(Y)$. But from $\delta_0^{(p)}(Y) \le -\lambda(Y)$ and \cite[Theorem 4.4]{bh}, it follows that
\[
0 \le \delta_j^{(p)}(-Y) + \delta_0^{(p)}(Y) \le \delta_j^{(p)}(-Y) -\lambda(Y),
\]
for all $j \ge 0$. Hence $\delta_j^{(p)}(-Y) \ge \lambda(Y)$.
\end{proof}
\section{Computation of $\theta^{(c)}(T_{a,b})$}\label{sec:tab}
Let $a,b,c > 1$ be coprime integers and suppose also that $c$ is prime. In this section we will prove that $\theta^{(c)}(T_{a,b}) = (a-1)(b-1)/2$, where $T_{a,b}$ is the $(a,b)$-torus knot. Since $\Sigma_c( T_{a,b} ) = \Sigma(a,b,c)$, we will be interested in computing the invariant $j^{(c)}( \Sigma(a,b,c) )$ of the Brieskorn sphere $\Sigma(a,b,c)$. Consequently we are interested in studying the structure of the Floer homology of $\Sigma(a,b,c)$. Combined with the results of Section \ref{sec:fdelta}, we will be able to carry out the computation of $\theta^{(c)}(T_{a,b})$.
\subsection{Floer homology of $-\Sigma(a,b,c)$}\label{sec:abc}
Let $1 < a < b < c$ be pairwise coprime integers and let $\Sigma(a,b,c)$ denote the Brieskorn homology $3$-sphere oriented so that it is the link of the singularity $x^a + y^b + z^c = 0$ in $\mathbb{C}^3$. The graded roots algorithm of N\'emethi can be used to compute the Floer homology $HF^+(-\Sigma(a,b,c))$ \cite{nem}. In this case the algorithm produces a $\tau$ function $\tau \colon\mathbb{Z}_{\ge 0} \to \mathbb{Z}$ from which a graded root $(R ,\chi)$ may be constructed \cite[Example 3.4 (3)]{nem}. The graded root $(R , \chi)$ gives rise to an associated $\mathbb{F}[U]$-module $\mathbb{H}(R,\chi)$, which up to a grading shift coincides with $HF^+(-\Sigma(a,b,c))$.
For a Brieskorn sphere $\Sigma(a,b,c)$, the $\tau$ function $\tau(i)$ is given as follows \cite[\textsection 11]{nem}, \cite{caka}. Let $e_0 , p_1, p_2, p_3$ be the unique integers with $0 < p_1 < a$, $0 < p_2 < b$, $0 < p_3 < c$ and
\[
a b c e_0 + p_1 bc + a p_2 c + ab p_3 = -1.
\]
Then
\begin{equation}\label{equ:tau1}
\tau(i) = \sum_{n=0}^{i-1} \Delta(n),
\end{equation}
where
\begin{equation}\label{equ:tau2}
\Delta(n) = 1 - e_0 n - \left\lceil \frac{p_1 n}{a} \right\rceil - \left\lceil \frac{p_2 n}{b} \right\rceil - \left\lceil \frac{p_3 n}{c} \right\rceil.
\end{equation}
By \cite[Theorem 1.3]{caka}, we have $\tau(n+1) \ge \tau(n)$ for $n > N$, where
\[
N = abc - bc - ac - ab.
\]
Note that $N > 0$ except when $(a,b,c) = (2,3,5)$. We will exclude this case from the discussion unless stated otherwise. For the purpose of computing the Floer homology, it suffices to consider the $\tau$ function up until the point where it is increasing. So we only need the restricted $\tau$ function $\tau \colon [0,N+1] \to \mathbb{Z}$, or equivalently, it suffices to determine the $\Delta$-function $\Delta \colon [0,N] \to \mathbb{Z}$. Furthermore, the function $\Delta$ is completely determined on $[0,N]$ as follows \cite[Theorem 1.3]{caka}. Let
\[
G = \{ g \in \mathbb{Z}_{\ge 0} \; | \; g = bc i + ac j + ab k \text{ for some } i,j,k \in \mathbb{Z}_{\ge 0} \}
\]
be the additive semigroup generated by $bc, ac, ab$. We have that $\Delta(n) \in \{ -1 , 0 , 1 \}$ for all $n \in [0,N]$. Moreover, for any $n \in [0,N]$, we have that
\[
\Delta(n) = \begin{cases} \; \; \, 1 & n \in G, \\ -1 & N-n \in G, \\ \; \; \, 0 & \text{otherwise}. \end{cases}
\]
It follow that for $n \in [0,N]$, we have
\begin{equation}\label{equ:taudiff}
\tau(n+1) = | G \cap [0,n] | - | G \cap [N-n,N] |
\end{equation}
where $|S|$ denotes the cardinality of a finite set $S$. From this expression it is clear that $\tau$ satisfies $\tau(N+1-n) = \tau(n)$ for all $n \in [0,N]$.
\begin{remark}\label{rem:crt}
Note that since $0 < N < abc$, it follows from the Chinese remainder theorem that any $n \in G \cap [0,N]$ can be written as $abc( i/a + j/b + k/c)$ for {\em uniquely determined} integers $i,j,k$ satisfying $0 < i < a$, $0 < j < b$, $0 < k < c$.
\end{remark}
We briefly explain how to obtain $HF^+(-\Sigma(a,b,c))$ from the $\tau$ function. First we recall the definition of a graded root:
\begin{definition}[\cite{nem}]
Let $R$ be an infinite tree with vertices $\mathcal{V}$ and edges $\mathcal{E}$. Denote by $[u,v]$ the edge joining $u$ and $v$. Edges are unordered, so $[u,v] = [v,u]$. We say that $R$ is a {\em graded root} with grading $\chi \colon \mathcal{V} \to \mathbb{Z}$ if
\begin{itemize}
\item[(a)]{$\chi(u) - \chi(v) = \pm 1$ for all $[u,v] \in \mathcal{E}$,}
\item[(b)]{$\chi(u) > \min \{ \chi(v) , \chi(w) \}$ if $[u,v], [u,w] \in \mathcal{E}$ and $v \neq w$,}
\item[(c)]{$\chi$ is bounded from below, $\chi^{-1}(k)$ is finite for any $k \in \mathbb{Z}$ and $|\chi^{-1}(k)| = 1$ for all sufficiently large $k$.}
\end{itemize}
\end{definition}
Define a partial relation $\succeq$ on $(R , \chi)$ by declaring $v \succeq w$ if there exists a sequence of vertices $w = w_0 , w_1 , \dots , w_k = v$ such that $[w_{i-1} , w_i] \in \mathcal{E}$ and $\chi(w_i) - \chi(w_{i-1}) = 1$ for $i = 1, \dots , k$. From the axioms (a)-(c), it is easily seen that for each $u \in \mathcal{V}$ there is a unique vertex $v \in \mathcal{V}$ with $[u,v] \in \mathcal{E}$ and $\chi(v) = \chi(u)+1$. From this and axiom (c), it is seen that for any two vertices $u,v \in \mathcal{V}$ there is a unique $\succeq$-minimal element $w \in \mathcal{V}$ with $w \succeq v$ and $w \succeq u$. We denote this element by $\sup(u,v)$.
For $v \in \mathcal{V}$, let $\delta_v$ be the number of edges which have $v$ as an endpoint. Clearly $\delta_v \ge 1$ and $\delta_v = 1$ if and only if $v$ is $\succeq$-minimal.
Given a graded root $(R , \chi)$ define $\mathbb{H}(R,\chi)$ to be the graded $\mathbb{F}[U]$-module constructed as follows. A basis for $\mathbb{H}(R,\chi)$ is given by $\{ e_v \; | \; v \in \mathcal{V} \}$ where $deg(e_v) = 2 \chi(v)$ and $U e_v = e_u$, where $u$ is the unique $u \in \mathcal{V}$ such that $[u,v] \in \mathcal{E}$ and $\chi(u) = \chi(v) + 1$. Note that our definition of $\mathbb{H}(R,\chi)$ differs slightly from the one in \cite{nem}. This is because we are using Floer cohomology rather than homology. The underlying graded abelian group of $\mathbb{H}(R,\chi)$ is the same as that in \cite{nem}, but our $\mathbb{F}[U]$-module action is the transpose of the one in \cite{nem}.
For a positive integer $n$ let $\mathcal{T}^+(n)$ denote the $\mathbb{F}[U]$-module $\mathbb{F}[U]/(U^n)$. For any integer $d$ and any $\mathbb{F}[U]$-module $M$, let $M_d$ be the degree-shift of $M$ defined by $(M_d)^j = M^{j-d}$. In particular, for any positive integer $n$ and any integer $d$, we have the $\mathbb{F}[U]$-module $\mathcal{T}^+_d(n) = (\mathcal{T}^+(n))_d$. We refer to $\mathcal{T}^+_d(n)$ as a {\em tower of length} $n$. The lowest degree in $\mathcal{T}^+_d(n)$ is $d$ and the highest degree is $d+2n-2$.
\begin{proposition}[\cite{nem}, Proposition 3.5.2]\label{prop:hfiso}
Let $(R , \chi)$ be a graded root. Set $I = \{ v \in \mathcal{V} \; | \; \delta_v = 1\}$. We choose an ordering on the set $I$ as follows. The first element $v_1 \in I$ is chosen so that $\chi(v_1)$ is the minimal value of $\chi$. If $v_1, \dots , v_k \in I$ are already determined and $J = \{v_1 , \dots , v_k\} \neq I$, then $v_{k+1}$ is chosen from $I \setminus J$ such that $\chi(v_{k+1}) = \min_{v \in I \setminus J} \chi(v)$. Let $w_{k+1} \in \mathcal{V}$ be the unique $\succeq$-minimal vertex of $R$ which dominates both $v_{k+1}$ and some element of $J$. Then we have an isomorphism of $\mathbb{F}[U]$-modules:
\[
\mathbb{H}(R , \chi) \cong \mathbb{F}[U]_{2\chi(v_1)} \oplus \bigoplus_{k \ge 2} \mathcal{T}^+_{2\chi(v_k)}( \chi(w_k) - \chi(v_k) ).
\]
In particular we have an isomorphism
\[
\mathbb{H}_{red}(R , \chi) \cong \bigoplus_{k \ge 2} \mathcal{T}^+_{2\chi(v_k)}( \chi(w_k) - \chi(v_k) ).
\]
\end{proposition}
Let $\tau \colon \{0 , 1, \dots , l\} \to \mathbb{Z}$ be any function and suppose there is an $l$ such that $\tau(i+1) \ge \tau(i)$ for all $i \ge l$. Following \cite[Example 3.4]{nem}, we construct a graded root $(R,\chi)$ from $\tau$ as follows. Start with the vertex set $\widetilde{\mathcal{V}} = \{ v_i^k \}$ where $0 \le i \le l$ and $k \ge \tau(i)$, edge set $\widetilde{\mathcal{E}} = [v_i^k , v_i^{k+1}]$ and define $\widetilde{\chi}$ by $\widetilde{\chi}(v_i^k) = k$. Now define $\mathcal{V} = \widetilde{\mathcal{V}}/\! \! \sim$ and $\mathcal{E} = \widetilde{\mathcal{E}}/\! \! \sim$ where for $i \le j$ we set $v_i^k \sim v_j^k$ and $[v_i^k , v_i^{k+1}] \sim [v_j^k , v_j^{k+1}]$ if $k \ge \max_{i \le u \le j} \tau(u)$. We define $\chi$ to be the function on $\mathcal{V}$ induced by $\widetilde{\chi}$. It is easily checked that $(\mathcal{V} , \mathcal{E} , \chi)$ is a graded root. We call it the {\em graded root associated to $\tau$}.
Let $1 < a < b < c$ be pairwise coprime integers and let $\tau$ be the $\tau$-function defined as in Equations (\ref{equ:tau1})-(\ref{equ:tau2}). Let $(R_\tau , \chi_\tau)$ be the associated graded root. Then from \cite{nem} it follows that up to a grading shift $HF^+(-\Sigma(a,b,c))$ is isomorphic to $\mathbb{H}( R_\tau , \chi_\tau)$:
\[
HF^+(-\Sigma(a,b,c)) \cong \mathbb{H}( R_\tau , \chi_\tau)[ u ].
\]
The precise value of the grading shift $u$ can be computed \cite[Proposition 4.7]{nem}, but we will not need this.
Observe that $\tau(0) = \tau(N+1) = 0$ and that $\tau(1) = \tau(N) = 1$. We will see in Proposition \ref{prop:taumax} that $\tau(n) \le 1$ for all $n \in [0,N+1]$, hence $1$ and $N$ are global maxima of $\tau$. Let $n \in [0,N+1]$ be a global maximum of $\tau$, that is, $\tau(n) = 1$. We will say $n$ is a {\em trivial maximum} if we either have that $\tau(i) = 1$ for all $1 \le i \le n$ or $\tau(i) = 1$ for all $n \le i \le N$.
\begin{proposition}\label{prop:topdegtau}
Suppose that all global maxima of $\tau$ are trivial. Let $\ell^+(-\Sigma(a,b,c))$ denote the highest non-zero degree in $HF^+_{red}(-\Sigma(a,b,c))$. Any non-zero element in the image of $U \colon HF^+_{red}(-\Sigma(a,b,c)) \to HF_{red}^+(-\Sigma(a,b,c))$ has degree strictly less than $\ell^+(-\Sigma(a,b,c))$.
\end{proposition}
\begin{proof}
Let $(R_\tau , \chi_\tau)$ be the graded root associated to $\tau$. From the definition of $(R_\tau , \chi_\tau)$, it is clear that $(R_\tau , \chi_\tau)$ depends only on the values of the local minima and local maxima of $\tau$. More precisely, let $S^{min}$ be the set of $m \in [0,N+1]$ such that $m=0$, $m = N+1$, or $\tau(m-1) > \tau(m)$ and there exists a $k \ge 0$ such that $\tau(i) = \tau(m)$ for $m \le i \le m+k$ and $\tau(m+k+1) > \tau(m)$. Similarly, let $S^{max}$ be the set of $M \in [1,N]$ such that $\tau(M) > \tau(M-1)$ and there exists a $k \ge 0$ such that $\tau(i) = \tau(M)$ for $M \le i \le M+k$ and $\tau(M+k+1) < \tau(M)$. Write the elements of $S^{min}$ and $S^{max}$ in increasing order as $S^{min} = \{ m_0 , m_1 , \dots , m_r \}$, $S^{max} = \{ M_1 , M_2 , \dots , M_r \}$ where $0 = m_0 < m_1 < \cdots < m_r = N+1$ and $1 = M_1 < M_2 < \cdots < M_r$. The local minima and maxima must occur in alternating order, so we have
\[
0 = m_ 0 < M_1 < m_1 < M_2 < \cdots < m_{r-1} < M_r < m_r = N+1.
\]
Furthermore, $M_1$ and $M_r$ correspond to the trivial global maxima. By assumption these are the only global maxima. Hence $\tau(M_j) \le 0$ for $1 < j < r$.
Recall that $(R_\tau , \chi_\tau)$ is constructed as follows. Start with $\widetilde{\mathcal{V}} = \{ v_i^k \}$ where $0 \le i \le N+1$ and $k \ge \tau(i)$, edge set $\widetilde{\mathcal{E}} = [v_i^k , v_i^{k+1}]$ and define $\widetilde{\chi}$ by $\widetilde{\chi}(v_i^k) = k$. Then $(R_\tau , \chi_\tau)$ is obtained by taking the quotient of this by the equivalence relation $\sim$, where for $i \le j$ we set $v_i^k \sim v_j^k$ and $[v_i^k , v_i^{k+1}] \sim [v_j^k , v_j^{k+1}]$ if $k \ge \max_{i \le u \le j} \tau(u)$.
It is easily seen that the minima of $\succeq$ on $(R_\tau , \chi_\tau)$ are precisely the elements $I = \{ v_{m_i}^{\tau(m_i)} \}_{0 \le i \le r}$. From Proposition \ref{prop:taumax} we have that $\tau(n) \le 1$ for all $n \in [0,N+1]$ and hence $\tau(m_i) \le 0$ for any local minimum. Hence we may choose a permutation $\sigma \colon \{ 0 , 1 , \dots , r \} \to \{ 0 , 1 , \dots , r\}$ such that $\tau( m_{\sigma(i)} ) \le \tau( m_{\sigma(i+1)})$ for $0 \le i \le r-1$ and $\sigma(r-1) = 0$, $\sigma(r) = r$. This gives an ordering of the set $I$ as in the statement of Proposition \ref{prop:hfiso}, namely $I = \{ v'_0 , v'_1 , v'_2 , \dots , v'_r \}$ where $v'_i = v_{m_{\sigma(i)}}^{\tau(m_{\sigma(i)})}$. For each $i \in \{1 , \dots , r\}$ and each $j \in \{0 , 1 , \dots , i-1\}$ one finds that $sup(v'_i , v'_j) = v_{m_\sigma(i)}^{K_{ij}}$, where $K_{ij}$ is the maximum of $\tau(M_a)$ for all $a$ such that $M_a$ lies in the interval joining $m_{\sigma(i)}$ and $m_{\sigma(j)}$. Now let $w_1 , w_2 , \dots , w_r$ be defined as in the statement of Proposition \ref{prop:hfiso}. Then it follows that $w_i = v_{m_{\sigma(i)}}^{K_i}$, where $K_i = \min_{0 \le j \le i-i} K_{ij}$. Now suppose that $i \neq r-1, r$. Then for all $j \in \{ 0 , \dots , i-1\}$ we have that $\sigma(i) \neq 0,r$ and $\sigma(j) \neq 0,r$. It follows that each maximum $M_a$ in the interval joining $m_{\sigma(i)}$ and $m_{\sigma(j)}$ is not trivial and hence $\tau(M_a) \le 0$. Hence $K_{ij} \le 0$ for all $i \le r-2$ and $j < i$. This also implies that $\chi(w_i) = K_{i} \le 0$ for all $i \le r-2$. On the other hand since $\tau(M_1) = \tau(M_{r}) = 1$, it follows that $K_{r-1} = K_r = 1$ and thus $\chi(w_{r-1}) = \chi(w_r) = 1$.
From Proposition \ref{prop:hfiso}, we have an isomorphism
\[
\mathbb{H}_{red}(R_\tau , \chi_\tau) \cong \bigoplus_{k = 1}^{r} \mathcal{T}^+_{2\chi(v'_k)}( \chi(w_k) - \chi(v'_k) ).
\]
The highest degree in the tower $\mathcal{T}^+_{2\chi(v'_k)}( \chi(w_k) - \chi(v'_k) )$ is $2\chi(w_k) - 2$. For $k \neq r-1, r$ we have $2\chi(w_k)-2 \le -2$ whereas for $k = r-1,r$ we have $2\chi(w_k)-2 = 0$. Thus the highest non-zero degree in $\mathbb{H}_{red}(R , \chi)$ is $0$ and only the towers for $k = r-1 , r$ attain this degree. Moreover the length of the tower $\mathcal{T}^+_{2\chi(v'_k)}( \chi(w_k) - \chi(v'_k) )$ is $\chi(w_k) - \chi(v'_k)$ which for $k=r-1 , r$ equals $1$, since $v'_{r-1} = v_{\sigma(r-1)}^{\tau(m_{\sigma(r-1)})} = v_0^{\tau(0)} = v_0^0$ and $v'_r = v_{\sigma(r)}^{\tau(m_{\sigma(r)})} = v_r^{\tau(m_r)} = v_r^0$, so $\chi(v'_{r-1}) = \chi(v'_{r}) = 0$. It follows that any non-zero element in the image of $U \colon \mathbb{H}_{red}(R_\tau , \chi_\tau) \to \mathbb{H}_{red}(R_\tau , \chi_\tau)$ has degree strictly less than $0$. Now since $HF_{red}^+(-\Sigma(a,b,c))$ is isomorphic to a grading shift of $\mathbb{H}_{red}(R_\tau , \chi_\tau)$, it also follows that any non-zero element in the image of $U \colon HF^+_{red}(-\Sigma(a,b,c)) \to HF_{red}^+(-\Sigma(a,b,c))$ has degree strictly less than $\ell^+(-\Sigma(a,b,c))$.
\end{proof}
\begin{proposition}\label{prop:taumax}
Let $1 < a < b < c$ be pairwise coprime integers and assume that $(a,b,c) \neq (2,3,5)$. Then $\tau(n+1) \le 1$ for all $n \in [0,N]$. Furthermore, all maxima of $\tau$ are trivial, except in the following cases: $(a,b,c) = (2,3,6n-1)$, $n \ge 2$ or $(a,b,c) = (2,3,6n+1)$, $n \ge 1$.
\end{proposition}
\begin{proof}
For $n \in [0,N]$, let $\alpha_n$ be the largest element of $G$ less than or equal to $N-n$. Suppose that $j \in G \cap [0,n] \setminus \{0\}$. We claim that $j + \alpha_n \in G \cap [N-n,N]$. Clearly $j + \alpha_n \in G$ because $j, \alpha_n \in G$. Since $j \le n$ and $\alpha_n \le N-n$, we also have $j + \alpha_n \le N$. Lastly, since $j > 0$, we have $j + \alpha_n \ge N-n$ for if not then $\alpha_n$ is not the largest element of $G$ less than or equal to $N-n$. Therefore $j + \alpha_n \in G \cap [N-n,N]$. So we have constructed an injective map $\phi_n : G \cap [0,n] \setminus \{0\} \to G \cap [N-n,N]$, given by $\phi(j) = j + \alpha_n$. Therefore $| G \cap [0,n] | \le 1 + |G \cap [N-n,N]$. Comparing with Equation (\ref{equ:taudiff}), we have shown that $\tau(n) \le 1$.
Now suppose that $n_0 \in [1,N]$ is a maximum of $\tau$, so $\tau(n_0) = 1$. By the symmetry $\tau(N+1+i) = \tau(i)$ of the $\tau$ function, it suffices to consider only maxima $n_0$ such that $n_0 \ge (N+1)/2$. Since $\tau$ can only change by $\pm 1$, there exists an $n \ge n_0$ such that $\tau(i) = 1$ for $n_0 \le i \le n$ and $\tau(n+1) = 0$. Since $\tau(N+1) = 0$, we see that $n \le N$. Also we have $n \ge n_0 \ge (N+1)/2$.
Since $\tau(n) =1$ and $\tau(n+1) = 0$, we have that $\Delta(n) = \tau(n+1) - \tau(n) = -1$. Hence $N - n \in G$ and $n \notin G$. Since $N-n \in G$, it follows that $\alpha_n = N-n$. Consider the map $\phi \colon G \cap [0,n] \to G \cap [N-n,N]$ given by $\phi(j) = j + N-n$ (we already saw that $\phi(j) \in G \cap [N-n ,N]$ for $j > 0$ and we also have $\phi(0) = N-n \in G \cap [N-n,N]$). Since $\tau(n+1) = 0$, it follows that $\phi$ is a bijection. Thus every $g \in G \cap [N-n,N]$ can be written uniquely as $g = (N-n) + g'$ for some $g \in G \cap [0,n]$. Since $N-n \in G$, we may write it in the form
\[
N-n = abc \left( \frac{i_0}{a} + \frac{j_0}{b} + \frac{k_0}{c} \right)
\]
for some $i_0,j_0,k_0 \ge 0$. It follows that any $g \in G \cap [N-n,N]$ has the form
\begin{equation}\label{equ:g}
g = abc \left( \frac{i}{a} + \frac{j}{b} + \frac{k}{c} \right)
\end{equation}
where $i \ge i_0$, $j \ge j_0$, $k \ge k_0$. Furthermore, by Remark \ref{rem:crt}, this is a necessary condition. That is, if $g \in G \cap [N-n,N]$ is written as in (\ref{equ:g}) then we must have $i \ge i_0$, $j \ge j_0$, $k \ge k_0$.
Suppose that $i_0 > 0$. There exists an integer $k$ such that $1/a$ lies in the interval $[(k-1)/c , k/c]$. Thus
\[
\frac{1}{a} \le \frac{k}{c} \le \frac{1}{a} + \frac{1}{c}.
\]
Consider
\[
g = (N-n) + abc\left( -\frac{1}{a} + \frac{k}{c} \right) = abc\left( \frac{i_0-1}{a} + \frac{j_0}{b} + \frac{k_0+k}{c} \right).
\]
Then $g \in G$ and $g \ge N-n$ since $k/c - 1/a \ge 0$. But $g$ is not of the form $abc(i'/a + j'/b + k'/c)$ with $i' \ge i_0$, $j' \ge j_0$, $k' \ge k_0$, hence $g \notin G \cap [N-n,N]$. The only way this can happen is that $g > N$. Thus
\[
g = (N-n) + abc\left( -\frac{1}{a} + \frac{k}{c} \right) > N,
\]
which implies that $n < abc( -1/a + k/c )$. But $k/c \le 1/a + 1/c$, so $n < abc/c = ab$. Note that since $a<b<c$, $ab$ is the smallest positive element of $G$. Thus $i \notin G$ for all $1 \le i \le n$. So $\Delta(i) \le 0$ for $1 \le i \le n$. But we also have that
\[
\tau(n) = 1 = \sum_{i=0}^{n-1} \Delta(i) = 1 + \sum_{i=1}^{n-1} \Delta(i).
\]
Since $\Delta(i) \le 0$ for $1 \le i \le n$, the only way we can have equality is that $\Delta(i) = 0$ for $1 \le i \le n$ and hence $\tau(i) = 1$ for $1 \le i \le n$. This means that $n$ is a trivial maximum.
Next, suppose that $i_0 = 0$ and $j_0 > 0$. Then by a similar argument to the $i_0 > 0$ case, there exists a $k$ such that $1/b \le k/c \le 1/b + 1/c$. Consider
\[
g = (N-n) + abc\left( -\frac{1}{b} + \frac{k}{c} \right).
\]
Arguing as in the $i_0 > 0$ case, we see that $n$ is again a trivial maximum.
Now consider the case that $i_0 = j_0 = 0$. Hence $N-n = abc(k_0/c)$ for some $k_0 \ge 0$. We will assume that $k_0 > 0$ for if $k_0 = 0$, then $n = N$ is a trivial maximum. Furthermore, recall that we are assuming $n \ge (N+1)/2 > N/2$. Therefore
\begin{equation}\label{equ:ink0}
\frac{k_0}{c} < \frac{N}{2abc} = \frac{1}{2} - \frac{1}{2a} - \frac{1}{2b} - \frac{1}{2c}.
\end{equation}
Recall that each element of $G \cap [N-n,N]$ has the form given by (\ref{equ:g}) with $i \ge 0$, $j \ge 0$, $k \ge k_0$. It follows that there does not exist a solution to
\begin{equation}\label{equ:jb}
\frac{k_0}{c} \le \frac{j}{b} + \frac{k_0-1}{c} \le 1 - \frac{1}{a} - \frac{1}{b} -\frac{1}{c}
\end{equation}
with $j \ge 0$. For if such a $j$ exists, we would have
\[
N-n = abc \left( \frac{k_0}{c} \right) < abc\left( \frac{j}{b} + \frac{k_0-1}{c} \right) \le N
\]
and thus $g = abc( j/b + (k_0-1)/c )$ would be an element of $G \cap [N-n,N]$ not of the form $abc(i'/a + j'/b + k'/c)$ with $i' \ge i_0$, $j' \ge j_0$, $k' \ge k_0$. In particular, $j=1$ is not a solution. But since $b < c$, we have $1/b + (k_0-1)/c > k_0/c$. So it must be the second inequality in (\ref{equ:jb}) that is violated. That is, we must have
\[
\frac{j}{b} + \frac{k_0-1}{c} > 1 - \frac{1}{a} - \frac{1}{b} - \frac{1}{c}.
\]
Rearranging and using (\ref{equ:ink0}), we find that
\begin{equation}\label{equ:cond1}
\frac{1}{a} + \frac{3}{b} - \frac{1}{c} > 1.
\end{equation}
If this condition is not satisfied, then all maxima of $\tau$ are trivial.
If $a \ge 4$, then $b \ge 5$ and
\[
\frac{1}{a} + \frac{3}{b} - \frac{1}{c} < \frac{1}{4} + \frac{3}{5} < 1.
\]
So (\ref{equ:cond1}) implies $a < 4$, hence $a = 2$ or $3$.
If $a=3$, then since $1/3 +3/5 < 1$, (\ref{equ:cond1}) implies that $b \le 5$. So $b = 4$ and then (\ref{equ:cond1}) implies that $c > 12$.
If $a=2$, then since $1/2 + 3/6 - 1/c < 1$, we must have $b < 6$. Hence $b = 3$ or $5$.
If $a=2, b=5$, then (\ref{equ:cond1}) is satisfied for any $c > 10$.
If $a=2, b=3$, then (\ref{equ:cond1}) is satisfied for any $c > 3$. But we are excluding $(2,3,5)$ so $c > 5$.
To summarise, (\ref{equ:cond1}) is satisfied only in the following cases:
\begin{itemize}
\item[(1)]{$(a,b,c) = (3,4,c)$, $c > 12$.}
\item[(2)]{$(a,b,c) = (2,5,c)$, $c > 10$.}
\item[(3)]{$(a,b,c) = (2,3,c)$, $c > 5$.}
\end{itemize}
It remains to show in cases (1) and (2) we still have that all maxima of $\tau$ are trivial. We do this by a direct computation of the $\tau$ function.
In case (1) there are four subcases:
\begin{itemize}
\item[(1a)]{$(a,b,c) = (3,4,12k+1)$, $k \ge 1$.}
\item[(1b)]{$(a,b,c) = (3,4,12k+5)$, $k \ge 1$.}
\item[(1c)]{$(a,b,c) = (3,4,12k+7)$, $k \ge 1$.}
\item[(1d)]{$(a,b,c) = (3,4,12k+11)$, $k \ge 1$.}
\end{itemize}
Similarly in case (2) there are four subcases:
\begin{itemize}
\item[(2a)]{$(a,b,c) = (2,5,10k+1)$, $k \ge 1$.}
\item[(2b)]{$(a,b,c) = (2,5,10k+3)$, $k \ge 1$.}
\item[(2c)]{$(a,b,c) = (2,5,10k+7)$, $k \ge 1$.}
\item[(2d)]{$(a,b,c) = (2,5,10k+9)$, $k \ge 1$.}
\end{itemize}
Case (1a): $N = 60k-7$, $G$ is generated by $\{ 12 , 36k+3 , 48k+4 \}$. By the symmetry of the $\tau$ function, it suffices to only look for maxima of $\tau(n+1)$ with $n \in [0,N/2]$. We consider the intersection of $G$ with $[0 , N/2]$. Since $36k+3 > N/2$, all elements of $G \cap [0,N]$ have the form $12j$ for some $j$. Thus
\[
G \cap [0,N/2] = \{ 12j \}_{0 \le j \le \lfloor N/24 \rfloor}.
\]
Now we consider the intersection of $N-G$ with $[0,N]$. Elements of $N-G$ in the range $[0,N]$ have the form $N - 12u$, $N - (36k+3)-12u$ or $N - (48k+4)-12u$ for some $u \ge 0$. In the first case, we have
\[
N - 12u = 12(5k-1-u)+5.
\]
In the second case, we have
\[
N -(36k+3)-12u = 12(2k-1-u)+2
\]
and in the third case, we have
\[
N - (48+4) - 12u = 12(k-1-u)+1.
\]
Therefore
\[
(N-G) \cap [0,N] = \{ 12j + 5\}_{0 \le j \le 5k-1} \cup \{ 12j+2\}_{0 \le j \le 2k-1} \cup \{ 12j+1 \}_{0 \le j \le k-1}.
\]
Partition $[0,N]$ into subintervals $I_j = [12j , 12j+11]$, $0 \le j \le 5k-2$ and $I_{5k-1} = [60k-12 , 60k-7]$. We are only interested in the subintervals which intersect with $[0,N/2]$, so we can assume $j \le \lfloor N/24 \rfloor$. Under this condition, we have $G \cap I_j \cap G = \{12j\}$ and
\[
(N-G) \cap I_j = \begin{cases} 12j+1, 12j+2, 12j+5 & 0 \le j \le k-1, \\ 12j+2, 12j+5 & k \le j \le 2k-1, \\ 12j+5 & 2k \le j. \end{cases}
\]
It follows easily that $\tau$ has only trivial maxima.
Case (1b): $N = 60k+13$, $G$ is generated by $12, 36k+15, 48k+20$. Arguing similarly to case (1a), we find
\[
G \cap [0,N/2] = \{ 12j \}_{0 \le j \le \lfloor N/24 \rfloor}
\]
and
\[
(N-G) \cap [0,N] = \{ 12j + 1\}_{0 \le j \le 5k+1} \cup \{ 12j+10\}_{0 \le j \le 2k-1} \cup \{ 12j+5 \}_{0 \le j \le k-1}.
\]
By a similar argument, we see that $\tau$ has only trivial maxima.
Case (1c): $N = 60k+23$, $G$ is generated by $12, 36k+21, 48k+28$. We find
\[
G \cap [0,N/2] = \{ 12j \}_{0 \le j \le \lfloor N/24 \rfloor}
\]
and
\[
(N-G) \cap [0,N] = \{ 12j + 11\}_{0 \le j \le 5k+1} \cup \{ 12j+2\}_{0 \le j \le 2k} \cup \{ 12j+7 \}_{0 \le j \le k-1}.
\]
We then see that $\tau$ has only trivial maxima.
Case (1d): $N = 60k+43$, $G$ is generated by $12, 36k+33, 48k+44$. We find
\[
G \cap [0,N/2] = \{ 12j \}_{0 \le j \le \lfloor N/24 \rfloor}
\]
and
\[
(N-G) \cap [0,N] = \{ 12j + 7\}_{0 \le j \le 5k+3} \cup \{ 12j+10\}_{0 \le j \le 2k} \cup \{ 12j+11 \}_{0 \le j \le k-1}.
\]
We then see that $\tau$ has only trivial maxima.
Case (2a): $N = 30k-7$, $G$ is generated by $10, 20k+2, 50k+5$. We find
\[
G \cap [0,N/2] = \{ 10j \}_{0 \le j \le \lfloor N/20 \rfloor}
\]
and
\[
(N-G) \cap [0,N] = \{ 10j + 3\}_{0 \le j \le 3k-1} \cup \{ 10j+1\}_{0 \le j \le k-1}
\]
We then see that $\tau$ has only trivial maxima.
Case (2b): $N = 30k-1$, $G$ is generated by $10, 20k+6, 50k+15$. We find
\[
G \cap [0,N/2] = \{ 10j \}_{0 \le j \le \lfloor N/20 \rfloor}
\]
and
\[
(N-G) \cap [0,N] = \{ 10j + 9\}_{0 \le j \le 3k-1} \cup \{ 10j+3\}_{0 \le j \le k-1}
\]
We then see that $\tau$ has only trivial maxima.
Case (2c): $N = 30k+11$, $G$ is generated by $10, 20k+14, 50k+35$. We find
\[
G \cap [0,N/2] = \{ 10j \}_{0 \le j \le \lfloor N/20 \rfloor}
\]
and
\[
(N-G) \cap [0,N] = \{ 10j + 1\}_{0 \le j \le 3k+1} \cup \{ 10j+7\}_{0 \le j \le k-1}
\]
We then see that $\tau$ has only trivial maxima.
Case (2d): $N = 30k+17$, $G$ is generated by $10, 20k+18, 50k+45$. We find
\[
G \cap [0,N/2] = \{ 10j \}_{0 \le j \le \lfloor N/20 \rfloor}
\]
and
\[
(N-G) \cap [0,N] = \{ 10j + 7\}_{0 \le j \le 3k+1} \cup \{ 10j+9\}_{0 \le j \le k-1}
\]
We then see that $\tau$ has only trivial maxima.
\end{proof}
\begin{proposition}\label{prop:topdeg}
Let $1 < a < b < c$ be pairwise coprime integers and assume $(a,b,c) \neq (2,3,5)$. Let $Y = -\Sigma(a,b,c)$ and let $\ell^+(Y)$ denote the highest non-zero degree in $HF^+_{red}(Y)$. Then $\ell^+(Y) = 2\delta(Y) - 2\min\{\tau\}$. Moreover, any non-zero element in the image of $U \colon HF_{red}^+(Y) \to HF_{red}^+(Y)$ has degree strictly less than $\ell^+(Y)$.
\end{proposition}
\begin{proof}
Recall that $HF^+(Y)$ is isomorphic to $\mathbb{H}(R_\tau , \chi_\tau)$ up to an overall grading shift. The lowest degree in $HF^+(-\Sigma(a,b,c))$ is $2\delta(Y)$ and the lowest degree in $\mathbb{H}(R_\tau , \chi_\tau)$ is $2\min\{ \tau\}$, hence the grading shift is $2\delta(Y) - 2\min\{ \tau \}$. The highest non-zero degree in $\mathbb{H}(R_\tau , \chi_\tau)$ is easily seen to be $0$, hence $\ell^+(Y) = 2\delta(Y) - 2\min\{ \tau \}$.
If $(a,b,c) \neq (2,3,6n \pm 1)$ for any $n$, then all global maxima of $\tau$ are trivial by Proposition \ref{prop:taumax}. Then by Proposition \ref{prop:topdegtau} we have that any non-zero element in the image of $U \colon HF_{red}^+(Y) \to HF_{red}^+(Y)$ has degree strictly less than $\ell^+(Y)$.
If $(a,b,c) = (2,3,6n \pm 1)$ for some $n$, then it is easily seen that $HF_{red}^+(Y)$ is concentrated in a single degree and hence $U$ acts trivially on $HF_{red}^+(Y)$.
\end{proof}
\subsection{Computation of $j^{(c)}(T_{a,b})$ and $\theta^{(c)}(T_{a,b})$}\label{sec:jtheta}
Following \cite{caka}, define $\kappa(a,b,c)$ to be the cardinality of $G \cap [0,N]$.
\begin{lemma}\label{lem:kappa}
Let $\tau_1(a,b,c)$ denote the number of integers $x,y,z$ with $0 < x < a$, $0 < y < b$, $0< z < c$ and $x/a + y/b + z/c < 1$. Then $\tau_1(a,b,c) = \kappa(a,b,c)$.
\end{lemma}
\begin{proof}
By Remark \ref{rem:crt}, any $x \in G$ with $x \le N$ has a unique representation $x bc + y ac + zab$ with $0 \le x < a$, $0 \le y < b$, $0 \le z <c$. Thus $\kappa(a,b,c)$ is the number of points $(x,y,z) \in \mathbb{Z}^3_{\ge 0}$ such that
\[
x bc + y ac + z ab \le abc - bc - ac - ab.
\]
From \cite[Theorem 1.3]{caka} it follows that $N \notin G$, so $\kappa(a,b,c)$ is also the number of $x,y,z \ge 0$ such that
\[
x bc + y ac + z ab < abc - bc - ac - ab.
\]
Dividing through by $abc$, this is equivalent to
\[
\frac{x}{a} + \frac{y}{b} + \frac{z}{c} < 1 - \frac{1}{a} - \frac{1}{b} - \frac{1}{c}
\]
which can be rewritten as
\[
\frac{x+1}{a} + \frac{y+1}{b} + \frac{z+1}{q} < 1.
\]
Setting $x' = x+1, y' = y+1, z' = z+1$, we see that $\kappa(a,b,c)$ is equal to $\tau_1(a,b,c)$.
\end{proof}
Recall that $\lambda$ denotes the Casson invariant. From \cite[\textsection 19]{sav}, we have $8\lambda( \Sigma(a,b,c)) = -(a-1)(b-1)(c-1) + 4\tau_1(a,b,c)$. Thus Lemma \ref{lem:kappa} gives:
\begin{equation}\label{equ:kap}
8\lambda( \Sigma(a,b,c) ) = -(a-1)(b-1)(c-1) + 4\kappa(a,b,c).
\end{equation}
From \cite{fs}, \cite{cs}, we also have that
\[
8 \lambda(\Sigma(a,b,c)) = \sum_{j=1}^{c-1} \sigma_{T_{b,c}}(j/c) = \sigma^{(c)}(T_{a,b}).
\]
\begin{theorem}\label{thm:theta}
Let $a,b,c > 1$ be coprime integers and suppose that $c$ is a prime number. Then
\[
j^{(c)}( -T_{a,b} ) = \begin{cases} \kappa(a,b,c) & \text{if } c \text{ is odd}, \\ 2\kappa(a,b,c) & \text{if } c=2 \end{cases}
\]
and
\[
\theta^{(c)}( T_{a,b} ) = (a-1)(b-1)/2.
\]
\end{theorem}
\begin{proof}
Let $Y = -\Sigma(a,b,c)$. If $(a,b,c)$ is a permutation of $(2,3,5)$, then $HF^+_{red}(Y) = 0$ and $\kappa(2,3,5) = 0$. Then from \cite[Proposition 3.16]{bh}, it follows that $j^{(c)}(-T_{a,b}) = 0$ and $\theta^{(c)}(T_{a,b}) = -\sigma^{(c)}(T_{a,b})/(c-1)$. But since $\kappa(2,3,5) = 0$, Equation (\ref{equ:kap}) gives $\sigma^{(c)}(T_{a,b}) = 8 \lambda(\Sigma(a,b,c)) = -(a-1)(b-1)(c-1)$. Hence $\theta^{(c)}(T_{a,b}) = (a-1)(b-1)/2$.
Henceforth we assume that $(a,b,c)$ is not a permutation of $(2,3,5)$ and hence $HF^+_{red}(Y) \neq 0$. Recall that $\Sigma(a,b,c)$ is the boundary of a negative definite plumbing \cite{nr} whose plumbing graph has only one bad vertex in the sense of \cite{os2}. Then from \cite[Corollary 1.4]{os2}, we have that $HF^+(Y)$ is concentrated in even degrees. As explained in \cite[\textsection 7]{bh} we have $\Sigma(a,b,c) = \Sigma_c(T_{a,b})$ and the generator of the $\mathbb{Z}_c$-action on $\Sigma(a,b,c)$ is isotopic to the identity. Therefore in the spectral sequence $\{ E_r^{p,q} , d_r \}$ for the equivariant Seiberg--Witten--Floer cohomology of $Y$, we have $E_2^{0,q} = H^0( \mathbb{Z}_c ; HSW^q(Y)) \cong HSW^q(Y)$. Furthermore, the graded roots algorithm implies that $HF_{red}^+(Y)$ is concentrated in degrees $d(Y)$ and above. It follows that there can be no differentials in the spectral sequence and hence $\delta_{\mathbb{Z}_c , S^0}(Y) = \delta(Y)$. Recall that $\delta^{(c)}_j( K ) = -\sigma^{(c)}(K)/2$ for all sufficiently large $j$. Then since $Y = -\Sigma(a,b,c) = \Sigma_c( -T_{a,b})$, it follows that $$\delta_{\mathbb{Z}_c , S^j}(Y) = -\sigma^{(c)}(-T_{a,b})/8 = \lambda( \Sigma(a,b,c) ) = -\lambda(Y)$$ for sufficiently large $j$. But since $HF^*_{red}(Y)$ is concentrated in even degrees, we also have $dim_{\mathbb{F}}(HF^*_{red}(Y)) = \delta(Y) + \lambda(Y)$. Hence for large enough $j$, we have $\delta_{\mathbb{Z}_c , S^0}(Y) - \delta_{\mathbb{Z}_c , S^j}(Y) = \delta(Y) + \lambda(Y) = dim_{\mathbb{F}}(HF^*_{red}(Y))$. From Proposition \ref{prop:topdeg}, it follows that any element in the image of $U \colon HF_{red}^+(Y , \mathfrak{s}) \to HF_{red}^+(Y , \mathfrak{s})$ has degree strictly less than $\ell^+(Y,\mathfrak{s})$. Thus conditions (1)-(4) of Proposition \ref{prop:jinv} are met. Therefore (since $c$ is an odd prime and $Y$ is an integral homology sphere) we have
\[
j'(Y) = \ell^+(Y)/2 - \delta_{\mathbb{Z}_c , S^0}(Y) + \delta(Y) + \lambda(Y) = \ell^+(Y)/2 + \lambda(Y),
\]
where $j'(Y) = j^{(c)}(Y)$ if $c$ is odd and $j'(Y) = j^{(2)}(Y)/2$ if $c=2$. Moreover, Proposition \ref{prop:topdeg} also gives $\ell^+(Y) = 2( \delta(Y) -\min(\tau))$. From \cite[\textsection 5]{caka}, we also have that $-\min(\tau) + \delta(Y) = \kappa(a,b,c) - \lambda(Y)$. Thus $\ell^+(Y) = 2(\kappa - \lambda(Y))$. Hence
\[
j'(-T_{a,b}) = j^{(c)}(Y) = \ell^+(Y)/2 + \lambda(Y) = \kappa(a,b,c),
\]
which gives $j^{(c)}(-T_{a,b}) = \kappa(a,b,c)$ if $c$ is odd and $j^{(2)}(-T_{a,b}) = 2 \kappa(a,b,c)$ if $c=2$. We also have $\sigma^{(c)}( T_{a,b} ) = -8 \lambda(Y) = 8 \lambda( \Sigma(a,b,c) )$ and thus we have
\begin{align*}
\theta^{(c)}(T_{a,b}) &= \max\left\{ 0 , \frac{2j'(Y)}{(c-1)} - \frac{\sigma^{(c)}(T_{a,b})}{2(c-1)} \right\} \\
&= \max\left\{ 0 , \frac{ 2\kappa(a,b,c) -4\lambda( \Sigma(a,b,c) )}{(c-1)} \right\} \\
&= \frac{1}{2}(a-1)(b-1)
\end{align*}
where the last line follows from Equation (\ref{equ:kap}).
\end{proof}
\section{Branched covers}\label{sec:branched}
Let $Y = \Sigma( a'_1 , a'_2 , \dots , a'_r)$ be a Brieskorn homology sphere and $p$ a prime such that $p$ divides $a'_1 a'_2 \cdots a'_r$. Without loss of generality we may assume $p$ divides $a'_1$. We set $a_1 = a'_1/p$ and $a_j = a'_j$ for $j > 1$. So $Y = \Sigma(pa_1 , a_2 , \dots , a_r)$. Then $\mathbb{Z}_p$ acts on $Y$ with quotient space $Y_0 = Y/\mathbb{Z}_p = \Sigma(a_1, \dots , a_r)$ and the quotient map $Y \to Y_0$ is a $p$-fold cyclic branched cover.
\begin{theorem}\label{thm:branch}
We have that
\[
\delta(-Y) - \delta^{(p)}_\infty(-Y) \ge {\rm rk}( HF_{red}^+(Y)) - p \, {\rm rk}( HF_{red}^+(Y_0) ).
\]
\end{theorem}
\begin{proof}
Let $W_0$ be the negative definite plumbing bounded by $Y_0$. Since $Y_0$ is an integral homology sphere it follows that $H_1(W_0 ; \mathbb{Z}) = 0$. Let $k \subset Y_0$ denote the branch locus of $Y \to Y_0$. Then $k$ is a knot in $Y_0$. Let $\Sigma \subset W_0$ be the pushoff of a Seifert surface for $k$, so $\Sigma$ is a properly embedded surface in $W_0$ which meets $\partial W_0 = Y_0$ in $k$. Let $W \to W_0$ be the $p$-fold cyclic cover of $W_0$ branched over $\Sigma$. Then $W$ has boundary $Y$ and the $\mathbb{Z}_p$-action on $Y$ extends to $W$. By \cite[Proposition 2.5]{bar}, for any characteristic $c \in H^2(W_0 ; \mathbb{Z})$ there exists a $\mathbb{Z}_p$-invariant spin$^c$-structure $\mathfrak{s}$ on $W$ such that $c_1(\mathfrak{s}) = \pi^*(c)$ in $H^2(W ; \mathbb{Q})$. Now we apply the equivariant Froyshov inequality \cite[Theorem 5.3]{bh} to $W$ giving $\delta_\infty(-Y) + \delta(W , \mathfrak{s}) \le 0$, where
\begin{align*}
\delta(W , \mathfrak{s}) &= \frac{ c_1(\mathfrak{s})^2 - \sigma(W) }{8} \\
&= \frac{ p c^2 - \sigma(W) }{8} \\
&= p \left( \frac{ c^2 - \sigma(W_0) }{8} \right) + \frac{ p\sigma(W_0) - \sigma(W)}{8}.
\end{align*}
The maximum of $(c^2 -\sigma(W_0))/8$ over all characteristics of $H^2( W_0 ; \mathbb{Z})$ equals $\delta(Y_0)$ \cite[Theorem 8.3]{nem}. Hence we obtain
\[
\delta^{(p)}_\infty(-Y) + p \delta(Y_0) + \frac{ p\sigma(W_0) - \sigma(W)}{8} \le 0,
\]
which we may rewrite as
\[
\delta(-Y) - \delta^{(p)}_\infty(-Y) \ge -(\delta(Y) - p \delta(Y_0) ) + \frac{ p\sigma(W_0) - \sigma(W)}{8}.
\]
Next, we claim that $(p \sigma(W_0) - \sigma(W) )/8 = p \lambda(Y_0) - \lambda(Y)$, where $\lambda(Y_0), \lambda(Y)$ are the Casson invariants of $Y_0$ and $Y$. Assuming this claim for the moment, our inquality becomes
\[
\delta(-Y) - \delta^{(p)}_\infty(-Y) \ge -(\delta(Y)+\lambda(Y)) + p( \delta(Y_0) + \lambda(Y_0) ).
\]
But since $HF_{red}^+(Y)$ and $HF_{red}^+(Y_0)$ are concentrated in odd degrees, \cite[Theorem 1.3]{os} gives
\[
{\rm rk}( HF_{red}^+(Y)) = -\delta(Y) - \lambda(Y), \quad {\rm rk}( HF_{red}^+(Y_0)) = -\delta(Y_0) - \lambda(Y_0),
\]
and hence we obtain
\[
\delta(-Y) - \delta^{(p)}_\infty(-Y) \ge {\rm rk}( HF_{red}(-Y)) - p \, {\rm rk}( HF_{red}(-Y_0) ).
\]
It remains to prove the claim that $(p \sigma(W_0) - \sigma(W) )/8 = p \lambda(Y_0) - \lambda(Y)$. From \cite[Theorem 2]{cosa}, we have that
\[
\lambda^{\mathbb{Z}/p}(Y) - p \lambda(Y_0) = \frac{\sigma(W) - p \sigma(W_0)}{8}
\]
where $\lambda^{\mathbb{Z}/p}(Y)$ is the equivariant Casson invariant of $Y$ with respect to the $\mathbb{Z}_p$-action \cite{cosa}. Furthermore \cite[Theorem 3]{cosa} implies that $\lambda^{\mathbb{Z}/p}(Y) = \lambda(Y)$, because $k \subset Y_0 = \Sigma(a_1 , \dots , a_r)$ is a fibre of the Seifert fibration on $Y_0$, so it is a graph knot in the terminology of \cite[\textsection 5]{cosa}. This proves the claim that $\lambda(Y) - p \lambda(Y_0) = (\sigma(W) - p \sigma(W_0))/8$.
\end{proof}
Let $Y,Y_0$ be as in Theorem \ref{thm:branch}. Then by \cite[Theorem 1.1]{kali}, we have an inequality ${\rm rk}( HF_{red}^+(Y)) \ge p \, {\rm rk}( HF_{red}^+(Y_0) )$. Combined with Theorem \ref{thm:branch}, this gives:
\[
p \, {\rm rk}( HF_{red}^+(Y_0) ) \le {\rm rk}( HF_{red}^+(Y)) \le p \, {\rm rk}( HF_{red}^+(Y_0) ) + (\delta(-Y) - \delta^{(p)}_\infty(-Y) ).
\]
In particular, the equality $\delta^{(p)}_\infty(-Y) = \delta(-Y)$ can only happen if ${\rm rk}( HF_{red}^+(Y)) = p \, {\rm rk}( HF_{red}^+(Y_0) )$. Hence we obtain:
\begin{corollary}
Let $Y,Y_0$ be as in Theorem \ref{thm:branch}. If ${\rm rk}( HF_{red}^+(Y)) > p \, {\rm rk}( HF_{red}^+(Y_0) )$ then $\delta^{(p)}_\infty(-Y) < \delta(-Y)$.
\end{corollary}
\section{Free actions}\label{sec:free}
Suppose $p$ does not divide $a_1 a_2 \cdots a_r$. Then the restriction of the $S^1$-action on $Y = \Sigma(a_1 , \dots , a_r)$ acts freely on $Y$. Let $Y_0 = Y/\mathbb{Z}_p$ be the quotient. Then $Y$ is a rational homology sphere. In fact if $Y = M( e_0 , (a_1 , b_1) , \dots , (a_r , b_r))$, then it is easily seen that $Y_0 = M( pe_0 , (a_1 , pb_1) , \dots , (a_r , pb_r) )$. It is easy to see that $H_1(Y ; \mathbb{Z}) \cong \mathbb{Z}_p$ and thus $Y_0$ has exactly $p$ spin$^c$-structures. Further, the pullback to $Y$ of any spin$^c$-structure on $Y_0$ must coincide with the unique spin$^c$-structure on $Y$.
\begin{theorem}\label{thm:free}
For any spin$^c$-structure $\mathfrak{s}_0$ on $Y_0$, we have
\[
\delta^{(p)}_\infty(Y) - \delta(Y) = {\rm rk}( HF_{red}^+(Y) ) - {\rm rk}( HF_{red}^+(Y_0 , \mathfrak{s}_0 ) ).
\]
\end{theorem}
\begin{proof}
Set $G = \mathbb{Z}_p$ and $H^*_G = H^*_G( pt ; \mathbb{F})$. By choosing a $G$-invariant metric on $Y$, we may construct a $S^1 \times G$-equivariant Conley index for $(Y,\mathfrak{s})$, which we denote by $I(Y , \mathfrak{s})$. As shown in \cite[\textsection 3]{lima}, the Conley index $I(Y_0 , \mathfrak{s}_0)$ for $(Y_0 , \mathfrak{s}_0)$ can be identified with the $\mathbb{Z}_p$-fixed point set $I(Y , \mathfrak{s})^{\mathbb{Z}_p}$ of $I(Y , \mathfrak{s})$. Thus we have isomorphisms
\begin{align*}
\widetilde{H}^*_{S^1 \times G}( I(Y , \mathfrak{s})^{\mathbb{Z}_p} ; \mathbb{F} ) & \cong \widetilde{H}^*_{S^1 \times G}( I(Y_0 , \mathfrak{s}_0) ; \mathbb{F}) \\
& \cong \widetilde{H}^*_{S^1}( I(Y_0 , \mathfrak{s}_0) , \mathbb{F}) \otimes_{\mathbb{F}} H^*_G \\
& \cong HSW^*( Y_0 , \mathfrak{s}_0) \otimes_{\mathbb{F}} H^*_G.
\end{align*}
It should be noted that the above isomorphisms only preserve relative gradings. Recall that $H^*_G \cong \mathbb{F}[Q]$, $deg(Q) = 1$ if $p=2$ and $H^*_G \cong \mathbb{F}[R,S]/(R^2)$, $deg(R) = 1$, $deg(S) = 2$ if $p$ is odd. In the case $p=2$, define $S = Q^2$. Let $\mathcal{S} = \{ 1 , S , S^2 , \dots \}$. Then $\mathcal{S}$ is a multiplicative subset of $H^*_{S^1 \times G} = H^*_{S^1 \times G}(pt ; \mathbb{F}) \cong H^*_G[U]$. The localisation theorem in equivariant cohomology \cite[III, Theorem 3.8]{die} applied to the pair $( I(Y , \mathfrak{s} ) , I(Y , \mathfrak{s})^{\mathbb{Z}_p})$ and multiplicative set $\mathcal{S}$ implies that the inclusion $I(Y , \mathfrak{s})^{\mathbb{Z}_p} \to I(Y , \mathfrak{s})$ induces an isomorphism
\begin{equation}\label{equ:lociso}
S^{-1} \widetilde{H}^*_{S^1 \times G}( I(Y , \mathfrak{s}) ; \mathbb{F} ) \to S^{-1} \widetilde{H}^*_{S^1 \times G}( I(Y , \mathfrak{s})^{\mathbb{Z}_p} ; \mathbb{F} ).
\end{equation}
Since $\widetilde{H}^*_{S^1 \times G}( I(Y , \mathfrak{s}) ; \mathbb{F} ) \cong HSW^*_{\mathbb{Z}_p}$ and as we have shown above, $\widetilde{H}^*_{S^1 \times G}( I(Y , \mathfrak{s})^{\mathbb{Z}_p} ; \mathbb{F} ) \cong HSW^*( Y_0 , \mathfrak{s}_0) \otimes_{\mathbb{F}} H^*_G$, we get an isomorphism
\[
S^{-1} HSW^*_{\mathbb{Z}_p}( Y , \mathfrak{s} ) \cong HSW^*(Y_0 , \mathfrak{s}_0) \otimes_{\mathbb{F}} S^{-1} H^*_G.
\]
This is an isomorphism of relatively graded $H^*_{S^1 \times G}$-modules.
For the rest of the proof we restrict to the case that $p$ is odd. The proof in the case that $p=2$ is similar. We have
\[
HSW^*(Y_0 , \mathfrak{s}_0) \cong HF^+_*(Y_0 , \mathfrak{s}_0) \cong \mathbb{F}[U] \theta_0 \oplus HF_{red}^+(Y_0 , \mathfrak{s}_0)
\]
for some $\theta_0$. Combined with (\ref{equ:lociso}), we have an isomorphism
\begin{equation}\label{equ:lociso1}
S^{-1} HSW^*_{\mathbb{Z}_p}( Y , \mathfrak{s} ) \cong \frac{\mathbb{F}[U,R,S,S^{-1}]}{(R^2)} \theta_0 \oplus \frac{HF_{red}^+(Y_0, \mathfrak{s}_0)[R,S,S^{-1}]}{(R^2)}.
\end{equation}
On the other hand, the proof of Proposition \ref{prop:deltaprop} gives an isomorphism
\[
HSW^*_{\mathbb{Z}_p}(Y , \mathfrak{s}) \cong E_\infty(Y) \cong \frac{\mathbb{F}[U,R,S]}{(R^2)} \theta \oplus E_\infty(Y)_{red}
\]
under which the $\mathbb{F}[U]$-module structure is given by an endomorphism of the form $\widehat{U} = U_{(0,2)} + U_{(1,1)} + \cdots $ with $U_{(0,2)} = U$. Localising with respect to $S$ gives an isomorphism
\[
S^{-1} HSW^*_{\mathbb{Z}_p}(Y , \mathfrak{s}) \cong E_\infty(Y) \cong \frac{\mathbb{F}[U,R,S,S^{-1}]}{(R^2)} \theta \oplus S^{-1} E_\infty(Y)_{red}.
\]
Define $W \subseteq S^{-1} HSW^*_{\mathbb{Z}_p}( Y , \mathfrak{s} ) )$ to be the set of $x \in S^{-1} HSW^*_{\mathbb{Z}_p}( Y , \mathfrak{s} ) )$ such that for each $j \ge 0$, there exists a $k \ge 0$ for which $U^k x \in \mathcal{F}_j(Y)$. The proof of Proposition \ref{prop:deltaprop} demonstrates that $W \cong S^{-1} E_\infty(Y)_{red}$. On the other hand, the isomorphism (\ref{equ:lociso1}) clearly shows that $W \cong HF_{red}^+(Y_0, \mathfrak{s}_0)[R,S,S^{-1}]/(R^2)$. Combining these, we have an isomorphism
\[
S^{-1} E_\infty(Y)_{red} \cong HF_{red}^+(Y_0, \mathfrak{s}_0)[R,S,S^{-1}]/(R^2).
\]
In any fixed degree $j$, the rank of $(HF_{red}^+(Y_0, \mathfrak{s}_0)[R,S,S^{-1}]/(R^2))^j$ is equal to ${\rm rk}( HF_{red}^+(Y_0 , \mathfrak{s}_0))$, hence the same is true of $S^{-1} E_\infty(Y)_{red}$. From \cite[Lemma 5.7]{bh}, we have that $S \colon E_{r}^{p,q}(Y) \to E_r^{p+2,q}(Y)$ is an isomorphism for all large enough $p$. Hence for large enough $p$, $E_r^{2p,*}(Y)$ is independent of $p$ and we denote the resulting group by $M_r^*$. We similarly define $M_\infty^*$. Clearly the rank of $(S^{-1} E_\infty(Y)_{red})^j$ for any $j$ equals ${\rm rk}( M_\infty )_{red}$. So we have proven that
\[
{\rm rk}(HF_{red}^+(Y_0 , \mathfrak{s}_0) ) = {\rm rk}(M_\infty)_{red}.
\]
Set $s_r = {\rm rk}(M_r)_{red}$. It follows from \cite[Lemma 5.8]{bh} that $(M_{r+1})_{red}$ is a subquotient of $(M_r)_{red}$. Hence the sequence $s_2 , s_3, \dots $ is decreasing and equals ${\rm rk}(M_\infty)_{red}$ for sufficiently large $r$. Furthermore, $M_2 \cong HF^+(Y)$, so $s_2 = {\rm rk}( HF_{red}^+(Y) )$.
Recall from the proof of Proposition \ref{prop:deltaprop} that
\[
E_r(Y) \cong \frac{\mathbb{F}[U,R,S]}{(R^2)} U^{m_r}\theta \oplus E_r(Y)_{red}
\]
for some increasing sequence $0 = m_0 \le m_1 \le \dots$. Suppose $m_{r+1} > m_r$. Then $d_r( U^{m_r + j} \theta ) \neq 0$ for $0 \le j \le m_{r+1} - m_r$. Notice that $U^{m_r+j}\theta$ has bi-degree $(0 , a)$ where $a = 2m_r + 2j + d(Y)$ is even, hence $d_r( U^{m_r + j} \theta )$ has bi-degree $( r , a + (1-r) )$. But all elements of $E_r(Y)_{red}$ have bi-degree $(u,v)$ with $v$ odd (because $E_r(Y)_{red}$ is a subquotient of $HF_{red}^+(Y)[ R , S ]/(R^2)$ and $HF_{red}^+(Y)$ is concentrated in odd degrees). Hence $a + (1-r)$ must be odd, which means that $r$ is even. So we can regard $d_r$ as mapping into $(M_r)_{red}$. Hence the rank of $(M_{r+1})_{red}$ is at least $m_{r+1} - m_{r}$ less than the rank of $(M_r)_{red}$, that is,
\begin{equation}\label{equ:srmr}
s_{r} - s_{r+1} \ge m_{r+1} - m_r.
\end{equation}
We claim that this inequality is actually an equality. Equivalently $M_{r+1}$ is the quotient of $M_r$ by the span of $\{ d_r( U^{m_r + j} ) \}$, $0 \le j \le m_{r+1} - m_r$. This is also equivalent to saying that $d_r( x ) = 0$ for all $r \ge 2$ and all $x \in E_r(Y)_{red}$. Consider a non-zero homogeneous element $x \in E_r(Y)_{red}^{a,b}$. Since $E_r(Y)_{red}$ is a subquotient of $HF_{red}^+(Y)[ R , S ]/(R^2)$ and $HF_{red}^+(Y)$ is concentrated in odd degrees, we have that $b$ is odd. Then $d_r(x)$ has bi-degree $(a+r , b+1-r)$. If $d_r(x) \neq 0$, then $b+1-r$ must be off and so $r$ must be even. Thus in order to prove the claim, it is sufficient to show that $d_r = 0$ for all even $r$.
Consider the equivariant Seiberg--Witten--Floer cohomology of $Y$ with integral coefficients $HSW^*_{\mathbb{Z}_p}( Y ; \mathbb{Z})$. This is a module over $H^*_{\mathbb{Z}_p}( pt ; \mathbb{Z} ) \cong \mathbb{Z}[ S ]/( pS )$ where $deg(S) = 2$. The key point to observe here is that $H^*_{\mathbb{Z}_p}( pt ; \mathbb{Z} )$ is concentrated in even degrees. There is a spectral sequence $(E_r^{p,q}(Y ; \mathbb{Z}) , d_r)$ and a filtration $\{ \mathcal{F}_j(Y ; \mathbb{Z}) \}$ such that $E_\infty(Y ; \mathbb{Z})$ is the associated graded module of the filtration. The mod $p$ reduction map $\mathbb{Z} \to \mathbb{Z}_p = \mathbb{F}$ induces a morphism of $HSW^*( Y ; \mathbb{Z}) \to HSW^*(Y ; \mathbb{F})$. There is also a reduction map of the equivariant Floer cohomologies, the filtrations and the spectral sequences. The results of \cite{os2} also hold with integer coefficients so
\[
HF^+(Y ; \mathbb{Z}) \cong \mathbb{Z}[U]_{d(Y)} \oplus HF_{red}^+(Y ; \mathbb{Z})
\]
and $HF_{red}^+(Y ; \mathbb{Z})$ is concentrated in odd degrees. By the universal coefficient theorem $HF_{red}^+(Y ; \mathbb{F}) \cong HF_{red}^+(Y ; \mathbb{Z}) \otimes_{\mathbb{Z}} \mathbb{F}$. Furthermore, we have that
\begin{align*}
E_2(Y ; \mathbb{Z}) &\cong H^*( \mathbb{Z}_p ; HSW^*(Y ; \mathbb{Z} )) \\
& \cong \frac{\mathbb{Z}[U , S]}{(pS)} \theta \oplus \frac{HF_{red}^+(Y ; \mathbb{Z})[S]}{(pS)}.
\end{align*}
Clearly every homogeneous element in $\mathbb{Z}[U , S]/(pS) \theta$ has bi-degree $(a,b)$ where $a$ and $b$ are even and every homogeneous element in $HF_{red}^+(Y ; \mathbb{Z})[S]/(pS)$ has bi-degree $(a,b)$ where $a$ is even and $b$ is odd. From this and the fact that the image of $d_r \colon E_r(Y ; \mathbb{Z} ) \to E_r(Y ; \mathbb{Z})$ is contained in $E_r(Y ; \mathbb{Z})_{red}$ it follows easily that $d_r \colon E_r(Y ; \mathbb{Z} ) \to E_r( Y ; \mathbb{Z})$ is zero for even $r$.
Now using induction on $r$ one shows the following properties hold: (1) any $x \in E_r^{a,b}(Y)$ with $a$ even is in the image of the reduction map $E_r^{a,b}(Y ; \mathbb{Z}) \to E_r^{a,b}(Y)$, (2) every $x \in E_r^{a,b}(Y)$ with $a$ odd is of the form $x = Ry$ where $y$ is in the image of the reduction map $E_r^{a,b}(Y ; \mathbb{Z}) \to E_r^{a,b}(Y)$ and (3) $d_r \colon E_r(Y) \to E_r(Y)$ is zero for even $r$. In particular, this proves the claim that $d_r \colon E_r(Y) \to E_r(Y)$ is zero for all even $r$ and hence the inequality (\ref{equ:srmr}) is actually an equality:
\begin{equation}\label{equ:srmr2}
s_{r} - s_{r+1} = m_{r+1} - m_r.
\end{equation}
Let $s = {\rm rk}( (M_\infty)_{red} ) = {\rm rk}( HF_{red}^+(Y_0 , \mathfrak{s}_0) )$ and let $m = \lim_{r \to \infty} m_r$. Summing (\ref{equ:srmr2}) from $r=2$ to infinity gives $s_2 - s = m - m_2$. But $m_2 = 0$ and $s_2 = {\rm rk}( HF_{red}^+(Y))$, so we get $m = {\rm rk}( HF_{red}^+(Y)) - {\rm rk}( HF_{red}^+(Y_0 , \mathfrak{s}_0))$. But recall from the proof of Proposition \ref{prop:deltaprop} that $\delta_\infty^{(p)}(Y) = \delta(Y) + m$, hence we get
\[
\delta_\infty^{(p)}(Y) = \delta(Y) + {\rm rk}( HF_{red}^+(Y)) - {\rm rk}( HF_{red}^+(Y_0 , \mathfrak{s}_0)).
\]
\end{proof}
We will take $\mathfrak{s}_0$ to be the restriction to $Y_0$ of the canonical spin$^c$-structure on the negative definite plumbing which $Y_0$ bounds. With this choice of spin$^c$-structure, the computation of $HF^+(-Y_0 , \mathfrak{s}_0)$ is easily obtained through the graded roots algorithm \cite{nem}. Let $\Delta_p(n)$ denote the delta function for $(Y_0 , \mathfrak{s}_0)$. Then (since $e_0 < 0$) we have
\begin{equation}\label{equ:deltap}
\Delta_p(n) = 1 + np e_0 - \sum_{j=1}^{r} \left\lceil \frac{npb_j}{a_j} \right\rceil
\end{equation}
\begin{proposition}\label{prop:deltap}
We have that $\Delta_p(n) \ge 0$ for $n > N/p$. In particular, if $p > N$, then $HF_{red}^+(-Y_0 , \mathfrak{s}_0 ) = 0$.
\end{proposition}
\begin{proof}
Following \cite[\textsection 3]{caka}, we let $f(x) = \lceil x \rceil - x$. Then from Equations (\ref{equ:integral}) and (\ref{equ:deltap}), we see that
\[
\Delta_p(n) = 1 + \frac{np}{a_1 \cdots a_r} - \sum_{j=1}^r f\left( \frac{ npb_j}{a_j} \right).
\]
For any integer $m$ we have $f(m/a_j) \le 1-1/a_j$, hence
\begin{equation}\label{equ:deltaest}
\Delta_p(n) \ge 1 + \frac{np}{a_1 \cdots a_r} - r + \sum_{j=1}^{r} \frac{1}{a_j}.
\end{equation}
Now suppose that $pn > N$. Thus
\[
pn > a_1 \cdots a_n \left( (r-2) - \sum_{j=1}^r \frac{1}{a_j} \right).
\]
Re-arranging, we get
\[
(1-r) + \frac{np}{a_1 \cdots a_r} + \sum_{j=1}^{r} \frac{1}{a_j} > -1.
\]
Combined with (\ref{equ:deltaest}), we get $\Delta_p(n) > -1$. But $\Delta_p$ is integer-valued, so $\Delta_p(n) \ge 0$.
\end{proof}
\begin{lemma}\label{lem:rk}
We have that ${\rm rk}( HF_{red}^+(-Y) ) > {\rm rk}( HF_{red}^+( -Y_0 , \mathfrak{s}_0 ))$ unless $r=3$ and up to reordering $(a_1,a_2,a_3)$ is one of the following:
\[
(2,3,5), (2,3,7), (2,3,11), (2,3,13), (2,3,17), (2,5,7), (2,5,9), (3,4,5), (3,4,7).
\]
\end{lemma}
\begin{proof}
We assume that $Y$ is not $\Sigma(2,3,5)$ and hence $N > 0$. Let $N_p = \lfloor N/p \rfloor$. By Proposition \ref{prop:deltap} we only need to consider $\Delta_p(x)$ for $x \in [0 , N_p]$.
In the proof we make use of abstract delta sequences \cite[\textsection 3]{kali}. Let $P_p = G \cap [ 0 , N_p]$. Consider the map $\phi \colon P_p \to [0,N]$ given by $\phi(n) = np$. From Equation (\ref{equ:deltap}) it is immediately clear that $\Delta(np) = \Delta_p(n)$, hence $\phi$ identifies the delta sequence $( P_p , \Delta_p)$ with a subsequence of the delta sequence $( P , \Delta)$. This gives the inequality ${\rm rk}( HF_{red}^+(-Y) ) \ge {\rm rk}( HF_{red}^+( -Y_0 , \mathfrak{s}_0 )$ by \cite[Proposition 3.5]{kali}. To refine this to a strict inequality we will show that (except for the cases listed in the statement of the lemma) there exists an $x \in P \setminus \phi(P_p)$ such that $2x \le N$. In this case $\{ x , N -x \}$ defines a delta subsequence of $( P , \Delta )$ disjoint from $\phi(P_p)$ and then by \cite[Corollary 3.12]{kali} we obtain a strict inequality ${\rm rk}( HF_{red}^+(-Y) ) > {\rm rk}( HF_{red}^+( -Y_0 , \mathfrak{s}_0 )$, as the module $\mathbb{H}_{red}$ corresponding to $\{ x , N - x \}$ has rank $1$.
We seek an $x \in P \setminus \phi(P_p)$ such that $2x \le N$. Equivalently $x \in P$, $p$ does not divide $x$ and $2x \le N$. Reorder $a_1, \dots ,a_r$ such that $a_1 > a_2 > \cdots > a_r$. Consider $x = a_2 a_3 \cdots a_r$. It follows that $p$ does not divide $x$ as $p$ is coprime to $a_2, \dots , a_r$ and we also have that $x \in P$. Hence we obtain a strict rank inequality provided that $2x \le N$. Suppose on the contrary that $2x > N$, that is,
\[
2a_2 \cdots a_r > a_1 \cdots a_r \left( (r-2) - \sum_{j=1}^{r} \frac{1}{a_j} \right).
\]
Re-arranging, this is equivalent to
\begin{equation}\label{equ:condition}
\frac{3}{a_1} + \frac{1}{a_2} + \frac{1}{a_3} + \cdots + \frac{1}{a_r} \ge (r-2).
\end{equation}
Since $a_j \ge 2$ for all $j$ and at most one can be equal to $2$, we get that
\[
\frac{3}{a_1} + \frac{1}{a_2} + \cdots + \frac{1}{a_r} < \frac{3}{a_1} + \frac{(r-1)}{2},
\]
hence
\[
\frac{3}{a_1} + \frac{(r-1)}{2} > (r-2).
\]
From $a_1 > a_2 > \cdots > a_r \ge 2$, we see that $a_r \ge r+1$, hence
\[
\frac{3}{r+1} + \frac{(r-1)}{2} > (r-2)
\]
which simplifies to $(r+1)(r-3) < 6$. If $r \ge 5$, then $(r+1)(r-3) \ge 12$, so this leaves only the cases $r=3,4$.
If $r=4$, then since $a_4 \ge 2$, $a_3 \ge 3$, $a_2 \ge 4$, $a_1 \ge 5$, we have
\[
\frac{3}{a_1} + \frac{1}{a_2} + \frac{1}{a_3} + \frac{1}{a_4} \le \frac{3}{5} + \frac{1}{4} + \frac{1}{3} + \frac{1}{2} = \frac{101}{60} < 2.
\]
Hence (\ref{equ:condition}) can not be satisfied with $r=4$.
Now suppose that $r=3$. If $a_3 \ge 4$, then $a_2 \ge 5$, $a_1 \ge 6$ and so
\[
\frac{3}{a_1} + \frac{1}{a_2} + \frac{1}{a_3} \le \frac{3}{6} + \frac{1}{5} + \frac{1}{4} = \frac{19}{20} < 1
\]
and so (\ref{equ:condition}) can only be satisfied if $a_3 = 2$ or $3$.
If $a_3 = 3$ and $a_4 \ge 5$, then $a_1 \ge 7$ (since $a_1$ must be coprime to $a_3 = 3$) and so
\[
\frac{3}{a_1} + \frac{1}{a_2} + \frac{1}{a_3} \le \frac{3}{7} + \frac{1}{5} + \frac{1}{3} = \frac{101}{105} < 1.
\]
So (\ref{equ:condition}) implies that $a_4 = 4$. Then since $3/8 + 1/4 + 1/3 = 23/24 < 1$, it follows that $a_1 < 8$. Hence if $a_3 = 3$, then $(a_1, a_2, a_3) = (5,4,3)$ or $(7,4,3)$.
Lastly, suppose that $a_3 = 2$. If $a_2 > 5$, then $a_2 \ge 7$ and $a_1 \ge 9$ (since $a_1,a_2$ must be coprime to $a_3$). So
\[
\frac{3}{a_1} + \frac{1}{a_2} + \frac{1}{a_3} \le \frac{3}{9} + \frac{1}{7} + \frac{1}{2} = \frac{41}{42} < 1.
\]
So $a_2 \le 5$, hence $a_2 = 3$ or $5$.
If $a_3 = 2$ and $a_2 = 5$, then (\ref{equ:condition}) implies $a_1 \le 10$, hence $(a_1, a_2, a_3) = (7, 5, 2)$ or $(9, 5, 2)$.
If $a_3 = 2$ and $a_2 = 3$, then (\ref{equ:condition}) implies $a_1 \le 18$, hence $(a_1,a_2,a_3)$ is one of $(5, 3, 2), (7, 3, 2), (11, 3, 2), (13, 3, 2), (17, 3, 2)$.
\end{proof}
\begin{table}
\begin{equation*}
\renewcommand{\arraystretch}{1.4}
\begin{tabular}{|c|c|c|c|c|}
\hline
$(a_1,a_2,a_3)$ & $N$ & ${\rm rk}( HF_{red}^+(-Y) )$ & $p$ & ${\rm rk}( HF_{red}^+(-Y_0) )$ \\
\hline
$(2,3,7)$ & $1$ & $1$ & none & \\
\hline
$(2,3,11)$ & $5$ & $1$ & $5$ & $1$ \\
\hline
$(2,3,13)$ & $7$ & $2$ & $5$ & $0$ \\
& & & $7$ & $1$ \\
\hline
$(2,3,17)$ & $11$ & $2$ & $5$ & $1$ \\
& & & $7$ & $0$ \\
& & & $11$ & $1$ \\
\hline
$(2,5,7)$ & $11$ & $2$ & $3$ & $0$ \\
& & & $11$ & $1$ \\
\hline
$(2,5,9)$ & $17$ & $2$ & $7$ & $1$ \\
& & & $11$ & $0$ \\
& & & $13$ & $0$ \\
& & & $17$ & $1$ \\
\hline
$(3,4,5)$ & $13$ & $2$ & $2$ & $0$ \\
& & & $7$ & $0$ \\
& & & $11$ & $0$ \\
& & & $13$ & $1$ \\
\hline
$(3,4,7)$ & $23$ & $2$ & $2$ & $1$ \\
& & & $5$ & $0$ \\
& & & $11$ & $1$ \\
& & & $13$ & $0$ \\
& & & $17$ & $0$ \\
& & & $19$ & $0$ \\
& & & $23$ & $1$ \\
\hline
\end{tabular}
\end{equation*}
\caption{}\label{fig}
\end{table}
\begin{theorem}\label{thm:free2}
We have that ${\rm rk}( HF_{red}^+(-Y) ) > {\rm rk}( HF_{red}^+( -Y_0 , \mathfrak{s}_0 ))$ except in the following cases:
\begin{itemize}
\item[(1)]{$Y = \Sigma(2,3,5)$ and $p$ is any prime.}
\item[(2)]{$Y = \Sigma(2,3,11)$ and $p=5$.}
\end{itemize}
In case (1) we have ${\rm rk}( HF_{red}^+(-Y) ) = {\rm rk}( HF_{red}^+( -Y_0 , \mathfrak{s}_0 )) = 0$ and in case (2) we have ${\rm rk}( HF_{red}^+(-Y) ) = {\rm rk}( HF_{red}^+( -Y_0 , \mathfrak{s}_0 )) = 1$.
\end{theorem}
\begin{proof}
By Lemma \ref{lem:rk}, we only need to consider the case that $r=3$ and up to reordering $(a_1,a_2,a_3)$ is one of:
\[
(2,3,5), (2,3,7), (2,3,11), (2,3,13), (2,3,17), (2,5,7), (2,5,9), (3,4,5), (3,4,7).
\]
In the case $Y = \Sigma(2,3,5)$ we have that ${\rm rk}( HF_{red}^+(-Y) ) = 0$ and $N < 0$. Hence we also have ${\rm rk}( HF_{red}^+(-Y_0) ) = 0$ by Proposition \ref{prop:deltap}.
In the remaining cases we have ${\rm rk}( HF_{red}^+(-Y) ) > 0$ and $N > 0$. Hence if $p > N$ then ${\rm rk}( HF_{red}^+(-Y) ) > {\rm rk}( HF_{red}^+(-Y_0) ) = 0$, by Proposition \ref{prop:deltap}. So we only need to consider primes such that $p \le N$ and coprime to $a_1, a_2, a_3$. This leaves only finitely many cases of $4$-tuples $\{ (a_1,a_2,a_3, p) \}$ to consider. We check each of these cases by directly computing the ranks of $HF_{red}^+(-Y)$ and $HF_{red}^+(-Y_0)$. The results are shown in Table \ref{fig}. By inspection we see that ${\rm rk}( HF_{red}^+(-Y) ) > {\rm rk}( HF_{red}^+( -Y_0 , \mathfrak{s}_0 ))$ in all cases except for $Y = \Sigma(2,3,11)$ and $p=5$.
\end{proof}
\begin{theorem}
Let $Y = \Sigma(a_1 , a_2 , \dots , a_r)$ be a Brieskorn homology sphere and let $p$ be a prime not dividing $a_1 \cdots a_r$. Then $\delta_\infty^{(p)}(Y) > \delta(Y)$ except in the following cases:
\begin{itemize}
\item[(1)]{$Y = \Sigma(2,3,5)$ and $p$ is any prime.}
\item[(2)]{$Y = \Sigma(2,3,11)$ and $p=5$.}
\end{itemize}
In both cases we have $\delta_\infty^{(p)}(Y) = \delta(Y) = 1$.
\end{theorem}
\begin{proof}
Theorems \ref{thm:free} and \ref{thm:free2} imply that $\delta_\infty^{(p)}(Y) \ge \delta(Y)$ with equality only in the cases listed. In these cases, we have $\delta_\infty^{(p)}(Y) = \delta(Y)$ and $\delta(Y) = 1$ \cite[\textsection 8]{os}.
\end{proof}
\bibliographystyle{amsplain}
|
\section{Introduction} In this text we establish three main results.
Firstly, we establish a distributional version of Sobolev inequality written in terms of Hardy-Littlewood submajorization (see Theorem \ref{dist sobolev critical}). We demonstrate that our version of this inequality is optimal in the sense that it cannot be improved in terms of the distribution function (see Proposition \ref{estimate from below lemma}).
Secondly, we show that our version of Sobolev inequality yields the optimal Sobolev embedding theorem (see Proposition \ref{sobolev critical all dimension} and Corollary \ref{optimal embedding theorem}). These results enhance earlier results due to Hansson \cite{Hansson}, Brezis and Wainger \cite{Brezis-Wainger}, and Cwikel and Pustylnik \cite{Cwikel-Pustylnik}.
Thirdly, we show that Solomyak's version of Cwikel-Solomyak estimate given in \cite{Solomyak1995} is optimal in the class of Orlicz spaces (see Corollary \ref{optimality within orlicz}).
We explain the details of our contribution in Section \ref{known section} below after explaining the notations.
\section{Preliminaries}
\subsection{Symmetric function spaces}
For detailed information about decreasing rearrangements and symmetric spaces briefly discussed below, we refer the reader to the standard textbook \cite{KPS} (see also \cite{LT2, LSZ-book}). Let $(\Omega,\nu)$ be a measure space. Let $S(\Omega,\nu)$ be the collection of all $\nu$-measurable functions on $\Omega$ such that, for some $n\in\mathbb{N},$ the function $|f|\chi_{\{|f|>n\}}$ is supported on a set of finite measure. For every $f\in S(\Omega,\nu)$ one can define the notion of decreasing rearrangement of $f$ (denoted by $\mu(f)$). This is a positive decreasing function on $\mathbb{R}_+$ equimeasurable with $|f|.$
Let $E(\Omega,\nu)\subset S(\Omega,\nu)$ and let $\|\cdot\|_E$ be Banach norm on $E(\Omega,\nu)$ such that
\begin{enumerate}[{\rm (1)}]
\item if $f\in E(\Omega,\nu)$ and $g\in S(\Omega,\nu)$ be such $|g|\leq|f|,$ then $g\in E(\Omega,\nu)$ and $\|g\|_E\leq\|f\|_E;$
\item if $f\in E(\Omega,\nu)$ and $g\in S(\Omega,\nu)$ be such $\mu(g)=\mu(f),$ then $g\in E(\Omega,\nu)$ and $\|g\|_E=\|f\|_E;$
\end{enumerate}
We say that $(E(\Omega,\nu),\|\cdot\|_E)$ (or simply $E$) is a symmetric Banach function space (symmetric space, for brevity). The classical example of such spaces is given by Lebesgue $L_p$-spaces,
($L_p(\Omega,\nu), \|\cdot\|_p)$, $1\leq p\leq \infty.$
If $\Omega=\mathbb{R}_+,$ then the function
$$t\to\|\chi_{(0,t)}\|_E,\quad t>0,$$
is called the {\it fundamental} function of $E.$ Similar definition is available when $\Omega$ is an interval or an arbitrary $\sigma$-finite measure space. The concrete examples of measure spaces $(\Omega, \nu)$ considered in this paper are $d$-dimensional torus $\mathbb{T}^d$ (equipped with Haar measure), $\mathbb{R}_+,$ $\mathbb{R}^d$ (equipped with Lebesgue measure), their measurable subsets and compact $d$-dimensional Riemannian manifolds $(X,g).$
Among concrete symmetric spaces used in this paper are $L_p$-spaces, Orlicz, Lorentz spaces and Marcinkiewicz spaces. Given a convex function $M$ on $\mathbb{R}_+$ (continuous at $0$) such that $M(0)=0,$ Orlicz space $L_M(\Omega,\nu)$ is defined by setting
$$L_M(\Omega,\nu)=\Big\{f\in S(\Omega,\nu):\ M(\frac{|f|}{\lambda})\in L_1(\Omega,\nu)\mbox{ for some }\lambda>0\Big\}.$$
We equip it with a Banach norm
$$\|f\|_{L_M}=\inf\Big\{\lambda>0:\ \Big\|M(\frac{|f|}{\lambda})\Big\|_1\leq 1\Big\}.$$
We refer the reader to \cite{KrasRut, KPS, LT2} for further information about Orlicz spaces.
Given a concave increasing function $\psi,$ Lorentz space $\Lambda_{\psi}$ is defined by setting
$$\Lambda_{\psi}(\Omega,\nu)=\Big\{f\in S(\Omega,\nu):\ \int_0^{\infty}\mu(s,f)d\psi(s)<\infty\Big\}.$$
Marcinkiewicz space $\mathtt{M}_{\psi}$ is defined by setting
$$\mathtt{M}_{\psi}(\Omega,\nu)=\Big\{f\in S(\Omega,\nu):\ \sup_{t>0}\frac1{\psi(t)}\int_0^t\mu(s,f)ds<\infty\Big\}.$$
Spaces $L_{2,1}$ and $L_{2,\infty}$ are Lorentz and Marcinkiewicz space with $\psi(t)=t^{\frac12},$ $t>0$ respectively (see e.g. \cite{LT2}).
{\color{blue} All mentioned examples are closed with respect to the Hardy-Littlewood submajorization. The latter is a pre-order defined on $L_1+L_{\infty}$ by setting
$$\int_0^t\mu(s,y)ds\leq\int_0^t\mu(s,x)ds,\quad t>0,$$
and is denoted by $y\prec\prec x.$}
Let us fix throughout this paper concave functions
$$\psi(t)=\frac1{\log(\frac{e}{t})},\quad {\rm and}\quad \phi(t)=t\log(\frac{e}{t}), \quad t\in(0,1)$$ and consider Orlicz functions $$M(t)=t\log(e+t),\quad{\rm and}\quad G(t)=e^t-1,\quad t>0.$$
Throughout this text, the interplay between Lorentz spaces $\Lambda_{\psi}(0,1)$ and $\Lambda_{\phi}(0,1)$, Orlicz spaces $L_M(0,1)$ and $L_G(0,1)$, and Marcinkiewicz spaces $\mathtt{M}_{\psi}(0,1)$ and $\mathtt{M}_{\phi}(0,1)$ play a fundamental role. The space $\mathtt{M}_{\psi}$ (respectively, the space $\Lambda_{\phi}$) is the largest (respectively, the smallest) symmetric Banach function space with the fundamental function $\phi.$
First of all, we observe that the spaces $\mathtt{M}_{\phi}(0,1)$ and $L_G(0,1)$ coincide (the corresponding norms are equivalent). Indeed, the fundamental functions of those spaces coincide, hence $L_G(0,1)\subset\mathtt{M}_{\phi}(0,1).$ On the other hand, $\phi'\in L_G(0,1)$ and, therefore, $\mathtt{M}_{\phi}(0,1)\subset L_G(0,1).$ This fact was first observed in \cite{AstSukIsrJ}.
By invoking K\"othe duality (see \cite{KPS} and \cite{LT2}), we infer that the spaces $\Lambda_\phi(0,1)$ and $L_M(0,1)$ also coincide and (the corresponding norms are equivalent).
In the statement of our main result, we use the notion of $2$-convexification of the symmetric space. Given a symmetric space $E(\Omega,\nu),$ we set
$$E^{(2)}(\Omega,\nu)=\{f\in S(\Omega,\nu):\ |f|^2\in E(\Omega,\nu)\},$$
$$\|f\|_{E^{(2)}}=\big\||f|^2\big\|_E^{\frac12},\quad f\in E^{(2)}(\Omega,\nu).$$
More details on the $2$-convexification of a function space may be found in the book \cite{LT2}.
We also need a definition of dilation operator $\sigma_u,$ $u>0,$ which acts on $S(\mathbb{R})$ by the formula
$$(\sigma_uf)(t)=f(\frac{t}{u}),\quad f\in S(\mathbb{R}).$$
We frequently use the mapping $r_d:\mathbb{R}^d\to\mathbb{R}_+$ defined by the formula
$$r_d(t)=|t|^d,\quad t\in\mathbb{R}^d.$$
\section{Connection with earlier results}\label{known section}
\subsection{Connection with Cwikel-type estimates}
Let $d\in\mathbb{N}$ and let $f$ be a measurable function on $d$-dimensional torus $\mathbb{T}^d$ and let $M_f$ be an (unbounded) multiplication operator by $f$ on $L_2(\mathbb{T}^d).$ Let $\Delta_{\mathbb{T}^d}$ be the torical Laplacian. In Proposition \ref{torus boundedness} we show that (symmetrized) Cwikel operator
$$(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_f(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}$$
is bounded for $f\in\mathtt{M}_{\psi}(\mathbb{T}^d).$
Symmetrized Cwikel estimate in the weak-trace ideal $\mathcal{L}_{1,\infty}$ as well as in the ideal $\mathcal{M}_{1,\infty}$ are important in Non-commutative Geometry and in Mathematical Physics. For the background concerning weak trace ideals $\mathcal{L}_{p,\infty}$, $1\leq p<\infty$ and Marcinkie\-wicz ideal $\mathcal{M}_{1,\infty}$ we refer to \cite{LSZ-book} (see also subsection \ref{optimality-subsection} below).
Symmetrized Cwikel estimate in $\mathcal{L}_{1,\infty}$ (on the torus) for {\it even} $d$ appeared in the foundational papers \cite{Solomyak1994,Solomyak1995}. The estimate there was given in terms of the Orlicz norm $\|\cdot\|_{L_M}$ (which is equivalent to $\|\cdot\|_{\Lambda_{\phi}}$). Recently, symmetrized Cwikel estimates in the ideal $\mathcal{M}_{1,\infty}$ were established (on the Euclidean space) in \cite{LSZ-last-kalton}. The estimate was given in terms of the Lorentz norm
$$f\to\|\mu(f)\chi_{(0,1)}\|_{\Lambda_{\phi}(0,1)}+\|f\|_{L_1(\mathbb{R}^d)}.$$
The surprising fact that $\Lambda_\phi(0,1)=L_M(0,1)$ demonstrates the convergence of those totally unrelated approaches.
The results in this paper complement the results cited in the preceding paragraph. Indeed, Cwikel operator belongs to $\mathcal{L}_{\infty}$ for $f\in\mathtt{M}_{\psi}$ and to $\mathcal{L}_{1,\infty}$ for $f\in\Lambda_{\phi}.$ This opens an avenue (using interpolation methods) to determine the least ideal to which the Cwikel operator belongs.
\subsection{Connection with Sobolev-type estimates}
The following Sobolev inequality is customarily credited to \cite{Hansson} and \cite{Brezis-Wainger}. Actually, none of those papers contain a complete proof or even a clear-cut statement. The proof is available in \cite{Cwikel-Pustylnik}. For a notion of Sobolev space on $\Omega\subset\mathbb{R}^d$ we refer the reader to Chapter 7 in \cite{Adams}.
\begin{thm}\label{thm 1} Let $d$ be even and let $\Omega$ be a bounded domain in $\mathbb{R}^d$ (conditions apply). There exists a constant $c_{\Omega}$ such that
$$\|u\|_{\Lambda_{\psi}^{(2)}}\leq c_{\Omega}\|u\|_{W^{\frac{d}{2},2}},\quad u\in W^{\frac{d}{2},2}(\Omega).$$
\end{thm}
This result is optimal due to the following theorem (proved in \cite{Cwikel-Pustylnik}).
\begin{thm}\label{thm 2} Let $d$ be even and let $\Omega$ be a bounded domain in $\mathbb{R}^d$ (conditions apply). If $X$ be a Banach symmetric function space on $\Omega$ such that
$$\|u\|_X\leq c_{\Omega}\|u\|_{W^{\frac{d}{2},2}},\quad u\in W^{\frac{d}{2},2}(\Omega)$$
for some constant $c_{\Omega}$,
then $\Lambda_{\psi}^{(2)}\subset X.$
\end{thm}
In the next sections, we prove the assertions which substantially strengthen theorems above and extend them to arbitrary dimensions $d\ge 1$. In the course of the proof, we also recast Sobolev inequality using distribution functions. Those results are very much inspired by \cite{Cwikel-Pustylnik} {\color{blue} and our technique is a substantial improvement of that in \cite{Cwikel-Pustylnik}.}
\section{Distributional Sobolev-type inequality}
In this section, we introduce Hardy-type operator $T$ (see e.g. \cite{KPS}) and employ it as a technical tool for our distributional version of Sobolev inequality. In the second subsection, we show that our result is optimal.
\subsection{Distributional Sobolev-type inequality}
Let $T:L_2(0,1)\to L_2(0,1)$ be the operator defined by the formula
$$(Tx)(t)=t^{-\frac12}\int_0^tx(s)ds+\int_t^1\frac{x(s)ds}{s^{\frac12}},\quad x\in L_2(0,1).$$
The boundedness of $T:L_2(0,1)\to L_2(0,1)$ is guaranteed by the fact that $T$ is actually a Hilbert-Schmidt operator. Indeed, its integral kernel is given by the formula
$$(t,s)\to t^{-\frac12}\chi_{\{s<t\}}+s^{-\frac12}\chi_{\{t<s\}},\quad 0<s,t<1,$$
which is, obviously, square-integrable.
The gist of (the critical case of the) Sobolev inequality on $\mathbb{R}^d$ may be informally understood as the boundedness of the operator $(1-\Delta)^{-\frac{d}{4}}$ from $L_2(\mathbb{R}^d)$ into a symmetric Banach function space on $\mathbb{R}^d$ which is sufficiently close to $L_{\infty}(\mathbb{R}^d).$ We suggest to view this result as a statement concerning distribution functions of elements $(1-\Delta)^{-\frac{d}{4}}x$ and $T(\mu(x)\chi_{(0,1)})$ where $x\in L_2(\mathbb{R}^d).$
Our distributional version of Sobolev inequality is as follows:
\begin{thm}\label{dist sobolev critical} Let $d\in\mathbb{N}.$ For every $x\in L_2(\mathbb{R}^d),$ we have
$$\mu\big((1-\Delta)^{-\frac{d}{4}}x\big)\chi_{(0,1)}\prec\prec c_dT(\mu(x)\chi_{(0,1)}).$$
Here, $\prec\prec$ denotes the Hardy-Littlewood submajorization.
\end{thm}
We need the following lemma stated in terms of convolutions.
\begin{lem}\label{oneil lemma} Let $x\in L_2(\mathbb{R}^d)$ and $g\in (L_{2,\infty}\cap L_1)(\mathbb{R}^d).$ We have
$$\mu(x\ast g)\chi_{(0,1)}\prec\prec 4\|g\|_{L_{2,\infty}\cap L_1}T(\mu(x)\chi_{(0,1)}).$$
\end{lem}
\begin{proof} Lemma 1.5 in \cite{oneil} states that
$$\int_0^t\mu(s,x\ast g)ds\leq \int_0^t\mu(s,x)ds\cdot \int_0^t\mu(s,g)ds+t\int_t^{\infty}\mu(s,x)\mu(s,g)ds.$$
Obviously,
$$\int_0^t\mu(s,x)ds\leq \|g\|_{2,\infty}\cdot \int_0^ts^{-\frac12}ds=2t^{\frac12}\|g\|_{2,\infty}\leq 2\|g\|_{L_{2,\infty}\cap L_1}\cdot t^{\frac12}.$$
For $t\in(0,1),$ we have
$$t\int_t^{\infty}\mu(s,x)\mu(s,g)ds=t\int_t^1\mu(s,x)\mu(s,g)ds+t\int_1^{\infty}\mu(s,x)\mu(s,g)ds\leq$$
$$\leq t\|g\|_{2,\infty}\int_t^1\mu(s,x)\frac{ds}{s^{\frac12}}+t\mu(1,x)\|g\|_1\leq$$
$$\leq \|g\|_{L_{2,\infty}\cap L_1}\cdot \Big(t\int_t^1\mu(s,x)\frac{ds}{s^{\frac12}}+t\mu(1,x)\Big).$$
We, therefore, have
\begin{equation}\label{oneil eq0}
\int_0^t\mu(s,x\ast g)ds\leq 2\|g\|_{L_{2,\infty}\cap L_1}\cdot F(t),\quad t\in(0,1),
\end{equation}
where
$$F(t)=t^{\frac12}\int_0^t\mu(s,x)ds+t\int_t^1\mu(s,x)\frac{ds}{s^{\frac12}}+t\mu(1,x),\quad t>0.$$
Computing the derivative, we obtain
$$F'(t)=\frac12t^{-\frac12}\int_0^t\mu(s,x)ds+\int_t^1\mu(s,x)\frac{ds}{s^{\frac12}}+\mu(1,x)\leq $$
$$\leq (T\mu(x)\chi_{(0,1)})(t)+\mu(1,x).$$
Next,
$$(T\mu(x)\chi_{(0,1)})(t)\geq t^{-\frac12}\int_0^t\mu(1,x)ds+\int_t^1\frac{\mu(1,x)ds}{s^{\frac12}}=$$
$$=\mu(1,x)\cdot (2-t^{\frac12})\geq\mu(1,x),\quad t\in(0,1).$$
Thus,
$$F'(t)\leq 2(T\mu(x)\chi_{(0,1)})(t),\quad t\in(0,1).$$
It is now immediate that
$$F(t)\leq 2\int_0^t\Big((T\mu(x)\chi_{(0,1)})(s)\Big)ds,\quad t\in(0,1).$$
Combining this equation with \eqref{oneil eq0}, we obtain
$$\int_0^t\mu(s,x\ast g)ds\leq 4\|g\|_{L_{2,\infty}\cap L_1}\int_0^t\Big((T\mu(x)\chi_{(0,1)})(s)\Big)ds,\quad t\in(0,1).$$
This is exactly the required assertion.
\end{proof}
\begin{proof}[Proof of Theorem \ref{dist sobolev critical}] We rewrite $(1-\Delta)^{-\frac{d}{4}}$ as a convolution operator. Namely,
$$(1-\Delta)^{-\frac{d}{4}}x=x\ast g,\quad x\in L_2(\mathbb{R}^d),$$
where $g$ is the Fourier transform of the function
$$t\to (1+|t|^2)^{-\frac{d}{4}},\quad t\in\mathbb{R}^d.$$
Precise expression for the function $g$ involves Macdonald function $K_{\frac{d}{4}}$ and is given in \cite{AronszajnSmith} (see formulae (2.7) and (2.10) there) as follows
$$g(t)=c_d|t|^{-\frac{d}{4}}K_{\frac{d}{4}}(|t|),\quad t\in\mathbb{R}^d.$$
Let $\mathbb{B}^d$ be the unit ball in $\mathbb{R}^d.$ If $d\neq0{\rm mod}4,$ it follows from formulae (9.6.2) and (9.6.10) in \cite{AbramovitzStegun} that $g\chi_{\mathbb{B}^d}\in L_{2,\infty}(\mathbb{B}^d)\subset L_1(\mathbb{B}^d).$ If $d=0{\rm mod}4,$ formula (9.6.11) in \cite{AbramovitzStegun} yields that $g\chi_{\mathbb{B}^d}\in L_{2,\infty}(\mathbb{B}^d)\subset L_1(\mathbb{B}^d).$ Also, $g\chi_{\mathbb{R}^d\backslash\mathbb{B}^d}\in (L_1\cap L_{\infty})(\mathbb{R}^d\backslash\mathbb{B}^d)$ by formula (9.7.2) in \cite{AbramovitzStegun}. Thus, $g\in (L_{2,\infty}\cap L_1)(\mathbb{R}^d).$ The assertion follows from Lemma \ref{oneil lemma}.
\end{proof}
\subsection{Optimality of the distributional Sobolev-type inequality }
We start with the following easy observation.
\begin{lem}\label{spherical rearrangement lemma} For $x\in S(0,\infty),$ we have
$$\mu(x\circ r_d)=\sigma_{\omega_d}\mu(x),\quad r_d(t)=|t|^d,\quad t\in\mathbb{R}^d,$$
where $\omega_d$ is the volume of the unit ball in $\mathbb{R}^d.$
\end{lem}
\begin{proof} Indeed, for every interval $(a,b),$ we have
$$m(r_d^{-1}(a,b))=m(\{t\in\mathbb{R}^d:\ a<|t|^d<b\})=$$
$$=m(\{t\in\mathbb{R}^d:\ |t|<b^{\frac1d}\})-m(\{t\in\mathbb{R}^d:\ |t|<a^{\frac1d}\})=\omega_d\cdot (b-a).$$
Hence, for every set $A\subset(0,\infty),$ we have
$$m(r_d^{-1}(A))=\omega_dm(A).$$
In other words, the mapping $r_d:\mathbb{R}^d\to\mathbb{R}$ preserves a measure (modulo a constant factor $\omega_d$). This suffices to conclude the argument.
\end{proof}
The following proposition shows (with the help of Lemma \ref{spherical rearrangement lemma}) that Theorem \ref{dist sobolev critical} is optimal.
\begin{prop}\label{estimate from below lemma} There exists a strictly positive constant $c_d'$ depending only on $d$ such that
$$(1-\Delta)^{-\frac{d}{4}}(x\circ r_d)\geq c_d'(Tx)\circ r_d$$
for every $0\leq x\in L_2(0,\infty)$ supported in the interval $(0,1).$
\end{prop}
\begin{proof} As in the proof of Theorem \ref{dist sobolev critical} above, we rewrite $(1-\Delta)^{-\frac{d}{4}}$ as a convolution operator. Namely,
$$(1-\Delta)^{-\frac{d}{4}}z=z\ast g,\quad z\in L_2(\mathbb{R}^d)$$
where $g$ is given by the formula
$$g(t)=c_d|t|^{-\frac{d}{4}}K_{\frac{d}{4}}(|t|),\quad t\in\mathbb{R}^d.$$
By formula (9.6.23) in \cite{AbramovitzStegun}, this is a strictly positive function.
Let $\mathbb{B}^d$ be the unit ball in $\mathbb{R}^d.$ Since $g$ is strictly positive, it follows from the formulae (9.6.2) and (9.6.10) in \cite{AbramovitzStegun} (when $d\neq0{\rm mod}4$) or from the formula (9.6.11) in \cite{AbramovitzStegun} (when $d=0{\rm mod}4$) that
$$g(t)\geq c_d'|t|^{-\frac{d}{2}},\quad 0<|t|<2.$$
Since $x\geq0,$ it follows that
$$\big((1-\Delta)^{-\frac{d}{4}}(x\circ r_d)\big)(t)=((x\circ r_d)\ast g)(t)\geq c_d'\int_{|t-s|<2}|t-s|^{-\frac{d}{2}}x(|s|^d)ds.$$
When $|t|<1,$ we have
$$\Big\{s\in\mathbb{R}^d:\quad |s|<1\Big\}\subset \Big\{s\in\mathbb{R}^d: |t-s|<2\Big\}.$$
Thus,
$$\big((1-\Delta)^{-\frac{d}{4}}(x\circ r_d)\big)(t)\geq c_d'\int_{|s|<1}|t-s|^{-\frac{d}{2}}x(|s|^d)ds,\quad |t|<1.$$
Obviously,
$$|t-s|\leq|t|+|s|\leq2\max\{|t|,|s|\}$$
and, therefore,
$$|t-s|^{-\frac{d}{2}}\geq 2^{-\frac{d}{2}}\min\{|t|^{-\frac{d}{2}},|s|^{-\frac{d}{2}}\}.$$
It follows that
$$\big((1-\Delta)^{-\frac{d}{4}}(x\circ r_d)\big)(t)\geq 2^{-\frac{d}{2}}c_d'\int_{|s|<1}\min\{|t|^{-\frac{d}{2}},|s|^{-\frac{d}{2}}\}x(|s|^d)ds,\quad |t|<1.$$
Passing to spherical coordinates, we obtain
$$\int_{|s|<1}\min\{|t|^{-\frac{d}{2}},|s|^{-\frac{d}{2}}\}x(|s|^d)ds=$$
$$=\int_0^1\min\{|t|^{-\frac{d}{2}},r^{-\frac{d}{2}}\}r^{d-1}x(r^d)dr\cdot\int_{\mathbb{S}^{d-1}}ds.$$
Making the substitution $r^d=u,$ we write
$$\int_0^1\min\{|t|^{-\frac{d}{2}},r^{-\frac{d}{2}}\}r^{d-1}x(r^d)dr=\frac1d\int_0^1\min\{|t|^{-\frac{d}{2}},u^{-\frac12}\}x(u)du=$$
$$=\frac1d\Big(\int_0^{|t|^d}|t|^{-\frac{d}{2}}x(u)du+\int_{|t|^d}^1u^{-\frac12}x(u)du\Big)=\frac1d(Tx)(|t|^d).$$
Combining the last $2$ equations, we complete the proof.
\end{proof}
\section{Sobolev embedding theorem in arbitrary dimension}
In this section, we extend Theorems \ref{thm 1} (see Proposition \ref{sobolev critical all dimension}) and \ref{thm 2} (see Corollary \ref{optimal embedding theorem}) to Euclidean spaces of arbitrary dimension. We provide new proofs of both theorems in their original setting (for both even and odd $d$).
In what follows, we extend $\psi$ to a concave increasing function on $(0,\infty)$ by setting
$$
\psi(t)=
\begin{cases}
\frac1{\log(\frac{e}{t})},& t\in(0,1)\\
t,& t\in[1,\infty).
\end{cases}
$$
\begin{prop}\label{sobolev critical all dimension} Let $d\in\mathbb{N}.$ For every $x\in L_2(\mathbb{R}^d),$ we have
$$\|(1-\Delta)^{-\frac{d}{4}}x\|_{\Lambda_{\psi}^{(2)}}\leq c_d\|x\|_2.$$
\end{prop}
With some effort, the crucial lemma below can be inferred from the proof of Theorem 5.1 in \cite{Cwikel-Pustylnik-JFA}. We provide a direct proof. For the definition of real interpolation method employed below we refer the reader to \cite{BerghLoefstrom}.
\begin{lem}\label{real interpolation lemma} We have
$$[\mathtt{M}_{\phi},L_{\infty}]_{\frac12,2}\subset\Lambda_{\psi}^{(2)}.$$
\end{lem}
\begin{proof} Let $t\in(0,1)$ and set $s=\psi^{-1}(t).$
Let $x\in\mathtt{M}_{\phi}+L_{\infty}$ and let $x=f+g,$ where $f\in\mathtt{M}_{\phi}$ and where $g\in L_{\infty}.$ Set
$$A=|x|^{-1}([\mu(s,x),\infty)),\quad B=|f|^{-1}([\frac12\mu(s,x),\infty)),\quad C=|g|^{-1}([\frac12\mu(s,x),\infty)).$$
Note that $m(A)\geq s.$
We have
$$|f|(u)<\frac12\mu(s,x),\quad |g|(u)<\frac12\mu(s,x),\quad u\in B^c\cap C^c.$$
Thus,
$$|x|(u)\leq |f|(u)+|g|(u)<\mu(s,x),\quad u\in B^c\cap C^c.$$
So, $u\in B^c\cap C^c$ implies that $u\in A^c.$ That is, $B^c\cap C^c\subset A^c$ or, equivalently, $A\subset B\cup C.$
If $m(B)\geq\frac12 m(A),$ then
$$\|f\|_{M_{\phi}}+t\|g\|_{L_{\infty}} \geq \|f\chi_B\|_{M_{\phi}} \geq \frac12\mu(s,x) \|\chi_B\|_{M_{\phi}} =$$
$$= \frac12\mu(s,x) \psi(m(B)) \geq \frac14\mu(s,x)\phi(m(A)) \geq \frac14 t\mu(s,x).$$
If $m(C)\geq \frac12 m(A),$ then
$$\|f\|_{M_{\phi}}+t\|g\|_{L_{\infty}} \geq t\|g\chi_C\|_{L_{\infty}} \geq \frac12t\mu(s,x).$$
In either case, we have
$$K(t,x,\mathtt{M}_{\phi},L_{\infty})\geq \frac14t\mu(s,x).$$
Thus,
$$\|x\|_{[\mathtt{M}_{\phi},L_{\infty}]_{\frac12,2}}^2=\int_0^1 (\frac1tK(t,x,\mathtt{M}_{\phi},L_{\infty}))^2 dt \geq$$
$$\geq\frac1{16}\int_0^1 \mu^2(s,x)dt \stackrel{t=\psi(s)}{=}\frac1{16} \int_0^1\mu^2(s,x)d\psi(s).$$
\end{proof}
The following lemma shows that the receptacle of the operator $T$ is strictly smaller than $\exp(L_2)$ (the space suggested by the Moser-Trudinger inequality).
\begin{lem}\label{t lemma} We have $T:L_2(0,1)\to\Lambda_{\psi}^{(2)}(0,1).$
\end{lem}
\begin{proof} Let $x\in L_{2,1}(0,1).$ It is immediate that
$$|(Tx)(t)|\leq t^{-\frac12}\int_0^t|x(s)|ds+\int_t^1|x(s)|\frac{ds}{s^{\frac12}}\leq$$
$$\leq t^{-\frac12}\int_0^t\mu(s,x)ds+\int_0^1|x(s)|\frac{ds}{s^{\frac12}}\leq t^{-\frac12}\int_0^t\mu(s,x)ds+\int_0^1\mu(s,x)\frac{ds}{s^{\frac12}}=$$
$$=t^{-\frac12}\int_0^t\mu(s,x)ds+2\|x\|_{2,1}\leq t^{-\frac12}\|x\|_{2,\infty}\int_0^t\frac{ds}{s^{\frac12}}+2\|x\|_{2,1}=$$
$$=2\|x\|_{2,\infty}+2\|x\|_{2,1}\leq c_{{\rm abs}}\|x\|_{2,1}.$$
Thus,
$$\|T\|_{L_{2,1}\to L_{\infty}}\leq c_{{\rm abs}}.$$
Let $x\in L_{2,\infty}(0,1).$ It is immediate that
$$|(Tx)(t)|\leq\int_0^1|x(s)|\cdot\min\{t^{-\frac12},s^{-\frac12}\}ds\leq$$
$$\leq\int_0^1\mu(s,x)\cdot\min\{t^{-\frac12},s^{-\frac12}\}ds\leq\|x\|_{2,\infty}\int_0^1s^{-\frac12}\cdot\min\{t^{-\frac12},s^{-\frac12}\}ds.$$
Obviously,
$$\int_0^1s^{-\frac12}\cdot\min\{t^{-\frac12},s^{-\frac12}\}ds=t^{-\frac12}\int_0^t\frac{ds}{s^{\frac12}}+\int_t^1\frac{ds}{s}=\log(\frac{e^2}{t}).$$
That is,
$$|(Tx)(t)|\leq\|x\|_{2,\infty}\log(\frac{e^2}{t}),\quad t\in(0,1).$$
Since the mapping $t\to\log(\frac{e^2}{t}),$ $t\in(0,1),$ falls into $\exp(L_1)(0,1),$ it follows that
$$\|T\|_{L_{2,\infty}\to \exp(L_1)}\leq c_{{\rm abs}}.$$
By real interpolation, we have
$$T:[L_{2,\infty},L_{2,1}]_{\frac12,2}\to [\exp(L_1),L_{\infty}]_{\frac12,2}$$
is a bounded mapping. By Lemma \ref{real interpolation lemma}, we have
$$[L_{2,\infty},L_{2,1}]_{\frac12,2}=L_2,\quad [\exp(L_1),L_{\infty}]_{\frac12,2}\stackrel{L.\ref{real interpolation lemma}}{\subset}\Lambda_{\psi}^{(2)}.$$
Thus, $T:L_2(0,1)\to\Lambda_{\psi}^{(2)}(0,1)$ is a bounded mapping.
\end{proof}
\begin{proof}[Proof of Proposition \ref{sobolev critical all dimension}] By Theorem \ref{dist sobolev critical} and Lemma \ref{t lemma}, we have
$$\Big\|\mu\big((1-\Delta)^{-\frac{d}{4}}x\big)\chi_{(0,1)}\Big\|_{\Lambda_{\psi}^{(2)}}\leq c_d\|x\|_2.$$
In other words,
$$\Big\|(1-\Delta)^{-\frac{d}{4}}x\Big\|_{\Lambda_{\psi}^{(2)}+L_{\infty}}\leq c_d\|x\|_2.$$
On the other hand, we have
$$\Big\|(1-\Delta)^{-\frac{d}{4}}x\Big\|_2\leq \|x\|_2.$$
Thus,
$$\Big\|(1-\Delta)^{-\frac{d}{4}}x\Big\|_{(\Lambda_{\psi}^{(2)}+L_{\infty})\cap L_2}\leq c_d\|x\|_2.$$
Obviously,
$$(\Lambda_{\psi}^{(2)}+L_{\infty})\cap L_2=\Lambda_{\psi}^{(2)}$$
and the assertion follows.
\end{proof}
The next assertion should be compared with Theorem 4 in \cite{Cwikel-Pustylnik} (it is proved in the companion paper \cite{Cwikel-Pustylnik-JFA} and constitutes the key part of the proof of Theorem 5.7 in that paper). Note that our $T$ is different from that in \cite{Cwikel-Pustylnik}. This difference is the reason why our proof is so much simpler.
\begin{thm}\label{t theorem} For every $z=\mu(z)\in \Lambda_{\psi}^{(2)}(0,1),$ there exists $x=\mu(x)\in L_2(0,1)$ such that
$$\mu(z)\leq Tx\mbox{ and }\|x\|_2\leq c_{{\rm abs}}\|z\|_{\Lambda_{\psi}^{(2)}}.$$
\end{thm}
The technical part of the proof of Theorem \ref{t theorem} is concentrated in the next lemma.
\begin{lem}\label{cwikel-pustylnik lorentz lemma} Let $z\in\Lambda_{\psi}^{(2)}(0,1)$ and let
$$y(t)=\sup_{0<s<t}\psi(s)\mu(s,z),\quad 0<t<1.$$
We have
$$\|y\|_{L_2((0,1),\frac{dt}{t})}\leq 3^{\frac12}\|z\|_{\Lambda_{\psi}^{(2)}}.$$
\end{lem}
\begin{proof} We claim that (here $C$ is the classical Cesaro operator)
$$\int_0^1\sup_{0<s<t}\psi^2(s)(C\mu(w))(s)\frac{dt}{t}\leq 3\|w\|_{\Lambda_{\psi}},\quad w\in\Lambda_{\psi}.$$
The functional on the left hand side is normal, subadditive and positively homogeneous. By Lemma II.5.2 in \cite{KPS}, it suffices to prove the inequality for the indicator functions.
Let $w=\chi_A$ and let $m(A)=u.$ Obviously, $\mu(w)=\chi_{(0,u)}$ and
$$(C\mu(w))(s)=\frac{\min\{u,s\}}{s},\quad 0<s<1.$$
Thus,
$$\int_0^1\sup_{0<s<t}\psi^2(s)(C\mu(w))(s)\frac{dt}{t}=\int_0^u\sup_{0<s<t}\psi^2(s)\frac{dt}{t}+$$
$$+\int_u^1\max\Big\{\sup_{0<s<u}\psi^2(s),\sup_{u<s<t}\psi^2(s)\frac{u}{s}\Big\}\frac{dt}{t}=$$
$$=\int_0^u\psi^2(t)\frac{dt}{t}+\int_u^1\max\Big\{\psi^2(u),u\sup_{u<s<t}s^{-1}\psi^2(s)\Big\}\frac{dt}{t}.$$
The function $s\to s^{-1}\psi^2(s),$ $s\in(0,1),$ decreases on the interval $(0,e^{-1})$ and increases on the interval $(e^{-1},1).$ Thus,
$$\sup_{u<s<t}s^{-1}\psi^2(s)=
\begin{cases}
u^{-1}\psi^2(u),& t\in(0,e^{-1})\\
t^{-1}\psi^2(t),& u\in(e^{-1},1)\\
\frac{e}{4},& u<e^{-1}<t
\end{cases}
$$
For $u\in(e^{-1},1),$ we have
$$\int_0^1\sup_{0<s<t}\psi^2(s)(C\mu(w))(s)\frac{dt}{t}=\int_0^u\psi^2(t)\frac{dt}{t}+\int_u^1\max\Big\{\psi^2(u),ut^{-1}\psi^2(t)\Big\}\frac{dt}{t}=$$
$$=\int_0^u\psi^2(t)\frac{dt}{t}+\int_u^1\psi^2(u)\frac{dt}{t}=\psi(u)+\psi^2(u)\log(\frac1u)\leq 2\psi(u).$$
For $u\in(0,e^{-1}),$ we have
$$\int_0^1\sup_{0<s<t}\psi^2(s)(C\mu(w))(s)\frac{dt}{t}=\int_0^u\psi^2(t)\frac{dt}{t}+\int_u^{e^{-1}}\psi^2(u)\frac{dt}{t}+$$
$$+\int_{e^{-1}}^1\max\Big\{\psi^2(u),\frac{e}{4}u\Big\}\frac{dt}{t}=\psi(u)+\psi^2(u)\log(\frac1{eu})+\max\Big\{\psi^2(u),\frac{e}{4}u\Big\}\leq 3\psi(u).$$
This yields the claim.
Let $w=\mu^2(z)\in\Lambda_{\psi}.$ We have
$$\int_0^1\sup_{0<s<t}\psi^2(s)\mu(s,w)\frac{dt}{t}\leq \int_0^1\sup_{0<s<t}\psi^2(s)(C\mu(w))(s)\frac{dt}{t}\leq 3\|w\|_{\Lambda_{\psi}}.$$
The assertion follows immediately.
\end{proof}
The following lemma affords a very substantial simplification and streamlining of the arguments employed in \cite{Cwikel-Pustylnik}.
\begin{lem}\label{l2 to l2 lemma} Let $y\in L_2((0,1),\frac{dt}{t})$ and let $x(t)=t^{-\frac12}y(t),$ $0<t<1.$ If $y$ is positive and increasing, then
$$\psi(t)\cdot (Tx)(t)\geq \frac1{4e}y(\frac{t}{2}),\quad 0<t<1.$$
\end{lem}
\begin{proof} Suppose $t\in(0,e^{-1}).$ We have
$$\psi(t)\cdot (Tx)(t)\geq\psi(t)\int_t^1\frac{y(s)ds}{s}\geq y(t)\cdot \psi(t)\int_t^1\frac{ds}{s}\geq\frac12 y(t)\geq \frac1{4e}y(\frac{t}{2}).$$
Suppose $t\in(e^{-1},1).$ We have
$$\psi(t)\cdot (Tx)(t)\geq \frac12(Tx)(t)\geq \frac1{2t^{\frac12}}\int_0^ty(s)ds\geq\frac12\int_{\frac{t}{2}}^ty(s)ds\geq \frac1{4e}y(\frac{t}{2}).$$
\end{proof}
\begin{proof}[Proof of Theorem \ref{t theorem}] Let $y$ be as in Lemma \ref{cwikel-pustylnik lorentz lemma} and let $x$ be as in Lemma \ref{l2 to l2 lemma}. It follows from Lemma \ref{cwikel-pustylnik lorentz lemma} that $\|x\|_2\leq 3^{\frac12}\|z\|_{\Lambda_{\psi}^{(2)}}.$ Obviously, $y$ is positive and increasing. It follows from Lemma \ref{l2 to l2 lemma} that
$$\psi(t)\cdot (Tx)(t)\geq \frac1{4e}y(\frac{t}{2})\geq \frac1{4e}\psi(\frac{t}{2})\mu(\frac{t}{2},z)\geq \frac1{8e}\psi(t)\mu(t,z),\quad 0<t<1.$$
Thus,
$$T\mu(x)\geq Tx\geq \frac1{8e}\mu(z).$$
\end{proof}
The next corollary shows that the result of Proposition \ref{sobolev critical all dimension} is optimal.
\begin{cor}\label{optimal embedding theorem} Let $E(\mathbb{R}^d)$ be a symmetric Banach function space on $\mathbb{R}^d.$ If
$$(1-\Delta)^{-\frac{d}{4}}:L_2(\mathbb{R}^d)\to E(\mathbb{R}^d),$$
then $\Lambda_{\psi}^{(2)}(\mathbb{R}^d)\subset (E+L_{\infty})(\mathbb{R}^d).$
\end{cor}
\section{Degenerate case of Cwikel estimate}
In this case, we consider operators
$$(1-\Delta)^{-\frac{d}{4}}M_f(1-\Delta)^{-\frac{d}{4}}\mbox{ and }(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_f(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}$$
which act, respectively, on $L_2(\mathbb{R}^d)$ and $L_2(\mathbb{T}^d)$ and evaluate their uniform norms. We show that the maximal (symmetric Banach function) space $E$ such that the operators above are bounded for every $f\in E$ is the Marcinkiewicz space $\mathtt{M}_{\psi}.$ For Euclidean space, this follows from Propositions \ref{space boundedness} and \ref{space boundedness optimal} below. For torus, this follows from Propositions \ref{torus boundedness} and \ref{torus boundedness optimal} below.
\subsection{Estimates for Euclidean space}
\begin{prop}\label{space boundedness} Let $d\in\mathbb{N}.$ We have
$$\Big\|(1-\Delta)^{-\frac{d}{4}}M_f(1-\Delta)^{-\frac{d}{4}}\Big\|_{\infty}\leq c_d\|f\|_{\mathtt{M}_{\psi}},\quad f\in\mathtt{M}_{\psi}(\mathbb{R}^d).$$
\end{prop}
\begin{proof} Without loss of generality, $f$ is real-valued and positive. We have
$$\Big\|(1-\Delta)^{-\frac{d}{4}}M_f(1-\Delta)^{-\frac{d}{4}}\Big\|_{\infty}=\sup_{\|x\|_2\leq1}\Big|\langle (1-\Delta)^{-\frac{d}{4}}M_f(1-\Delta)^{-\frac{d}{4}}x,x\rangle\Big|=$$
$$=\sup_{\|x\|_2\leq1}\Big|\langle f\cdot (1-\Delta)^{-\frac{d}{4}}x,(1-\Delta)^{-\frac{d}{4}}x\rangle\Big|=$$
$$=\sup_{\|x\|_2\leq1}\Big\|f\cdot \Big|(1-\Delta)^{-\frac{d}{4}}x\Big|^2\Big\|_1\leq \sup_{\|x\|_2\leq1}\|f\|_{\mathtt{M}_{\psi}}\Big\|(1-\Delta)^{-\frac{d}{4}}x\Big\|_{\Lambda_{\psi}^{(2)}}^2.$$
The assertion follows now from Proposition \ref{sobolev critical all dimension}.
\end{proof}
The converse inequality follows from Proposition \ref{estimate from below lemma} and Theorem \ref{t theorem}.
\begin{prop}\label{space boundedness optimal} Let $d\in\mathbb{N}.$ Let $f=\mu(f)\in\mathtt{M}_{\psi}(0,\infty).$ We have
$$\Big\|(1-\Delta)^{-\frac{d}{4}}M_{f\circ r_d}(1-\Delta)^{-\frac{d}{4}}\Big\|_{\infty}\geq c_d\|f\|_{\mathtt{M}_{\psi}},.$$
\end{prop}
\begin{proof} We have
$$\Big\|(1-\Delta)^{-\frac{d}{4}}M_{f\circ r_d}(1-\Delta)^{-\frac{d}{4}}\Big\|_{\infty}=$$
$$=\sup_{\|\xi\|_2\leq1}\Big|\langle (1-\Delta)^{-\frac{d}{4}}M_{f\circ r_d}(1-\Delta)^{-\frac{d}{4}}\xi,\xi\rangle\Big|=$$
$$=\sup_{\|\xi\|_2\leq1}\Big|\langle (f\circ r_d)\cdot (1-\Delta)^{-\frac{d}{4}}\xi,(1-\Delta)^{-\frac{d}{4}}\xi\rangle\Big|=$$
$$=\sup_{\|\xi\|_2\leq1}\Big\|(f\circ r_d)\cdot \Big|(1-\Delta)^{-\frac{d}{4}}\xi\Big|^2\Big\|_1.$$
Let us now restrict the supremum to the radial $\xi.$ That is, let $\xi=x\circ r_d,$ where $x\in L_2(0,\infty)$ and $\|x\|_2\leq\omega_d^{-\frac12}.$ We have
$$\Big\|(1-\Delta)^{-\frac{d}{4}}M_{f\circ r_d}(1-\Delta)^{-\frac{d}{4}}\Big\|_{\infty}\geq$$
$$\geq\sup_{\|x\|_2\leq\omega_d^{-\frac12}}\Big\|(f\circ r_d)\cdot \Big|(1-\Delta)^{-\frac{d}{4}}(x\circ r_d)\Big|^2\Big\|_1.$$
Let us further assume that $x=\mu(x)$ is supported on the interval $(0,1).$ By Proposition \ref{estimate from below lemma} we have
$$\Big\|(1-\Delta)^{-\frac{d}{4}}M_{f\circ r_d}(1-\Delta)^{-\frac{d}{4}}\Big\|_{\infty}\geq$$
$$\geq\sup_{\substack{x=\mu(x)\\ \|x\|_2\leq\omega_d^{-\frac12}\\ x=0\mbox{ on }(1,\infty)}}\Big\|(f\circ r_d)\cdot \Big|(1-\Delta)^{-\frac{d}{4}}(x\circ r_d)\Big|^2\Big\|_1\geq$$
$$\geq c_d\cdot\sup_{\substack{x=\mu(x)\\ \|x\|_2\leq\omega_d^{-\frac12}\\ x=0\mbox{ on }(1,\infty)}}\Big\|(f\circ r_d)\cdot \Big|Tx\circ r_d\Big|^2\Big\|_1=c_d\omega_d\cdot\sup_{\substack{x=\mu(x)\\ \|x\|_2\leq\omega_d^{-\frac12}\\ x=0\mbox{ on }(1,\infty)}}\big\|f\cdot |Tx|^2\big\|_1.$$
By Theorem \ref{t theorem}, we have
$$\Big\|(1-\Delta)^{-\frac{d}{4}}M_{f\circ r_d}(1-\Delta)^{-\frac{d}{4}}\Big\|_{\infty}\geq$$
$$\geq c_d\omega_d\cdot\sup_{\substack{z=\mu(z)\in\Lambda_{\psi^{(2)}}(0,1)\\ \|z\|_{\Lambda_{\psi^{(2)}}}\leq c_{{\rm abs}}^{-1}\omega_d^{-\frac12}}}\big\|f\cdot z^2\big\|_1=c_dc_{{\rm abs}}^{-2}\cdot\sup_{\substack{z=\mu(z)\in\Lambda_{\psi}(0,1)\\ \|z\|_{\Lambda_{\psi}}\leq 1}}\big\|f\cdot z\big\|_1.$$
It is clear that
$$\sup_{\substack{z=\mu(z)\in\Lambda_{\psi}(0,1)\\ \|z\|_{\Lambda_{\psi}}\leq 1}}\big\|f\cdot z\big\|_1=\|f\chi_{(0,1)}\|_{\mathtt{M}_{\psi}}.$$
Since $f=\mu(f)$ and since $\psi$ is linear on $(1,\infty),$ it follows that
$$\|f\chi_{(0,1)}\|_{\mathtt{M}_{\psi}}=\|\mu(f)\chi_{(0,1)}\|_{\mathtt{M}_{\psi}}\approx_d\|f\|_{\mathtt{M}_{\psi}+L_{\infty}}\approx_d\|f\|_{\mathtt{M}_{\psi}}.$$
Combining three last equations, we complete the proof.
\end{proof}
\subsection{Estimates for the torus}
The following lemma is taken from \cite{DAO} (see Lemmas 4.5 and 4.6 there).
\begin{lem}\label{non-compact to compact lemma} Let $h$ be a measurable function on $[-1,1]^d.$ We have
$$M_h(1-\Delta)^{-\frac{d}{2}}M_h\Big|_{L_2([-1,1]^d)}=M_ha(\nabla_{\mathbb{T}^d})M_h\Big|_{L_2([-1,1]^d)},$$
where
$$a(n)=(1+|n|^2)^{-\frac{d}{2}}+b(n),\quad b(n)=O((1+|n|^2)^{-\frac{d+1}{2}}),\quad n\in\mathbb{Z}^d.$$
\end{lem}
The following lemma relies on the post-critical Sobolev inequality:
$$\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d+1}{4}}\Big\|_{L_2(\mathbb{T}^d)\to L_{\infty}(\mathbb{T}^d)}\leq c_d.$$
The validity of this inequality follows immediately from the fact
$$\Big\{(1+|n|^2)^{-\frac{d+1}{2}}\Big\}_{n\in\mathbb{Z}^d}\in l_1(\mathbb{Z}^d).$$
\begin{lem}\label{postcritical torus lemma} Let $h\in L_1(\mathbb{T}^d).$ We have
$$\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d+1}{4}}M_h(1-\Delta_{\mathbb{T}^d})^{-\frac{d+1}{4}}\Big\|_{\infty}\leq c_d\|h\|_1.$$
\end{lem}
\begin{proof} Without loss of generality, $h$ is real-valued and positive. We have
$$\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d+1}{4}}M_h(1-\Delta_{\mathbb{T}^d})^{-\frac{d+1}{4}}\Big\|_{\infty}=\sup_{\|x\|_2\leq1}\Big|\langle (1-\Delta_{\mathbb{T}^d})^{-\frac{d+1}{4}}M_h(1-\Delta_{\mathbb{T}^d})^{-\frac{d+1}{4}}x,x\rangle\Big|=$$
$$=\sup_{\|x\|_2\leq1}\Big|\langle h\cdot (1-\Delta_{\mathbb{T}^d})^{-\frac{d+1}{4}}x,(1-\Delta_{\mathbb{T}^d})^{-\frac{d+1}{4}}x\rangle\Big|=$$
$$=\sup_{\|x\|_2\leq1}\Big\|h\cdot \Big|(1-\Delta_{\mathbb{T}^d})^{-\frac{d+1}{4}}x\Big|^2\Big\|_1\leq \sup_{\|x\|_2\leq1}\|h\|_1\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d+1}{4}}x\Big\|_{\infty}^2.$$
The assertion follows now from the post-critical Sobolev inequality.
\end{proof}
It is of crucial importance that the estimate in the preceding lemma is given in terms of $\|h\|_1$ rather than $\|h\|_{\mathtt{M}_{\psi}}.$
The following proposition is a version of Proposition \ref{space boundedness} for $\mathbb{T}^d.$ Observe that Marcinkiewicz space $\mathtt{M}_{\psi}$ strictly contains the Orlicz space $L_M,$ $M(t)=t\log(e+t),$ $t>0. $ The result below is the best possible criterion for the boundedness of the symmetrized Cwikel operator
$$(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_f(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}.$$
The sharpness of the result will be demonstrated below in Proposition \ref{torus boundedness optimal}.
\begin{prop}\label{torus boundedness} Let $d\in\mathbb{N}.$ Let $f\in\mathtt{M}_{\psi}(\mathbb{T}^d).$ We have
$$\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_f(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}\Big\|_{\infty}\leq c_d\|f\|_{\mathtt{M}_{\psi}}.$$
\end{prop}
\begin{proof} Without loss of generality, $f$ is real-valued, positive and supported on $[-1,1]^d.$
In this proof, we frequently use the following property: let $H_0\subset H$ be a Hilbert subspace and let $A:H\to H_0.$ If $A$ vanishes on the orthogonal complement of $H_0,$ then $\|A\|_{\infty}=\|A|_{H_0}\|_{\infty}.$
Firstly, note that
$$\Big\|(1-\Delta)^{-\frac{d}{4}}M_f(1-\Delta)^{-\frac{d}{4}}\Big\|_{\infty}=\Big\|M_{f^{\frac12}}(1-\Delta)^{-\frac{d}{2}}M_{f^{\frac12}}\Big\|_{\infty}=$$
$$=\Big\|M_{f^{\frac12}}(1-\Delta)^{-\frac{d}{2}}M_{f^{\frac12}}\Big|_{L_2([-1,1]^d)}\Big\|_{\infty}.$$
By Lemma \ref{non-compact to compact lemma}, we have
$$M_{f^{\frac12}}(1-\Delta)^{-\frac{d}{2}}M_{f^{\frac12}}\Big|_{L_2([-1,1]^d)}=$$
$$=M_{f^{\frac12}}(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{2}}M_{f^{\frac12}}\Big|_{L_2([-1,1]^d)}+M_{f^{\frac12}}b(\nabla_{\mathbb{T}^d}) M_{f^{\frac12}}\Big|_{L_2([-1,1]^d)}.$$
By triangle inequality, we have
$$\Big\|M_{f^{\frac12}}(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{2}}M_{f^{\frac12}}\Big|_{L_2([-1,1]^d)}\Big\|_{\infty}\leq$$
$$\leq\Big\|M_{f^{\frac12}}(1-\Delta)^{-\frac{d}{2}}M_{f^{\frac12}}\Big|_{L_2([-1,1]^d)}\Big\|_{\infty}+\Big\|M_{f^{\frac12}}b(\mathbb{T}^d)M_{f^{\frac12}}\Big|_{L_2([-1,1]^d)}\Big\|_{\infty}=$$
$$=\Big\|M_{f^{\frac12}}(1-\Delta)^{-\frac{d}{2}}M_{f^{\frac12}}\Big\|_{\infty}+\Big\|M_{f^{\frac12}}b(\mathbb{T}^d)M_{f^{\frac12}}\Big\|_{\infty}\leq $$
$$\leq \Big\|(1-\Delta)^{-\frac{d}{4}}M_f(1-\Delta)^{-\frac{d}{4}}\Big\|_{\infty}+c_d\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d+1}{4}}M_f(1-\Delta_{\mathbb{T}^d})^{-\frac{d+1}{4}}\Big\|_{\infty}.$$
The assertion follows from Proposition \ref{space boundedness} and Lemma \ref{postcritical torus lemma}.
\end{proof}
The following proposition is a version of Proposition \ref{space boundedness optimal} for $\mathbb{T}^d.$
\begin{prop}\label{torus boundedness optimal} Let $d\in\mathbb{N}.$ Let $f=\mu(f)\in\mathtt{M}_{\psi}(0,1).$ We have
$$\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_{f\circ r_d}(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}\Big\|_{\infty}\geq c_d\|f\|_{\mathtt{M}_{\psi}}.$$
\end{prop}
\begin{proof} Firstly, note that
$$\Big\|(1-\Delta)^{-\frac{d}{4}}M_{f\circ r_d}(1-\Delta)^{-\frac{d}{4}}\Big\|_{\infty}=\Big\|M_{f^{\frac12}\circ r_d}(1-\Delta)^{-\frac{d}{2}}M_{f^{\frac12}\circ r_d}\Big\|_{\infty}=$$
$$=\Big\|M_{f^{\frac12}\circ r_d}(1-\Delta)^{-\frac{d}{2}}M_{f^{\frac12}\circ r_d}\Big|_{L_2([-1,1]^d)}\Big\|_{\infty}.$$
By Lemma \ref{non-compact to compact lemma}, we have
$$M_{f^{\frac12}\circ r_d}(1-\Delta)^{-\frac{d}{2}}M_{f^{\frac12}\circ r_d}\Big|_{L_2([-1,1]^d)}=$$
$$=M_{f^{\frac12}\circ r_d}(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{2}}M_{f^{\frac12}\circ r_d}\Big|_{L_2([-1,1]^d)}+M_{f^{\frac12}\circ r_d}b(\nabla_{\mathbb{T}^d}) M_{f^{\frac12}\circ r_d}\Big|_{L_2([-1,1]^d)}.$$
By triangle inequality, we have
$$\Big\|M_{f^{\frac12}\circ r_d}(1-\Delta)^{-\frac{d}{2}}M_{f^{\frac12}\circ r_d}\Big|_{L_2([-1,1]^d)}\Big\|_{\infty}\leq$$
$$\leq\Big\|M_{f^{\frac12}\circ r_d}(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{2}}M_{f^{\frac12}\circ r_d}\Big|_{L_2([-1,1]^d)}\Big\|_{\infty}+$$
$$+\Big\|M_{f^{\frac12}\circ r_d}b(\mathbb{T}^d)M_{f^{\frac12}\circ r_d}\Big|_{L_2([-1,1]^d)}\Big\|_{\infty}=$$
$$=\Big\|M_{f^{\frac12}\circ r_d}(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{2}}M_{f^{\frac12}\circ r_d}\Big\|_{\infty}+\Big\|M_{f^{\frac12}\circ r_d}b(\mathbb{T}^d)M_{f^{\frac12}\circ r_d}\Big\|_{\infty}.$$
Recall that
$$|b(n)|\leq c_d(1+|n|^2)^{-\frac{d+1}{2}},\quad n\in\mathbb{Z}^d.$$
Thus,
$$\Big\|(1-\Delta)^{-\frac{d}{4}}M_{f\circ r_d}(1-\Delta)^{-\frac{d}{4}}\Big\|_{\infty}\leq$$
$$\leq\Big\|M_{f^{\frac12}\circ r_d}(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{2}}M_{f^{\frac12}\circ r_d}\Big\|_{\infty}+\Big\|M_{f^{\frac12}\circ r_d}a(\mathbb{T}^d)M_{f^{\frac12}\circ r_d}\Big\|_{\infty}\leq$$
$$\leq\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_{f\circ r_d}(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}\Big\|_{\infty}+c_d\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d+1}{4}}M_{f\circ r_d}(1-\Delta_{\mathbb{T}^d})^{-\frac{d+1}{4}}\Big\|_{\infty}.$$
By Proposition \ref{space boundedness optimal} and Lemma \ref{postcritical torus lemma}, we have
$$c_d'\|f\|_{\mathtt{M}_{\psi}}\leq\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_{f\circ r_d}(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}\Big\|_{\infty}+c_d''\|f\|_1.$$
In other words, we have
\begin{equation}\label{tbo eq1}
\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_{f\circ r_d}(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}\Big\|_{\infty}\geq c_d'\|f\|_{\mathtt{M}_{\psi}}-c_d''\|f\|_1.
\end{equation}
On the other hand, we have
$$\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_{f\circ r_d}(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}\Big\|_{\infty}\geq\langle (1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_{f\circ r_d}(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}1,1\rangle=$$
$$=\langle (1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_{f\circ r_d}1,1\rangle=\langle M_{f\circ r_d}1,(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}1\rangle=\langle M_{f\circ r_d}1,1\rangle.$$
Thus,
\begin{equation}\label{tbo eq2}
\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_{f\circ r_d}(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}\Big\|_{\infty}\geq \|f\|_1.
\end{equation}
Combining \eqref{tbo eq1} and \eqref{tbo eq2}, we obtain
$$(1+c_d'')\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_{f\circ r_d}(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}\Big\|_{\infty}\geq c_d'\|f\|_{\mathtt{M}_{\psi}}.$$
This completes the proof.
\end{proof}
\subsection{Optimality of Cwikel-Solomyak estimates within the class of Orlicz spaces}\label{optimality-subsection}\label{optimality subsection}
The following material is standard; for more details we refer the reader to \cite{LSZ-book,Simon-book}.
Let $H$ be a complex separable infinite dimensional Hilbert space, and let $B(H)$ denote the set of all bounded operators on $H$, and let $K(H)$ denote the ideal of compact operators on $H.$ Given $T\in K(H),$ the sequence of singular values $\mu(T) = \{\mu(k,T)\}_{k=0}^\infty$ is defined as:
\begin{equation*}
\mu(k,T) = \inf\{\|T-R\|_{\infty}:\quad \mathrm{rank}(R) \leq k\}.
\end{equation*}
Let $p \in (0,\infty).$ The weak Schatten class $\mathcal{L}_{p,\infty}$ is the set of operators $T$ such that $\mu(T)$ is in the weak $L_p$-space $l_{p,\infty}$, with quasi-norm:
\begin{equation*}
\|T\|_{p,\infty} = \sup_{k\geq 0} (k+1)^{\frac1p}\mu(k,T) < \infty.
\end{equation*}
Obviously, $\mathcal{L}_{p,\infty}$ is an ideal in $B(H).$
Our next lemma follows from Proposition \ref{torus boundedness optimal} and provides one of the two ingredients in the proof of our third main result.
\begin{lem}\label{simple uniform lemma} Let $E(\mathbb{T}^d)$ be a symmetric Banach function space on $\mathbb{T}^d.$ Suppose that
$$\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_f(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}\Big\|_{\infty}\leq c_{d,E}\|f\|_E,\quad f\in E(\mathbb{T}^d).$$
It follows that $E\subset \mathtt{M}_{\psi}.$
\end{lem}
\begin{proof} Take $h=\mu(h)\in E(0,1)$ and let $f=h\circ r_d.$ We have
$$\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_{h\circ r_d}(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}\Big\|_{\infty}\leq c_{d,E}\omega_d\|h\|_E,\quad h=\mu(h)\in E(0,1).$$
By Proposition \ref{torus boundedness optimal}, we have
$$c_d\|h\|_{\mathtt{M}_{\psi}}\leq c_{d,E}\omega_d\|h\|_E,\quad h=\mu(h)\in E(0,1).$$
This completes the proof.
\end{proof}
The next lemma demonstrates efficiency of general theory of symmetric function spaces in the study of Cwikel estimates.
\begin{lem}\label{orlicz in marcinkiewicz} If $N$ is an Orlicz function such that $L_N(0,1)\subset\mathtt{M}_{\psi}(0,1),$ then $L_N(0,1)\subset L_M(0,1),$ where $M(t)=t\log(e+t),$ $t>0.$
\end{lem}
\begin{proof} We have
$$\|\chi_{(0,t)}\|_{\mathtt{M}_{\psi}}\leq c_N\|\chi_{(0,t)}\|_{L_N},\quad 0<t<1.$$
Thus,
$$t\log(\frac{e}{t})\leq \frac{c_N}{N^{-1}(\frac1t)},\quad t\in(0,1).$$
Setting $u=t^{-1},$ we write
$$u^{-1}\log(eu)\leq \frac{c_N}{N^{-1}(u)},\quad u>1.$$
Setting $v=N^{-1}(u),$ we write
$$N(v)^{-1}\log(eN(v))\leq \frac{c_N}{v},\quad v>N^{-1}(1).$$
Equivalently,
$$N(v)\geq c_N^{-1}v\log(eN(v)),\quad v>N^{-1}(1).$$
By convexity,
$$N(v)\geq \frac{v}{N^{-1}(1)},\quad v>N^{-1}(1).$$
Thus,
$$N(v)\geq c_N^{-1}v\log(\frac{ev}{N^{-1}(1)}),\quad v>N^{-1}(1).$$
Thus,
$$N(v)\geq c_N'M(v),\quad v>c_N''.$$
\end{proof}
The following corollary is our third main result. It demonstrates that Cwikel inequality proved by Solomyak (for even $d$) cannot be improved within the class of Orlicz spaces. Observe that a version of Solomyak inequality for an arbitrary $d$ is established in \cite{SZ-solomyak}.
It is interesting to compare the result of the following corollary with Theorem 9.4 in \cite{Shar14} which is proved under an artificial condition on Orlicz function $N.$ The result below holds for an arbitrary Orlicz function.
\begin{cor}\label{optimality within orlicz} If $N$ is an Orlicz function such that
$$\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_f(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}\Big\|_{\infty}\leq c_{d,N}\|f\|_{L_N},\quad f\in L_N(\mathbb{T}^d),$$
then $L_N(0,1)\subset L_M(0,1).$
\end{cor}
\begin{proof} By Lemma \ref{simple uniform lemma}, $L_N\subset\mathtt{M}_{\psi}.$ By Lemma \ref{orlicz in marcinkiewicz}, $L_N\subset L_M.$
\end{proof}
\subsection{Cwikel-Solomyak estimate in Lorentz ideals}
We have the uniform norm estimate for Cwikel operator. Solomyak proved the $\|\cdot\|_{1,\infty}$-quasi-norm estimate for Cwikel operator. The next natural step is to involve real interpolation to obtain Lorentz norm estimates for Cwikel operator.
\begin{cor} Let $1<p<\infty$ and $1\leq q\leq \infty.$ We have
$$\Big\|(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_f(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}\Big\|_{p,q}\leq c_{p,q,d}\|f\|_{[\Lambda_{\phi},\mathtt{M}_{\psi}]_{1-\frac1p,q}},\quad f\in [\Lambda_{\phi},\mathtt{M}_{\psi}]_{1-\frac1p,q}(\mathbb{T}^d).$$
\end{cor}
\begin{proof} Let $A:\mathtt{M}_{\psi}(\mathbb{T}^d)\to \mathcal{L}_{\infty}$ be a bounded operator defined by the setting
$$A:f\to (1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}M_f(1-\Delta_{\mathbb{T}^d})^{-\frac{d}{4}}.$$
We have that $A:\Lambda_{\psi}(\mathbb{T}^d)\to\mathcal{L}_{1,\infty}$ is bounded. By real interpolation, we have
$$A:[\Lambda_{\phi},\mathtt{M}_{\psi}]_{1-\frac1p,q}(\mathbb{T}^d)\to [\mathcal{L}_{1,\infty},\mathcal{L}_{\infty}]_{1-\frac1p,q}.$$
It is well-known that
$$[\mathcal{L}_{1,\infty},\mathcal{L}_{\infty}]_{1-\frac1p,q}=\mathcal{L}_{p,q}.$$
This completes the proof.
\end{proof}
|
\section{Introduction}
\label{introduction}
The study of flavor changing neutral currents (FCNCs) has long been a topic of great interest both theoretically and experimentally. This class of effects is considerably suppressed according to the standard model (SM), where they arise up to the one-loop level. Therefore FCNC transitions could provide signals of new physics and shed light on any possible SM extension. In the experimental side, the advent of the LHC offers a great potential to search for signals of various rare FCNC top quark decays, such as $t\to c V$ ($V=\gamma,g, Z$), $t\to cH$, $t\to c\ell^-\ell^+$, $t\to c \gamma\gamma$, $t\to c gg$, $t\to c\gamma H$, and $t\to c \gamma Z$. While the two-body decays $t\to c V$ and $t\to cH$ have been largely studied in the context of the SM and several of its extensions \cite{Eilam:1990zc,Lu:1996ji,Wang:1994qd,Lu:1998gm,Lu:2003yr,Couture:1994rr,Li:1993mg,Lopez:1997xv,Yang:1997dk,Frank:2005vd,GonzalezSprinberg:2007zz,CorderoCid:2005kp,CorderoCid:2004vi,Cortes-Maldonado:2013rca}, less attention has been paid to the three-body decays as they are expected to be more suppressed, but also because they involve lengthy and cumbersome calculations.
The decay $t\to c\gamma$ can only arise at the one-loop level due to electromagnetic gauge invariance and is further suppressed due to the GIM mechanism, so its branching ratio is considerably small, of the
order of $10^{-10}$ \cite{DiazCruz:1989ub}. However, in other SM extensions such a decay is not GIM-suppressed and its branching ratio $Br(t\to c\gamma)$ can be enhanced by several orders of magnitude, ranging from values of the order of $ 10^{-7}$ in two-Higgs doublet models \cite{Eilam:1990zc} up to $10^{-5}$ in supersymmetric models \cite{Couture:1994rr,Li:1993mg,Lopez:1997xv,Yang:1997dk}. Furthermore, some time ago, it was pointed out that there are some new physics scenarios where the three-body decay $t\to c\gamma\gamma$ could have a larger branching ratio than those of the two-body decays $t\to c\gamma$ \cite{Diaz-Cruz:1999wcs}. Thus, it is worth studying the rare three-body FCNC top quark decays in extension models despite the complexity involved in the respective calculation in order to assess if they could be at the reach of experimental detection. Along this line, the decay $t\to c\gamma\gamma$ has been studied in the framework of the little Higgs model with T-parity \cite{Han:2011xd,Han:2016hef} and also in a top-color assisted technicolor theory \cite{Yue:2001cy}.
In this work we will present a calculation of the $t\to c\gamma\gamma$ decay in the framework of leptoquark (LQ) models.
Two-body FCNC top quark decays have already been calculated in the context of this class of models: a calculation of the contribution of a model with an $SU(2)$ scalar LQ doublet to the two-body decays $t\to cX$ ($X=\gamma, Z, H, g$) and also to the three-body decay $t\to c\ell^-\ell^+$ was presented in Ref. \cite{Bolanos:2019dso} along with a comprehensive analysis of the parameter space of the model consistent with the then current constraints from direct LQ searches at the LHC, the Higgs boson coupling modifiers, the muon $g-2$ anomaly, and the lepton flavor violating decay (LFV) $\tau\to \mu \gamma$. To our knowledge there is no previous calculation of the contribution of LQs to the three-body FCNC top quark rare decay $t\to c\gamma\gamma$.
LQs are hypothetical particles carrying both lepton and color number that were proposed long ago in the context of Grand Unification theories \cite{Pati:1974yy,Georgi:1974sy,Fritzsch:1974nn,Georgi:1974yf} though they also appear naturally in other beyond the SM theories. A shortcoming of some of these models is that they could allow dangerous LQ diquark couplings that would induce proton decay at the tree-level, so additional symmetries must be invoked to preserve proton stability.
In addition, the LQ couplings to the first two fermion generations were found to be tightly constrained by low energy data \cite{Davidson:1993qk}. However, LQ couplings to the second and third fermion families are not strongly constrained yet and very recently LQ particles have become the source of renewed attention in the literature \cite{Dorsner:2016wpm} since they can explain some lepton flavor universality violating (LFUV) effects hinted at semi-leptonic $B$ decays, namely, the $R_{K^*}$ and $R_{D^*}$ anomalies \cite{Sakaki:2013bfa,Becirevic:2017jtw,
Crivellin:2017zlb,Becirevic:2018afm,Angelescu:2018tyl,Buttazzo:2017ixm,Marzocca:2021azj,Angelescu:2021lln}. Furthermore, LQs can also provide a possible explanation to the muon $g-2$ discrepancy and can be accommodated in models where neutrino mass is generated radiatively \cite{Saad:2020ihm}. It is thus interesting to calculate LQ contributions to FCNC rare top quark decays.
The rest of the presentation is organized as follows. In Section \ref{model} we present an overview
of LQ models and focus on a minimal renormalizable scalar LQ model with no proton decay, where there are potential sources of flavor change in the quark sector induced by scalar LQs. Section \ref{calculation} is devoted to discuss
the calculation of the $t\to c\gamma\gamma$ decay amplitude in our LQ model: for the sake of completeness, the invariant amplitude for the general fermion decay $f_i\to f_j\gamma\gamma$ is obtained via the Passarino-Veltman reduction method and the corresponding form factors are presented in \ref{FormFactors}. From these expressions the decay width for the $t\to c\gamma\gamma$ process follows straightforwardly. The numerical analysis of the LQ parameter space consistent with the current experimental constraints, considering two potential scenarios for the LQ couplings along with the numerical evaluation of the $t\to c\gamma\gamma$ branching ratio are presented in Sec. \ref{paramcons}. Finally, Sec. \ref{conclusions} is devoted to the conclusions and outlook.
\section{A renormalizable scalar LQ model with proton stability}
\label{model}
We now present the theoretical framework required for the calculation of the rare top quark decay $t\to c\gamma\gamma$ focusing on a model where there is no dangerous contribution to proton decay.
As already mentioned, LQs were proposed circa 1970 in the context of Grand Unification theories \cite{Pati:1974yy,Georgi:1974sy,Fritzsch:1974nn,Georgi:1974yf}, though such class of particles also arises naturally in theories with composite fermions \cite{Schrempp:1984nj,Buchmuller:1985nn,Gripaios:2009dq}, superstring-inspired $E_6$ models \cite{Witten:1985xc,Hewett:1988xc}, technicolor models \cite{Ellis:1980hz,Farhi:1980xs,Hill:2002ap}, etc. LQs are hypothetical particles carrying both lepton and color number, thereby coupling simultaneously to lepton and quarks. A systematic classification of all $SU(3)_C\times SU(2)_L\times U(1)_Y$ LQ representations and their renormalizable couplings to the SM fields was presented in Ref. \cite{Buchmuller:1986zs} via effective Lagrangians. According to all possible representations of the SM gauge group, it was found that LQs can be accommodated in ten representations: five scalar ones and five vector ones. The vector LQ couplings are tightly constrained by experimental data and recently more attention has been paid to scalar LQs. For the purpose of this work, we are interested in
the scalar LQ representation with the following quantum numbers under the SM gauge group $(3,2,7/6)$, which is usually denoted as $R_2$ in the literature \cite{Buchmuller:1986zs}. This $SU(2)$ doublet representation of LQs provides simple renormalizable models that conserve baryon number \cite{Arnold:2013cva}, thereby forbidding dangerous contributions to proton decay.
The phenomenology of LQ models with scalar doublets $R_2$ has been extensively studied in the literature. Very recently it has been shown that models with three LQs doublets $R_2^\ell$, dubbed $Y_{2\ell}$ in Ref. \cite{Crivellin:2021ejk}, can successfully explain some LFVU anomalies, such as the tensions between the SM predictions and the corresponding experimental measurements of the processes $(g-2)_\mu$, $b\to s \ell^+\ell^-$, and $b\to c\tau\nu$ \cite{Crivellin:2022mff,Cheung:2001ip,ColuccioLeskow:2016dox,Crivellin:2020tsz}. In addition, models with a scalar LQ doublet can also provide a mechanism of neutrino mass generation via the mixing of $R_2 $ with an extra LQ singlet $S_1$ through radiative corrections \cite{Zhang:2021dgl,AristizabalSierra:2007nf}.
In this work we will consider LQ models where the SM is augmented with one or more $SU(2)$ scalar doublets $R_2$ that induce the FCNC decay $t\to c\gamma\gamma$. We thus present the theoretical framework for a lone scalar doublet $R_2$ and the extension for models with multiple scalar LQ doublets $R_2$ will follow straightforwardly. As already mentioned, the presence of additional scalar LQ doublets $R_2$ can be necessary to explain the LFUV anomalies in $b$-hadron decays though not necessarily gives additional contributions to the FCNC top quark decays since some extra symmetries may be necessary to fulfil the experimental constraints.
The $R_2$ LQ doublet has hypercharge $7/6$, thereby giving rise to two scalar LQs with electric charges $2/3$ and $5/3$, which we denote by $\Omega_{2/3}$ and $\Omega_{5/3}$, where the subscripts denote the electric charge of each LQ. Both of them predict rich phenomenology and can give new contributions to several observables, which tightly constrain the parameter space.
In particular, the charge $5/3$ LQ couples to left- and right-handed fermions simultaneously and can induced new physics effects in the LFV Higgs decays $H\to \mu\tau$ \cite{Baek:2015mea} and $\ell_i\to \ell_j\gamma$ as well as the FCNC top quark decays $t\to c\gamma,cZ$, and $cH$ \cite{Bolanos:2019dso}. There could also be contributions to the $t\to c\gamma\gamma$ decay, which is the topic of interest of the present work.
The $R_2$ LQ doublet has the following general Yukawa lagrangian
\begin{equation}
\mathcal{L}_{F=0}=Y^{RL}_{ij} R_2^T\bar{u}^i_R i\tau_2 L_L^j+Y^{LR}_{ij}\bar{Q}^i_Le^j_R R_2+\text{H.c}.,
\end{equation}
where as usual $L_L^i$ and $Q_L^i$ are $SU(2)_L$ left-handed lepton and quark doublets, whereas $e_R^i$ and $q_R^i$ are $SU(2)$ singlets, with $i$ and $j$ being generation indices. For the Yukawa couplings we use the notation of Refs. \cite{Crivellin:2021ejk,Crivellin:2022mff} for comparison purposes. After electroweak symmetry breaking we expand the $SU(2)_L$ LQ doublet into its components: $R_2^T= \left(\Omega_{5/3}\;\; \Omega_{2/3}\right)$. The flavor eigenstates of the up quark $u$ can be chosen as the mass eigenstates $u'$, whereas the flavor eigenstates of the down quarks $d$ are rotated to the mass eigenstates $d^{'i}$ as $d^i\to V^{ik} d^{'i}$, with $V^{ik}$ the Cabibbo-Kobayashi-Maskawa mixing matrix element. We then obtain (dropping the prime for simplicity) the following interactions of the mass eigenstates
\begin{align}
\mathcal L_{F=0}&= \left(Y^{RL}_{ij}\bar u^i P_L e^j+Y^{LR}_{ij}\bar u^i P_R e^j\right)\Omega_{5/3}\nonumber\\
&+V^{ki*}Y^{LR}_{kj}\bar d^iP_R e^j\Omega_{2/3}-Y^{RL}_{ij}\bar u^iP_L \nu_j \Omega_{2/3}+\text{H.c.},
\end{align}
where $P_L$ and $P_R$ are the chiral projection operators.
In the last expression, the LQs have already been rotated to the mass eigenstates.
We will consider that both scalar LQs only couple to the fermions of the second and third generations as the couplings to the fermions of the first generation are strongly constrained by low-energy data.
As for the LQ couplings to the photon, they can be obtained from the kinetic lagrangian
\begin{equation}
\label{LQKinetic}
{\cal L}_{\text{kin}}=\frac{1}{2}(D_\mu R_2)^\dagger D^\mu R_2,
\end{equation}
where the $SU(2)_L\times U(1)_Y$ covariant derivative is given by
\begin{equation}
D_\mu R_2= \left(\partial_\mu+ig\frac{\tau^i}{2} W^i+i g'\frac{7}{6}B^\mu\right)R_2.
\end{equation}
Thus, the LQ couplings to one and two photons can be simply written as
\begin{align}
\mathcal{L}_{\text{kin}}&\supset\sum_{Q=2/3,5/3}\Big( iQ_Q A^\mu\left(\Omega_Q^*\partial_\mu \Omega_Q-\Omega_Q\partial_\mu \Omega_Q^* \right)\nonumber\\
&+Q_Q^2A_\mu A^\mu\Omega_Q^*\Omega_Q\Big),
\end{align}
where the subscript $Q$ denotes the LQ electric charge.
Finally, we consider the following renormalizable effective LQ interactions to the SM Higgs doublet $\Phi$
\begin{equation}\label{HiggsLQ}
\mathcal{L} = \left( M_{R_2}^2+\lambda_{R_2} \Phi^\dagger \Phi\right)\left(R_2^\dagger R_2 \right),
\end{equation}
where $M_{R_2}$ is the LQ mass matrix. After rotating to the mass eigenstates the LQ masses become non-degenerate at the lowest order in $\upsilon^2$. We can also obtain the Higgs boson coupling to the LQs:
\begin{equation}
\label{ScaLagI}
{\mathcal L}\supset \sum_{Q=2/3,5/3} \lambda_{\Omega_{Q}} v H \Omega_Q^* \Omega_Q,
\end{equation}
which can be useful to constrain the LQ couplings to the SM Higgs boson from the Higgs coupling modifiers $\kappa_\gamma$ and $\kappa_{g}$.
Apart from the usual SM Feynman rules, the remaining ones necessary for our calculation can be obtained from the above Lagrangians and are presented in Fig. \ref{FeynmanRules}. A complete set of Feynman rules for all the $SU(3)_c\times SU(2)_L\times U(1)_Y$ gauge invariant scalar LQ representations are presented in Ref. \cite{Crivellin:2021ejk}.
\begin{figure}[hbt!]
\centering
\includegraphics[width=8cm]{FeynRules}
\caption{ Feynman rules for the $R_2$ scalar LQ interactions to the fermions and photon necessary for the calculation of the $f_i\to f_j\gamma\gamma$ decay along with the usual SM Feynman rules. All the four-momenta are incoming.\label{FeynmanRules}}
\end{figure}
Below we present the calculations of the three-body decay $t\to c\gamma\gamma$.
\section{One-loop scalar LQ contribution to the $t\to c\gamma\gamma$ decay}
\label{calculation}
For the sake of completeness we have obtained general results for the contribution of a scalar LQ with charge $Q_S$ to the decay $f_i\to f_j\gamma\gamma$, where $f_{i}$($f_j$) can be a lepton or quark. These expressions are useful to calculate both the FCNC top quark decay $t\to c\gamma\gamma$ and the LFV decay $\ell_i\to \ell_j\gamma\gamma$ as well.
It is worth noting that our results are also valid for the contribution of other scalar LQs, such as the weak $SU(2)$ singlet $\chi_{1/3}$. Although the interaction of such a scalar LQ to a fermion pair involves Majorana-type Feynman rules that require special treatment, unlike the ones corresponding to the LQ $\Omega_{5/3}$ and $\Omega_{2/3}$ interactions, it can be shown that after some algebra the $\chi_{1/3}$ contributions turn out to be identical to those of $\Omega_{5/3}$ and can be obtained from the latter after replacing the electric charge $Q_{5/3}\to Q_{1/3}$ and the respective coupling constants to fermion pairs. A similar situation arises in the calculation of the LQ contribution to the two-body decay $f_i\to f_j\gamma$, where the results for $\chi_{1/3}$ can be obtained from the contribution of $\Omega_{5/3}$ once the corresponding electric charge and coupling constants are replaced.
For our calculation we use the following convention for the particle four-momenta:
\begin{equation}
f_i(p)\to f_j(p')\gamma(p_1) \gamma(p_2).
\end{equation}
with $\mu$ and $\nu$ the Lorentz indices of the photon four-momenta $p_1$ and $p_2$. Hence, the mass-shell conditions are given by $p_1^2=p_2^2=0$, $p^2=m_i^2$ and $p'^2=m_j^2$. However we will use the $m_j\to 0$ limit as a good approximation since in all $f_i\to f_j \gamma\gamma$ decays of phenomenological interest the mass of the outgoing fermion is always negligible as compared to the mass of the ingoing fermion. Also, due to the transversality conditions of the photon fields, i.e., $p_1^\mu \epsilon_\mu(p_1)=p_2^\nu \epsilon_\nu(p_2)=0$, any terms proportional to $p_1^\mu$ and $p_2^\nu$ can be dropped from the invariant amplitude before contracting with the respective photon polarization vectors. By the same reason, the replacement $p^\nu\to p_1^\nu+p^{\prime \nu}$ can also be done throughout the calculation.
At the one-loop level, the contribution from $\Omega_{5/3}$ to the rare three body decay $f_i\to f_j\gamma\gamma$ arises from the box diagrams shown in Fig. \ref{BoxDiagrams} as well as the bubble and triangle diagrams of Fig. \ref{FeynDiagrams1}.
\begin{figure}[h!]
\centering
\includegraphics[width=8cm]{BoxDiagrams}
\caption{Box diagrams that contribute to the decay $f_i\to f_j\gamma\gamma$ in the LQ model. There are three additional diagrams that are obtained by exchanging the photons.}\label{BoxDiagrams}.
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=8cm]{FeynDiagrams1}
\caption{Bubble and triangle Feynman diagrams for the decay $f_i\to f_j\gamma\gamma$ in the LQ model, where $f_i$ and $f_j$ are quarks (charged leptons) and $f_i$ is a lepton (quark). The crossed diagrams that are obtained by exchanging the photons are not shown.}\label{FeynDiagrams1}.
\end{figure}
After writing out the invariant amplitude for each Feynman diagram, the loop integrals were worked out with the Passarino-Veltman method
\cite{Passarino:1978jh}. This task was performed with the aid of the Mathematica package FeynCalc and a cross-check was done via Package-X \cite{Patel:2015tea}. We verified that the amplitude is free of ultraviolet divergences and obeys both Bose symmetry and gauge invariance under the $U(1)_\text{em}$ group. It is worth mentioning that ultraviolet divergencies cancel out separately in the amplitude of each set of Feynman diagrams of Figs. \ref{BoxDiagrams} and \ref{FeynDiagrams1}, whereas gauge invariance is only achieved after adding up all of the amplitudes.
In the $m_j\to 0$ limit, the invariant amplitude for the decay $f_i\to f_j\gamma\gamma$ can be conveniently written in the following way:
\begin{equation}
\label{invamplitude}
\mathcal M= \epsilon^*_\mu(p_1) \epsilon^*_\nu(p_2)T^{\alpha\mu}(p_1)T^{\beta\nu}(p_2)\bar{f}_j(\mathcal{M}_LP_L+\mathcal{M}_RP_R)f_i,
\end{equation}
where the tensor $T^{\alpha\beta}(p_i)$ is given by
\begin{align}
T^{\alpha\beta}(p_i)&=\frac{1}{m_i^2}\left((p_i\cdot p')g^{\alpha\beta}-p_i^\alpha p'^\beta\right).
\end{align}
which clearly obeys
\begin{align}
T^{\alpha\mu}(p_1)p_{1\mu}&=T^{\alpha\nu}(p_2)p_{2\nu}=0,
\end{align}
and thus electromagnetic gauge invariance is manifest.
As far as $\mathcal M_{L}$ and $\mathcal M_{R}$ are concerned, they are given in terms of six independent form factors $F_n(\hat{s},\hat{t})$
\begin{align}
\label{ML}
\mathcal M_L&=\frac{\alpha N_c}{4\pi m_i}\Bigg[F_1(\hat{s},\hat{t})\, \gamma_\alpha \gamma_\beta+\frac{1}{m_i}F_2(\hat{s},\hat{t})\, \gamma_\alpha p_{1\beta}\nonumber\\&+\frac{1}{m_i^2}F_3(\hat{s},\hat{t})\, p_{2\alpha} p_{1\beta}+\frac{1}{m_i}\Big(F_4(\hat{s},\hat{t})\, \gamma_\alpha \gamma_{\beta}\nonumber\\&+\frac{1}{m_i}F_5(\hat{s},\hat{t})\, \gamma_\alpha p_{1\beta}
+\frac{1}{m_i^2}F_6(\hat{s},\hat{t})\, p_{2\alpha} p_{1\beta}\Big)(\slashed{p_1}-\slashed{p_2})\Bigg]
\nonumber\\&+\left(\begin{array}{c}p_{1\mu}\leftrightarrow p_{2\nu}\\\hat{s}\leftrightarrow\hat{t}\end{array}\right),
\end{align}
where we introduced the Mandelstam-like scaled variables
\begin{align*}
\hat{s}&=\frac{1}{m_i^2}(p-p_1)^2,\\
\hat{t}&=\frac{1}{m_i^2}(p-p_2)^2.
\end{align*}
A similar expression to Eq. \eqref{ML} holds for $\mathcal M_R$ but with $F_n$ replaced by $\widetilde{F}_n$ ($n=1,\ldots 6$), which is obtained from $F_n$ as follows
\begin{equation}
\widetilde{F}_n=F_n\left(\begin{array}{c}Y^{RL}_{i k}\leftrightarrow Y^{LR}_{i k}\\Y^{RL}_{j k}\leftrightarrow Y^{LR}_{j k}\end{array}\right).
\end{equation}
From the above expressions, it is easy to show that Bose symmetry is obeyed.
The contributions of our LQ model to the $F_n(\hat{s},\hat{t})$ form factors are presented in \ref{FormFactors} in terms of Passarino-Veltman integral coefficients.
The decay width is given by
\begin{equation}
\label{decaywidth}
\Gamma(f_i\to f_j\gamma\gamma )=\frac{m_i}{256\pi^3}\int_0^1 d\hat{s}\int_0^{1-\hat{s}}d\hat{t} \left|\overline{\mathcal{M}}\right|^2,
\end{equation}
where the average square amplitude is presented for completeness in \ref{squareampli}.
From our general expressions for the decay $f_i\to f_j\gamma\gamma$, we can easily obtain the invariant amplitude and decay width of the $t\to c\gamma\gamma$ process, in which case the internal fermion $f_k$ in the Feynman diagrams of Figs. \ref{BoxDiagrams} and \ref{FeynDiagrams1} is a charged lepton. We thus set the mass of the decaying and internal fermions as $m_i\to m_t$ and $m_k=m_\ell$ with $\ell=\mu,\tau$, whereas for the final fermion we use $m_c\simeq 0$.
\section{LQ parameter space constraints}
\label{paramcons}
We now discuss the constraints from experimental data on the parameter space of the model and examine two scenarios: one with a single LQ doublet and another with multiple LQ doublets. The latter model is particularly interesting as allows one to explain the LFUV anomalies in $B$ decays and the muon anomaly.
Below we will concentrate on the bounds on the LQ masses and the couplings constants $Y^{RL,LR}_{i\ell_j}$, which are required to obtain an estimate for the decay $t\to c\gamma\gamma$.
\subsection{Bounds on LQ mass}
\label{LQmasscons}
The most up-to-date constraints on the masses of various kinds of vector and scalar LQs have been obtained from the data of direct searches for LQs at the CERN LHC by the CMS and ATLAS collaborations.
Scalar LQs have been searched for through singly production $pp\to S \bar\ell\to q\ell\bar\ell$ or pair production $pp\to S^\dagger S\to q\bar{q}\ell\bar\ell, q\bar q \nu\bar \nu$, thereby yielding bounds on the LQ masses, which depend on the LQ decay channels and the size of their couplings to fermions. Most of these bounds rely on the assumption that LQs can only couple to one generation of fermions and have a dominant decay channel, though very recently a more general scenario where LQs can couple simultaneously to fermions of distinct generations was analyzed \cite{ATLAS:2020dsk,ATLAS:2020xov,CMS:2018oaj}. Since a scalar LQ doublet with non-degenerate mass components can give dangerous contributions to the oblique parameters, we consider that both components of the scalar LQ doublet $R_2$ are mass degenerate, which indeed is true at the lowest order in $\upsilon$. We thus need to consider the current bounds on the masses of LQs of electric charges $2/3$ and $5/3$.
For the mass of a charge $5/3$ LQ, the most stringent bound was obtained by the CMS collaboration \cite{CMS:2018svy} by using data collected in 2016 at $\sqrt{s}=13$ TeV and assuming that the main LQ decay mode is that into a top quark and a tau lepton. Such an analysis has excluded a charge $5/3e$ scalar LQ with a mass below 900 GeV. As for the bounds on other types of scalar LQs, they are more stringent, slightly above 1 TeV. For instance, for the mass of a third-generation charge $2/3e$ scalar LQ decaying into $b\tau/t\nu_\tau$, the ATLAS collaboration \cite{ATLAS:2021jyv} has set a lower bound of about 1.2 TeV. As already mentioned, such a charge $2/3$ scalar LQ can be identified with the second component of the $R_2$ doublet and thus the mass constraint applies to both $\Omega_{2/3}$ and $\Omega_{5/3}$.
There are also theoretical analysis where the parameter space of scalar LQs have been constrained via the LHC data \cite{Schmaltz:2018nls}. If one considers models where LQs provide an explanation for the LFUV anomalies in $B$ meson decays, LQ couplings slightly larger than $O(1)$ are required. In this scenario, more stringent constraints on the LQ masses arise from LQ pair production \cite{Angelescu:2018tyl,Angelescu:2021lln}, ranging from 1 to 2 TeVs. In our analysis below we consider LQ masses above 1 TeV and impose constraints on the LQ couplings to fermions such that no dangerous LQ-mediated contributions to other observable quantities can be induced.
\subsection{Bounds on LQ couplings}
\label{LQcoupcons}
In order to discuss the constraints on the LQ couplings we will consider the following two scenarios for our model, which can allow us to asses the order of magnitude of the $t\to c\gamma\gamma$ branching ratio.
{\bf Scenario I)}There is only one LQ doublet $R_2$ that has both left- and right-handed couplings to fermions of the second and third generations only, namely, $Y^{RL,LR}_{i\ell}$, where $\ell=\mu,\tau$ and $i=2,3$ stands for the quark generation. In this scenario there are LFV transitions between the muon and the tau lepton and there is indeed an explanation for the muon $g-2$ anomaly, though an explanation for the LFUV anomalies in $B$ meson decays is disfavored. This model was considered in our previous work on the two-body top quark decays $t\to cX$ ($X=\gamma,g,H,Z$) \cite{Bolanos:2019dso} and constraints on the parameter space were obtained from the muon $g-2$ anomaly and the LFV decay $\tau\to \mu\gamma$, together with extra constraints to avoid large contributions to other observable quantities.
{\bf Scenario II)}There are multiple LQ doublets $R_2$: for instance in a particular realization recently proposed, the SM is augmented with three LQ doublets $R_2^\ell$ $(\ell=e,\mu,\tau)$, such that each one of them can only couple to the leptons of one generation, though the three of them can couple to quarks of the second and third generations \cite{Crivellin:2022mff}. This scenario does provides an explanation for the LFUV anomalies in $B$ meson decays as well as the $a_\mu$ anomaly, but forbids LFV effects, which can be dangerous since large LQ couplings to the fermions are required. An analysis on the constraints on the parameter space of this model was presented in Ref. \cite{Crivellin:2022mff}.
Below we analyze the constraints on the LQ couplings in the above scenarios and focus on the allowed parameter space region most promising for the $t\to c\gamma\gamma$ branching ratio.
\subsubsection{Scenario I}
In our analysis we will follow our previous work \cite{Bolanos:2019dso}, where we assumed that the $\Omega_{5/3}$ scalar LQ is responsible for the muon $g-2$ anomaly and considered the bounds on the $Y^{LR,RL}_{i\mu }$ and $Y^{LR,RL}_{i\tau }$ couplings obtained from the LFV decay $\tau\to \mu\gamma$.
The analytical expressions for the contribution of a scalar LQ to the muon anomalous magnetic dipole moment and the LFV decay $\ell_i\to \ell_j\gamma$ were obtained long ago and were also reproduced in Ref. \cite{Bolanos:2013tda,Bolanos:2019dso} in terms of Feynman parameter integrals and Passarino-Veltman scalar functions. For the sake of completeness we present such results in \ref{LFVformulas}. Note that $a_\mu$ contains a chirality flipping term proportional to $m_{q_i}\times {\rm Re}\left(Y_{i\mu}^{RL} Y_{i\mu}^{LR*}\right)$, which gives the dominant contribution for a heavy internal quark. Since such contribution requires that the scalar LQ has both left- and right-handed couplings to the fermions, it is absent for chiral LQs. Thus the contribution to $a_\mu$ from the chiral LQ $\Omega_{2/3}$ via an internal $b$ quark is expected to be much smaller than the contribution from the non-chiral LQ $\Omega_{5/3}$.
As far as the experimental constraints are concerned, for the muon AMDM $a_\mu$ we consider the current average of its experimental measurements \cite{Muong-2:2006rrc,Muong-2:2021ojo}, whereas for the SM theoretical prediction we consider the estimate obtained by the muon $g-2$ theory initiative \cite{Aoyama:2020ynm}. This yields the following $4.2\sigma$ discrepancy between theory and experiment:
\begin{equation}
\Delta a_\mu=251 (59)\times 10^{-11}.
\end{equation}
For the decay $\tau\to \mu\gamma$, the expected future experimental sensitivity \cite{Belle-II:2018jsg} puts strong constraints on LFV processes between the second and third generations
\begin{equation}
{\rm Br}(\tau\to \mu\gamma)\le 1.0\times 10^{-9},
\end{equation}
but we will consider the current experimental constraint \cite{BaBar:2009hkt}
\begin{equation}
{\rm Br}(\tau\to \mu\gamma)\le 4.4\times 10^{-8}.
\end{equation}
Note that $\Omega_{5/3}$ and $\Omega_{2/3}$ can also give dangerous contributions to several observables through their couplings to the $\bar b\tau$ and $t \nu_\tau$ pairs both at the tree and one-loop level. Therefore one must verify that the values of the LQ couplings consistent with the muon $g-2$ discrepancy and the constraint on the $\tau\to\mu\gamma$ decay do not introduce tension between the theory predictions and experimental data. For instance at the LHC, double (single) tau lepton production $\tau^- \tau^+$ ($\tau \nu$) can arise at the tree level via single LQ production $pp\to S\ell^+\to q\ell^-\ell^+$ ($pp\to S\ell^+\to q\nu\ell^+$), with $q$ being identified with a single jet. On the other hand, among those observables sensitive to large LQ couplings at the one-loop level there are the decays $Z\to \tau^-\tau^+$ and $Z\to \nu\nu$. We will thus impose the additional constraints $|Y^{LR,RL}_{i\mu}|<1$, which ensures that no tension will arise between the theory predictions and the experimental data \cite{Angelescu:2021lln,Crivellin:2022mff}. This is a more stringent constraint than $|Y^{LR,RL}_{i\mu}|<4\pi$, which is usually imposed to avoid the breakdown of perturbation theory.
Rather than making any a priori assumption about the LQ coupling constants to leptons and quarks, we consider non-vanishing left- and right-handed couplings to the leptons and quarks of both the second and third generations: $Y^{LR,RL}_{i\mu}$, where $i$ stands for the generation quark. Without losing generality we consider purely real LQ couplings and randomly scan for a few thousands of 4-tuples $\{Y_{2\mu }^{RL}$, $Y_{2\mu }^{LR}$, $Y_{3\mu }^{RL}$, $Y_{3\mu }^{LR}\}$ consistent with the discrepancy of the muon $g-2$ anomaly at 95\% C.L. The allowed region is shown in Fig. \ref{amubounds} on the ${\rm Re}\left(Y_{2\mu }^{RL} Y_{2\mu }^{LR}\right)$ vs ${\rm Re}\left(Y_{3\mu }^{RL} Y_{3\mu}^{LR}\right)$ plane for a charge $5/3$ scalar LQ with a mass of 1 TeV and $1.5$ TeV.
\begin{figure*}[!h]
\includegraphics[width=17cm]{amubounds.pdf}
\caption{Allowed area on the ${\rm Re}\left(Y_{2\mu }^{RL} Y_{2\mu }^{LR}\right)$ vs ${\rm Re}\left(Y_{3\mu }^{RL} Y_{3\mu}^{LR}\right)$ plane consistent with the muon $g-2$ anomaly at 95\% C.L. for a charge $5/3$ scalar LQ with mass of 1 and 1.5 TeVs in three scenarios of the signs of the ${\rm Re}\left(Y_{i\mu}^{RL} Y_{i\mu}^{LR}\right)$ products, which determine the signs of the partial contributions. As explained in the text, we also impose the extra constraints $|Y^{LR,RL}_{i\mu}|<1$ to be consistent with other constraints from experimental data. \label{amubounds}}
\end{figure*}
In our analysis we have considered three scenarios for the relative signs of the products of the left- and right-handed couplings ${\rm Re}\left(Y_{i\mu}^{RL} Y_{i\mu}^{LR}\right)$ as they determine the sign of the contribution to $a_\mu$ from the quark of generation $i$. We observe that the largest values for the coupling products are allowed when they have opposite signs (red and green points) as large partial contributions can cancel each other out to give the required negative total contribution. On the other hand, in the scenario when both couplings are negative (blue points), the partial contributions add up, thereby imposing a tighter constraint of the LQ coupling products.
Unless an additional flavor symmetry is introduced, $\Omega_{5/3}$ can also couple to the $\tau$ lepton and induce LFV processes.
Thus the decay $\tau\to \mu\gamma$ imposes an extra constraint on the $Y_{i\mu}^{RL,LR}$ and $Y_{i\tau}^{RL,LR}$ couplings, which must be combined with that arising from the muon $g-2$ anomaly. Again we do not impose an a priori condition for the coupling constants and randomly scan for set of points $\{Y_{2\ell}^{RL}$, $Y_{2\ell}^{LR}$, $Y_{3\ell}^{RL}$, $Y_{3\ell}^{LR}\}$ ($\ell=\mu,\tau$) consistent with the discrepancy of the muon $g-2$ anomaly and the experimental upper bound on the $\tau\to \mu\gamma$ decay, along with the extra upper bounds $|Y_{i\ell}^{RL,LR}|\le 1$. It turns out that the $t\to c\gamma\gamma$ invariant amplitude of Eq. \eqref{invamplitude} is given in terms of form factors of the form (see \ref{FormFactors})
\begin{align}
F_n
&=\sum_{\ell_k=\mu,\tau}\Big(f_n^{LL} Y^{RL}_{i\ell_k} Y^{RL}_{j\ell_k}+f_n^{RR} Y^{LR}_{i\ell_k} Y^{LR}_{j\ell_k}+f_n^{RL} Y^{LR}_{i\ell_k} Y^{RL}_{j\ell_k}
\nonumber\\& +f_n^{LR}Y^{RL}_{i\ell_k} Y^{LR}_{j\ell_k}\Big).
\label{Fn}
\end{align}
From our numerical analysis of the allowed values for the LQ coupling constants, we infer that from all the products of coupling constants of Eq. \eqref{Fn}, the ones that can reach the largest allowed values are $Y^{RL}_{3\mu} Y^{RL}_{2\mu}$ and $Y^{LR}_{3\tau} Y^{LR}_{2\tau}$, whereas the remaining products are much more constrained. Therefore the $t\to c\gamma\gamma$ decay width will depend mainly on the values of this pair of coupling products and it will reach its maximal value when one of them reaches its largest allowed values. We thus show the allowed region on the ${\rm Re}\left(Y_{2\tau }^{RL} Y_{3\tau}^{RL}\right)$ vs ${\rm Re}\left(Y_{2\tau }^{LR} Y_{3\tau }^{LR}\right)$ plane in Fig. \ref{lfvbounds} and on the ${\rm Re}\left(Y_{2\mu }^{RL} Y_{3\mu}^{RL}\right)$ vs ${\rm Re}\left(Y_{2\mu }^{LR} Y_{3\mu }^{LR}\right)$ plane in Fig. \ref{lfvbounds1} for a scalar LQ $\Omega_{5/3}$ with mass of 1 TeV (top plots) and 1.5 TeV (bottom plots) in the three scenarios of the signs of the products ${\rm Re}\left(Y_{i\mu}^{RL} Y_{i\mu}^{LR}\right)$ of Fig. \ref{amubounds}.
\begin{figure*}[!h]
\includegraphics[width=17cm]{LFVbounds.pdf}
\caption{Allowed area on the ${\rm Re}\left(Y_{2\tau }^{RL} Y_{3\tau}^{RL}\right)$ vs ${\rm Re}\left(Y_{2\tau }^{LR} Y_{3\tau }^{LR}\right)$ plane consistent with the muon $g-2$ anomaly at 95\% C.L. and the LFV decay $\tau\to\mu\gamma$ for a scalar LQ $\Omega_{5/3}$ with mass of 1 TeV (top plots) and 1.5 TeV (bottom plots) in the three scenarios of the signs of the products ${\rm Re}\left(Y_{i\mu}^{RL} Y_{i\mu}^{LR}\right)$ considered in Fig \ref{amubounds}. The constraint $|Y_{i \ell }^{RL,LR}\le| 1$ is also imposed to be consistent with other experimental constraints. \label{lfvbounds}}
\end{figure*}
\begin{figure*}[!h]
\includegraphics[width=17cm]{LFVbounds1.pdf}
\caption{The same as in Fig. \ref{lfvbounds} but for the allowed area on the ${\rm Re}\left(Y_{2\mu }^{RL} Y_{3\mu}^{RL}\right)$ vs ${\rm Re}\left(Y_{2\mu }^{LR} Y_{3\mu }^{LR}\right)$ plane. \label{lfvbounds1}}
\end{figure*}
We can conclude that the largest allowed values correspond to $Y_{2\tau }^{LR} Y_{3\tau }^{LR}$ products, which are obtained when both the $c$ and $t$ contributions to $a_\mu$ have opposite signs (red and green points). The products of LQ couplings to the muon are more restricted, which is due to the condition imposed by the muon $g-2$ anomaly.
For illustration purpose, in Table \ref{allowpoints} we show a few sets of values in which the product $Y^{LR}_{2\tau} Y^{LR}_{3\tau}$ reaches its largest allowed values, which are slightly below the unity. These sets of points will yield the maximal values of the $t\to c\gamma\gamma$ branching ratio in scenario I.
Also, as expected we observe that the constraints on the LQ couplings are relaxed when the LQ mass increases and the allowed area enlarges slightly when the LQ increases from 1 to 1.5 TeVs. However, although the LQ couplings could be less restricted, a possible enhancement of the $t\to c\gamma\gamma$ branching ratio may be suppressed by the larger value of the LQ mass.
\subsubsection{Scenario II}
Since LFUV anomalies in $B$ meson decays cannot be addressed in models with a lone $R_2$ doublet \cite{Angelescu:2021lln}, quite recently scenarios with multiple scalar LQ multiplets have been examined, such as the one of Ref. \cite{Crivellin:2022mff}, where a LQ doublet $R_2^\ell$ for each lepton generation was introduced. In order to avoid dangerous LFV effects induced by large LQ couplings, an additional flavor symmetry must be imposed such that each scalar LQ doublet $R_2^\ell$ ($\ell=\mu,\tau$) can only couple to one lepton generation, though they still have couplings to quarks of the second and third generations. Therefore, although no LQ-mediated LFV processes arise, FCNC top quark decays could still be allowed at the one-loop level.
A comprehensive analysis of the bounds on LQ $R_2^\tau$ couplings obtained from several observables was presented in Ref. \cite{Crivellin:2022mff}.
It was shown that one scalar LQ doublet $R_2^\tau$ with a mass of 1.7 TeVs can explain the $R_D$ and $R_{D^*}$ anomalies, provided that the values of the coupling constants $Y_{3\tau}^{LR}$ and $Y_{2\tau}^{RL*}$ are slightly larger than $O(1)$ and there is a large complex phase in the product of both coupling constants. This is still consistent with the data on $\tau \nu$ and $\tau^- \tau^+$ production at the LHC, but some tension could arise in the electroweak fit for the data of the decays $Z\to \tau^-\tau^+$ and $Z\to \bar{\nu}\nu$ \cite{Crivellin:2022mff}. Also, the constraint $Y_{2\tau}^{LR} Y_{3\tau}^{LR*}\lesssim 0.25$ is obtained from $B_s-\bar{B}_s^*$ mixing \cite{Crivellin:2022mff}.
The $R_2^\tau$ doublet alone gives no explanation for the $R_{K^*}$ anomaly and the muon $g-2$ discrepancy, thus a second doublet LQ doublet $R_2^\mu$ is necessary \cite{Crivellin:2022mff}. It was shown that the presence of two scalar doublets improves the fit of the $b\to s\ell^-\ell^+$ data and relax the tension in the electroweak data as smaller $R_2^\tau$ couplings are required to explain the $R_D$ and $R_{D^*}$ anomalies for $M_{R_2^\mu}=2$ TeV and $M_{R_2^\tau}=1.7$ TeV. A summary of the values of the coupling constants consistent with the LFUV anomalies in $B$ decays and the $a_\mu$ anomaly, taken from Ref. \cite{Crivellin:2022mff}, is shown in the last row of Table \ref{allowpoints}. Note that the LQ doublet $R_2^\tau$ alone could yield an enhanced branching ratio for FCNC top quark decays due to the large LQ couplings, though an additional suppression due to the heavier value of the LQ mass is expected too.
\begin{table*}[!h]
\begin{center}
\caption{Sample sets of values of LQ couplings to fermions consistent with the muon $g-2$ anomaly and
the experimental constraint on the LFV decay $\tau\to\mu\gamma$ for $m_{R_2}=1$ TeV (rows 1 through 3) and $m_{R_2}=1.5$ TeV (rows 4 through 7) in scenario I. The last row shows the allowed values consistent with an explanation for the $R_{D,D^*}$ and $R_{K^*}$ anomalies, the muon $g-2$ discrepancy, and other experimental constraints for $m_{R_2^\mu}=1.7$ TeV and $m_{R_2^\mu}=2$ TeV in scenario II (results extracted from Ref. \cite{Crivellin:2022mff}).\label{allowpoints}}
\begin{tabular}{cccccccccc}
\hline
\hline
Scenario&LQ mass (TeV) &$Y_{2\mu}^{RL}$& $Y_{2\mu}^{LR}$&$Y_{3\mu}^{RL}$& $Y_{3\mu}^{LR}$&$Y_{2\tau}^{RL}$& $Y_{2\tau}^{LR}$
&$Y_{3\tau}^{RL}$& $Y_{3\tau}^{LR}$\\
\hline
\hline
I&1.0&-0.167 & 0.246 & 0.011 & 0.018 & 0.059 & 0.92 & 0.015 & 0.88\\
I&1.0& 0.949 & 0.013 & -0.073 & 0.018 & 0.011 & 0.917 & 0.018 & 0.798 \\
I&1.0& 0.193 & 0.019 & -0.014 & 0.044 & 0.024 & 0.978 & 0.012 & 0.888 \\
I&1.5&0.431 & 0.041 & -0.025 & 0.011 & 0.037 & 0.994 & 0.016 & 0.831 \\
I&1.5& 0.424 & 0.012 & -0.02 & 0.021 & 0.023 & 0.957 & 0.041 & 0.917 \\
I&1.5&-0.236 & 0.155 & 0.012 & 0.038 & 0.016 & 0.958 & 0.013 & 0.995 \\
II&$1.7$ $(R^\tau)$, $2.0$ $(R^\mu)$&--&--&2.4-2.6&$-7\times 10^{-3}$&2.4-2.55&0.3-0.35&--&0.9-1.1\\
\hline
\end{tabular}
\end{center}
\end{table*}
Following the notation of Ref. \cite{Angelescu:2018tyl} we present in Table \ref{modelfea} the main features of the scenarios just discussed. It seems that scenario II is the most promising for a less suppressed $t\to c\gamma\gamma$ branching ratio as LQ couplings to the $\tau$ lepton of the order of $O(1)$ are allowed, however, a larger mass is also necessary to fulfil the constraints from experimental data, which may result in an additional suppression. Below we present an estimate for the $t\to c\gamma\gamma$ branching ratio in these scenarios.
\begin{table}[!h]
\begin{center}
\caption{Scenarios discussed in the text along with the anomalies addressed and other predicted new physics effects.\label{modelfea}}
\begin{tabular}{lccccc}
\hline
\hline
Scenario &$a_\mu$&$R_D$, $R_{D^*}$&$R_{K^*}$&LFV&FCNC\\
\hline
\hline
I&\Checkmark&\XSolidBrush&\XSolidBrush&\Checkmark&\Checkmark\\
II with $R_2^\tau$ alone&\Checkmark&\Checkmark&\XSolidBrush&\XSolidBrush&\Checkmark\\
II with both $R_2^\mu$ and $R_2^\tau$&\Checkmark&\Checkmark&\Checkmark&\XSolidBrush&\Checkmark\\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Estimate of the $t\to c\gamma\gamma$ branching ratio}
For the numerical evaluation of the invariant amplitude \eqref{invamplitude} and to achieve a best numerical stability in the evaluation of the double integral of Eq. \eqref{decaywidth}, we decomposed the Passarino-Veltman integrals of Eqs. \eqref{f1LL} through \eqref{f6RR} into scalar functions (the results are too lengthy to be presented in this work), which then were evaluated through the LoopTools \cite{vanOldenborgh:1990yc,Hahn:1998yk} package. Also, an independent evaluation was performed via the Collier package \cite{Denner:2016kdg}, which showed a good agreement with the LoopTools evaluation. To obtain de $t\to c\gamma\gamma$ decay width we impose the kinematic cuts $E_\gamma>5$ GeV for both photons.
As far as scenario I is concerned, from the above analysis we can conclude that the contribution to the $t\to c\gamma\gamma$ decay width from the loops with an internal muon is much smaller than that from the loops with an internal tau lepton, which is due to the small allowed values of the LQ-muon couplings. In fact we can neglect the muon contribution to the $t\to c\gamma\gamma$ invariant amplitude as it is more than two orders of magnitude below than the one of the tau lepton. Therefore, the largest values of the $t\to c\gamma\gamma$ decay width for a specific value of the LQ mass are reached in the region where the product of couplings $Y^{LR}_{2\tau} Y^{LR}_{3\tau}$ reaches its largest allowed values. In this region the behavior of the $t\to c\gamma\gamma$ branching ratio as a function of the LQ coupling can be roughly approximated as
\begin{equation}
Br(t\to c\gamma\gamma)\simeq f(m_S,m_c,m_t,m_\tau) |Y^{LR}_{2\tau} Y^{LR}_{3\tau}|^2,
\end{equation}
with $f(m_S,m_c,m_t,m_\tau)$ of the order of $10^{-11}$--$10^{-10}$ at most for $m_S\simeq 1$ TeV.
In Fig. \ref{branchingplot} we show the behavior of the LQ contribution to $Br(t\to c\gamma\gamma)$ in scenario I as a function of the LQ mass around the region where the $Y^{LR}_{2\tau} Y^{LR}_{3\tau}$ product reaches its largest allowed values and thus $Br(t\to c\gamma\gamma)$ reaches its largest values. For illustration purpose we also show the case where the product $Y^{RL}_{2\tau} Y^{RL}_{3\tau}$ would dominate over $Y^{LR}_{2\tau} Y^{LR}_{3\tau}$, though in this case $Br(t\to c\gamma\gamma)$ does not reach its largest values as $Y^{RL}_{2\tau} Y^{RL}_{3\tau}$ is allowed to be of the order of $10^{-1}$ at most.
\begin{figure}[!ht]
\includegraphics[width=8cm]{Brs.pdf}
\caption{LQ contribution to the branching ratio of the $t\to c\gamma\gamma$ decay in scenario I as a function of the LQ mass in the region of the parameter space where the product $Y^{LR}_{2\tau} Y^{LR}_{3\tau}\sim O(1)$ reaches its largest allowed values, which corresponds to the largest possible value of $Br(t\to c \gamma\gamma)$. We also include the case where the product $Y^{RL}_{2\tau} Y^{RL}_{3\tau}$ dominates over $Y^{LR}_{2\tau} Y^{LR}_{3\tau}$ for $Y^{RL}_{2\tau} Y^{RL}_{3\tau}\simeq O(1)$, though this is not allowed by the constraints from experimental data. \label{branchingplot}}
\end{figure}
We can conclude that the LQ contribution to $Br(t\to c\gamma\gamma)$ can be of the order of $10^{-11}$ at most for $Y^{LR}_{2\tau} Y^{LR}_{3\tau}$ of the order of $O(1)$, though it would decrease considerably if the LQ couplings decrease by one order of magnitude. Also, we note that there is little dependence on the LQ mass in the interval from 1 to 2 TeVs.
As for scenario II, according to the allowed values of the LQ couplings presented in Table \ref{allowpoints} from Ref. \cite{Crivellin:2022mff}, $Br(t\to c\gamma\gamma)$ would be dominated by the $R_2^\tau$ contribution since any products of $R_2^\mu$ couplings that enter into Eq. \eqref{Fn} would be considerably suppressed. Furthermore, the dominant term arises from the $Y^{RL}_{2\tau} Y^{LR}_{3\tau}$ product. In this case we obtain for $m_{R_2^\tau}=1.7$ TeV
\begin{equation}
Br(t\to c\gamma\gamma)\simeq 9.11\times 10^{-12}\times| Y^{RL}_{2\tau} Y^{LR}_{3\tau}|^2,
\end{equation}
where we have neglected all other contributions. For $Y^{RL}_{2\tau}\simeq 2$ and $Y^{LR}_{3\tau}\simeq 1$ we obtain again the estimate $Br(t\to c\gamma\gamma)\simeq 10^{-11}$. Therefore, in both scenario I and scenario II, the $t\to c\gamma\gamma$ branching ratio would reach values as high as $10^{-12}$--$10^{-11}$. These are the largest possible values that one can expect for the contribution to the $t\to c \gamma\gamma$ decay from the $R_2$ doublet scalar LQs since experimental constraints severely constraint the LQ coupling constants. Unless an extraordinary cancellation in the LQ contribution to experimental observables (fine tunning) occurs in a more sophisticated framework with additional LQ multiplets, the LQ contribution to the $t\to c\gamma\gamma$ decay is expected to be beyond the observable level.
In Table \ref{branchings} we present a summary of the one-loop contributions to two- and three-body FCNC top quark decays from a scalar $SU(2)$ doublet from Ref. \cite{Bolanos:2019dso} and this work. We have considered the largest estimate consistent with the current constraints from experimental data.
\begin{table}[!htb]
\begin{center}
\caption{Largest estimated values of the one-loop contributions to two- and three-body FCNC top quark decays from models with scalar $SU(2)$ doublets consistent with the current constraints from experimental data \cite{Bolanos:2019dso}.\label{branchings}}
\begin{tabular}{ccc}
\hline
\hline
Decay channel&Branching ratio &Branching ratio \\
&$m_{R_2}=1$ TeV&$m_{R_2}=1.5$ TeV\\
\hline
\hline
$t\to c\gamma$&$10^{-9}$&$10^{-9}$\\
$t\to cg$&$10^{-9}$&$10^{-10}$\\
$t\to cZ$&$10^{-8}$&$10^{-9}$\\
$t\to cH$&$10^{-9}$&$10^{-10}$\\
$t\to c\mu^-\mu^+$&$10^{-6}$&$10^{-7}$\\
$t\to c\tau^-\tau^+$&$10^{-7}$&$10^{-8}$\\
$t\to c\gamma\gamma$&$10^{-11}$&$10^{-12}$\\
\hline
\end{tabular}
\end{center}
\end{table}
\section{Final remarks}
In this work we have presented a calculation of the rare three-body FCNC top quark decay $t\to c\gamma\gamma$ in the framework of a renormalizable model where the SM is augmented with one or three $SU(2)$ scalar LQ doublets with hypercharge $7/6$, which gives rise to two scalar LQs with electric charge of $5/3$ and $2/3$. We considered two particular scenarios: a minimal model with a lone scalar LQ doublet (scenario I) and another model with three scalar LQ doublets $R_2^\ell$ ($\ell=\mu,\tau$) (scenario II). While scenario I can address the muon $g-2$ anomaly, scenario II was proposed recently \cite{Crivellin:2022mff} to also explain the LFUV anomalies in $b$-hadron decays. The general aspects and the generic Lagrangian and Feynman rules for this class of models are discussed, and analytical expressions for the one-loop LQ contribution to the invariant amplitude of the general decay $f_i\to f_j\gamma\gamma$ are presented in terms of Passarino-Veltman integral coefficients, from which the corresponding invariant amplitude for the decay $t\to c\gamma\gamma$ follows easily.
A discussion of the current bounds on the LQ masses and the LQ couplings to leptons and quarks is also presented. In scenario I the region of allowed values of LQ couplings is found by requiring that the charge $5/3$ is responsible for the muon $g-2$ discrepancy and obeys the constraint on the LFV decay $\tau\to \mu \gamma$ as well as the extra constraint on the LQ couplings $|Y^{LR,RL}_{i\ell}|\le 1$, which is imposed to avoid tension with experimental data of processes sensitive to LQ contributions. It is found that the LQ couplings can be as large as $0.1$--$1$ for a LQ mass around 1 TeV, which results in a branching ratio for the $t\to c\gamma\gamma$ decay of the order of $10^{-11}$--$10^{-12}$. When the mass of the LQ increases up to around $1.5$ TeV, this branching ratios decreases slowly but there is a high dependence on the magnitude of the LQ coupling constants.
As far as scenario II is concerned, we consider the bounds obtained in the analysis of Ref. \cite{Crivellin:2022mff}, where the parameter space of the model was constrained by requiring a solution to the LFUV anomalies and the muon $g-2$ discrepancy, along with constraints from experimental data. For a LQ doublet $R_2^\tau$ with a mass of $1.7$ TeV, LQ couplings to the $\tau t$ pair with values slightly larger than $O(1)$ are still allowed, whereas the couplings of the $R_2^\mu$ doublet would be one or two orders of magnitude below for a LQ mass of 2 TeV. In this scenario the branching ratio of the $t\to c\gamma\gamma$ decay is also of the order of $10^{-11}$--$10^{-12}$.
In conclusion the LQ contribution to the branching ratio of the three body decay $t\to c\gamma\gamma$ is about two or three orders of magnitude below than the one for the two-body decay $t\to c\gamma$, which is below the expected experimental reach.
\label{conclusions}
\begin{acknowledgements}
The work of R. S\'anchez-V\'elez was partially supported by Consejo Nacional
de Ciencia y Tecnolog\'ia under grant A1-S-23238. G. Tavares-Velasco acknowledges partial support from Sistema Nacional de Investigadores (Mexico) and Vicerrector\'ia de Investigaci\'on y Estudios de Posgrado de la Ben\'emerita
Universidad Aut\'onoma de Puebla.
\end{acknowledgements}
\onecolumn
|
\section{Introduction}
Cleft lip and palate (CLP)~\cite{Kummer2016Evaluation}, is the most common congenital deformity in the oral and maxillofacial region, with the incidence of 1.4 in 1000 live births in human. In CLP patients, they usually have inability to pronounce the normal sound due to the incomplete closure of their soft palate (\emph{i.e.}, velopharyngeal dysfunction), and thus results in the production of hypernasality~\cite{Vijayalakshmi2007Acoustic,Guo2006Voice}. To treat this illness, CLP palates are usually required to conduct a series of palate surgery and subsequent speech therapy. Therefore, how to estimate the CLP severity is a critical element to determine the final treatment for patients.
Hypernasality is regarded as one of the primary symptoms in cleft palate and it is usually characterized as excessive nasal resonance in pronouciation. As a result, the CLP patients with hypernasality cannot pronounce vowel and consonant as clear as normal people. Therefore, to a certain degree, the severity of hypernasality reflects the degree of the opening and closing of a velopharyngeal passageway between the oral cavity and nasal cavity. How to estimate the severity of hypernasality has been considered as an important metric to evaluate the outcome of primary cleft palate repair, and to determine the need for the further treatment, such as pharyngoplasty and speech therapy. In clinical examinations, hypernasality rating is usually evaluated by speech-language pathologists (SLPs). However, availability of expert SLPs is usually limited, and the diversity of intra-rater and inter-rater reliablities will also affect the result of subjective hypernasality evaluation. Accordingly, some scientists attempted to investigate the possibility of detecting hypernasality by using machines or instruments. For example, nasometer~\cite{Bettens2014Instrumental} is a widely used instrument to measure nasalance with a rating of 0-100. Nevertheless, these instrumental methods do not demonstrate high correlation with clinical perception of hypernasality and these instruments also require experienced clinicians to operate~\cite{WATTERSON199313}.
Recently, machine learning based automatic hypernasality estimation has drawn enormous attention. Existing methods are mainly based on speech signal processing as cleft palate speech exhibits abnormal nasal resonance in the spectrum features. The training pipeline of machine learning based methods is to first extract acoustic features, like Mel-frequency cepstral coefficients (MFCC) or filter banks, and then perform classification using support vector machine and Gaussian mixture models. For example, \cite{Maier2009Automatic} applied voice intonation features, MFCC and Teager energy operator to detect hypernasality. \cite{He2014Automatic} investigated the effect of energy shifts to the low-frequency bands to estimate hypernasality. \cite{Mohammad2020Single} introduced a single frequency filter bank based long-term average spectral for hypernasality estimation. \cite{Golabbakhsh2017acoustic} analyzed multiple different acoustic features for automatic identification of hypernasality. \cite{Dubey2019Hypernasality,Saxon2019Objective} also designed some novel acoustic features to better extract hypernasality-related semantics. These works mainly focused on extracting or designing advanced acoustic features for hypernasal speech detection. Inspired by the success of deep neural network in speech processing tasks, some works~\cite{Vikram2021Attention,Xiyue2019HypernasalityNet,Vikram2021Deep,Wang2019cnn} also tried to use recurrent neural network, convolutional network or attention model to detect hypernasality in an end-to-end manner.
Although significant progresses have been achieved in evaluating hypernasality, there still remain a few limitations in the current systems. First, existing systems for hypernasality estimation only leverage cleft palate datasets. However, different from common speech processing tasks, cleft palate datasets are usually extremely difficult to collect, due to privacy and its annotations that require expert SLPs labeling. In other words, CLP datasets are usually in a low-resource scenario, which results in a poor generalization for model capacity. Secondly, previous work~\cite{bechet12Consonantal,Nikitha2017Vowel} have analyzed that hypernasality has obviously abnormal area in acoustic space (\emph{e.g.}, vowel or consonant), especially in the spectrum dimension~\cite{Xiyue2019HypernasalityNet}. Thus, it can be concluded that learning high quality acoustic features is beneficial for predicting hypernasality. However, existing systems (especially neural network based models) for hypernasality estimation only build the connection between the CLP speech and its hypernasality rating, without modeling acoustic features explicitly or implicitly. How to address the above two items for hypernasality estimation is the target of our paper.
Different from CLP datasets, many speech processing tasks in general domain usually have rich resource. Specifically, automatic speech recognition (ASR)~\cite{James1990ASR} is one of the most representative tasks in speech processing. The objective of ASR is to recognize the content of the input speech and return its corresponding text. By leveraging additional text annotations, the model is able to understand the content of speech semantic (\emph{e.g.}, vowel or consonant) and extract high quality representations for acoustic features. In addition, many ASR datasets~\cite{Vassil2015Librispeech,Hui2017Aishell} usually have enormous speech data (\emph{e.g.}, Librispeech~\cite{Vassil2015Librispeech} includes 960 hrs data), and do not suffer from the low-resource issue. Considering these characteristics of ASR task, we have raised a hypothesis: is it possible to improve hypernasality estimation by leveraging the advantages of the ASR objective.
As aforementioned, hypernasality derives abnormal expressions in the acoustic space (\emph{e.g.}, vowel and consonant). The existing advanced methods for hypernasality estimation are mainly using neural network~\cite{Vikram2021Attention,Xiyue2019HypernasalityNet,Vikram2021Deep,Wang2019cnn}. But all of them just simply stacked multiple neural network layers over the acoustic features (like MFCCs) to conduct classification, and cannot enable model to further extract better acoustic semantics since these methods only focused on learning hypernasality-related features and ignored to understand acoustic semantics. However, it is worthy to note that ASR models can naturally extract high-quality representations of acoustic features (especially in identifying vowel and consonant), since large-scale text annotations encourage model to thoroughly understand the content of acoustic features (\emph{e.g.}, for text ``am" and its phonation /\textschwa m/, the model needs to identify vowel /\textschwa/ and consonant /m/, and then returns the answer). Inspired by success of transfer learning in other tasks~\cite{Huh2016imagenet,Jacob2019BERT}, we deem that using ASR model for initialization is beneficial for estimating hypernasality.
Therefore, in this paper, we introduce a novel approach to leverage ASR model for hypernasality estimation from the perspective of transfer learning. Specifically, we first pre-train an encoder-decoder framework by using the ASR objective on large-scale ASR corpus, and then apply the pre-trained ASR encoder to conduct hypernasality estimation on CLP corpus. Such design can take advantages of ASR model in learning hypernasality-related semantics: 1) ASR tasks usually includes enormous audio data that enables model to obtain a better generalization; 2) Benefiting from text annotations, the neural network based encoder with ASR objective is able to extract better phonetic representations for hypernasality estimation. Experimental results on two different CLP datasets also indicate that our model with ASR initialization is superior to the base model without using initialization.
\begin{figure}[!t]
\centering
\includegraphics[width=0.45\textwidth]{figs/asr.pdf}
\caption{The training pipeline of our approach. The first row is using ASR dataset for model training, and then use ASR-trained encoder for an initialization to conduct hypernasality assessment on cleft palate dataset.}
\label{fig_asr}
\end{figure}
\section{Method}
\label{sec2:method}
\subsection{ASR Formulation}
Automatic speech recognition (ASR) is to identify the content of human voices and then return the corresponding text. Assuming ASR corpus as ($\mathcal{S}, \mathcal{T}$), where $\mathcal{S}$ and $\mathcal{T}$ correspond to audios and the corresponding language descriptions, respectively. For audio data, they are first processed as a sequence of acoustic features (\emph{e.g.}, Mel filter-bank features) at the frame level, and then feed into the model. Generally, we adopt an encoder-decoder framework~\cite{Sutskever2014sequence} to handle ASR task where the decoder is formulated as an auto-regressive manner, and the parameters are defined as $\theta = \{\theta_{enc}, \theta_{dec}\}$. The objective of ASR system is to predict the most possible corresponding text from the given speech as $\arg \max P(\cal{T}|\cal{S})$. The objective function is optimized by maximum likelihood estimation as:
\begin{equation}
\label{eq1}
\mathcal{L}_{ASR} = - \sum_{(s, t)}^{(\mathcal{S}, \mathcal{T})} \sum_{i=1}^{|t|} \log P(t_i | t_{<i}, s; \theta_{enc}, \theta_{dec}),
\end{equation}
where $(s, t)$ is a paired speech and text, $t_{<i}$ means tokens before position $i$ and $\{\theta_{enc}, \theta_{dec}\}$ represent the parameters of encoder and decoder, and $P(t_i | t_{<i}, s; \theta_{enc}, \theta_{dec})$ is to predict the $i$-th token in the text sequence based on previous position and the given speech. Therefore, by optimizing the Eqn.~\ref{eq1}, we can obtain a well-trained ASR model in the general domain, and its encoder can extract deep representations from acoustic features so that the decoder is able to predict the corresponding text accurately.
\subsection{Hypernasality Estimation Formulation}
The target of hypernasality estimation is to evaluate the level of hypernasality from the speech of CLP patients. More specifically, by adopting different evaluation criteria, hypernasality estimation can be subdivided as hypernasality detection and hypernasality assessment. The former is to detect whether the cleft palate speech includes hypernasality or not, which can be simplified as a binary classification task. The latter is to predict the hypernasality rating of the corresponding cleft palate speech to determine the hypernasality severity, which is a more challenging multi-class classification task. In this paper, the rating of hypernasality severity is formulated as a range of 0$-$3 (0-normal, 1-mild, 2-moderate, 3-severe), each of which represents one class in the multi-class classification task.
The CLP corpus can be denoted as a 2-tuple ($\mathcal{X}, \mathcal{Y}$), where $\mathcal{X}$ and $\mathcal{Y}$ represent the datasets of CLP speech and the corresponding hypernasality rating, respectively. For CLP speech, we adopt the same preprocessing step to obtain the similar acoustic features (\emph{i.e.}, Mel filter-bank feature), like ASR task, to guarantee the consistency. To conduct hypernasality estimation, we employ an encoder network with a classifier layer, where the parameters are defined as $\tilde{\theta} = \{\tilde{\theta}_{enc}, \tilde{\theta}_{cls}\}$. Therefore, the objective of hypernasality estimation is formulated as following:
\begin{equation}
\label{eq2}
\mathcal{L}_{CLS} = - \sum_{(x, y)}^{(\mathcal{X}, \mathcal{Y})} \log P(y |x; \tilde{\theta}_{enc}, \tilde{\theta}_{cls}),
\end{equation}
where $\tilde{\theta}_{enc}$ and $\tilde{\theta}_{cls}$ represent the parameters of the encoder and classifier respectively. The classifier is a linear layer to scale the dimension of the encoder outputs as the number of categories $|C|$, with a connected softmax function to predict probability. $|C|$ is respectively set as 2 and 4 for hypernasality detection and hypernasality assessment.
\subsection{Transfer Learning}
As mentioned above, with the help of large-scale speech data and the labeled text, the ASR encoder can be viewed as a powerful acoustic feature extractor. Consequently, we deem that the encoder with ASR training is better for hypernasality estimation due to its ability in learning high quality acoustic representation. To fulfill this target, we require the architecture of encoder used in the ASR task and hypernasality estimation to be completely identical, and same as the input acoustic features (Mel filter-bank). More specifically, we first train an encoder-decoder framework in an ASR objective, and then conduct hypernasality estimation task based on ASR encoder. Therefore, the objective function of our method is as follows: $$\mathcal{L}_{CLS} = - \sum_{(x, y)}^{(\mathcal{X}, \mathcal{Y})} \log P(y |x; \theta_{enc}, \tilde{\theta}_{cls}),$$ where $\theta_{enc}$ is the parameters of ASR encoder in Eqn.~\ref{eq1}. Figure~\ref{fig_asr} also presents a simple description about our training pipeline.
\section{Experiments}
\label{sec3:exp}
\begin{table}[!t]
\centering
\begin{tabular}{l|r|r}
\toprule
Rating & \# NMCPC & \# CNH \\
\midrule
Normal & 25 & 239 \\
Mild & 11 & 190 \\
Moderate & 14 & 509 \\
Severe & 16 & 108 \\
\bottomrule
\end{tabular}
\caption{Statistics of NMCPC and CNH cleft palate datasets. `` \# NMCPC'' and ``\# CNH" column represent the patient number of each hypernasality severity.}
\label{tab:my_label}
\end{table}
\subsection{CLP dataset}
\subsubsection{NMCPC-CLP}
NMCPC-CLP is a dataset collected by New Mexico Cleft Palate Center~\cite{Mohammad2020Single}, which is mainly composed of English speakers with cleft palate. This dataset includes 41 CLP patients and we sample 25 normal speakers as the control group. The average age of NMCPC-CLP is 9.2$\pm$3.3 years. Each patient is required to record a random subset of sentences from the candidate sentences. Based on the sampled audios, each patient will be assigned a score from 0 to 3 (0 stands normal and 3 means severe). For hypernasality detection, the patients with hypernasality are defined as who has a score between 1 and 3. Refer to Table~\ref{tab:my_label} for more details about the NMCPC-CLP dataset.
\subsubsection{CNH-CLP}
CNH-CLP is a dataset collected by a Chinese hospital, which includes cleft palate patients with a range from children to adults. All patients in CNH-CLP dataset are Chinese speakers. The procedure of audio collection for each patient is similar as NMCPC-CLP dataset. Table~\ref{tab:my_label} introduces the detailed information about the CNH dataset.
\subsection{ASR Datasets}
\subsubsection{Librispeech}
Librispeech~\cite{Vassil2015Librispeech} is a large-scale speech recognition dataset in English domain. Librispeech includes 960 hrs speech data, sampled at 16,000 Hz with the corresponding text for training. During the ASR training, we select dev-clean/dev-other as the development set. We use Librispeech to train an ASR model, and then apply it to NMCPC-CLP dataset since these two datasets are both in English domain.
\subsubsection{Aishell}
Aishell-1~\cite{Hui2017Aishell} is a widely used speech recognition dataset in Chinese domain. Aishell-1 dataset includes 150/10 hrs audios for the training/dev set. Each audio is sampled at 16,000 Hz, with the corresponding text. We use Aishell-1 for ASR training and then use it for CNH-CLP dataset as these two tasks belong to the same language.
\begin{table}[t]
\centering
\begin{tabular}{l l |rr}
\toprule
& & ASR & CLS \\
\midrule
\multirow{3}{*}{CNN} & stride & \multicolumn{2}{c}{2, 2} \\
& kernel width & \multicolumn{2}{c}{5, 5} \\
& channel & \multicolumn{2}{c}{1024} \\
\midrule
\multirow{5}{*}{Transformer} & layer & \multicolumn{2}{c}{12} \\
& hidden size & \multicolumn{2}{c}{512} \\
& filter size & \multicolumn{2}{c}{2048}\\
& dropout & \multicolumn{2}{c}{0.1} \\
& heads & \multicolumn{2}{c}{8} \\
\midrule
\multirow{3}{*}{Optimization} & batch & 256 & 32 \\
& learning rate & 2e-3 & 2e-4 \\
& Epoch & 100 & 30 \\
\bottomrule
\end{tabular}
\caption{Hyper-parameters of experimental setup. The ``ASR'' and ``CLS'' column means the setting of automatic speech recognition, and hypernasality estimation.}
\label{tab:hyper}
\end{table}
\subsection{Setup}
For audio inputs, we first resample audios at 16,000 Hz and then extract 80-channel log mel filter-bank features (25 ms window size and 10 ms shift) for all ASR datasets and CLP datasets. We choose Fairseq-S2T~\cite{wang-etal-2020-fairseq} as the development toolkit. For ASR task, the encoder is composed of two convolutional layers for sub-sampling and a stack of transformer layers~\cite{Vaswani2017transformer}, and the decoder is a stack of transformer layers with cross-attention modules. During the ASR training stage, we use SpecAugment~\cite{Daniel2020SpectAug} for data augmentation. We adopt Adam~\cite{Kingma2015Adam} as the default optimizer. The detailed hyperparameters are reported in Table~\ref{tab:hyper}. In addition, to reduce variance, we use five-fold cross-validation to measure our classification accuracy. All evaluations are tested at the speaker level.
\section{Results}
\label{sec:results}
\subsection{Hypernasality Estimation Accuracy}
We adopt the precision as the metric to evaluate the performance of our model in hypernasality assessment. The results are shown in Table~\ref{tab:results}. Our baseline is the base model without using ASR encoder. We also list some baselines~\cite{Mohammad2020Single}, which use statistical features for reference. From Table~\ref{tab:results}, we have the following observations: 1) when configured with ASR encoder, our method can outperform the baseline method by a large margin, including hypernasality detection and assessment, in both NMCPC and CNH cleft palate datasets; 2) compared with the statistical methods, our method also achieves significant improvement, especially in hypernasality detection. These improvements also demonstrate the effectiveness of our method. Besides, our method is capable of generality and can be applied to any neural networks used in previous works~\cite{Dubey2019Hypernasality,Vikram2021Attention}.
\begin{table}[h]
\centering
\begin{tabular}{l|c c| c c}
\toprule
& \multicolumn{2}{c|}{NMCPC} & \multicolumn{2}{c}{CNH} \\
Method & \multicolumn{1}{c}{HD} & \multicolumn{1}{c|}{HA} & \multicolumn{1}{c}{HD} & \multicolumn{1}{c}{HA} \\
\midrule
Baseline & {\scriptsize 90.1 $\pm$ 0.3} & {\scriptsize 77.8 $\pm$ 0.6} & {\scriptsize 93.1 $\pm$ 0.3} & {\scriptsize 73.1 $\pm$ 0.3} \\
+ ASR & {\scriptsize 93.4 $\pm$ 0.2} & {\scriptsize 83.6 $\pm$ 0.2} & {\scriptsize 96.5 $\pm$ 0.2} & {\scriptsize 79.8 $\pm$ 0.5} \\
\midrule
MFCC~\cite{Mohammad2020Single} & 84.07 & 62.64 & {\scriptsize 88.1 $\pm$ 0.2} & {\scriptsize 69.1 $\pm$ 0.1} \\
CQCC~\cite{Mohammad2020Single} & 84.07 & 70.05 & - & - \\
SFFB~\cite{Mohammad2020Single} & 89.00 & 82.10 & - & - \\
\bottomrule
\end{tabular}
\caption{Results of hypernasality estimation on NMCPC and CNH cleft palate datasets. The ``+ ASR'' row means using ASR encoder. ``HD'' and ``HA" represent hypernasality detection and hypernasality assessment respectively.}
\label{tab:results}
\vspace{-0.3cm}
\end{table}
\subsection{Analysis}
\begin{figure}[h]
\centering
\includegraphics[width=0.4\textwidth]{figs/cm.pdf}
\caption{Confusion matrix of our method for hypernasality classification on CNH dataset.}
\label{fig_cm}
\end{figure}
\subsubsection{Confusion Matrix}
To demonstrate the generalization of our model in predicting hypernasality, we also visualize the confusion matrix of our model by using ASR encoder in hypernasality classification, and the results are shown in Figure~\ref{fig_cm}. From Figure~\ref{fig_cm}, we observe that our model is significant in predicting each label, especially in normal, mild and moderate patients. Considering that CNH is an unbalanced dataset (moderate and severe patients occupy 50\% and 10\% respectively), the accuracy of our model in predicting severe cases is still acceptable. Overall, these results indicate that our model can obtain better generalization performance and avoid overfitting effectively.
\subsubsection{Visualization}
To better explain the advantages of our methods in identifying acoustic features, we also visualize the hidden unit of encoder output. More specifically, we assume the output of encoder as $h \in \mathbb{R}^{L \times D}$, where $L$ and $D$ represent the frame length and the hidden size, respectively. Therefore, we first calculate the average value of $h$, and then use its absolute value as the activated value of $i$-th frame (\emph{i.e.}, $\frac{1}{D} |\sum_{j=1}^D {h_{i,j}}|$). The result and its corresponding mel spectrum are shown in Figure~\ref{fig_attn}. We can find that the outputs of our model is more sensitive to acoustic features and demonstrate higher activated value, while the model without using ASR is too smooth to distinguish the semantic of acoustic features. This phenomenon also validates our hypothesis that ASR model can indeed help to extract high quality acoustic features.
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{figs/attn.jpg}
\caption{Comparisons between our method and baseline without using ASR in the activated value of encoder output.}
\vspace{-0.2cm}
\label{fig_attn}
\vspace{-0.2cm}
\end{figure}
\section{Conclusion}
In this paper, we introduce a simple and effective approach to improve hypernasality estimation from the perspective of using ASR model. More specifically, we note that existing neural network based methods only stack multiple neural network layers for classification, and cannot extract high quality representation for acoustic features, which is useful for hypernasality estimation. To address this deficiency, we propose to fine tune the encoder, which is pre-trained with the ASR objective, for hypernasality estimation. Such design allows our model to enjoy the benefits of ASR model from two aspects: 1) ASR corpus usually includes more audio data, which enables better generalization; 2) the labeling text of ASR corpus guides model to better extract acoustic features. Experimental results on two cleft palate datasets also demonstrate the effective of our methods in hypernasality assessment. In the future, we expect to focus on two research directions: 1) is it possible to utilize more powerful pre-trained speech models (like wav2vec~\cite{Steffen2019wav2vec}) to conduct hypernasality estimation in cleft palate speech; 2) traditional methods for hypernasality estimation usually designed some advanced statistical acoustic features, and thus we want to explore the potential of combining neural network based features and statistical features to better estimate hypernasality.
\label{sec:conclusion}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec_intro}
High-order finite element (FE) methods are becoming increasingly relevant in
computational science and engineering disciplines due to their potential for better simulation accuracy and
favorable scaling on modern architectures \cite{Fischer2002,fischer2020scalability,MFEM2020, CEED2021}.
A vital component of these methods is high-order computational meshes for discretizing the geometry.
Such meshes are essential for achieving optimal convergence rates on domains
with curved boundaries/interfaces, symmetry preservation, and alignment with
the key features of the flow in moving mesh simulations
\cite{Shephard2011, Dobrev2012, Boscheri2016}.
To fully benefit from high-order geometry representation, one must control the quality of a high-order mesh while avoiding geometrical approximation errors. However, constructing an optimal high-order mesh is not trivial and can often become a bottleneck
for curvilinear domains. This issue is further exacerbated in time-dependent
problems when remeshing might be needed.
The goal of the current work is to address this issue using
implicit meshing where we take easy-to-generate
meshes and align them to the target interface and/or boundary with
minimal user intervention. In particular, our focus is on the case when the target surface is prescribed implicitly by a smooth discrete function without an analytic parameterization.
Common examples of this scenario include use of level set functions to represent curvilinear domains as a combination of geometric primitives in
Constructive Solid Geometry (CSG) \cite{requicha1977constructive},
material interfaces in multimaterial configurations \cite{Osher1994},
and evolving parameterized geometries in shape and topology optimization applications \cite{sokolowski1992introduction,allaire2014shape,hojjat2014vertex}, amongst other applications.
Boundary conforming high-order meshes are typically generated by starting with a conforming linear mesh that is projected to a higher order space before the mesh faces are curved to fit the boundary \cite{xie2013generation,fortunato2016high,moxey2015isoparametric,gargallo2016distortion,toulorge2016optimizing,poya2016unified,ruiz2022automatic}.
Some notable approaches for directly generating high-order boundary fitted meshes are Rangarajan's method based on Universal meshes \cite{rangarajan2019algorithm} and the Vorocrust method for polyhedral meshes \cite{abdelkader2020vorocrust}. We also note Mittal's distance function-based approach for approximate tangential relaxation \cite{mittal2019mesh} and the
\textit{DistMesh} algorithm for meshing geometries specified by implicit functions \cite{persson2004simple}.
For generating interface fitted meshes, existing methods mainly rely on topological operations
where the input mesh is split across the interface to generate an
interface conforming mesh \cite{chen2017interface,mmgplatform}.
Some exceptions are Ohtake's method for adapting linear triangular surface meshes to align with domains with sharp features \cite{ohtake2002dual}
and Le Goff's method for aligning meshes to interfaces prescribed implicitly using volume fractions \cite{le2017volume}.
Barring \cite{persson2004simple,rangarajan2019algorithm,abdelkader2020vorocrust,ohtake2002dual,le2017volume},
existing methods mainly rely on an initial conforming meshing for boundary fitting and
on topological operations such as splitting for interface fitting.
In contrast to existing work, we propose a high-order boundary and interface fitting method that is predominantly algebraic and seldom requires topological operations,
extends to different element types (quadrilaterals/triangles/hexahedrons/tetrahedrons)
in 2D and 3D, and works for implicit parameterization of the target surface using discrete finite element (FE) functions.
We formulate the implicit meshing challenge as a mesh optimization
problem where the objective function is the sum of a
chosen mesh-quality metric using the Target-Matrix Optimization Paradigm (TMOP) \cite{TMOP2019SISC, Knupp2012} and a penalty
term that weakly forces selected faces of a mesh to align with
the target surface prescribed as the zero level set of a discrete
function. Additionally, we use an adaptive strategy to mark the faces for alignment and set the penalization weight, to ensure robustness of the method for nontrivial curvilinear boundaries/interfaces.
We also introduce the notion of a source mesh that can be used to accurately represent the level set and its gradient with sufficient accuracy.
This mesh is decoupled from the mesh being optimized, which allows to
represent the domain of intereset with higher level of detail.
This approach is crucial for cases where the target boundary is
beyond the domain of the mesh being optimized or the input
mesh does not have sufficient resolution around the zero level set.
The remainder of the paper is organized as follows.
In Section \ref{sec_prelim} we review the basic TMOP components and
our framework to represent and optimize high-order meshes.
The technical details of the proposed method for surface fitting and tangential
relaxation are described in Section \ref{sec_bif}.
Section \ref{sec_results} presents several academic tests that
demonstrate the main features of the methods,
followed by conclusions and direction for future work in Section \ref{sec_concl}.
\section{Preliminaries}
\label{sec_prelim}
In this section, we describe the key notation and our prior work
that is relevant for understanding our newly developed boundary-
and interface-fitting method.
\subsection{Discrete Mesh Representation}
\label{sec_mesh}
In our finite element based framework, the domain $\Omega \in \mathbb{R}^d$, $d=\{2,3\}$, is
discretized as a union of $N_E$ curved mesh elements, each of order $p$. To obtain a
discrete representation of these elements, we select a set of scalar basis
functions $\{ \bar{w}_i \}_{i=1}^{N_p}$, on the reference element $\bar{E}$.
For example, for tensor-based elements (quadrilaterals in 2D, hexahedra in 3D),
we have $N_p = (p+1)^d$, and the basis spans the space
of all polynomials of degree at most $p$ in each variable, denoted by $Q_p$.
These $p$th-order basis functions are typically chosen to be Lagrange
interpolation polynomials at the Gauss-Lobatto nodes of the reference element.
The position of an
element $E$ in the mesh $\mathcal{M}$ is fully described by a matrix
$\mathbf{x}_E$ of size $d \times N_p$ whose columns represent the coordinates
of the element {\em control points} (also known as {\em nodes} or element {\em degrees of freedom}).
Given $\mathbf{x}_E$ and the positions $\bar{x}$ of the
reference element $\bar{E}$, we introduce the map
$\Phi_E:\bar{E} \to \mathbb{R}^d$ whose image is the geometry of the physical element $E$:
\begin{equation}
\label{eq_x}
x(\bar{x}) =
\Phi_E(\bar{x}) \equiv
\sum_{i=1}^{N_p} \mathbf{x}_{E,i} \bar{w}_i(\bar{x}),
\sum_{i=1}^{N_p} \mathbf{x}_{E,i} \bar{w}_i(\bar{x}),
\qquad \bar{x} \in \bar{E}, ~~ x=x(\bar{x}) \in E,
\end{equation}
where $\mathbf{x}_{E,i}$ denotes the $i$-th column of $\bx_E$, i.e.,
the $i$-th node of element $E$.
To represent the whole mesh, the coordinates of the control points of every
element are arranged in a global vector $\bx$ of size $d \times N_x$ that
stores the coordinates of all node positions and ensures continuity at the
faces of all elements. Here $N_x$ is the global number of control points for
a principal direction.
Throughout the manuscript $x$ will denote the position function defined
by \eqref{eq_x}, while bold $\bx$ will denote the global vector of all
node coordinates.
\subsection{Geometric Optimization and Simulation-Based $r-$adaptivity\ with TMOP}
\label{sec_ra}
The input of TMOP is the user-specified transformation matrix $W$, from
reference-space $\bar{E}$ to target element $E_t$, which represents the
ideal geometric properties desired for every mesh point. Note that after
discretization, there will be multiple input transformation matrices $W$ -- one
for every quadrature point in every mesh element.
The construction of this transformation is guided by the fact that any Jacobian
matrix can be written as a composition of four components:
\begin{equation}
\label{eq_W}
W = \underbrace{\zeta}_{\text{[volume]}} \underbrace{R}_{\text{[rotation]}}
\underbrace{Q}_{\text{[skewness]}} \underbrace{D}_{\text{[aspect-ratio]}}.
\end{equation}
A detailed discussion on how TMOP's target construction methods encode geometric information into the target
matrix $W$ is given by Knupp in \cite{knupp2019target}, and various
examples of target construction for different mesh adaptivity goals are given in
\cite{TMOP2019SISC,TMOP2020CAF,TMOP2021EWC}.
Using \eqref{eq_x}, the Jacobian of the mapping $\Phi_E$ at any reference point
$\bar{\bx} \in \bar{E}$ from the reference-space coordinates to the current
physical-space coordinates is defined as
\begin{equation}
\label{eq_A}
A(\bar{x}) = \frac{\partial \Phi_E}{\partial \bar{x}} =
\sum_{i=1}^{N_w} \mathbf{\bx}_{E,i} [ \nabla \bar{w}_i(\bar{x}) ]^T \,.
\end{equation}
Combining \eqref{eq_A} and \eqref{eq_W}, the transformation from the
target coordinates to the current physical coordinates (see Fig. \ref{fig_tmop})
is defined as
\begin{equation}
\label{eq_T}
T = AW^{-1}.
\end{equation}
\begin{figure}[tb!]
\centerline{
\includegraphics[width=0.5\textwidth]{figures/tmop.jpeg}}
\caption{Schematic representation of the major TMOP matrices.}
\label{fig_tmop}
\end{figure}
With the target transformation $W$ defined in the domain, we next specify a
mesh quality metric $\mu(T)$ that compares the
transformations $A$ and $W$ in terms of the geometric parameters of interest
specified in \eqref{eq_W}. For example, $\mu_2=\frac{\mid T \mid^2}{2\tau}-1$
is a \emph{shape} metric\footnote{We follow the metric numbering in \cite{Knupp2020,knupp2022geometric}.} that depends on the skew and aspect-ratio components, but
is invariant to orientation/rotation and volume. Similarly,
$\mu_{77}=\frac{1}{2}(\tau-\frac{1}{\tau})^2$ is a
\emph{size} metric that depends only on the volume of the element.
We also have \emph{shape}$+$\emph{size} metrics such as $\mu_{80} =\gamma \mu_2 + (1-\gamma) \mu_{77}$
that depend on volume, skew and aspect-ratio, but are invariant to orientation/rotation.
Note that the mesh quality
metrics are defined such that they evaluate to 0 for an identity
transformation, i.e. $\mu(T) = 0$ when $T=I$ ($A=W$).
This allows us to pose the mesh optimization problem as minimization of $\mu(T)$,
amongst other advantages \cite{Knupp2020}.
The quality metric $\mu(T)$ is used to define the TMOP objective
function for $r-$adaptivity
\begin{equation}
\label{eq_F_full}
F(\bx) = \sum_{E \in \mathcal{M}} F_E(\bx_E)
= \sum_{E \in \mathcal{M}} \int_{E_t} \mu(T(x)) dx_t,
\end{equation}
where $F$ is a sum of the TMOP objective function for each element in the mesh ($F_E$),
and $E_t$ is the target element corresponding to the element $E$.
In \eqref{eq_F_full}, the integral is computed as
\begin{equation}
\label{eq_vm}
\sum_{E \in \mathcal{M}} \int_{E_t} \mu(T(x_t)) dx_t =
\sum_{E \in \mathcal{M}} \sum_{x_q \in E_t}
w_q\,\det(W(\bar{x}_q))\, \mu(T(x_q)),
\end{equation}
where $\mathcal{M}$ is the current mesh with $N_E$ elements, $w_q$ are the
quadrature weights, and the position $x_q$ is the image of the reference
quadrature point location $\bar{x}_q$ in the target element $E_t$.
Optimal node locations are determined by minimizing the TMOP objective function
\eqref{eq_F_full}.
This is performed by solving $\partial F(\bx) / \partial \bx = 0$
using the Newton's method where we
improve the mesh degrees-of-freedom (nodal positions) as
\begin{eqnarray}
\label{eq_r_adaptivity_solve}
\bx_{k+1} = \bx_k - \alpha \mathcal{H}^{-1}(\bx_k)\mathcal{J}(\bx_k).
\end{eqnarray}
Here, $\bx_k$ refers to the nodal positions at the $k$-th Newton iteration during
$r-$adaptivity, $\alpha$ is a scalar determined by a line-search procedure, and
$\mathcal{H}(\bx_k)$ and $\mathcal{J}(\bx_k)$ are the
Hessian ($\partial^2 F(\bx_k) / \partial \bx$)
and the gradient ($\partial F(\bx_k) / \partial \bx$),
respectively, associated with the TMOP objective function. The line-search
procedure requires that $\alpha$ is chosen such that $F(\bx_{k+1}) < 1.2 F(\bx_k)$,
$|\mathcal{J}(\bx_{k+1})| < 1.2 |\mathcal{J}(\bx_k)|$, and
that the determinant of the Jacobian of the transformation at each quadrature point
in the mesh is positive, $\det(A(\bx_{k+1})) > 0$.
These line-search constraints have been tuned using many numerical experiments
and have demonstrated to be effective in improving mesh quality.
For Newton's method, we solve the problem $\mathcal{H}^{-1}(\bx_k)\mathcal{J}(\bx_k)$
using a Krylov subspace method such as the Minimum Residual (MINRES) method
with Jacobi preconditioning;
more sophisticated preconditioning techniques can be found in \cite{Roca2022}. Additionally,
the optimization solver iterations \eqref{eq_r_adaptivity_solve} are
done until the relative $l_2$ norm of the gradient of the objective function
with respect to the current and original mesh nodal positions is below a certain
tolerance $\varepsilon$, i.e., $|\mathcal{J}(\bx)|/|\mathcal{J}(\bx_0)| \leq \varepsilon$.
We set $\varepsilon = 10^{-10}$ for the results presented in the current work.
Using the approach described in this section, we have demonstrated $r-$adaptivity\ with TMOP
for geometry and simulation-driven optimization; see Fig. \ref{fig_blade}
for example of high-order mesh optimization for a turbine blade using $W=I$ with
the shape metric $\mu_2(T)$.
\begin{figure}[t!]
\begin{center}
$\begin{array}{cc}
\includegraphics[height=0.3\textwidth]{figures/blade_init_mesh} &
\includegraphics[height=0.3\textwidth]{figures/blade_opt_mesh}
\vspace{-2mm} \\
\textrm{(a)} & \textrm{(b)} \\
\end{array}$
\end{center}
\vspace{-7mm}
\caption{(a) Original and (b) optimized 4th order meshes for a turbine blade.}
\label{fig_blade}
\end{figure}
\section{Boundary \& Interface Fitting}
\label{sec_bif}
Our goal for boundary and interface fitting is to enable alignment of a selected set of mesh nodes to some target surface of interest prescribed as the zero isocontour of a smooth discrete level set function, $\sigma(x)$.
Figure \ref{fig_square_circle}(a) and (b) show a simple example of a circular interface represented using a level set function and a triangular mesh with multi-material interface that is to be aligned to the circular interface.
To effect alignment with the zero isocontour of $\sigma(x)$, we modify
the TMOP objective function \eqref{eq_F_full} in the following way:
\begin{equation}
\label{eq_F_full_sigma}
F(\bx) = \underbrace{\sum_{E \in \mathcal{M}} \int_{E_t}
\mu(T(x)) dx_t}_{F_{\mu}} +
\underbrace{w_{\sigma} \sum_{s \in S} \sigma^2(x_s)}_{F_{\sigma}}.
\end{equation}
Here, $F_{\sigma}$ is a penalty-type term that depends on the penalization
weight $w_{\sigma}$, the set of nodes $\mathcal{S}$ marked for fitting (e.g., the mesh
nodes lying on the material interface in Fig. \ref{fig_square_circle}(b)),
and the level set function $\sigma(x)$, evaluated at the positions $x_s$
of all nodes $s \in \mathcal{S}$.
The $F_{\sigma}$ term penalizes the nonzero values of $\sigma(x_s)$ for all
$s \in \mathcal{S}$.
Minimizing this term represents weak enforcement of $\sigma(x_s) = 0$,
only for the nodes in $\mathcal{S}$, while ignoring the values
of $\sigma$ for the nodes outside $\mathcal{S}$.
Minimizing the full nonlinear objective function, $F = F_{\mu} + F_{\sigma}$,
would produce a balance between mesh quality and surface fitting.
Note that all nodes are treated together, i.e., the nonlinear solver would
make no explicit separation between volume nodes of $\mathcal{M}$ and the nodes $s \in \mathcal{S} \subseteq \mathcal{M}$ marked for fitting.
Furthermore, as there is no pre-determined unique target position for
each node of $\mathcal{S}$, the method naturally allows tangential
relaxation along the interface of interest, so that mesh quality
can be improved while maintaining a good fit to the surface.
Figure \ref{fig_square_circle}(c) shows an example of a triangular mesh
fit to a circular interface using \eqref{eq_F_full_sigma}.
In this example, we use the shape metric $\mu_2$ with equilateral targets
and a constant penalization weight, $w_{\sigma}=10^3$.
\begin{figure}[tb!]
\begin{center}
$\begin{array}{ccc}
\includegraphics[height=0.3\textwidth]{figures/circle_level_set} &
\includegraphics[height=0.3\textwidth]{figures/square_tri2_material} &
\includegraphics[height=0.3\textwidth]{figures/square_tri2_material_fit}
\vspace{-2mm}\\
\textrm{(a)} &
\textrm{(b)} &
\textrm{(c)} \\
\end{array}$
\end{center}
\vspace{-7mm}
\caption{(a) Level set function $\sigma(\bx)$, (b) a Cartesian mesh with material
interface nodes marked for fitting, and (c) mesh optimized to fit the zero level set
of $\sigma(\bx)$.}
\label{fig_square_circle}
\end{figure}
The first step in our method is to use a suitable strategy for representing the
level set function with sufficient accuracy (Section \ref{sec_level_set_rep}).
Next, we choose the set of nodes $\mathcal{S}$ that will be aligned to the zero level set of $\sigma$ such that we can get the best possible fit for the given mesh (Section \ref{sec_marking}).
Finally, we set the penalization weight $w_{\sigma}$\
such that an adequate fit is achieved to the target surface while optimizing the mesh with respect to the quality metric $\mu$ (Section \ref{sec_adaptive_sigma}). The adaptive strategy for setting $w_{\sigma}$\ requires us to modify the line-search and convergence criterion of the Newton's method (Section \ref{sec_converge}).
For completeness, the derivatives of $F_{\sigma}$ are discussed in Section \ref{sec_derivatives}. Using various examples, we demonstrate in Section \ref{sec_results} that our method extends to both simplices and hexahedrals/quadrilaterals of any order, and is robust in adapting a mesh interface and/or boundary to nontrivial curvilinear shapes.
\subsection{Level Set Representation}
\label{sec_level_set_rep}
The following discussion is related to the case when $\sigma(x)$ is a discrete
function so that its values and derivatives can't be computed analytically.
Then the first step in our implicit meshing framework is to ensure that
the level set function $\sigma(x)$ is defined with sufficient accuracy.
A drawback of discretizing $\sigma(x)$ on the mesh being optimized,
$\mathcal{M}$, is that the resulting mesh will be sub-optimal in terms of
mesh quality and interface/boundary fit if
(a) the mesh does not have enough resolution to represent $\sigma(x)$ and
its gradient with sufficient accuracy, especially near the zero level set, or
(b) the target level set is outside the initial domain of the mesh.
For the latter, it's impossible to compute the necessary values and
derivatives of the level set function accurately at the boundary nodes that
we wish to fit, once the mesh moves outside the initial domain;
see \autoref{fig_boundary}(a) for an example.
\begin{figure}[b!]
\begin{center}
$\begin{array}{cc}
\includegraphics[height=0.25\textwidth]{figures/boundary_1} &
\includegraphics[height=0.25\textwidth]{figures/boundary_2} \vspace{-2mm}\\
\textrm{(a)} &
\textrm{(b)} \\
\end{array}$
\end{center}
\vspace{-7mm}
\caption{Using a background mesh to fit discretely prescribed
domain boundaries. (a) Initial unfitted mesh and target boundary
curve (orange). (b) The boundary curve is prescribed implicitly as the zero level set of a discrete function on an adaptively refined nonconforming background mesh.}
\label{fig_boundary}
\end{figure}
To address these issues, we introduce the notion of a background/source
mesh ($\mathcal{M}_B$) for discretizing $\sigma(x_B)$, where $x_B$ represents
the positions of the source mesh nodes, as in \eqref{eq_x}.
Since $\mathcal{M}_B$ is independent of $\mathcal{M}$,
we can choose $\mathcal{M}_B$ based on the desired accuracy.
We achieve this by using adaptive nonconforming mesh refinements
\cite{cerveny2019nonconforming} around the zero level set of $\sigma(x_B)$;
see Figure \ref{fig_boundary}(b) for an example.
Using a source mesh for $\sigma(x_B)$, however, requires transfer of the
level set function and its derivatives from $\mathcal{M}_B$ to the
nodes $\mathcal{S} \in \mathcal{M}$ prior to each Newton iteration.
This transfer between the source mesh and
the current mesh is done using \emph{gslib}, a high-order interpolation
library \cite{gslibrepo}:
\begin{equation}
\label{eq_interpolation}
\sigma(x) = \mathcal{I}(x, x_B, \sigma(x_B)), ~~
\partial \sigma(x) = \mathcal{I}(x, x_B, \partial \sigma(x_B)), ~~
\partial^2 \sigma(x) = \mathcal{I}(x, x_B, \partial^2 \sigma(x_B)),
\end{equation}
where $\mathcal{I}$ represent the interpolation operator that depends on the
current mesh nodes ($x$), the source mesh nodes ($x_B$), and the source
function $\sigma(x_B)$ or its gradients.
A detailed description of how high-order functions
can be transferred from a mesh to an arbitrary set of points in physical space
using \emph{gslib} is described in Section 2.3 of \cite{mittal2019nonconforming}.
\begin{figure}[tb!]
\begin{center}
$\begin{array}{ccc}
\includegraphics[height=0.29\textwidth]{figures/reactor_design_level_set} \hspace{2mm} &
\includegraphics[height=0.29\textwidth]{figures/reactor_design_amr} \hspace{2mm} &
\includegraphics[height=0.3\textwidth]{figures/reactor_design_sigma} \hspace{2mm} \vspace{-2mm}\\
\textrm{(a)} &
\textrm{(b)} &
\textrm{(c)} \\
\end{array}$
\end{center}
\vspace{-7mm}
\caption{Multimaterial domain for a target application.
(a) Domain modeled using geometric primitives with $\mathcal{G}(\bx_B)= 1$ (orange) and $-1$ (blue), (b)
background mesh with adaptive mesh refinements around the zero level set of
$\mathcal{G}(\bx_B)$, and (c) distance function calculated on the background mesh
to be used as $\sigma(\bx)$ in \eqref{eq_F_full_sigma}.}
\label{fig_nrd_ls}
\end{figure}
While surfaces such as the circular interface in Fig. \ref{fig_square_circle} and sinusoidal boundary in Fig. \ref{fig_boundary} are straightforward to define using smooth
level-set functions, more intricate domains are typically defined
as a combination of nonsmooth functions for different geometric primitives, which are not well suited for
our penalization-based formulation \eqref{eq_F_full_sigma}.
Consider for example a two-material application problem in
Fig. \ref{fig_nrd_ls}(a) where the domain is prescribed as a combination of geometric primitives for
a circle, rectangle, parabola, and a trapezium.
The resulting step function $\mathcal{G}(x_B)$ is 1 inside one material and -1 inside the other material.
For such cases, we start with a coarse
background mesh and use adaptive mesh refinement
around the zero level set of $\mathcal{G}(x_B)$, see Fig \ref{fig_nrd_ls}(b).
Then we compute a discrete distance function \cite{belyaev2015variational}
from the zero level set of $\mathcal{G}(x_B)$, see Fig \ref{fig_nrd_ls}(c).
The advantage of using this distance function in \eqref{eq_F_full_sigma} is that
it is (i) generally smoother and (ii) maintains the location of the zero level set.
\subsection{Marking for Interface Fitting}
\label{sec_marking}
For a given level set function, the choice of $\mathcal{S}$ impacts the ability
of the penalization-based approach to adapt the mesh nodes to the target level set function.
A naive approach for marking is to partition the mesh into
two fictitious materials as:
\begin{equation}
\label{eq_eta}
\eta_E=
\begin{cases}
0,& \text{if } \int_E \sigma(x)\geq 0\\
1, & \text{otherwise}
\end{cases},
\end{equation}
where $\eta_E$ is the material indicator for element $E$, and then
mark the nodes at the material interface for fitting.
To demonstrate the limitations of this approach, consider the triangular mesh shown in Figure \ref{fig_square_circle_tri}(a)
that is marked for fitting to the circular level set shown in Fig. \ref{fig_square_circle}.
Marking using \eqref{eq_eta} leads to elements that have adjacent faces
marked for fitting (highlighted in red). Such a configuration can lead to vanishing
Jacobians at the vertex shared by adjacent faces, as the mesh deforms to align
to the level set of interest. Additionally, the quality of the elements surrounding the mesh
is sub-optimal as evident in Figure \ref{fig_square_circle_tri}(b).
\begin{figure}[bt!]
\begin{center}
$\begin{array}{cccc}
\includegraphics[height=0.25\textwidth]{figures/square_tri_material_bad} &
\includegraphics[height=0.25\textwidth]{figures/square_tri_material_fit_bad} &
\includegraphics[height=0.25\textwidth]{figures/square_tri_material_good} &
\includegraphics[height=0.25\textwidth]{figures/square_tri_material_fit_good} \vspace{-2mm}\\
\textrm{(a)} & \textrm{(b)} & \textrm{(c)} & \textrm{(d)} \\
\end{array}$
\end{center}
\vspace{-7mm}
\caption{(a) Triangular mesh marked using the marking strategy in
\eqref{eq_eta} and the (b) optimized mesh. (c) Mesh marked using
\eqref{eq_eta2} and the resulting (d) optimized mesh.}
\label{fig_square_circle_tri}
\end{figure}
The fundamental issue here is that whenever we have multiple adjacent faces of an element marked for fitting,
the fit could be sub-optimal depending on the complexity of the target shape/geometry.
To address this challenge we update the
material indicator for each element in the mesh as
\begin{equation}
\label{eq_eta2}
\tilde{\eta}_E=
\begin{cases}
\eta_E,& \text{if } N_{E,M} <= 1 \\
1-\eta_E,& \text{else if } N_{E,M} = N_{E,F}-1 \\
\eta_E, {\tt (optionally)\,mark\,E\,for\,splitting} & \text{otherwise}
\end{cases},
\end{equation}
where $N_{E,M}$ is the number of marked faces of element $E$,
and $N_{E,F}$ is the total number of faces of element $E$. Note that \eqref{eq_eta2}
is formulated as a two pass approach where we first loop over all elements with $\eta_E=0$
and then with $\eta_E=1$. This ensures that conflicting decisions are not
made for adjacent elements surrounding the interface.
For triangular meshes, the approach \eqref{eq_eta2} usually guarantees
that there is at most one face marked per element, without
the need to split/refine any element in the mesh.
Figures \ref{fig_square_circle_tri}(c, d) show an example of marking using \eqref{eq_eta2},
which avoids the issue of vanishing Jacobian at the vertex shared by
adjacent faces.
In quadrilateral elements, when $1 < N_{E,M} < N_{E,F}-1$, we can optionally
do conforming splits on each element to bisect the vertex connecting adjacent
faces that have been marked for fitting. The conforming split increases the computational cost associated with the mesh due to increased number of degrees-of-freedom,
but leads to better mesh quality. Figure \ref{fig_square_circle_quad}
shows an example of a comparison of a quadrilateral mesh fit to the circular level set
using the marking strategy \eqref{eq_eta} and \eqref{eq_eta2}.
\begin{figure}[tb!]
\begin{center}
$\begin{array}{cccc}
\includegraphics[height=0.25\textwidth]{figures/square_quad_material} &
\includegraphics[height=0.25\textwidth]{figures/square_quad_material_fit} &
\includegraphics[height=0.25\textwidth]{figures/square_quad_material_split} &
\includegraphics[height=0.25\textwidth]{figures/square_quad_material_split_fit} \vspace{-2mm} \\
\textrm{(a)} & \textrm{(b)} & \textrm{(c)} & \textrm{(d)} \\
\end{array}$
\end{center}
\vspace{-7mm}
\caption{(a) Quadrilateral mesh with material indicators using the marking strategy in \eqref{eq_eta} and the (b) optimized mesh. (c) Mesh marked using
\eqref{eq_eta2} with conforming split introduced for elements with 2 faces marked
for fitting, and (d) the resulting optimized mesh.}
\label{fig_square_circle_quad}
\end{figure}
For tetrahedrons and hexahedrons however, conforming splits independent of adjacent
elements are not possible, and we are currently exploring nonconforming refinement
strategies.
\subsection{Adaptive Penalization Weight}
\label{sec_adaptive_sigma}
Recall that the balance between mesh quality and node fitting is controlled
by the penalization weight $w_{\sigma}$\ in \eqref{eq_F_full_sigma}.
Numerical experiments show that use of a constant $w_{\sigma}$\ requires tweaking on
a case-by-case basis, and can result in a sub-optimal fit if $w_{\sigma}$\ is
too small, in which case the objective function is dominated by
the mesh quality metric term, or if $w_{\sigma}$\ is too large, in which case the conditioning of the Hessian matrix is poor.
Let the maximum surface fitting error be defined as
\[
|\sigma|_{\mathcal{S},\infty} := \max_{s \in {\mathcal{S}}} |\sigma(x_s)|.
\]
Figure \ref{fig_adapt_weight_sigma_circle} demonstrates
how $|\sigma|_{\mathcal{S},\infty}$
varies for the uniform quad mesh fit to a circular interface
(recall Figure \ref{fig_square_circle_quad}(a, b))
for different fixed values of $w_{\sigma}$.
As we increase $w_{\sigma}$\ from 1 to $10^{4}$, the fitting error decreases.
However, the error worsens if we increase $w_{\sigma}$ further.
\begin{figure}[bt!]
\centerline{
\includegraphics[height=0.5\textwidth]{figures/adaptiveweightcirclequad}
}
\caption{Impact of $w_{\sigma}$\ on the surface fitting error.}
\label{fig_adapt_weight_sigma_circle}
\end{figure}
To address this issue we use an adaptive approach for setting $w_{\sigma}$\, where
we monitor $|\sigma|_{\mathcal{S},\infty}$, and scale $w_{\sigma}$\
by a user-defined constant ($\alpha_{\sigma}=10$ by default) if the relative decrease in the
maximum nodal fitting error between subsequent Newton iterations is below a
prescribed threshold ($\epsilon_{\Delta \sigma}=0.001$ by default). That is,
\begin{equation}
\label{eq_w_adaptive}
w_{k+1,\sigma} =
\begin{cases}
\alpha_{\sigma}\cdot w_{k,\sigma} & \text{ if $\frac{|\sigma|_{k,\mathcal{S},\infty}-|\sigma|_{k+1,\mathcal{S},\infty}}{|\sigma|_{k+1,\mathcal{S},\infty}} < \epsilon_{\Delta \sigma}$} \\
w_{k,\sigma} & \text{ otherwise }
\end{cases},
\end{equation}
where we use the subscript $k$ to denote a quantile at the $k$th Newton iteration.
Figure \ref{fig_adapt_weight_sigma_circle}
shows that this adaptive approach for setting $w_{\sigma}$\ significantly
improves the quality of the mesh fit to the desired level set.
Updating the value of $w_{\sigma}$\ changes the definition of the objective function
\eqref{eq_F_full_sigma}, which requires some modifications of the line seach
and convergence criterion of the Newton iterations to
achieve overall convergence, compared to the constant $w_{\sigma}$\ case;
details will follow in Section \ref{sec_converge}.
Nevertheless, our numerical tests suggest that this impact is negligent
in comparison to the improvement of the fitting error.
\subsection{Convergence \& Line-Search Criterion}
\label{sec_converge}
Recall that in the general TMOP approach, the line-search and convergence
criteria for the Newton's method are based on the magnitude and the
derivatives of the objective function, see Section \ref{sec_ra}.
In the penalization-based formulation \eqref{eq_F_full_sigma},
the current criteria do not suffice because
the magnitude and derivatives of the objective function depend on
the penalization weight $w_{k,\sigma}$, which can change between
subsequent Newton iterations due to \eqref{eq_w_adaptive}.
We modify our line-search criteria by adding two additional inequalities,
namely, $\alpha$ in \eqref{eq_r_adaptivity_solve} is chosen to ensure:
\begin{equation}
\label{eq_lsearch_sigma}
|\sigma|_{k+1,\mathcal{S},\infty} < 1.2~|\sigma|_{k,\mathcal{S},\infty},
\end{equation}
\begin{equation}
\label{eq_lsearch_detA}
\det(A(\bx_{k+1})) > 0.001 \det(A(\bx_0)).
\end{equation}
The inequality \eqref{eq_lsearch_sigma} prevents sudden
jumps of the fitting error.
The constraint \eqref{eq_lsearch_detA} is mostly applicable in regimes when
$w_{\sigma}$\ is big enough to make the quality term $F_{\mu}$ effectively inactive.
Such regimes represent a situation when one is willing to sacrifice mesh
quality for more accurate fitting.
When the quality term $F_{\mu}$ is relatively small, it may be unable to
prevent the appearance of infinitesimally small positive Jacobians;
the constraint \eqref{eq_lsearch_detA} is used to alleviate this situation.
The convergence criterion is also modified to utilize the fitting error, i.e.,
the Newton's method is used until $|\sigma|_{\mathcal{S},\infty}$ is
below a certain user-specified threshold ($\epsilon_{\sigma}=10^{-5}$ by default). We also use an optional convergence criterion based on a limit on the maximum number of consecutive Newton iterations through which the penalization weight $w_{\sigma}$\ is adapted using \eqref{eq_w_adaptive}; this limit is $N_{\sigma}=10$ by default. This latter criterion avoids excessive computations in cases where the mesh topology does not allow the fitting error to reduce beyond a certain limit.
\subsection{Derivatives}
\label{sec_derivatives}
As our default choice for nonlinear optimization is the Newton's method, we
must compute first and second order derivatives of $F_{\sigma}$ with respect to the mesh nodes.
Let the FE position function be $x = (x_1 \dots x_d)^T$ where $d$ is the
space dimension; each component can be written as
$x_a(\bar{x}) = \sum_i x_{a,i} \bar{w}_i(\bar{x})$, where $x=x(\bar{x})$,
see Section \ref{sec_mesh}.
Then the Newton's method solves for the full vector
\[
\bx = (x_{1,1} \dots x_{1,N_x}, x_{2,1} \dots x_{2,N_x} \dots x_{d,N_x})^T
\]
that contains the positions of all mesh nodes.
The formulas for the first and second derivatives of $F_{\sigma}$
are the following:
\begin{equation}
\label{eq_derivatives}
\begin{split}
\frac{\partial F_{\sigma}(\bx)}{\partial x_{a,i}}
&= 2 \omega_{\sigma} \sum_{s \in \mathcal{S}}
\sigma(x_s) \frac{\partial \sigma(x_s)}{\partial x_{a}}
\frac{\partial x_a(\bar{x}_s)}{\partial x_{a,i}}
= 2 \omega_{\sigma} \sum_{s \in \mathcal{S}}
\sigma(x_s) \frac{\partial \sigma(x_s)}{\partial x_{a}}
\bar{w}_i(\bar{x}_s), \\
\frac{\partial^2 F_{\sigma}(\bx)}{\partial x_{b,j} \partial x_{a,i}}
&= 2 \omega_{\sigma} \sum_{s \in \mathcal{S}}
\left(
\frac{\partial \sigma(x_s)}{\partial x_{b}}
\frac{\partial \sigma(x_s)}{\partial x_{a}} +
2 \omega_{\sigma} \frac{\partial^2 \sigma(x_s)}{\partial x_{b} \partial x_{a}}
\right) \bar{w}_i(\bar{x}_s) \bar{w}_j(\bar{x}_s), \\
a,b &= 1 \dots d, \quad i,j = 1 \dots N_x.
\end{split}
\end{equation}
The above formulas require the spatial gradients of $\sigma$ at the current
positions $\{x_s\}_{s \in \mathcal S}$ of the marked nodes.
These gradients can be closed-form expressions, when $\sigma$ is prescribed
analytically, or the gradients are obtained from the background mesh
$\mathcal{M}_B$ (see Section \ref{sec_level_set_rep}),
when $\sigma$ is a discrete function.
\subsection{Algorithm/Summary}
\label{sec_algo}
In this section, we summarize our penalization-based method for boundary and
interface fitting with TMOP.
The inputs to our method are the active/current mesh $\mathcal{M}$ that is
defined through the global vector $\bx$ of nodal positions,
a user-selected target construction option to form $W$ as in \eqref{eq_W},
a mesh quality metric $\mu$,
a source/background mesh with nodal coordinates $\bx_B$ along with the level set
function $\sigma(x_B)$ as explained in Section \ref{sec_level_set_rep},
initial penalization weight $w_{\sigma}$,
parameters for adaptive penalization-weight ($\alpha_{\sigma}$ and
$\epsilon_{\Delta \sigma}$ as in Section \ref{sec_adaptive_sigma}),
and parameters for convergence criterion $\epsilon_{\sigma}$ and
$N_{\sigma}$ as in Section \ref{sec_converge}.
Algorithm \ref{algo_fit} summarizes our penalization-based method where we
use subscript $k=0\dots N_{opt}$ to denote different quantiles at the
$k$th Newton iteration.
\begin{algorithm}[h!]
\label{algo_fit}
\SetAlgoLined
\KwIn{$\bx$, $\mu$, $\bx_B$, $\sigma(x_B)$, $\alpha_{\sigma}$, $\epsilon_{\Delta \sigma}$, $w_{\sigma}$, $\epsilon_{\sigma}$, $N_{opt}$}
\KwOut{$\bx_s$}
$n_{\sigma}:=0$,\,\,$k=0$\,\,,$\bx_{0}=\bx$\\
Determine $\mathcal{S}$, the set of nodes for fitting (Section \ref{sec_marking})\\
$\sigma(\bx_0) = \mathcal{I}(\bx_0, \bx_B, \sigma(\bx_B))$\\
\While{$|\sigma|_{k,\mathcal{S},\infty}>\epsilon_{\sigma}$ \&\& $n_{\sigma} < N_{\sigma}$ \&\& $k < N_{opt}$}{
$W_i = I$ for each quadrature point $i$ ~\cite{knupp2019target}.\\
${\Delta \bx} = \mathcal{H}^{-1}(\bx_k)\mathcal{J}(\bx_k)$ $\rightarrow$ solve using MINRES (Section \ref{sec_ra}). \\
$\bx_{k+1} = \bx_{k} - \alpha {\Delta \bx}$, with $\alpha$ determined using line-search (Section \ref{sec_ra}). \\
\eIf{$\frac{|\sigma|_{k,\mathcal{S},\infty}-|\sigma|_{k+1,\mathcal{S},\infty}}{|\sigma|_{k+1,\mathcal{S},\infty}} < \epsilon_{\Delta \sigma}$} {
$w_{k+1,\sigma} \rightarrow \alpha_{\sigma}\cdot w_{k,\sigma}$ \\
$n_{\sigma}=0$
}
{
$n_{\sigma} = n_{\sigma}+1$
}
$\sigma(\bx_{k+1}) = \mathcal{I}(\bx_{k+1}, \bx_B, \sigma(\bx_B))$\\
$k = k+1$\\
}
$\bx = \bx_{k}$ \label{algo_end}
\caption{Implicit Meshing}
\end{algorithm}
\section{Numerical Results}
\label{sec_results}
In this section, we demonstrate the main properties of the method
using several examples. The presented tests use $W=I$ as the target matrix, and the following shape metrics:
\begin{equation}
\begin{split}
\mu_2 &= \frac{|T|^2}{2\tau} - 1, \\
\mu_{303} &= \frac{|T|^2}{3\tau^{2/3}} - 1,
\end{split}
\end{equation}
where $|T|$ and $\tau$ are the Frobenius norm and determinant of $T$, respectively.
Both metrics are polyconvex in the sense of \cite{Garanzha2010, Garanzha2014},
i.e., the metric integral $F_{\mu}$ in \eqref{eq_F_full} theoretically has a minimizer.
Exploring the effect of the smoothness of sigma on the convexity of the objective function \eqref{eq_F_full_sigma}
will be the subject of future studies.
Our implementation utilizes MFEM, a finite element library that supports arbitrarily high-order spaces and meshes \cite{MFEM2020}. This implementation is freely available at \url{https://mfem.org}.
\subsection{Fitting to a Spherical Interface}
As a proof of concept, we adapt 3rd order hexahedral (hex) and tetrahedral (tet) meshes to align to a spherical surface; see Figure \ref{fig_3D_sphere_mesh}(a) and (b). The domain is a unit-sized cube, $\Omega \in [0, 1]^3$, and the level set function $\sigma$ is defined such that its zero isosurface is located at a distance of 0.3 from the center of the domain ($x_c = (0.5, 0.5, 0.5)$).
Although this level set is simple enough to be defined analytically,
the presented computations
represent and use $\sigma$ as a discrete finite element function.
Additionally, no background mesh is used in this example, i.e., $\bx_B = \bx$.
\begin{figure}[b!]
\begin{center}
$\begin{array}{cccc}
\includegraphics[height=0.23\textwidth]{figures/3D-hex-sphere} &
\includegraphics[height=0.23\textwidth]{figures/3D-tet-sphere} &
\includegraphics[height=0.22\textwidth]{figures/3D-hex-sphere-opt} &
\includegraphics[height=0.22\textwidth]{figures/3D-tet-sphere-opt} \vspace{-2mm}\\
\textrm{(a)} & \textrm{(b)} & \textrm{(c)} & \textrm{(d)} \\
\end{array}$
\end{center}
\vspace{-7mm}
\caption{Initial (a) hexahedral and (b) tetrahedral meshes showing the interfaces for the 3D surface fitting tests. Optimized (c) hexahedral and (d) tetrahedral mesh obtained using \eqref{eq_F_full_sigma}.}
\label{fig_3D_sphere_mesh}
\end{figure}
The initial hex mesh is an $8\times8\times8$ Cartesian-aligned mesh.
The material indicators are setup such that there are a total of 32 elements that have more than one face marked for fitting.
The tet mesh is obtained by taking a $4\times4\time4$ Cartesian-aligned hex mesh and splitting each hex into 24 tetrahedrons (4 tets-per-hex face). The material indicators are setup such that all faces marked for fitting belong to different elements. The optimized meshes, shown in Fig. \ref{fig_3D_sphere_mesh}(c) and (d), have a maximum error of $O(10^{-10})$ at the interface with respect to the zero level set, which is achieved in 76 and 44 Newton iterations for the hexahedral and tetrahedral mesh, respectively. The hexahedral mesh requires more iterations as compared to tetrahedral mesh partly because there are multiple elements with adjacent faces marked for fitting, which requires more work from the adaptive weight mechanism to force those mesh faces to align with the spherical interface while ensuring positive Jacobians everywhere in the mesh.
\subsection{Boundary Fitting for Geometric Primitive-Based Domain in 3D}
This next example demonstrates implicit
high-order meshing using boundary fitting for a nontrivial curvilinear geometry specified as a combination of simple geometric primitives.
Here, the target domain is represented as an intersection of a cube
(\emph{side}=0.5, centered around $x_c = (0.5, 0.5, 0.5)$) with a
sphere (\emph{radius}=0.3, centered around $x_c$) that has three Cartesian-aligned cylinders (\emph{radius}=0.15, \emph{length}=0.5) removed from it. Figure \ref{fig_3D_geom_prim} shows the CSG tree
with geometric primitives and the boolean operations that are used to construct the target geometry.
\begin{figure}[b!]
\begin{center}
$\begin{array}{c}
\includegraphics[height=0.35\textwidth]{figures/geometric_primitive_3d.jpeg}
\end{array}$
\end{center}
\vspace{-7mm}
\caption{CSG tree with geometric primitives used to define the target geometry.}
\label{fig_3D_geom_prim}
\end{figure}
Using the approach outlined in Section \ref{sec_level_set_rep}, we
combine the geometric primitives on a third-order
source mesh that is adaptively refined five times around the zero level set of the target domain. We then compute the distance function on this background mesh using the $p-$Laplacian solver \cite{belyaev2015variational}. This distance function is used as the level set function in \eqref{eq_F_full_sigma}. Figure \ref{fig_3D_geom_prim_bg}(a) shows a slice-view of the background mesh along with the zero iso-surface of the level set function, and Fig. \ref{fig_3D_geom_prim_bg}(b) shows a slice-view of the level set function computed as the distance from the zero level set of the geometric primitive-based description of the domain. Our numerical experiments showed that using a third order source mesh with five adaptive refinements was computationally cheaper than using a lower order mesh with more adaptive refinements to obtain the same level of accuracy for capturing the the curvilinear boundary using the distance function.
\begin{figure}[tb!]
\begin{center}
$\begin{array}{cc}
\includegraphics[height=0.40\textwidth]{figures/cgs_3d_bg_mesh} &
\includegraphics[height=0.40\textwidth]{figures/cgs_3d_bg_levelset} \vspace{-2mm}\\
\textrm{(a)} & \textrm{(b)} \\
\end{array}$
\end{center}
\vspace{-7mm}
\caption{(a) Adaptively refined background mesh used to model the target domain, and (b)
level set function computed using the distance from the zero level set of the
geometric primitive-based geometry.}
\label{fig_3D_geom_prim_bg}
\end{figure}
\begin{figure}[tb!]
\begin{center}
$\begin{array}{cc}
\includegraphics[height=0.40\textwidth]{figures/cgs_3d_origmesh} &
\includegraphics[height=0.40\textwidth]{figures/cgs_3d_optmesh}\vspace{-2mm} \\
\textrm{(a)} & \textrm{(b)} \\
\end{array}$
\end{center}
\vspace{-7mm}
\caption{(a) Initial trimmed mesh and (b) final fitted mesh for the primitive-based geometry test.}
\label{fig_3D_geom_prim_opt}
\end{figure}
The input mesh to be fit for this problem is a uniform second-order $128 \times 128 \times 128$ Cartesian-aligned mesh for $\Omega \in [0, 1]^3$. This mesh is trimmed based on the location of the zero level set and is then optimized using \eqref{eq_F_full_sigma}. Figure \ref{fig_3D_geom_prim_opt} shows the input and the optimized mesh, where the achieved fitting error is
$|\sigma|_{\mathcal{S},\infty} = \mathcal{O}(10^{-6})$.
\subsection{Interface Fitting for Shape Optimization Application}
This example serves to demonstrate the applicability of the proposed approach to setup the initial multimaterial domain to be used in a shape optimization problem.
Figure \ref{fig_reactor_domain} shows the cross section of a tubular reactor that consists of a highly conductive metal (\emph{red}) and a low conductivity heat generating region (\emph{blue}). The design optimization problem is formulated such that the energy production in the system is maximized while keeping the overall volume of the red region fixed.
To achieve this, the shape of the \emph{red} subdomain is morphed using a gradient-based approach in our in-house design optimization framework that requires an interface fitted mesh as an input.
This initial fitted mesh is generated using the method described in this manuscript.
\begin{figure}[tb!]
\begin{center}
$\begin{array}{c}
\includegraphics[width=0.6\textwidth]{figures/reactor_design_setup}
\end{array}$
\end{center}
\vspace{-7mm}
\caption{Reactor domain to be meshed for shape optimization, along with the
symmetric portion and its primitive decomposition.
The internal material interfaces must be fitted in the final mesh.}
\label{fig_reactor_domain}
\end{figure}
Exploiting the cyclic symmetry, we discretize a portion of the domain via a triangular mesh. Since this initial mesh does not need to align with the material interface, it can be conveniently generated using an automatic mesh generator \cite{blacker2016cubit}. We then use an approach similar to the previous section for interface fitting where the multimaterial domain is realized as a combination of geometric primitives (annulus, parabola, and trapezium,
as highlighted in Figure \ref{fig_reactor_domain}).
The finite element distance function from the interface is computed, and
used as the level-set function $\sigma(x)$ in \eqref{eq_F_full_sigma}.
Figure \ref{fig_reactor_meshes}(a) shows the input non-fitted mesh, colored by
the adaptively assigned a material indicator, such that there is at most one
face per triangle marked for fitting.
Figure \ref{fig_reactor_meshes}(b) shows the adapted mesh which
aligns with the material interface of the domain. The maximum interface fitting
error of the optimized mesh is $\mathcal{O}(10^{-7})$.
Finally, Figures \ref{fig_reactor_heat}(a)-(d) show the initial
interface-fitted domain and the shape optimized domain along with
the corresponding temperature fields.
\begin{figure}[b!]
\begin{center}
$\begin{array}{cc}
\includegraphics[width=0.35\textwidth]{figures/reactor_design_mesh_orig} &
\includegraphics[width=0.35\textwidth]{figures/reactor_design_mesh_opt} \vspace{-2mm} \\
\textrm{(a)} & \textrm{(b)} \\
\end{array}$
\end{center}
\vspace{-7mm}
\caption{Original mesh (a) and interface fitted mesh (b)
for the reactor design problem.}
\label{fig_reactor_meshes}
\end{figure}
\begin{figure}[b!]
\begin{center}
$\begin{array}{ccc}
\includegraphics[width=0.3\textwidth]{figures/initialDesign_attribute} &
\includegraphics[width=0.3\textwidth]{figures/optimalDesign_attribute} &
\\
\text{(a) Initial fitted domain} &
\text{(c) Shape optimized domain} \\
\includegraphics[width=0.3\textwidth]{figures/initialDesign_temperature} &
\includegraphics[width=0.3\textwidth]{figures/optimalDesign_temperature} & \hspace{-10mm}
\includegraphics[width=0.12\textwidth,height=0.2\textwidth]{figures/reactorlegend} \\
\text{(b) Temperature in the initial domain} &
\text{(d) Temperature in the shape optimized domain}
\end{array}$
\end{center}
\vspace{-7mm}
\caption{Interface fitted domain and shape optimized domain along
with corresponding temperature fields from our in-house design optimization framework.}
\label{fig_reactor_heat}
\end{figure}
\section{Conclusion \& Future Work}
\label{sec_concl}
We have presented a novel method for implicit high-order meshing which can
align easy-to-generate meshes with the interfaces and boundaries of the
domain of interest.
We formulate the mesh optimization problem as a variational minimization of the sum of a chosen mesh-quality metric and a penalty term that weakly forces the selected faces of the mesh to align with the target surface.
The key technical components of the proposed method are the use of a source
mesh to represent the level set function with sufficient accuracy,
adaptively setting the penalization weight, and careful marking of the faces of
the mesh to be fit to the target isocontour of the level set field.
We have demonstrated that the method is effective for generating boundary-
and interface-fitted meshes for nontrivial curvilinear geometries.
In future work, we will improve the method by developing mesh refinement strategies for hexahedral and tetrahedral meshes, which are required when the mesh topology limits how well the mesh can fit to the target surface (Section \ref{sec_marking}).
We will also explore ways for aligning meshes to domains
with sharp features \cite{ohtake2002dual,zahr2020implicit}, as we
currently assume that the level set function $\sigma$ used in
\eqref{eq_F_full_sigma} is sufficiently smooth around its zero level set.
Finally, we will also look to optimize our method and leverage
accelerator-based architectures by utilizing partial assembly and
matrix-free finite element calculations \cite{camier2022accelerating}.
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
\end{document}
\section{Real-world Intervention}
\label{sec:design}
We conducted our reminder email intervention in a CS1 course lasting 13 weeks. Students were randomly assigned to a subject line condition each week. The messages were sent on the same day and time every week (i.e. Wednesday evenings). There were initially $1119$ students enrolled in the course, with a final count of $1025$ by the end of the term.
These emails are aimed at the goal of building time management skills for students, we sent weekly reminder emails with three different versions of subject lines. The email content nudged students to start their weekly homework early. Our end goal is to find better strategies to motivate students to interact with the email content, which we cannot do if they do not open these emails. Hence, we wanted to determine whether one subject line would captivate students more, thereby motivating them to engage with our prompts.
As it is difficult to fully observe student engagement to asynchronous emails, we chose to collect data on the open rates of the messages as our response measurement of engagement in this intervention. To assign subject lines, we used variations of adaptive multi-armed bandit algorithms to give one of three different versions of the email subject line each week. Data from the previous week is then used to further allocate email subject lines for the coming week, as described in Section 3. This allows us to investigate which might be more effective in leading students to open the reminder email and pivot towards sending this subject line more under the assumption its impact is identical from week to week.
To design the subject lines, we used various psychological and marketing theories.
\begin{adjustwidth}{-1em}{}
\begin{enumerate}
\item[] \textbf{Subject Line 1 (SL1)}: \textit{Hey \{name\}, when are you going to make a plan for your {{course code}} homework assignments?} This subject line was designed using implementation intentions to help students transform the goal of planning when to start their homework into an action \cite{gollwitzer1999implementation}. Using the student’s first-name shows email personalization and can catch their attention easily \cite{sahni_wheeler_chintagunta_2018}.
\item[] \textbf{Subject Line 2 (SL2)}: \textit{[{\{course code\}} Action Required] Hey \{name\}, start planning your week now!} \includegraphics[scale=0.8]{sections/alarm_clock.PNG}: This subject line adds a sense of urgency to the email, highlighting the importance of planning their homework, and has an emoji at the end to make it stand out in their inbox \cite{valenzuela2022boost}. This subject line also includes the student’s first name \cite{sahni_wheeler_chintagunta_2018}.
\item[] \textbf{Subject Line 3 (SL3)}: \textit{Reminder: These are your {{course code}} tasks for next week.} This subject line only mentions the content of the email and does not include any personalization or action prompt i.e. does not use any of the theories described above.
\end{enumerate}
\end{adjustwidth}
\section{Analysis and Results}
\label{sec:results}
Figure \ref{fig:experiment} represents the behaviour of each of the arms in the experiment per allocation algorithm. Each plot represents a different algorithm from UR, TS, $\text{TS}^\dagger$. Note that since all algorithms draw from the same burn-in period, the first 4 weeks are duplicated across plots. For TS and $\text{TS}^\dagger$, an additional week of duplication is shared due to initialization of priors from $\text{TS}^\dagger$.
The left vertical axis represents the cumulative reward mean per week, i.e. the proportion of students who opened the email each week (with 95\% confidence interval in shades), whereas the right vertical axis shows the assignment proportion to every arm week-wise. Furthermore, to understand the output of the experiment in absolute values, Table \ref{tab:empirical_data} displays the final cumulative empirical mean, and the final number of students allocated to each arm throughout the experiment.
\begin{figure*} \label{fig:experiment}
\includegraphics[width=\textwidth]{sections/cumulative2.png}
\caption{Cumulative sample arm means (hard lines) and weekly allocation proportions (dashed lines) for each policy (blue: UR, red: TS, green: $\text{TS}^\dagger$). Error bar gradients represent 95\% confidence intervals.}
\end{figure*}
\begin{table*}
\centering
\begin{tabular}{c|c|c|c|c|c|c|c|c|c}
\toprule
& \multicolumn{3}{c|}{UR}
& \multicolumn{3}{c|}{TS}
& \multicolumn{3}{c}{$\text{TS}^\dagger$}\\
Arm
& \multicolumn{1}{c|}{Mean}
& \multicolumn{1}{c|}{SD}
& \multicolumn{1}{c|}{Observations}
& \multicolumn{1}{c|}{Mean}
& \multicolumn{1}{c|}{SD}
& \multicolumn{1}{c|}{Observations}
& \multicolumn{1}{c|}{Mean}
& \multicolumn{1}{c|}{SD}
& \multicolumn{1}{c}{Observations}\\
\midrule
1 & 60.61 & 0.87 & 3130 & 60.07 & 0.86 & 3217 & 61.21 & 0.81 & 3618\\
2 & 57.96 & 0.89 & 3094 & 60.60 & 1.07 & 2086 & 61.99 & 1.13 & 1852\\
3 & 58.52 & 0.89 & 3036 & 60.36 & 1.15 & 1825 & 61.44 & 1.20 & 1653\\
\bottomrule
\end{tabular}
\caption{Empirical cumulative average reward (email open rate), standard deviation, and total allocation per arm, per policy, at the end of the real-world deployment.}
\label{tab:empirical_data}
\end{table*}
Post experiment, we checked to what degree the average reward (mean open rate) differed across arms under each policy to determine whether we assigned to optimal conditions. We used Wald \textit{z}-tests to compare if the arms' means have a statistically significant difference, taking into account their respective standard errors \cite{wald1943tests}. Since we have three arms and three policies, by fixing each policy we are able to compare pairs of arms under each. The results can be found in Table \ref{tab:wald_test_allocations}. To avoid making type I errors in this preliminary work, we controlled for a family-wise false positive rate of $0.05$ using a conservative Bonferroni Correction for every set of three—relevant—hypothesis test.
Table \ref{tab:wald_test_allocations} also shows the \textit{p}-value for such tests, where we can see that all of them are greater than the adjusted-threshold of $0.017$, failing to reject the null hypothesis of all real means being equal. This result indicates within this adaptive policy allocation, it is unlikely that the arm means were different across all weeks.
\begin{table*}
\centering
\begin{tabular}{c|c|c|c|c|c|c}
\toprule
& \multicolumn{2}{c|}{UR}
& \multicolumn{2}{c|}{TS}
& \multicolumn{2}{c}{$\text{TS}^\dagger$}\\
& \textit{p}-value
& Wald statistic
& \textit{p}-value
& Wald statistic
& \textit{p}-value
& \multicolumn{1}{c}{Wald statistic}\\
\midrule
Arm 1 vs. Arm 2 & 0.033 & 2.129 & 0.701 & 0.384 & 0.573 & 0.564\\
Arm 2 vs. Arm 3 & 0.653 & 0.449 & 0.880 & 0.151 & 0.738 & 0.335\\
Arm 3 vs. Arm 1 & 0.095 & 1.668 & 0.840 & 0.202 & 0.872 & 0.161\\
\bottomrule
\end{tabular}
\caption{\textit{P}-values and statistics from the Wald \textit{z}-test by comparing the empirical cumulative arms' rewards within policies. Bonferroni adjusted threshold $p$=0.017.}
\label{tab:wald_test_allocations}
\end{table*}
Finally, we conducted exploratory data analysis on the correlation between email open rate and their desired effect on student behaviour: earlier start times and higher homework attempt rate. We have not found any significance on their homework start times as of yet, but there is preliminary evidence that students who opened the emails have a higher attempt rate than those who did not.
\section{Methods: Traditional and Adaptive Experimentation}
\label{sec:methods}
To address the exploration versus exploitation problem in MAB settings, we use a set of three algorithms which all share data from a uniform burn-in period of 4 weeks. The uniform algorithm allocation is continued at half of the allocations through all the weeks post burn-in to ensure we are collecting enough data for all arms and to maintain enough samples from all arms. Pure TS—which selects an arm based on its prior Beta distribution with parameters $\alpha$ and $\beta$, representing measures for \textit{successes} and \textit{failures} respectively—accounts for another quarter of the allocations and is used to evaluate the effectiveness and problems of adaptive algorithms.
We introduce a new algorithm $\text{TS}^\dagger$ for the last quarter of allocations. This uses the same stochastic allocation as TS but draws half-weighted priors from both itself and the uniform allocation to explicitly maintain exploration. We hypothesize that this third allocation algorithm can balance the rapid choice of optimal arms and better adapt in the face of potentially non-stationary arm means.
Given that the adaptive portion of our experiment relies on the Thompson Sampling policy, we explicitly show how the priors $\left(\alpha_{k, t}^{\pi},\ \beta_{k, t}^{\pi}\right)$ change over time, both in equations and with Algorithm \ref{alg:method}. Note in the equations that the priors are updated in a weekly batch. Hence the TS algorithm was chosen for its stochastic nature to avoid assigning all students to the same condition each week. This weekly update ensures time is not an uncontrollable factor affecting students' willingness to open emails.
\subsection{Equations}
Let $k \in \{1,2,3\}$ represent the arm a particular student was assigned at any time $t \in \{1, 2, ..., 13\}$ in weeks. $r_{k,t}^\pi$ represents the binomial distributed cumulative reward from all the students assigned to arm $k$ for policy $\pi$ at time $t$, and $n_{k,t}^\pi$ is the total number of students assigned to arm $k$ for policy $\pi$ at time $t$.
\begin{gather*}
\left(\alpha_{k, t}^{\text{UR}},\ \beta_{k, t}^{\text{UR}}\right)=\left(1+\sum_{\tau=1}^t\left[r_{k, \tau}^{\text{UR}}\right], 1+\sum_{\tau=1}^t \left[n_{k,\tau}^{\text{UR}}-r_{k, \tau}^{\text{UR}}\right]\right) \\ \text{for } t \leq 13\\
\\
\left(\alpha_{k, t}^{\text{TS}},\ \beta_{k, t}^{\text{TS}}\right) = \\
\begin{cases}
\left(\alpha_{k, t}^{\text{UR}},\ \beta_{k, t}^{\text{UR}}\right) & \text{for } t \leq 5\\
\left(\alpha_{k, t-1}^{\text{TS}},\ \beta_{k, t-1}^{\text{TS}}\right)\\ \qquad + \left(r_{k, t-1}^{\text{TS}},\ n_{k, t-1}^{\text{TS}} - r_{k, t-1}^{\text{TS}}\right) & \text{for } 6 \leq t \leq 13\\
\end{cases}\\
\\
\left(\alpha_{k, t}^{\text{TS}^\dagger},\ \beta_{k, t}^{\text{TS}^\dagger}\right) = \\
\begin{cases}
\left(\alpha_{k, t}^{\text{UR}},\ \beta_{k, t}^{\text{UR}}\right) & \text{for } t \leq 5\\
\left(\alpha_{k, t-1}^{\text{TS}^\dagger},\ \beta_{k, t-1}^{\text{TS}^\dagger}\right)\\ \qquad + \frac{1}{2}\left(r_{k, t-1}^{\text{TS}},\ n_{k, t-1}^{\text{TS}} - r_{k, t-1}^{\text{TS}}\right)\\ \qquad + \frac{1}{2}\left(r_{k, t-1}^{\text{UR}},\ n_{k, t-1}^{\text{UR}} - r_{k, t-1}^{\text{UR}}\right)& \text{for } t = 6\\
\left(\alpha_{k, t-1}^{\text{TS}^\dagger},\ \beta_{k, t-1}^{\text{TS}^\dagger}\right)\\ \qquad + \frac{1}{2}\left(r_{k, t-1}^{\text{TS}^\dagger},\ n_{k, t-1}^{\text{TS}^\dagger} - r_{k, t-1}^{\text{TS}^\dagger}\right)\\ \qquad + \frac{1}{2}\left(r_{k, t-1}^{\text{UR}},\ n_{k, t-1}^{\text{UR}} - r_{k, t-1}^{\text{UR}}\right) & \text{for } 7 \leq t \leq 13\\
\end{cases}
\end{gather*}
\RestyleAlgo{ruled}
\SetKwComment{Comment}{/* }{ */}
\begin{algorithm}[hbt!]
\caption{Update rules per policy}\label{alg:method}
$\pi \in \{\text{UR}, \text{TS}, \text{TS}^\dagger\}$\\
$k \in \{1,2,3\}$\\
$\left(\alpha_{k, 0}^{\pi},\ \beta_{k, 0}^{\pi}\right) \gets \left(1,\ 1\right), \forall \pi, k$ \Comment*[r]{Initialize parameters}
\For{$t = 1, 2, ..., 5$}{
$n_{k,t}^\text{UR}$ \Comment*[r]{Allocate with UR}
$r_{k,t}^\text{UR}$ \Comment*[r]{Get total reward}
$\left(\alpha_{k, t}^{\pi}, \beta_{k, t}^{\pi}\right) \gets \left(\alpha_{k, t-1}{^{\pi}},\ \beta_{k, t-1}^\pi\right) + \left(r_{k,t}^\text{UR},\ n_{k,t}^\text{UR} - r_{k,t}^\text{UR}\right)$ \Comment*[r]{Update priors}
}
\For{$t = 6$}{
\For{$\pi' \in \{\text{UR}, \text{TS}\}$}{
$n_{k,t}^{\pi'}$ \Comment*[r]{Allocate with policy $\pi'$}
$r_{k,t}^{\pi'}$ \Comment*[r]{Get total reward per $\pi'$}
$\left(\alpha_{k, t}^{\pi'},\ \beta_{k, t}^{\pi'}\right) \gets \left(\alpha_{k, t-1}^{\pi'},\ \beta_{k, t-1}^{\pi'}\right) + \left(r_{k,t}^{\pi'},\ n_{k,t}^{\pi'} - r_{k,t}^{\pi'}\right)$ \Comment*[r]{Update priors}
}
$\left(\alpha_{k, t}^{\text{TS}^\dagger},\ \beta_{k, t}^{\text{TS}^\dagger}\right) \gets \left(\alpha_{k, t-1}^{\text{TS}^\dagger},\ \beta_{k, t-1}^{\text{TS}^\dagger}\right) + \frac{1}{2}\left(r_{k,t}^{\text{TS}},\ n_{k,t}^{\text{TS}} - r_{k,t}^{\text{TS}}\right) + \frac{1}{2}\left(r_{k,t}^{\text{UR}},\ n_{k,t}^{\text{UR}} - r_{k,t}^{\text{UR}}\right)$ \Comment*[r]{Update $\text{TS}^\dagger$ prior}
}
\For{$t = 7, 8, ..., 13$}{
$n_{k,t}^{\pi}$ \Comment*[r]{Allocate with policy $\pi$}
$r_{k,t}^{\pi}$ \Comment*[r]{Get total reward per policy $\pi$}
\eIf{$\pi = \text{TS}^\dagger$}{
$\left(\alpha_{k, t}^{\text{TS}^\dagger},\ \beta_{k, t}^{\text{TS}^\dagger}\right) \gets \left(\alpha_{k, t-1}^{\text{TS}^\dagger},\ \beta_{k, t-1}^{\text{TS}^\dagger}\right) + \frac{1}{2}\left(r_{k,t}^{\text{TS}^\dagger},\ n_{k,t}^{\text{TS}^\dagger} - r_{k,t}^{\text{TS}^\dagger}\right) + \frac{1}{2}\left(r_{k,t}^{\text{UR}},\ n_{k,t}^{\text{UR}} - r_{k,t}^{\text{UR}}\right)$ \Comment*[r]{Update $\text{TS}^\dagger$ prior}
}{
$\left(\alpha_{k, t}^{\pi},\ \beta_{k, t}^{\pi}\right) \gets \left(\alpha_{k, t-1}^{\pi},\ \beta_{k, t-1}^{\pi}\right) + \left(r_{k,t}^{\pi},\ n_{k,t}^{\pi} - r_{k,t}^{\pi}\right)$ \Comment*[r]{Update priors}
}
}
\end{algorithm}
\section{Introduction}
Traditional A/B randomized comparisons are widely used in different research and industry areas to determine if a particular condition is more effective. Similarly, adaptive experimentation is used to assign participants to the most effective current condition while keeping the ability to test the other conditions \cite{williams2018enhancing}. Using adaptive experimentation in education can help explore various conditions but also direct more students to more useful ones in a randomized experiment \cite{lomas2016interface, williams2016axis, williams2018enhancing}.
In this setting, we focus on the adaptive experimentation's problem regarding the exploration-exploitation trade-off. It is well studied that some adaptive policies might be exploiting an arm when no significant difference exists in that direction \cite{nogas2021algorithms, williams2016axis}. This might cause a lack of exploration of all conditions where the policy starts allocating more users to an arm that was favoured from high values of its random rewards, even if the true mean is higher. We study the impact of adding a fixed allocation of uniform exploration to the TS algorithm—which is employed to solve the MABs problem and has been used in A/B experiments to adaptively assign participants to conditions \cite{williams2018enhancing}. TS is a Bayesian algorithm that has shown promising results in maximizing the reward \cite{agrawal2012analysis, chapelle2011empirical}. It maintains a posterior distribution for each action based on how rewarding that action is expected to be. Actions are chosen proportional to their posterior probability, and the outcome of the chosen action is then used to update each posterior distribution. TS also performs well in cases where batches of actions are chosen, and the rewards and outcomes are updated only after each batch \cite{kalkanli2021asymptotic}.
This work aims to build onto the study of academic interventions targeted to enhance study habits and time management of college students, building upon \cite{zavaleta2021using} by studying the variation from injecting UR data to TS priors each week. Previous work focused on some methods that improve student's study habits and attitudes \cite{aquino2011study},
and using A/B comparisons in email interventions to students \cite{zavaleta2022can}.
We conducted a real-world adaptive experiment in an online synchronous CS1 course at a large publicly funded research-focused North American University. The motivation for this intervention is to show students that simply knowing how to program is only one part of the puzzle; building consistent study habits and depositions such as time management are also crucial to being a good programmer. These skills aim to impact longer-term success, in addition to capturing early improvement from keeping up with course content.
We explore the difference between implementing a smaller traditional experiment (i.e., allocating students to conditions uniformly at random) in addition to an adaptive experiment (i.e., TS) mixed on the same cohort. This approach is motivated by creating stability in the adaptive component and increasing the probability of assigning more students to a better condition. Even more, it could make it easier to draw conclusions from the experiment since we can look at the traditional part of the experiment or the combined one. In total, we deployed three policy learning algorithms. The first is simple uniform random (UR) sampling, which is assigned to half the cohort each week. The second one is a standard TS algorithm that was introduced after a burn-in period of four weeks. Lastly, the third group is introduced a week after the TS algorithm and uses a combination of uniformly sampled and TS sampled evidence to update its beliefs each week, which we refer to as $\text{TS}^\dagger$. This allows us to evaluate adaptive methods (TS) and uniform random side by side and evaluate how a combined version of these algorithms behaves. A more thorough explanation can be found in \S\ref{sec:design} and \S\ref{sec:methods}.
\section{Conclusion}
We present and evaluate a deployment of email reminders in a first-year computer science course which leverages Multi-Armed Bandits adaptive algorithms. The goal was to increase student engagement with our messages by manipulating subject lines relative to a uniform allocation. On a high level, these reminders are aimed at guiding computer science students to build the time management skills necessary for success in the field. Our work highlights the results from using adaptive algorithms in a real-world setting, and explains problems we encounter such as assigning higher allocation probability to an arm when there is no significant different among the arms' means. We recognize there could be outside factors that can affect the reward every iteration. Additional work should be done to address these and the evidence of non-stationarity in more depth.
\section{Acknowledgements}
This work was partially supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) (No. RGPIN-2019-06968), as well as by the Office of Naval Research (ONR) (No. N00014-21-1-2576)
\section{Discussion}
This work outlines a deployment framework for using adaptable multi-armed bandits for randomized testing. The goal was to maintain statistical power while giving students more optimal arm assignments. In our particular setting, multiple factors could have affected students' email engagement, such as midterm tests, coursework deadlines, etc. Moreover, student patterns may change across the semester, motivating us to explore ways of using an adaptive algorithm that has existing information about the probability of each arm being optimal, while getting updated current probabilities from the uniform data. We then formed the $\text{TS}^\dagger$ algorithm that learns from both the traditional TS and UR algorithms.
From our results, even though Table \ref{tab:wald_test_allocations} shows no statistically significant difference between arm means under all policies, there is an empirical mean increase of $4.6\%$ and $3.6\%$ from arm 1 compared to arms 2 and 3. This empirical difference causes TS and $\text{TS}^\dagger$ to extremely favour assigning to one arm, which only begins to change after multiple weeks. Their respective plots in Figure 1 illustrate this behaviour. These two points combined highlight the problem that MAB algorithms encounter: choosing an arm even when neither is more optimal than the other based only on average rewards. In other words, in the situation arm means are not significantly different, uniform random would be preferable as it provides more balanced data without erroneously determining an optimal arm to exploit.
Interestingly, Figure 1 shows that $\text{TS}^\dagger$ was able to hold on to the higher allocation of arm 1 for much longer than TS. However, this observation is counterintuitive, since our goal was to maintain more uniformity of allocations when conditions become equal. This further demonstrates that $\text{TS}^\dagger$ still suffers from the lack of this uniform guarantee, similar to regular TS, and instead seeks to allocate an arm rapidly, even without certainty of its optimality. Furthermore, one point of tension for adaptive experimentation is to know when to switch or adapt to a different allocation rather than hold on to a previously highly allocated condition which may no longer be optimal. In this environment, it seems that the TS algorithm can ``adapt" better than $\text{TS}^\dagger$, at the risk of greater uncertainty of allocation optimality.
Furthermore, sample estimates in Figure 1 hint at a non-stationary behaviour across weeks, meaning the optimal arm could change each week. This is a characteristic that can cause unexpected behaviour in adaptive algorithms to perform as expected, as bandit allocations such as TS assume constant arm means.
Our results highlight a lesson for researchers interested in deploying adaptive interventions in education: one must be certain of their student behaviour first. In the case that conditions are not significantly different, these algorithms lose statistical power compared to uniform random (i.e. traditional A/B testing) by favouring one arm over others. This can hinder instructors' ability to make beneficial choices to students based on experimental results.
In our case, we were not assigning students to a significantly poorer condition and the experiment had a simple manipulation. However, this was only able to be determined later in the experiment, not during the early deployments. In other settings with a higher risk for harm, for instance, most students could be assigned to an initially optimal arm—based on empirical reward—that turns out to be the least optimal arm in the long run. In a field such as computer science where topics build on top of previous lessons, it is important we only compound positive effects for students in our interventions.
\section{Limitations and Future Work}
As discussed, the convergence for adaptive methods is still too rapid, even for non-significant differences in a dynamic environment. This could indicate that traditional uniform algorithms could be useful in controlling the number of students assigned to an optimal condition in the face of changing optimal arms each week. However, different analyses should be made to test for other possibilities. For example, it could be that the subject lines were too general to cause a behavior change, or even too effective in doing so. It could also be related to having the email itself be effective, regardless of the subject line.
Furthermore, we recognize that unobserved factors in our study that could influence students' decision to open or not open an email, beyond the subject line. For instance, the engagement of a student with a course could vary throughout a semester. Also, this study was made on a single CS1 course. Thus, it may not be representative across domains.
Finally, our study encounters the problem of personalization, as the updates are made in a weekly batch to ensure time does not affect the decision of students' to open an email. However, this comes at the cost of the individualization of adaptive experimentation and does not allow for a finer-grained feedback structure beyond the weekly subject line.
Future work will focus on different approaches to addressing the problem of potentially non-stationary arm means. Even more, there is room to focus on adaptive frameworks that consider contextual factors to better predict student behaviour during, and not after, an experiment. Lastly, an additional investigation could design a personalized variant of bandits for email reminders that do not depend on a large batch size or reward collected from peers.
|
\section{Introduction}
The \textit{Legendre transformation} of a curve in two dimensions is a curve in another space. The second curve is created by the coefficients of the original curve’s tangent and supporting lines. For the curve $y=f(x)$, the lines are written as
$$y = mx - g(m)$$
Properly interpreted, the curve $y = f(x)$ in $x,y$-space has the same information as the curve $d = g(m)$ in $m,d$-space. The two spaces are said to be \textit{dual} to each other, and the curves
$$y = f(x)\text{ and }d = g(m)$$
are \textit{dual curves}, which form a unique pair. The transformation between the two curves is called the \textit{Legendre transformation} and is written
$$\L\{f(x)\}(m) = g(m).$$
The Legendre transformation, which relates curves in a pair of two-dimensional spaces, was introduced by Adrien-Marie Legendre (1752–1833). It is widely employed in classical and statistical mechanics and thermodynamics \cite{Arnold1978,doi:10.1119/1.3119512}, and in econometrics \cite{Blackorby1978,RePEc:eee:jetheo:v:1:y:1969:i:4:p:374-396}. In mathematics, it arises in Young’s inequality \cite{hardy_1967,Dragoslov1970}, in the Clairaut differential equation \cite{Ince1944,Sternberg1954}, and in polar
reciprocals of geometry \cite{Bix2006,Glaeser2016}.
A potential limitation on the use of this transformation may be the lack of a ready source of pairs of functions that are related by the transformation. In order to remedy that, we present an extensive table of such pairs. Because the transformation is involutive or reflexive, the table is even more expansive than it might initially appear, that is, it can be read right-to-left and left-to-right. For curves that are not amenable to inclusion in the table, parametric representations and approximating curves are available and discussed here. These suggest that the Legendre transformation could have even more widespread applicability.
\section{Finding the dual curve}
Given a differentiable function $y = f(x)$, the dual curve can be expressed with $x$ as the parameter. The equation of the tangent line at $x = a$ is
$$y = f'(a)x - (af'(a) - f(a)).$$
Replacing the parameter $a$ with $x$ gives the parametric form of the dual curve in $m,d$-space in terms of the parameter $x$:
\begin{align}
m &= f'(x), \label{mdef}\\
d &= xf'(x) - f(x).\label{ddef}
\end{align}
Substituting numerical values for $x$ into \eqref{mdef} and \eqref{ddef} gives corresponding points of the dual curve.
If \eqref{mdef} can be solved for $x$, so that
\begin{equation}
x = (f')^{-1}(m),\label{solveforx}
\end{equation}
then \eqref{solveforx} can be substituted into \eqref{ddef} yielding an explicit form $d = g(m)$ for the dual curve, which is Method 1 in Section 3.
\begin{example}[an example $y = f(x)$ for which \eqref{mdef} cannot be analytically solved for $x$]
Consider the composite function $y = f(x) = \sin(x^{2})$ for $x\in\bbR$. Then, $y' = f'(x) = m = 2x\cos(x^{2})$ for $m\in\bbR$ and $d = mx - y = 2x^{2}\cos(x^{2}) - \sin(x^{2})$. Thus, a parametric representation of the dual curve is
\begin{align*}
m &= 2x\cos(x^{2})\\
d &= 2x^{2}\cos(x^{2}) - \sin(x^{2}).
\end{align*}
\end{example}
Opting for a parametric portrayal can be very useful for sums of functions, as well.
\begin{example}[sum of functions]
Consider $y = f(x) = \sin(x^{2}) - x^{3} + e^{x}$ for $x\in\bbR$, then $y'=f'(x) = m = 2x\cos(x^{2}) - 3x^{2} + e^{x}$ for $m\in\bbR$ and $d=mx-y = 2x^{2}\cos(x^{2}) - \sin(x^{2}) - 2x^{3} + (x-1)e^{x}$.
\end{example}
In certain situations, which are described in Methods 1–4 in Section 3, explicit formulations of the dual curve are available and yield the pairs of functions, such as those in \ref{table}, which is in \ref{app}.
\section{Methodologies for finding and verifying the Legendre-transformation pairs in \ref{table}}
The following theorem contains three facts that are helpful. Theorem \ref{thmleg}\ref{thmslope} says that $x$ plays the role of slope in the dual space. Theorem \ref{thmleg}\ref{thmmaint} says that convexity is maintained by the transformation \cite{doi:10.1119/1.3119512,Rockafellar70convexanalysis}. Theorem \ref{thmleg}\ref{thmrefl} says that we can read \ref{table} in \ref{app} left-to-right and right-to-left \cite{Rockafellar70convexanalysis,Lay2007}.
\begin{theorem}\label{thmleg}
If $f(x)$ is a differentiable function and $\L\{f(x)\}(m) = g(m)$, then
\begin{enumerate}[label=\textup{(\alph*)}]
\item\label{thmslope} $x$ is slope in the dual curve: the independent variables $x$ and $m$ are slopes of the corresponding dual curves. In particular where both curves possess first derivatives and nonzero second derivatives,
$$m = f'(x)\text{ and }x = g'(m).$$
\item \label{thmmaint} Maintenance of strict convexity and strict concavity: corresponding portions of $f(x)$ and $g(m)$ have the same concavity, in particular, where both curves possess second derivatives,
$$g''(m) = 1/f''(x).$$
\item\label{thmrefl} Reflective or involutive property of the transformation:
$$\L\{\L\{f(x)\}(m)\}(a) = f(a).$$
\end{enumerate}
\end{theorem}
\begin{proof}
For \ref{thmslope}, by definition $m = y'(x) = f'(x)$. Let $\L\{f(x)\}(m) = g(m)$ be given by the parametric equations
\begin{align*}
m &= f'(a)\\
d &= g(m(a)) = af'(a) - f(a).
\end{align*}
Thus, $\frac{dm}{da}=f''(a)$. Then,
$$\frac{dg}{dm} = \frac{dg}{da}\frac{da}{dm} = (f'(a) + af''(a) - f'(a))\frac{1}{f''(a)} = a$$
for any value of $a$.
For \ref{thmmaint}, apply $\frac{d}{da}$ to $g'(m) = a$ to obtain $g''(m) \frac{dm}{da} = g''(m) f''(a) = 1$.
For \ref{thmrefl}, in $m,d$-space, for any given value of $a$, the equation of the corresponding tangent line is
$$d = a(m - f'(x)) + af'(a) - f(a) = am - f(a).$$
Therefore, $\L\{g(m)\}(a) = f(a)$ as claimed.
\end{proof}
Theorem \ref{thmleg}\ref{thmslope} provides the locations of points where first derivatives are zero, if any such points exist, in terms of values of the dual function. For $y = f(x)$, from $m = f'(x) = 0, x = g'(m) = g'(0)$ and $y = f(x(m)) = f(x(0)) = 0x - g(0)$. Any zeros of the derivative occur at $(x,y) = (g'(0),-g(0))$. Similarly, any zeros of the derivative of $d = g(m)$ occur at $(m,d) = (f'(0),-f(0))$.
\ref{table}, which is in \ref{app}, in contains pairs of curves in the two dual spaces that are related by the Legendre transformation. There are four main ways for determining the entries of the table. We illustrate each method with an example. For a given curve, differing methods might be applied to various portions of the curve. Usually, more than one method can be successfully employed.
From the equation of a tangent or supporting line $y = mx - g(m)$, the Legendre transformation of $y = f(x)$ is
\begin{equation}
\L\{f(x)\}(m) = g(m) = mx(m) = f(x(m)).\label{lif}
\end{equation}
\begin{method}[using lines of tangency and the inverse of the derivative of $y = f(x)$]\label{methinv}
For differentiable functions $f(x)$, whose derivatives have an inverse, from Theorem \ref{thmleg}\ref{thmslope},
\begin{equation}
x(m) = (f')^{-1}(m).\label{inv}
\end{equation}
Substituting \eqref{inv} into \eqref{lif} gives
\begin{eqnarray}
d = g(m) = \L\{f(x)\}(m) = mx(m) - f(x(m)) = m(f')^{-1}(m) - f((f')^{-1}(m)).\label{mff}
\end{eqnarray}
\end{method}
\begin{example}[using lines of tangency in Method \ref{methinv}]
Consider the function
\begin{equation}
y = f(x) = \frac{x^{3}}{3}\label{tangex}
\end{equation}
for $x\in\bbR^{+}$. Placing $x(m) = (f')^{-1}(m) = m^{1/2}$ into \eqref{mff} gives the dual curve
\begin{equation}
d = g(m) = \L\{f(x)\}(m) = m(m^{1/2}) - \frac{(m^{1/2})^{3}}{3} = \frac{2m^{3/2}}{3}\label{Lx33}
\end{equation}
for $m\in\bbR^{+}$. The convexity of \eqref{tangex} and \eqref{Lx33} illustrates Theorem \ref{thmleg}\ref{thmmaint}. Equations \eqref{tangex} and \eqref{Lx33} are Entry \ref{xpp} of \ref{table} in \ref{app} with $p=3$ and $q=3/2$.
\end{example}
\begin{method}[using lines of support]\label{methsupp}
For lines of support, the negation of the lines' $y$-intercept for each value of $m$ is the value of $g(m)$.
\end{method}
\begin{example}
Consider the convex function
\begin{equation}
y = f(x) = \begin{cases}
-x + 2 & x\leq 1,\\
2x - 1 & 1 < x.
\end{cases}\label{supportex}
\end{equation}
The transformation of the line segment $y = -x + 2$ for $x < 1$ is the point $(m,d) = (-1,-2)$. The transformation of the line segment $y = 2x - 1$ for $1 < x$ is the point $(m,d) = (2,1)$. At $x = 1$, the lines of support of \eqref{supportex} are $y = mx + (1 - m)$ for $-1 < m < 2$, so that the transformation of the point $(x,y) = (1,1)$ is the open line segment $d = m - 1$ with $-1 < m < 2$, whose end points are $(-1,-2)$ and $(2,1)$. Thus,
$$d = g(m) = m - 1\text{ for }-1\leq m\leq 2.$$
This example illustrates how line segments and points that are vertices transform into each other.
\end{example}
\begin{method}[using the integral form]\label{methint}
The integral form of the Legendre transformation $y = f(x)$, whose derivative has an inverse, is
\begin{equation}
g(m) = \L\{f(x)\}(m) = \int (f')^{-1}(m)dm + C,\label{gint}
\end{equation}
which can be confirmed by showing the equality of the derivatives of \eqref{mff} and \eqref{gint} with respect to
$m$. From Theorem \ref{thmleg}\ref{thmslope}, $m = f'(x) = f'(g'(m))$, and, thus, $(f')^{-1}(m) = g'(m)$. The constant of
integration can be determined from the slope and negation of the $y$-intercept of the tangent line at a point of $y = f(x)$. Selecting the point where $m = m_{0}$, \eqref{gint} can be expressed
\begin{equation}
g(m) = \int_{m_{0}}^{m}(f')^{-1}(t)dt + g(m_{0}).\label{defint}
\end{equation}
\end{method}
\begin{example}[using the integral form in Method \ref{methint}]
For $y = f(x) = \cos x$ with $-\pi/2 < x < \pi/2$, we have $f'(x) = -\sin x = m$ and $(f')^{-1}(m) = -\Sin m$ for $-1 < m < 1$. The integral of the inverse sine function is
$$\int(-\Sin m)dm = -m\Sin m - \sqrt{1 - m^{2}} + C.$$
From \eqref{defint}, employing the point $(0, 1)$ of $y = f(x)$, where $m_{0} = 0$ and the negation of the $y$-intercept of the tangent line is $g(m_{0}) = g(0) = -1$, obtain
\begin{align*}
g(m) = \L\{\cos x\}(m) &= \int_{0}^{m} (-\Sin t)dt + (-1)\\
&= \left(-m\Sin m - \sqrt{1 - m^{2}}\right) - \left(-0\Sin 0 - \sqrt{1 - 0^{2}}\right) - 1\\
&= -m\Sin m - \sqrt{1 - m^{2}}
\end{align*}
for $-1<m<1$, which is Entry \ref{cos} of \ref{table} in \ref{app}.
\end{example}
\begin{method}[using a limit of the negation of $y$-intercepts]\label{methsup}
For a convex function $y = f(x)$, or a portion of the function that is convex, the lines of support or tangency for each value of $m$ give
\begin{equation}
d = g(m) = \L\{f(x)\}(m) = \sup_{x}\{mx - f(x)\},\label{sup}
\end{equation}
where $\sup$ denotes supremum, which is also called the least upper bound. Recall that all convex functions are continuous \cite{Lay2007,Tiel1984ConvexAA,Bauschke2017}. This method determines the line with the largest $y$-intercept among all lines with slope $m$ and below the convex function. For the strictly concave function $y = f(x)$, write
$$d = g(m) = \L\{f(x)\}(m) = \inf_{x}\{mx - f(x)\},$$
where $\inf$ denotes infimum, which is also called the greatest lower bound.
\end{method}
\begin{example}[using the interpretation as the limit in Method 4]
To derive Entry \ref{ex} of \ref{table} in \ref{app}, where $y = f(x) = e^{x}$ is convex for $x\in\bbR$ and has positive slope, select any positive value for $m$ and set
$$\frac{d}{dx}(mx - e^{x}) = m - e^{x} = 0.$$
Then, $x = \ln m$. Substituting this into \eqref{sup} gives
$$d = g(m) = \L\{f(x)\}(m) = m(\ln m) - e^{\ln m} = m\ln m - m$$
for $m > 0$.
\end{example}
Method \ref{methsup} is effective for deriving properties in Part \ref{props} of the table. For example, for Entry \ref{fpa} of the table, for convex functions $y$,
$$\L\{y(x) + a\}(m) = \sup_{x}\{mx - (y(x) - a)\} = \sup_{x}\{mx - y(x)\} - a = \L\{y(x)\}(m) - a.$$
With most modern plotting software, such as Desmos, GeoGebra, and MATLAB, one can use the following test in order to rapidly verify that functions $f$ and $g$ are Legendre-transformation pairs on stated domains. From \eqref{ddef} and \eqref{mdef}, the functions are transformations of each other on the $x$-domain where
$$y = xf'(x) - f(x) - g(f'(x))$$
is the zero function. The $m$-domain for the pair is the range of
$$y = f'(x)$$
over the $x$-domain that was just confirmed. Many other options for similar authentications are available. These checks of the accuracy of the entries of the table do not replace this section's methods for creating entries.
\section{The table of Legendre-transformation pairs}
Because the purpose of \ref{table} in \ref{app} is to serve as a quick and easy source of
Legendre-transformation pairs, there is some redundancy. For example, Entry \ref{combo} is an amalgam of Entries \ref{scaleout}-\ref{shiftin}. Entry \ref{ex} is $\L\{e^{x}\}(m) = m\ln m - m$ and Entry \ref{xln}, reading right-to-left, is
$\L\{e^{x-1}\} = m\ln m$; Entry \ref{xln} can be obtained from Entry \ref{ex} using Entry \ref{shiftin} with $f(x) = e^{x}$
and $a = -1$. Some entries are special cases of others. For example, Entry \ref{ex} is Entry \ref{ax} with $a=e$.
The properties in Part \ref{props} are aids for expanding the applicability of the remainder of the
table to new functions that are related to table entries by the properties. For example, to obtain
the transformation of $\alpha x^{3} + \beta x^{2}$ for any nonzero $\alpha$ and $\beta$, consider Entry \ref{x3x2}, which gives the
transformation of $x^{3}/3 + x^{2}/2$. Use Entry \ref{scalein} with $a = 3\alpha/(2\beta)$ to obtain the transformation of $(3\alpha x/(2\beta))^{3}/3 + (3\alpha x/(2\beta))^{2}/2$, then apply Entry \ref{scaleout} with $a = (2\beta)^{3}/(3\alpha)^{2}$ to find
$$\frac{(2\beta)^{3}}{(3\alpha)^{2}}\left(\frac{1}{3}\left(\frac{3\alpha x}{2\beta}\right)^{3} + \frac{1}{2}\left(\frac{3\alpha x}{2\beta}\right)^{2}\right) = \alpha x^{3} + \beta x^{2}.$$
Part \ref{props} supplies relationships between operations on variables and functions in one space
and operations in the dual space. For example, Entry \ref{scaleout} shows that multiplying a function by $a >
0$ in one space corresponds to \textit{epi-multiplication} by $a$ in the other space; epi-multiplication by $a$
is dividing the independent variable by a and simultaneously multiplying the function by $a$. In
applications, the variables and functions can have physical, economic, or other meanings and the
properties describe how algebraic and other operations in one space influence variables and
functions, and hence their meanings, in the other space. Entries \ref{scaleout}-\ref{combo} reflect how changes of
coordinates in one space alter the transformed function.
The infimal convolution or epi-sum in Entries \ref{sumfs} and \ref{px2} is discussed and used in \cite{Bauschke2017}
and \cite{Tiel1984ConvexAA}. For differentiable functions, the infimum is the minimum that can be found by
differentiation with respect to $t$, so that implementation of those two entries can be relatively
straightforward.
Parts \ref{alg}, \ref{exp}, and \ref{trig} contain algebraic, logarithmic and exponential, and trigonometric functions,
respectively. The special functions in Part \ref{spe} are extensively used in many areas of physics and
other disciplines \cite{bell2004special,lebedev1972special,50831}.
The \textit{Lambert $W$ function}, which is the solution $w = w(x)$ of the transcendental equation $we^{w} = x$ and a subject of Part \ref{spe}, has become a much-used tool in physics \cite{Corless1996,doi:10.1080/07468342.2004.11922095} and other
disciplines \cite{10.1214/11-AOAS457,10.1214/14-AOAS790,10.2112/JCOASTRES-D-17-00181.1,doi:10.4169/074683410X480276}.
Its history is contained in \cite{Hayes2005}, and many properties are in \cite{Corless1996} and \cite{50831}.
Another name for the Lambert $W$ function is the \textit{product logarithm}. Numerical values are
available in \textit{Mathematica}, \textit{Maple}, \textit{Matlab}, and \textit{Wolfram Alpha}. The function $y = W(x)$ is defined on the complex plane and has two real branches. The principal branch is the function $y = W_{0}(x) =
W_{p}(x)$, which has domain $x\in (-1/e,\infty)$, contains the points $(-1/e,-1)$ and $(0,0)$ and is increasing,
concave, and positive for $x\in [0,\infty)$. The other branch is the function $y = W_{-1}(x) = W_{m}(x)$, which
has domain $x\in [-1/e,0)$ , contains the point $(-1/e,-1)$, is decreasing, convex for
$x\in (-1/e,-2/e^{2})$, concave for $x\in (-2/e^{2},0)$, and asymptotically approaches the negative $y$-axis.
\section{Using \ref{table}}
For the first part of Entry \ref{sin} of \ref{table} in \ref{app}, the full transformation for all $m\in\bbR$ is
\begin{equation}
\L\{\sin(x)\}(m) = \begin{cases}
(\pi/2)m-1 & \text{ for }m\leq 0,\\
m\Cos m - \sqrt{1 - m^{2}} & \text{ for }0<m<1,\\
0 & \text{ for }1\leq m,
\end{cases}\label{usageex}
\end{equation}
by using the supporting lines at the endpoints of the domain of $y = f(x)$. Method \ref{methsupp} is employed to find the extension in \eqref{usageex}, which is outside $0 < m < 1$. Many entries have the aspect that only the nonlinear portion is displayed in the table, but line segments to the left and/or right can be taken to complete the function, as in \eqref{usageex}. When they have no physical or other meaning, these line segments are ignored. For simplicity, they are omitted from the table.
Often, it is assumed that all functions whose Legendre transformations are sought are convex. Then, Theorem \ref{thmleg}\ref{thmmaint} implies that the Legendre transformation is convex, as well. Differentiable convex functions have derivatives that are invertible because the derivatives are monotonic and single valued \cite{Tiel1984ConvexAA}. Some entries of the table are split into parts, which depend on whether the portion of the function is convex or concave.
The transformation can be applied to functions that are part convex and part concave, but there will most likely be complications and, sometimes, unexpected results. Theorem \ref{thmleg}\ref{thmmaint} shows that, if a function has a second derivative that is zero at a point, the corresponding point on the dual curve has an undefined second derivative. Example \ref{conex} illustrates some of these ideas.
\begin{example}[function that has both concave and convex portions]\label{conex}
The function $y = f(x) = \sin x$ for $-\pi/2 < x < \pi/2$ is in Entry \ref{sin}. The function is convex on the left and concave on the right. The function $y = f(x)$ and its transformation are displayed in Figures \ref{sinfig} and \ref{Lsinfig}. As $y = f(x)$ is traced left to right, starting at $(x,y) = (-\pi/2,-1)$, the corresponding points of its transformation are traced downward, starting at $(m,d) = (0,1)$. For each value of $x\neq 0$, the tangent lines to the sine function at points with coordinates $x$ and $-x$ have the same slope $m$ and different intercepts, which makes the transformation in Figure \ref{Lsinfig} a curve, not a function. The point of inflection $(0,0)$ in Figure \ref{sinfig} corresponds to the cusp at $(1,0)$ in Figure \ref{Lsinfig}. The symmetry in Figure \ref{sinfig} about the $m$-axis illustrates Entry \ref{odd}.
\end{example}
The Legendre transformation changes the function $f$ into the function $g$ by changing the independent variable from $x$ to $m$. In physical settings, there may be many other variables present, but the underlying process is unchanged. The transformation’s appearance might change because partial derivatives are employed, in order that other variables are held constant \cite{Alford2019}.
Most domains are presented as open sets, but, in very many cases, boundary points can be
filled in simply by substitution or by a limiting process.
\begin{figure}[ht]
\begin{center}
\begin{tikzpicture}[scale=2]
\foreach \i in {-2,-1.75,...,2} {
\draw [very thin,gray] (\i,-.025) -- (\i,.025);
}
\foreach \i in {-1.25,-1,...,1.25} {
\draw [very thin,gray] (-.025,\i) -- (.025,\i);
}
\draw [thin,gray] (2,-.05) -- (2,.05) node[below, yshift=-.2cm, black] {$2$};
\draw [thin,gray] (1,-.05) -- (1,.05) node[below, yshift=-.2cm, black] {$1$};
\draw [thin,gray] (-1,-.05) -- (-1,.05) node[below, yshift=-.2cm, black] {$-1$};
\draw [thin,gray] (-2,-.05) -- (-2,.05) node[below, yshift=-.2cm, black] {$-2$};
\draw [thin,gray] (-.05,1) -- (.05,1) node[left, xshift=-.2cm, black] {$1$};
\draw [thin,gray] (-.05,-1) -- (.05,-1) node[left, xshift=-.2cm, black] {$-1$};
\draw[->] (-2.1, 0) -- (2.1, 0) node[right] {$x$};
\draw[->] (0, -1.35) -- (0, 1.35) node[above] {$y$};
\draw[scale=1, domain=-1.57:1.57, smooth, variable=\x, blue, <->] plot ({\x}, {sin(180*\x/pi)});
\end{tikzpicture}
\end{center}
\caption{$y = f(x) = \sin x$ for $-\pi/2 < x < \pi/2$}\label{sinfig}
\end{figure}
\begin{figure}[ht]
\begin{center}
\begin{tikzpicture}[scale=2]
\foreach \i in {-1.25,-1,...,1.25} {
\draw [very thin,gray] (\i,-.025) -- (\i,.025);
}
\foreach \i in {-1.25,-1,...,1.25} {
\draw [very thin,gray] (-.025,\i) -- (.025,\i);
}
\draw [thin,gray] (1,-.05) -- (1,.05) node[below, yshift=-.2cm, black] {$1$};
\draw [thin,gray] (-1,-.05) -- (-1,.05) node[below, yshift=-.2cm, black] {$-1$};
\draw [thin,gray] (-.05,1) -- (.05,1) node[left, xshift=-.2cm, black] {$1$};
\draw [thin,gray] (-.05,-1) -- (.05,-1) node[left, xshift=-.2cm, black] {$-1$};
\draw[->] (-1.5, 0) -- (1.5, 0) node[right] {$m$};
\draw[->] (0, -1.5) -- (0, 1.5) node[above] {$d$};
\draw[scale=1, domain=0:1, smooth, variable=\m, blue] plot ({\m}, {-pi/180*\m*asin(sqrt(1 - \m*\m)) + sqrt(1 - \m*\m)});
\draw[scale=1, domain=0:1, smooth, variable=\m, blue] plot ({\m}, {pi/180*\m*asin(sqrt(1 - \m*\m)) - sqrt(1 - \m*\m)});
\end{tikzpicture}
\end{center}
\caption{$d = g(m) = \L\{\sin(x)\}(m)$ from Entry \ref{sin} is the Legendre transformation of the function in Figure \ref{sinfig}}\label{Lsinfig}
\end{figure}
\section{Approximating polynomials}
The Lagrange inversion or Lagrange-B\"urmann theorem pertains to finding dual functions. It says
that, given an analytic function $y = r(x)$ with Taylor series about $x = x_{0}$ and nonzero derivative at
$x = x_{0}$, a Taylor series about $y_{0} = r(x_{0})$ of the inverse function $x = s(y)$ can be found and has a
non-zero radius of convergence. Contour integration and various other techniques can be
employed \cite{50831,https://doi.org/10.1002/zamm.19630430916,Abramowitz1972}. \cite{dienes1957taylor} performs inversion by the method of indeterminate
coefficients.
For the present application, $y = r(x) = f'(x)$ and $x = s(y) = g'(y)$, which are inverse
functions by Theorem \ref{thmleg}\ref{thmslope}. Because all derivatives may not exist for functions of interest, we
find approximating polynomials in the dual space. These polynomials would be terms of the
Taylor series. Interesting ideas, especially about this type of potentially limited use of Lagrange
inversion, can be found in \cite{Grossman2005}, \cite{Johnson2002}, \cite{Krattenthaler1988}, and \cite{Traub1962}.
Polynomials that approximate $d = g(m)$ near $m = m_{0}$ can be found without first finding the
function $g$. Each term of the approximating polynomial, which are the terms of the Taylor
expansion of $g$, is a function of the derivatives of $y = f(x)$ at $x = x_{0}$ up to the same degree as the
polynomial in $m, d$-space. The terms of the polynomial approximation of $g$ about $m=m_{0}$ are
determined by the terms of the polynomial of the same degree of approximation of $f(x)$ at $x = x_{0}$.
Besides differentiability of $f$ at $x = x_{0}$ to the degree desired, it is required that $f''(x_{0}) \neq 0$.
From the definition of $m = f'(x)$,
$$m_{0} = f'(x_{0}).$$
The defining equation of $d = g(m)$ is from the tangent line $y=mx - d$ so that $g = mx - f$ and
$$g(m_{0}) = x_{0}f'(x_{0}) - f(x_{0}).$$
From Theorem \ref{thmleg}\ref{thmslope}, $g'(m) = x$, so that
$$g'(m_{0}) = x_{0}.$$
From Theorem \ref{thmleg}\ref{thmmaint}, $g''(m) = 1/f''(x)$, so that
$$g''(m_{0}) = \frac{1}{f''(x_{0})}.$$
A recursion formula for $g^{(n+1)}(m)$ for integers $n\geq 2$ is available. Beginning with $g''(m) = 1/f''(x)$, differentiate with respect to $x$ to obtain
$$\frac{dg''(m)}{dm}\frac{dm}{dx} = g'''(m)f''(x) = \frac{d}{dx}\frac{1}{f''(x)} = -\frac{f'''(x)}{(f''(x))^{2}}$$
and
$$g'''(m) = \frac{1}{f''(x)}\frac{d}{dx}\frac{1}{f''(x)} = -\frac{f'''(x)}{(f''(x))^{3}},$$
so that
$$g'''(m_{0}) = -\frac{f'''(x_{0})}{(f''(x_{0}))^{3}}.$$
Continuing to differentiate similarly, we obtain
\begin{align*}
g^{(4)}(m) &= \frac{1}{f''(x)}\frac{d}{dx}\left(-\frac{f'''(x)}{(f''(x))^{3}}\right) = \frac{3(f'''(x))^{2} - f''(x)f^{(4)}(x)}{(f''(x))^{5}},\\
g^{(5)}(m) &= \frac{1}{f''(x)}\frac{d}{dx}\left(\frac{3(f'''(x))^{2} - f''(x)f^{(4)}(x)}{(f''(x))^{5}}\right) = \frac{10f''(x)f'''(x)f^{(4)}(x) - 15(f'''(x))^{3} - (f''(x))^{2}f^{(5)}(x)}{f''(x)^{7}},
\end{align*}
and so forth. The right-hand sides can be evaluated at $x=x_{0}$. To obtain the next higher derivative, and thus the next term, divide the derivative of the current derivative as a function of $x$ by $f''(x)$ according to the recursion formula
$$(g^{(n+1)}(m))(x) = \frac{1}{f''(x)}\frac{d}{dx}(g^{(n)}(m))(x).$$
\begin{example}[product of functions $y=x\sin x$]
The Taylor series for $y = f(x) = x\sin x$ about $x=0$ is
\begin{align*}
y = f(x) = \sum_{i=0}^{\infty} f^{(i)}(0)\frac{x^{i}}{i!} = x\sin x = x\sum_{i=0}^{\infty}(-1)^{i} \frac{x^{2i+1}}{(2i+1)!} = \sum_{i=1}^{\infty} (-1)^{i-1}\frac{x^{2i}}{(2i-1)!},
\end{align*}
which converges for all real numbers \cite{50831}. Thus, $m_{0} = f'(0)$, the odd derivatives of $f$ at $x=0$ are zero, and, for all nonnegative integers $n$, $f^{(2n)}(0) = (-1)^{n+1}(2n)$. Then, at $m = m_{0} = 0$, $g(0) = 0, g'(0) = 0, g''(0) = \frac{1}{2}, g'''(0) = 0,$ and
$$g^{(4)}(0) = \frac{3(f'''(0))^{2} - f''(0)f^{(4)}(0)}{(f''(0))^{5}} = \frac{3(0)^{2} - (2)(-4)}{(2)^{5}} = \frac{1}{4},$$
so the first five terms of the approximating polynomial at $m=0$ to the Legendre transformation of $x\sin x$ is
$$0 + 0m + \frac{1}{2}\frac{m^{2}}{2!} + 0\frac{m^{3}}{3!} + \frac{1}{4}\frac{m^{4}}{4!} = \frac{m^{2}}{4} + \frac{m^{4}}{96}.$$
\end{example}
\section{Closing comments}
Clairaut’s differential equation
\begin{equation}
y = xy' + h(y')\label{clair}
\end{equation}
is a expression of the Legendre transformation between two differentiable functions \cite{Ince1944,Sternberg1954}. This can be seen by replacing $y$ by $f$, $y'$ by $m$, and the function $h$ by the function $-g$. The singular solution of \eqref{clair} is the Legendre transformation of the function $g(m)$ and vice versa by the reflectivity property in Theorem \ref{thmleg}\ref{thmrefl}. The general solution of \eqref{clair}, which contains a constant of integration as a parameter, is the tangent lines to the singular solution \cite{Rainville1964,zill2008differential}. \ref{table} in \ref{app} is an integral table for Clairaut’s differential equation \cite{Sternberg1954}. Clairaut’s equation and the Legendre transformation are displayed in the formula for integration by parts with transformation are displayed in the formula for integration by parts with $\int f'(x)dx = mx - \int g'(m)dm$.
Besides $m,d$-space, there are many spaces that are dual to $x,y$-space. Each dual space is formulated in terms of the coefficients of a standard form for the equations of lines that serve as tangent and supporting lines to curves in $x,y$-space \cite{kilner2021tangent}. Here, $m$ and $d$ from $y = mx - g(m)$ yield curves $d = g(m)$ that are dual to $y = f(x)$, which has the lines as supporting or tangent lines. Hence, $f(x)$ and $g(m)$ are Legendre transformation pairs.
Using $y = mx + b(m)$ to give points in $m,b$-space is often called the Legendre transformation. This transformation is intuitive, because the $y$-intercept $b$ is a coordinate in the dual space, but the reflexivity property is lost. In order to use the table for this alternative definition, the adjustment can be made between $b$ and $d$ \cite{Alford2019,kilner2021tangent}.
Another example is $u,v$-space, where the coefficients $u$ and $v$ of
\begin{equation}
ux+vy = 1 \label{uv}
\end{equation}
in $x,y$-space give the dual curves $v = w(u)$ \cite{Thompson1996}. The pairs in the table can be converted to relationships for the $u$ and $v$ coefficients using the identities $m = –u/v$ and $d = 1/v$ \cite{kilner2021tangent}. Different lines in $x,y$-space may have no representations, thus losing various portions of curves. In $x,y$-space, vertical lines have undefined slope $m$ and, therefore, neither $m,d$-representation nor $m,b$-representation, and lines through the origin of $x,y$-space have no $u,v$-representation because the lines cannot be written as \eqref{uv}. Accommodations or adjustments in order to include omitted lines can be made by introducing points at infinity, but that can sometimes defeat the goal of simple or physical representations.
|
\section{Introduction}\label{sec:intro}
It is widely accepted that an
active galactic nucleus (AGN) is powered by
mass accretion onto a supermassive black hole (SMBH)
in the centre of the host galaxy.
The gravitational energy of accreting matter is converted
into radiation and/or jets.
Radio galaxies (RGs) are radio-loud AGNs characterised
by powerful synchrotron radiation
driven by relativistic jets on scales approximately
in the range
10--100 kiloparsec.
The interstellar medium (ISM) in the centre of the RGs
can play a key role in fuelling the SMBH.
Several research groups have suggested that
the different roles of hot and cool ISM accretion
can be related to a different mode of accretion
in RGs,
which can lead to different radio-loud AGN classifications
\citep{hardcastle07,buttiglione10,best12}.
Certain CO observations of RGs support the hypothesis
that RGs are fed by cold gas that probes
the circumnuclear disks
\citep[CNDs;][]{prandoni10,maccagni18,ruffa19}.
The CNDs of RGs
can serve as a reservoir of fuel for their SMBHs.
Thus,
by determining the molecular gas structure and kinematics of CNDs,
important clues can be obtained
regarding mass accretion in RGs.
High angular resolution imaging of the molecular gas
within the SMBH sphere of influence ($r_g$)
can also be a powerful tool for accurately measuring
SMBH masses \citep[e.g.][]{davis13}.
However, the distribution of low-$J$ CO lines appears to extend
to the edge of CNDs,
and the detection of strong CO emission
from within $r_g$ seems to be rare
for CNDs in early-type galaxies \citep[ETGs;][]{davis18,boizelle19,north19}.
Alternative lines may trace the molecular gas distribution
within $r_g$ better,
thereby enabling more accurate measurements of SMBH masses.
At present, emission lines other than CO have largely been overlooked,
and dense gas tracers
such as HCN and HCO$^+$ lines are expected to trace possibly
farther into $r_g$ than the optically thick low-$J$ CO lines.
Moreover, recent interferometric observations of
the dense gas emission-line tracer HCN
towards Seyfert galaxies (SGs)
have indicated
that the accretion onto SMBHs is triggered
by star formation and supernovae originating from within CNDs
\citep{izumi16}.
To date, star formation activities in CNDs
have been investigated in a limited number of RGs,
such as
NGC~5128 \citep{espada19},
NGC~1052 \citep{kameno20},
and NGC~1275 \citep{nagai21}.
These investigations of RGs
have been primarily conducted
using the distribution of CO as a molecular gas mass tracer.
Dense gas tracers can potentially
serve as better probes
for examining star formation activity in CNDs
because star formation is closely related to dense gas.
NGC~4261 (3C~270) is a nearby Fanaroff--Riley I RG
with a symmetric, kiloparsec-scale two-sided jet
\citep{birkinshaw85}.
Its AGN is classified as a type 2
low-ionisation nuclear emission-line
region (LINER) galaxy \citep{jaffe96, ho97}
with a low Eddington ratio,
$L_{\rm bol}/L_{\rm Edd}$, ranging
from $10^{-5.11}$ to $10^{-4.54}$
\citep{hernandez13, inayoshi20}
and
a low X-ray luminosity of $L_{\rm 2-10 keV} = 10^{41.51}$
erg s$^{-1}$ \citep{hernandez13}.
This galaxy is known to have a nuclear disk of dust and gas
with a radius of a few hundred parsecs
lying orthogonal to the jet,
as revealed by \textit{Hubble} Space Telescope (HST)
observations
\citep[][]{jaffe93,jaffe96}.
At radio frequencies,
H~$\textsc{i}$ absorption
has been detected at the systemic velocity of the galaxy
($V_{\rm sys}$) towards the core
via the Very Large Array \citep[VLA;][]{jaffe94},
and it has been confirmed
at a projected distance of approximately 2.5 pc
from the core
with the European
Very Long Baseline Interferometry (VLBI)
Network \citep{vanlangevelde00}.
The H~$\textsc{i}$ absorbing gas is interpreted to be
in the inner part of the disk of dust and gas
found in the HST image,
and it obscures the core and innermost counter jet.
Subsequent VLBI observations at multiple frequencies
revealed the presence of parsec-scale ionised absorbing gas,
which was likely at the inner parsec-scale radii of the HST disk
\citep[][]{jones97,jones00,jones01,haga15}.
Molecular lines were observed towards the centre
of NGC~4261 for emission
(CO $J$=2--1 and $J$=3--2; \citealt{boizelle21})
and absorption (CO $J$=1--0; \citealt{jaffe94}).
In this Letter, we report the first detection and
the interferometric emission-line maps of
HCN $J$=1--0 and HCO$^{+}$ $J$=1--0 transitions in NGC~4261,
which trace a 100-parsec rotating CND
perpendicular to the kiloparsec-scale radio jet.
We adopt a luminosity distance ($D_{\rm L}$) of 31.7 Mpc and
a $V_{\rm sys}$ of 2212 km s$^{-1}$
\citep[e.g.][]{babyk19,cappellari11}.
Hence, 1 arcsecond corresponds to 151 pc for the galaxy.
\section{Observations and data reduction} \label{sec:obs}
Observations were conducted on February 15, 2019,
with the NOrthern Extended Millimeter Array (NOEMA)
of the Institut de radioastronomie millim\'{e}trique (IRAM)
in A configuration with ten antennas.
The phase centre was set to the position of NGC~4261 at
RA(J2000)=$12^{\rm h}19^{\rm m}23^{\rm s}.220$ and
Dec(J2000)=05$^{\circ}$49$^{\prime}$30$^{\prime\prime}$.775.
The full width at half maximum of the NOEMA primary beam was
$55^{\prime\prime}$ at 88 GHz.
The projected baseline lengths ranged from 22 to 734 m
over the course of the observations.
The local oscillator frequency was set to 82.0 GHz,
with frequencies ranging from 70.398 to 78.115 GHz in the lower sideband
and from 85.886 to 93.603 GHz in the upper sideband.
The PolyFix correlator was configured with a frequency resolution of 2~MHz.
The nearby source 3C~273 was observed
as both the bandpass and gain calibrators.
The absolute flux calibration was performed using MWC 349 (1.07 Jy)
and LkH$\alpha$101 (0.22 Jy).
The absolute flux calibration uncertainty for NOEMA is
less than $10\%$ at Band 1 ($\lambda$3 mm)
\footnote{IRAM NOEMA Data Reduction CookBook,
https://www.iram.fr/IRAMFR/GILDAS/doc/html/pdbi-cookbook-html/pdbi-cookbook.html}.
The raw visibility data were first converted into
the Flexible Image Transport System (FITS) format data
through the GILDAS software \citep{pety05}.
Then, calibration and imaging were performed
by using the NRAO Astronomical Image Processing System (AIPS) package
\citep{greisen90}.
We applied uniform weighting to the images
to obtain a higher spatial resolution of $<1^{\prime\prime}$.
To create the continuum map,
all line-free spectral windows with frequency ranges
of 70.398--74.459, 74.461--78.115, and 89.948--93.603 GHz
were combined, resulting in a centre frequency of 80 GHz.
After the continuum emission was subtracted in the \textit{u-v} plane,
channel maps of lines were made every 20~MHz,
corresponding to a velocity resolution of 68 km~s$^{-1}$.
\section{Results}
\subsection{Continuum emission}
\begin{figure}[ht]
\begin{center}
\resizebox{\hsize}{!}{\includegraphics[bb=0 0 512 272]{figures/n4261noema_cnt.pdf}}
\caption{Continuum map of NGC~4261 at 80~GHz.
Contours begin at the $3\sigma$ level and
increase by factors of $\sqrt{3}$ to
$9\sigma$ and by factors of 3 thereafter,
where $\sigma=3.8\times10^{-2}$ mJy beam$^{-1}$.
The position offsets are measured from
RA(J2000)=$12^{\rm h}19^{\rm m}23^{\rm s}.220$ and
Dec(J2000)=05$^{\circ}$49$^{\prime}$30$^{\prime\prime}$.775.
The peak intensity, located at the centre,
is 0.36 Jy beam$^{-1}$.
The synthesised beam is
$1^{\prime\prime}.8\times0^{\prime\prime}.85$
at a PA of $25^{\circ}$,
as indicated by the cross-hatched ellipse
in the bottom-left corner.
}
\label{fig:n4261cnt}
\end{center}
\end{figure}
Figure~\ref{fig:n4261cnt} shows a bright compact source
at the phase centre
and weak jet features spanning 5 kpc
aligned along the east-west direction.
The position angle (PA) of the alignment is estimated
as $87\pm1^{\circ}$
via a linear-regression fit to the jet features.
This estimation is in agreement
with the 5 GHz radio jet PA of $88\pm2^{\circ}$
imaged with the VLA
\citep{birkinshaw85}.
The bright compact source is partially resolved into
a core
at $\Delta$RA = $0^{\prime\prime}.0$
along with the east and west nuclear jet components
at $1^{\prime\prime}.5$
and $-2^{\prime\prime}$.0, respectively.
The peak position of the continuum emission
coincides with that of the phase centre.
The flux density of the compact core
within the central $\pm3^{\prime\prime}$
is measured as $S_{\rm 80GHz}$ = 360 mJy
by means of a two-dimensional Gaussian fit
using the AIPS task JMFIT.
Together with literature flux measurements,
$S_{\rm 115GHz}$ = $326.34\pm0.82$ mJy
observed with the IRAM 30 m telescope
\citep{ocana10},
$S_{\rm 236GHz}$ = $253\pm25$ mJy,
and
$S_{\rm 348GHz}$ = $223\pm22$ mJy
from Atacama Large Millimeter/submillimeter Array projects
2017.1.00301.S and 2017.1.01638.S
(Boizelle, priv. communication),
we find a spectrum $S_\nu \propto \nu^\alpha$
with $\alpha = -0.34\pm 0.02$ in the 80-348 GHz range.
This core spectrum is very inconsistent with thermal emission,
although it is still somewhat shallower than
canonical synchrotron values measured in the extended jets,
suggesting a partially optically thick core environment.
\begin{figure*}[!ht]
\begin{center}
\includegraphics[width=17cm,bb=0 0 504 216]{figures/spc.pdf}
\end{center}
\caption{
Integrated spectrum over the region within $\pm1^{\prime\prime}$
of the core from 85.886 to 89.946 GHz,
after subtracting continuum emission.
The frequency resolution is 20~MHz,
and the rms noise is 0.28 mJy.
(a) Spectrum as functions of topocentric frequency in GHz.
The vertical dashed lines represent $V_{\rm sys}$
for each molecular line.
(b)
Spectra of HCN $J$=1--0 and HCO$^+$ $J$=1--0 emission lines
as functions of velocity with respect to $V_{\rm sys}$.
}
\label{fig:n4261spc}
\end{figure*}
\begin{figure*}[!ht]
\centering
\includegraphics[clip,width=15cm]{figures/moment.pdf}
\caption{
Integrated intensity (moment-0) maps of
(a) HCN $J$=1--0 and (b) HCO$^{+}$ $J$=1--0 lines
towards the central $\pm2^{\prime\prime}$.
Both moment-0 images use the same colour scaling
from 0 to 1.50 Jy beam$^{-1}$ km s$^{-1}$
and the same contour levels in steps of $4\sigma$,
where $\sigma=8.8\times10^{-2}$ Jy beam$^{-1}$ km s$^{-1}$.
The peak intensity is 1.50 and 0.87 Jy beam$^{-1}$ km s$^{-1}$
for HCN and HCO$^{+}$, respectively.
Intensity-weighted velocity (moment-1) maps of
(c) HCN $J$=1--0 and (d) HCO$^{+}$ $J$=1--0 lines
towards the central $\pm2^{\prime\prime}$,
with the same color scaling
from $-800$ to $+800$ km s$^{-1}$
in relative velocity with respect to $V_{sys}$.
All maps are restored with a circular Gaussian
beam of $0^{\prime\prime}.75$,
which is the typical minor-axis size of the synthesised beam.
The position offsets are measured from
RA(J2000)=$12^{\rm h}19^{\rm m}23^{\rm s}.220$ and
Dec(J2000)=05$^{\circ}$49$^{\prime}$30$^{\prime\prime}$.775.
}
\label{fig:moment}
\end{figure*}
\subsection{HCN and HCO$^{+}$ emission}
The spectral profiles of
the HCN $J$=1--0 and HCO$^+$ $J$=1--0 emission lines
integrated over the region
within the central $\pm1^{\prime\prime}$
are shown in Figure~\ref{fig:n4261spc}.
Both molecular lines are detected
above the $3\sigma$ level
within a velocity range of $\pm700$ km~s$^{-1}$
from $V_{\rm sys}$.
Furthermore, both lines are
below the $3\sigma$ level
in the channels around $V_{\rm sys}$
and
exhibit a nearly symmetrical double-peaked spectral profile.
The double-peaked spectra resemble the double-horned profile
expected from an inclined rotating disk
with a central depression or a cavity
\citep[e.g.][]{wiklind97}.
In addition, a possible absorption feature is detected
at the redshifted frequencies of C$_2$H $N$=1--0
with a significance of $3\sigma$
in the deepest absorption channel.
The detected feature could be a blend of six hyperfine components of C$_2$H $N$=1--0
($J$=3/2--1/2, $F$=1--1, 2--1, 1--0, and $J$=1/2--1/2, $F$=1--1, 0--1, 1--0).
The velocity-integrated intensity (moment-0) maps
of the HCN $J$=1--0 and HCO$^+$ $J$=1--0 emission lines
shown in Figures~\ref{fig:moment}(a) and (b)
reveal a single component,
which spatially coincides with the central continuum peak.
A faint feature can be seen $1^{\prime\prime}.5$ east
of the phase centre in the HCN moment-0 map,
but it does not reach the $4\sigma$ level.
A least-squares ellipse fit
to the regions defined by the
$4\sigma$ contour of the integrated intensity
for each moment-0 map
is listed in Table~\ref{tab:ellifit}.
The extent of the significant HCN emission
spans $1^{\prime\prime}.4$ (210 pc)
along the north-south direction (PA$=2^{\circ}$),
which is
slightly longer than
a beam size of $0^{\prime\prime}.75$.
The distribution of HCO$^{+}$ emission is more concentrated
at the centre
and the HCO$^+$ component is fainter.
Both of these molecular lines
originate from the same $1^{\prime\prime}.7$-diameter dust disk
found in the HST images \citep{jaffe93,jaffe96}
and are more centrally concentrated
compared to the CO $J$=2--1 emission
that spans $2^{\prime\prime}$
in prior interferometric observations \citep{boizelle21}.
The intensity-weighted velocity (moment-1) map of HCN
(Figure~\ref{fig:moment}(c))
tentatively
shows a velocity gradient
along the major axis,
perpendicular to the jet PA.
The distribution and velocity structure of the HCN line
are in agreement with those obtained for
CO $J$=2--1 and $J$=3--2 lines
\citep{boizelle21}.
Furthermore,
the moment-1 map of HCO$^+$ (Figure~\ref{fig:moment}(d))
roughly follows
the velocity gradient along the north-south direction,
although the velocity gradient is less evident
than that of HCN.
It should be noted that
the HCO$^+$ distribution appears to exhibit
a barely resolved disk structure.
Thus, multiple velocity features should be spatially unresolved.
The HCN $J$=1--0 to HCO$^{+}$ $J$=1--0 ratio ($R_{\rm HCN/HCO^{+}}$)
and the HCN $J$=1--0 to CO $J$=1--0 ratio ($R_{\rm HCN/CO}$)
are proposed to be good indicators of an AGN-dominated environment in SGs
\citep{kohno01, kohno05}.
Velocity-integrated flux densities of
HCN ($S_{\rm HCN} \Delta V$)
and HCO$^+$ ($S_{\rm HCO^{+}} \Delta V$)
within the central $\pm1^{\prime\prime}$ are
1.48 and 0.79 Jy km s$^{-1}$, respectively.
We derive the $R_{\rm HCN/HCO^{+}}$
= 1.87
on sub-kiloparsec scales,
which is consistent with the mean ratio of $1.84\pm0.43$
for a sample of AGN host galaxies \citep{privon15}.
Assuming the CO $J$=2--1 to $J$=1--0 intensity ratio $R_{\rm 21}$ = 0.79
from the xCOLD (extended CO Legacy Database) for GASS (GALEX Arecibo SDSS Survey)
sample of nearby galaxies \citep{koss21}
and the velocity-integrated flux density of CO $J$=2--1 of 3.06 Jy km s$^{-1}$
measured for NGC~4261 \citep{boizelle21},
we get $R_{\rm HCN/CO}$ = 0.38.
The resultant line ratios $R_{\rm HCN/HCO^{+}}$ of 1.87 and
$R_{\rm HCN/CO}$ of 0.38 in NGC~4261 are typical values
expected for `pure' AGNs
with the absence of any associated nuclear starburst
activity \citep{kohno05}.
\begin{table}
\caption{Elliptical fit parameters to integrated intensities of HCN and HCO$^+$}
\label{tab:ellifit}
\begin{tabular}{c c c c c}
\hline\hline
Line &
$\Delta{\rm RA}$ &
$\Delta{\rm Dec}$ &
$\theta_{\rm mj}\times\theta_{\rm mn}$ &
PA \\
& [$^{\prime\prime}$]
& [$^{\prime\prime}$]
& [$^{\prime\prime}\times^{\prime\prime}$]
& [$^\circ$] \\
(1) & (2) & (3) & (4) & (5) \\
\hline
HCN(1--0)
& $0.08\pm0.01$
& $0.02\pm0.01$
& $1.4\times0.8$
& 2 \\
HCO$^+$(1--0)
& $-0.04\pm0.02$
& $-0.07\pm0.02$
& $1.0\times0.7$
& 35 \\
\hline
\end{tabular}
\tablefoot{
Column(1): Line species.
(2): RA offset from the continuum source centroid
at RA(J2000)=$12^{\rm h}19^{\rm m}23^{\rm s}.220$.
(3): DEC offset from the continuum source centroid
at
DEC(J2000)=05$^{\circ}$49$^{\prime}$30$^{\prime\prime}$.775.
(4): Angular widths of the major and minor axes.
(5): Position angle.
}
\end{table}
\section{Discussions}
\begin{figure}[t]
\begin{center}
\includegraphics[width=7cm,bb=0 0 214 289]{figures/pv.pdf}
\caption{
Luminosity-weighted moment-1 measurements in declination for HCN and HCO$^+$ lines.
The dashed magenta line represents
the best-fit Keplerian rotation model
to the HCN data (red squares).
The dot-dashed cyan line indicates
a linear gradient fit to the HCO$^+$ data (blue triangles).
The straight dashed lines in magenta and cyan
correspond to a rotating gas ring with a radius of 66 and 92 pc, respectively.
}
\label{fig:pv}
\end{center}
\end{figure}
\subsection{Keplerian rotation of CNDs}
Luminosity-weighted moment-1 measurements
for $J$=1--0 emission lines of HCN and HCO$^+$
along the major axis (PA$=0^{\circ}$)
are shown in Figure~\ref{fig:pv}.
Data points are derived from the velocity slice
across the centre along PA$=0^{\circ}$
in the HCN and HCO$^+$ moment-1 maps
(Figure~\ref{fig:moment}(c)(d)).
We performed a linear fit to the HCN data points
at a position offset within $\pm0^{\prime\prime}.4$
and a Keplerian rotation fit to the HCN data
at $<-0^{\prime\prime}.4$ and $>+0^{\prime\prime}.4$.
The HCN data points and
their best-fit Keplerian rotation curves
indicate that
the HCN emission traces the rotation with
a radius in the range 66--130 pc.
The enclosed mass estimated from the Keplerian rotation fitting is
$(1.6\pm0.1)\times10^9$ $M_{\odot}$,
after adopting a disk inclination angle of $64^{\circ}$
\citep{ferrarese96}.
This is in good agreement with the black hole mass measurement
made using the CO lines
($1.67\times10^9 M_{\odot}$; \citealt{boizelle21}),
while it is three times larger
than the mass determination inside $0^{\prime\prime}.1$
(14.5 pc)
based on the ionised gas kinematics
($4.9\times10^8 M_{\odot}$; \citealt{ferrarese96}).
The HCO$^+$ data also appear to display a velocity gradient
along the major axis,
while the data points are more scattered
from the linear gradient.
This could be due to
the barely resolved HCO$^+$ multiple velocity features.
The best-fit rotational gas model
for the HCO$^+$ data is
consistent with the Keplerian rotational model
obtained from the HCN data.
This implies that
HCN and HCO$^+$ emission trace the same galaxy potential
for the same radii of approximately 60--130 pc.
\subsection{Mass of dense molecular gas in CNDs}
The HCN line luminosity
has been used to estimate the mass of the dense molecular gas,
by applying the HCN luminosity-to-mass conversion factor
$\alpha_{\rm HCN}$.
Using the HCN line luminosity
in accordance with \cite{solomon97} and \cite{tan18},
the dense molecular gas mass ($M_{\rm dg}$) is as follows:
\begin{eqnarray}
M_{\rm dg} &=& \alpha_{\rm HCN} L_{\rm HCN}^{\prime} \nonumber \\
&=& 3.25\times10^7 \alpha_{\rm HCN} S_{\rm HCN}\Delta V
\nu_{\rm obs}^{-2} D_{\rm L}^2 (1+z)^{-3}
~ M_{\odot},
\end{eqnarray}
where
$\alpha_{\rm HCN}$ denotes the HCN luminosity-to-dense-gas-mass
conversion factor in $M_{\odot}$ (K km s$^{-1}$ pc$^{2}$)$^{-1}$,
$L_{\rm HCN}^{\prime}$ denotes the HCN luminosity
in K~km~s$^{-1}$~pc$^2$,
$S_{\rm HCN} \Delta V$ denotes the velocity-integrated flux density
of HCN in Jy~km~s$^{-1}$,
$\nu_{\rm obs}$ denotes the observed line frequency in GHz
and $D_{\rm L}$ denotes the luminosity distance in Mpc.
We obtain $L_{\rm HCN}^{\prime}$ = $6.03\times10^6$ K~km~s$^{-1}$~pc$^2$.
The luminosity-to-mass conversion factor, $\alpha_{\rm HCN}$,
can vary from
0.24 to over 20 $M_{\odot}$ (K km~s$^{-1}$ pc$^2$)$^{-1}$
\citep{barcos18,evans20,jones21},
depending on the gas density and the line opacity
\citep{jones21,wang21}.
If we consider the standard extragalactic conversion factor
$\alpha_{\rm HCN}$ of
10 $M_{\odot}$ (K km~s$^{-1}$ pc$^2$)$^{-1}$
\citep{gao04},
$M_{\rm dg}$ is calculated to be
$6.03\times10^7$ $M_{\odot}$.
This value is over five times higher than the total gas mass
$M_{\rm gas}$ = $1.12\times10^7$ $M_{\odot}$
reported by \citet{boizelle21}
using a typical CO conversion.
The adopted conversion factor $\alpha_{\rm HCN}$ = 10 could be
an order of magnitude too large
because NGC~4261 has a higher $R_{\rm HCN/CO}$ compared with other galaxies.
The values of $M_{\rm dg}$ and $M_{\rm gas}$ ($10^{7}$ $M_{\odot}$)
are consistent with CND masses on 100 pc scales
in galaxies of various types, including SGs, ETGs, and RGs
\citep[e.g.][]{izumi16,boizelle17,garcia19,ruffa19}.
In contrast,
it is significantly more massive than
the molecular gas mass of the 100 pc CND
in other nearby RGs:
NGC~5128 ($2 \times 10^6$ $M_{\odot}$;
\citealt{mccoy17})
and NGC~1052
($5.3 \times 10^5$ $M_{\odot}$; \citealt{kameno20}).
Interestingly, these three RGs
show dissimilar CND characteristics to one another
despite many resemblances,
such as
the classification of a LINER AGN,
the bright and two-sided radio jets,
the presence of a surrounding torus,
the central condensation of the ionised gas,
and
the central high column density of hydrogen
\citep[e.g.][]{marconi00,kameno01,markowitz07,balokovic21}.
Further observations in a larger sample of RGs
are required
to understand the variety of the observed CND characteristics.
\subsection
{$M_{\rm dg}$--$\dot M_{\rm BH}$ correlation}
\citet{izumi16} report
a positive correlation between $M_{\rm dg}$ and
the black hole mass accretion rate, $\dot M_{\rm BH}$,
for SGs.
By applying an $M_{\rm dg}$ of $6.03\times10^7$ $M_{\odot}$
and
using our measurement
to the regression line for $M_{\rm dg}$ and $\dot M_{\rm BH}$
offered by \citet{izumi16},
the inferred $\dot M_{\rm BH}$ corresponds to
$10^{-2.48}$ $M_\odot$ yr$^{-1}$.
This value is comparable to
an $\dot M_{\rm BH}$ of $10^{-2.70}$ $M_\odot$ yr$^{-1}$,
which is obtained by using the
$L_{\rm bol}$--$\dot M_{\rm BH}$ relation \citep{alexander12}:
\begin{equation}
\dot M_{\rm BH} = 0.15
\biggl( \frac{0.1}{\eta} \biggr)
\biggl( \frac{L_{\rm bol}}{10^{45} {\rm erg~s}^{-1}} \biggr)
~ M_\odot {\rm yr}^{-1},
\end{equation}
where $\eta$ = 0.1 is a typical value
for mass--energy efficiency conversion
\citep{marconi04}
and $L_{\rm bol}$ is equal to $10^{42.6}$ erg s$^{-1}$ for NGC~4261,
as reported by \citet{hermosa22}.
The derived $M_{\rm dg}$ appears to be in agreement with
the positive correlation between $M_{\rm dg}$ and
$\dot M_{\rm BH}$ at the CND scale in NGC~4261.
It should be noted, however, that there are significant uncertainties
in $\alpha_{\rm HCN}$,
the $L_{\rm 2-10keV}$--$L_{\rm bol}$ relation \citep[e.g.][]{eracleous10},
and
the confidence interval in the correct $\eta$ to use
in general SGs.
\section{Conclusions}
We mapped
the central 5 kpc of NGC~4261 with NOEMA
in the HCN and HCO$^+$ $J$=1--0 lines and
the 80 GHz continuum.
The continuum image reveals a core-dominant
synchrotron jet structure,
which consists of
a bright central source and weak jet features
aligned along the east-west direction.
HCN and HCO$^{+}$ emission lines are detected in NGC~4261
for the first time,
covering a velocity range of
$\pm700$ km~s$^{-1}$ relative to $V_{\rm sys}$.
The molecular gas is distributed in
a rotating sub-kiloparsec disk structure,
which coincides with the bright central source
in position.
The Keplerian rotation model obtained from
the velocity fields of HCN and HCO$^+$
yields an enclosed mass of $(1.6\pm0.1)\times10^{9}$
$M_\odot$.
Using the HCN line luminosity
and the standard extragalactic luminosity-to-mass
conversion factor,
the dense gas mass, $M_{\rm dg}$,
associated with the CND is estimated
to be $6.03\times10^7$ $M_{\odot}$.
This value is comparable to a typical CND mass
measured in galaxies of various types,
including SGs, ETGs, and RGs.
The derived $M_{\rm dg}$ and $\dot M_{\rm BH}$ in NGC~4261
align with the positive correlation
between $M_{\rm dg}$ and $\dot M_{\rm BH}$
seen in SGs,
which supports the scenario that
star formation in CNDs drives mass accretion onto SMBHs,
although there are significant uncertainties in the parameters
of the correlation.
\begin{acknowledgements}
We acknowledge the anonymous referee for valuable comments that improved our manuscript.
This work is based on IRAM/NOEMA observations carried out under project number W18CK.
IRAM is supported by INSU/CNRS (France), MPG (Germany) and IGN (Spain).
S.S.-S. is supported by JSPS KAKENHI grant No. 21K03628.
S.K. is supported by JSPS KAKENHI grant No. 18K0371.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{INTRODUCTION}
\label{sec:intro}
Originating only 380,000 years after the Big Bang, the Cosmic Microwave Background (CMB) is a direct window into the early Universe. The polarization of the CMB contains a wealth of information relating to inflation, reionization, dark matter interactions, and more.\cite{wmapResults,plank2018params,bicep,actBi} The Cosmology Large Angular Scale Surveyor (CLASS) measures the CMB polarization at frequency bands centered near 40, 90, 150, and 220 GHz.\cite{TEH2014,KH2016} Each CLASS telescope consists of a novel, front-end polarization modulator, custom optics and cryogenics, and a focal plane of transition-edge sensor (TES) bolometers\cite{jeff2018, dahal2022}. CLASS has observed from the Atacama Desert in Chile since 2016 with the goal of measuring the largest scales accessible from the ground.
The first optical element in each CLASS telescope is a variable-delay polarization modulator (VPM), which consists of a polarizing wire grid in front of and parallel to a movable mirror.\cite{Chuss2012} This creates a phase delay between incident polarization parallel and perpendicular to the wires. The CLASS VPM mirrors move at a frequency of 10 Hz relative to the wire grid. This rapid modulation encodes the CMB linear polarization (Stokes $Q$/$U$) at a frequency above typical atmospheric and instrumental noise sources that would otherwise impact large angular-scale observations.\cite{miller2016} The VPM also provides sensitivity to circular polarization (Stokes $V$). The CLASS VPM design and 40 GHz performance have been described in Harrington et al. (2018)\cite{KH2018} and Harrington et al. (2021)\cite{KH2021}. In this proceeding we provide an analysis of an initial, representative data set at all four CLASS frequencies.
\section{Data \& Modeling}
\label{sec:data}
Most CMB polarimeters today utilize total power sensors (such as TESs) that are coupled (e.g., by probe antennas) to a single linear polarization. Thus these detectors measure a combination of the total intensity (Stokes $I$), linear polarization (Stokes $Q$/$U$), and circular polarization (Stokes $V$). Orthogonal detector pairs can be differenced to isolate the linear polarization signal. In principle, the pair-difference cancels the common, dominant unpolarized component and the circular polarization. In practice, the cancellation is imperfect and the pair-differenced data will also include spurious signal (i.e., intensity-to-polarization leakage) that could overwhelm the cosmological signal CLASS aims to measure. To suppress the spurious signal on the largest angular scales, CLASS modulates Stokes $U$ and $V$ at a frequency of 10 Hz. (Here, $+Q$ is defined parallel to the VPM wires.) The 10 Hz modulation frequency of the CLASS VPMs is much higher than the rate at which this spurious signal is expected to change, which can then be removed via high pass filtering. CLASS is then left with modulated $U$ and $V$. CLASS tracks the relative position of the VPM wire grid and mirror; thus, the modulation functions are known at all times. This allows the recovery of the incident $U$ and $V$ signals. To observe the $Q$ signal, CLASS performs daily boresight rotations that range between -45$^\circ$ and +45$^\circ$. For a more rigorous discussion of demodulation, see Harrington et al. (2021).\cite{KH2021}
We perform a similar analysis to Harrington et al. (2021),\cite{KH2021} focusing on a representative subset of data collected during November 2021, when CLASS observed at all four frequencies. CLASS observes by performing continuous 720$^{\circ}$ rotations in azimuth at constant elevation. These observations were broken into 3-hour segments, then corrected for glitches and other errors. Data segments from known poorly performing detectors and high glitch rates are excluded. We also excluded segments when the average wind speed at the CLASS site was above 5 m/s and the precipitable water vapor (PWV) was above 3 mm.\footnote[2]{PWV data were acquired from the APEX radiometer website, \url{https://archive.eso.org/wdb/wdb/asm/meteo_apex/form}} After pair differencing and demodulation, power spectral densities are calculated for the pair-differenced, demodulated-$U$, and demodulated-$V$ data. Once spectra have been produced for each data segment, a noise model is fit to each spectrum. This model, given in Equation \ref{eq:noiseModel}, consists of two components: a constant white noise and a power law rising to low frequencies (red noise).
\begin{equation}
\label{eq:noiseModel}
PSD(f) = w_n^2 \Bigg(1 + \Bigg( \frac{f}{f_k} \Bigg)^\alpha \Bigg)
\end{equation}
The model parameters are the power law slope $\alpha$, the white noise level $w_n$, and knee frequency $f_k$, which is the frequency at which the two components of the model are equal. This is demonstrated in Figure \ref{fig:fkExample}, which shows spectra of CLASS pair-differenced, demodulated-$U$, and demodulated-$V$ data from a single 3-hour segment with arbitrary scale and offset for visual clarity. The dashed lines in the $U$ spectrum represent the separate red and white noise components. The colored arrows indicate the knee frequency of each spectra. Using a CLASS azimuthal scan speed of $\omega=2^{\circ}$/s and elevation $\theta=45^{\circ}$, each knee frequency $f$ can be converted into an angular scale on the sky and a corresponding multipole $\ell \approx 360^{\circ} f/ \omega\sin{\theta} $. The multipoles are shown in the top axis of Figure \ref{fig:fkExample}.
\begin{figure}[ht]
\begin{center}
\includegraphics[height=7cm]{figures/spectra_w_arrows.png}
\end{center}
\caption[example]
{ \label{fig:fkExample}
Spectra of 40 GHz data from one 3-hour segment, arbitrarily scaled and offset for visual clarity. Spectra are modeled as white noise plus a power law (dashed lines for demodulated-$U$). The knee frequency is where these components are equal (colored arrows). Demodulation significantly lowers the knee frequency compared to pair differencing alone, enabling access to the largest scales (lowest $\ell$, top axis).}
\end{figure}
\section{Results}
\label{sec:results}
Data from all four observing bands were modeled to extract knee frequency values. Figure \ref{fig:dists} shows the distribution of knee frequencies for each band. The medians and 16$\%$ and 84$\%$ percentiles of each distribution are given in Table \ref{tab:fks}. For all four bands, the overall distribution of knee frequencies is significantly lower for demodulated compared to pair-differenced data. The median knee frequency decreases by approximately an order of magnitude between the pair-differenced and $U$ data. CLASS aims to constrain the CMB polarization down to multipoles $\ell < 10$. From Figures \ref{fig:fkExample} and \ref{fig:dists}, given the CLASS scan speed, this roughly corresponds to a knee frequency $f_k \lesssim 40\,\mathrm{mHz}$. The median knee frequency for all four CLASS bands is below 40 mHz for the $U$ data. The lower knee frequencies provided by demodulation significantly reduce the noise present in the CLASS data at low-$\ell$. The knee frequencies achieved by CLASS are necessary to measure the CMB on the largest angular scales. However, there are other systematic effects, such as scan-synchronous pick-up, present at low-$\ell$ that are the subject of separate studies.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=10cm]{figures/spie_distributions_w_ell.png}
\end{center}
\caption { \label{fig:dists}
Distribution of knee frequencies for all four CLASS observing bands, for pair-differenced (red), demodulated-$U$ (green), and demodulated-$V$ (yellow). Medians of each distribution as denoted by vertical dashed lines.}
\end{figure}
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
Band & \# segments & pair-diff (mHz) & demod-$U$ (mHz) & demod-$V$ (mHz)\\
\hline
40 GHz & 946 & 170$_{-78.9}^{+297}$ & 9.1$_{-5.4}^{+27}$ & 2.9$_{-2.0}^{+9.6}$ \\
\hline
90 GHz & 2429 & 503$_{-305}^{+752}$ & 29.1$_{-16.7}^{+41.9}$ & 7.1$_{-5.4}^{+12.1}$ \\
\hline
150 GHz & 3447 & 244$_{-134}^{+464}$ & 20.4$_{-13.3}^{+32.0}$ & 6.2$_{-4.7}^{+10.5}$ \\
\hline
220 GHz & 561 & 391$_{-243}^{+629}$ & 36.4$_{-22.1}^{+58.3}$ & 11.5$_{-7.1}^{+20.6}$ \\
\hline
\end{tabular}
\caption{\label{tab:fks} Median knee frequencies for each CLASS observing band and data set from November 2021. Error values give the 68-percentile range of the knee frequency distribution around the median. The second column gives the number of 3-hour data segments analyzed for each band.}
\end{table}
As is clear from Figure \ref{fig:dists} and Table \ref{tab:fks}, the $V$ knee frequencies are systematically lower than those for $U$. This implies that, compared to $U$, the $V$ data has a higher white noise level relative to the red noise component. This may be due to higher $V$ white noise in the presence of comparable red noise, or due to overall weaker red noise in the CLASS $V$ data. Sources of noise in the polarization data that are not modulated by the VPM have their amplitudes increased by correcting the signal and data for the VPM modulation efficiency.\cite{KH2018} White noise for the $V$ data is approximately double the $U$ white noise due to this correction.\cite{KH2021} Whether the higher $V$ white noise is responsible for the lower knee frequencies depends on whether the red noise is uncorrelated with the $V$ modulation function or if it represents on-sky signal.\footnote[2]{While CLASS has made the first detection of continuum circular polarization due to Zeeman splitting of oxygen in the atmosphere\cite{petroff2020, padilla2020}, we do not expect this component to have significant random fluctuations, and there are no other known sources of $V$ fluctuations on sky.}
\section{Conclusions}
\label{sec:conc}
CLASS observes the CMB polarization on the largest angular scales to constrain many cosmological phenomena such as inflation and reionization. Employing a VPM on each telescope provides CLASS with additional observational stability beyond that of a standard, pair-differenced-only polarization measurement. This is demonstrated here by the reduction in knee frequencies of demodulated CLASS data compared to pair differencing alone. For the data considered here, CLASS achieved median knee frequencies of 9.1, 29.1, 20.4, and 36.4 mHz for the $U$ data at 40, 90, 150, and 220 GHz, respectively. While red noise is not the sole factor in the final low-$\ell$ sensitivity of CLASS, these results provide a first look at the long-timescale stability of CLASS data across all four frequencies. On-going work will expand this analysis to include the whole multi-frequency data set.
\acknowledgments
We acknowledge the National Science Foundation Division of Astronomical Sciences for their support of CLASS under Grant Numbers 0959349, 1429236, 1636634, 1654494, 2034400, and 2109311. The CLASS project employs detector technology developed under several previous and ongoing NASA grants. Detector development work at JHU was funded by NASA cooperative agreement 80NSSC19M0005. Data analysis for CLASS is conducted using computational resources of the Advanced Research Computing at Hopkins (ARCH). We further acknowledge the very generous support of Jim and Heather Murren (JHU A$\&$S ’88), Matthew Polk (JHU A$\&$S Physics BS ’71), David Nicholson, and Michael Bloomberg (JHU Engineering ’64). R.R. is supported by the ANID BASAL projects ACE210002 and FB210003. Z.X. is supported by the Gordon and Betty Moore Foundation through grant GBMF5215 to the Massachusetts Institute of Technology. CLASS is located in the Parque Astronómico Atacama in northern
Chile under the auspices of the Agencia Nacional de Investigación y Desarrollo (ANID).
|
\section{Introduction}
The use of the mathematical concept of topology has proven very powerful to describe edge states in systems where the topological invariant inside the material differs from that of the outside~\cite{SSH_1979,Solitons,proof_1D,Haldane1988,QWZmodel,KaneQHSE2005,bulk_edge_XiaoLiang2006,TRSinvariant2007,Kane_Z2,ShouchengQSHE,TI3D_liang2007,TIindices_liang2007,QSHE3D_Roy2009,defect_Jeffrey2010,HasanRMP2010,EdgeDirac2011,topo_class,Exact_edge2018}. This concept has since been generalized to higher-order topological states following the work by Benalcazar, Bernevig, and Hughes (BBH). These include corner and hinge states, where a system in $d$-dimension can generate topological edge states in $(d-n)$-dimensions ($n>1$)~\cite{multipole_PRB_2017,multipole_science2017,Zhida2017,HOTI_science2018,HOTinversion2018,HOTISinversion2018,second_order2,HOTprinciple2019,Yafei2020,Changan2020,type2_2020,FermiArcs2020,Exact_localized2021}. In two-dimensions, a standard example of a system with higher order topological states is the BBH model, described by a simple Hamiltonian with only nearest-neighbor hopping terms. This has been realized in various different experimental setups, such as photonic\cite{photonic2019,photonic2020}, phononic~\cite{phononic2018}, acoustic\cite{Acoustic2020}, and microwave- and electrical-circuit systems\cite{microwave_circuit2018,circuit2018,circuit2019}. The Wannier band topology of the BBH model can be characterized by an off-diagonal quadrupole moment calculated through the nested-Wilson loop formalism~\cite{multipole_PRB_2017,multipole_science2017,Wilson2019}. In the original papers~\cite{multipole_science2017,multipole_PRB_2017}, the general role of the spatial symmetry in protecting the quadrupole moment was emphasized. However, recent mathematical studies of Toeplitz operators and $K$-theory have proven that there exist topologically protected corner states in the BBH model even without spatial symmetry as long as chiral symmetry is preserved \cite{Shin2019,Shin2021}. Also, by calculating the quadrupole moment using real space operators, it has become clear that chiral symmetry can also protect the quadrupole moment in a disordered BBH model \cite{2020DisEQI}.
In this paper, we address the following questions: firstly, can we calculate the quadrupole moment and characterize the Wannier band topology through the Wilson loop formalism in a more general chiral-symmetric model? Secondly, can we define a bulk quadrupole moment that does not require the rather complex nested-Wilson loop formalism and give a more physical interpretation to the quadrupole as defined by BBH?
The BBH model is built from one-dimensional (1D) Su-Schrieffer-Heeger (SSH) chains, with a suitable chosen cross-chain coupling.
Here, we study a generalized two-dimensional (2D) model by extending all 1D SSH chains of the BBH model to general two-band chiral-symmetric chains. We call this model the ``generalized BBH model'', and this has first been proposed in Refs.~\cite{SecondChiral2019,benalcazar2021chiral}. We calculate the Wannier centers and polarization by finding the exact expression for the Wilson loops along two orthogonal axes in the first Brillouin Zone (FBZ). Applying these results to the analysis of the Wannier bands, we show that their topology is indeed characterized by a $\mathbb{Z} \times\mathbb{Z}$ invariant and clarify the discontinuous displacement of the Wannier centers corresponding to the topological phase transition. Furthermore, we evaluate a simple expression for the quantized quadrupole moment through the nested-Wilson loop formalism. We find that the Wannier sector polarizations and quadrupole moment are determined by the fractional parts of the winding numbers of the constitutive chains in the corresponding direction. Thus, we conclude that chiral symmetry is the only symmetry required for the quadrupole moment of the generalized BBH model to be quantized.
We adapt the calculation of the electric dipole moment, which can be evaluated as the expectation value of the position operator in the Wannier functions of the central unit cell of the lattice \cite{RMPVanderbilt2012}, to define a 2D bulk quadrupole topological invariant. This quantity is related to the standard ``electric quadrupole moment". We will use this expression to characterize the number of corner modes and localized charges in generalized BBH model on a finite lattice \cite{defect_Jeffrey2010,multipole_PRB_2017,multipole_science2017,SecondChiral2019}. This idea is also discussed in Refs.~\cite{localizedWannier1997,Resta2011,RMPVanderbilt2012,EQM2020,EQM2021,Watanabe2019,Watanabe2020}, though not in the context of the same model. By calculating this novel expression for the bulk quadrupole of the generalized BBH model, we obtain the correct bulk-boundary correspondence and thus answer the second question posed above. This definition of the bulk quadrupole operator is gauge-dependent, which could mean we cannot draw a physical conclusion from this object \cite{Watanabe2019}. However, our numerical results lead to the conjecture that when one uses a common physical gauge for the bulk and the edge localized Wannier states, we can use this new topological invariant to analyze the corner charges.
\section{1D two-bands insulator with chiral symmetry}
\label{sec:1D}
We first review the model of a 1D two-band insulator with chiral symmetry in order to gain insight into the interplay between this symmetry and topology. The Bloch Hamiltonian of the 1D model can be written as
\begin{equation}\label{eq:1D_insulator_def}
h(k)=\vec{b}(k)\cdot\vec{\sigma}=b^1(k)\sigma_1+b^2(k)\sigma_2,
\end{equation}
where the vector $\vec{b}(k)$ will be called the ``characteristic vector" and $k\in[0,2\pi]$ is a 1D wavevector. The vanishing of the third component $b^3(k)$ is a consequence of chiral symmetry: $\{h(k),\sigma_3\}=0$. The eigen-energies of $h(k)$ are $\pm|\vec{b}(k)|$.
To define the electric polarization, one should solve for the eigenvalues of the position operator projected onto the occupied bands. This reduces to the calculation of the Wilson loop~\cite{multipole_PRB_2017,multipole_science2017}, as reviewed in the supplementary material (SM)~\cite{SM}. It can be proven that the electric polarization is determined by the Zak-Berry phase as an integral of the Berry connection of the occupied band $\vec{\mathcal{A}}(\vec{b})=i \matrixel{\psi_-}{\partial_{\vec{b}}}{\psi_-}$ in the characteristic vector space along a closed loop $\mathcal{C}:k\in[0,2\pi]\to\vec{b}(k)$ as
\begin{align} \label{eq:polarization_definition}
p=\frac{1}{2\pi}\int_{\mathcal{C}}\,d\vec{b}\cdot \vec{\mathcal{A}}(\vec{b})\, \text{mod}\,1=\frac{\mathcal{N}}{2}\, \text{mod}\,1.
\end{align}
The integer $\mathcal{N}$ is the winding number of $\mathcal{C}$ around the origin. Thus, the energy gap $2|\vec{b}(k)|$ must close when $\mathcal{N}$ changes. The topological phases of this 1D model correspond to the AIII topological class~\cite{topo_class}. We choose the eigenstates of Eq.~(\ref{eq:1D_insulator_def}) to be
\begin{equation}
\psi_\pm=(b^1-ib^2,\pm|\vec{b}|)^T/\sqrt{2|\vec{b}|^2}.\label{1d_gauge}
\end{equation}
In this gauge choice, the Zak phase in Eq.~(\ref{eq:polarization_definition}) gives the exact fractional winding number. The exact number of zero modes at each endpoint of a finite chain is also equal to this winding number $\mathcal{N}$~\cite{EdgeDirac2011,ShortTopological,proof_1D}. This is called the bulk-boundary correspondence~\cite{EdgeDirac2011,defect_Jeffrey2010,topo_class}, and hints that the chosen gauge is physically sufficient in characterizing the topological zero modes.
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{Fig1.pdf}
\caption{(a) A generalized BBH model with an additional long-range hopping in the $x$-direction. The dashed lines correspond to the negative hopping terms with minus signs. (b) Coupling of two BBH lattice layers with a stacking fault or AB stacking pattern.}\label{fig:2D_diagram1}
\end{figure}
\section{Generalized 2D Quadrupole insulator with Chiral Symmetry}
\label{sec:general}
In Refs.~\cite{multipole_PRB_2017,multipole_science2017} the authors propose a 2D crystal with vanishing bulk dipole moments and a non-zero quantized off-diagonal quadrupole moment. That model can be regarded as a network composed of perpendicular SSH chains arranged in a square lattice with a $\pi$ flux threading each square plaquette. The resulting spectrum consists of four bands, and the system is insulating at half-filling whenever the hopping parameters in the $x$ and $y$ directions are different. In our work, we generalize this model by replacing the SSH chains with 1D two-band insulators described in Sec.~\ref{sec:1D} above. The corresponding Bloch Hamiltonian is
\begin{align}
\mathcal{H}(k_x,k_y)=h_x(k_x)\otimes\tau_3 +\sigma_0\otimes h_y(k_y),\label{hBBH}
\end{align}
where $h_x(k_x)=\vec{b}_x(k_x)\cdot\vec{\sigma}$, $h_y(k_y)=\vec{b}_y(k_y)\cdot\vec{\tau}$. The four sub-lattice sites in one unit cell are denoted as $|\alpha_x,\beta_y\rangle$ ($\alpha_x,\beta_y=A,B$) and shown in Fig.~\ref{fig:2D_diagram1}(a). The Pauli matrix $\tau_3$ in the Hamiltonian $\mathcal{H}(\vec{k})$ reflects the presence of a $\pi$ flux. The characteristic vectors for each of the $h_\mu$ Hamiltonians ($\mu=x,y$) can be written as
\begin{align}
\vec{b}_\mu(k_\mu)&=|\vec{b}_\mu(k_\mu)|\big(\cos{\phi_\mu(k_\mu)},\sin{\phi_\mu(k_\mu)},0\big),
\end{align}
and the eigen-energies of $\mathcal{H}(\vec{k})$ are
\begin{equation}
E_\pm(\vec{k})=\pm\sqrt{|\vec{b}_x(k_x)|^2+|\vec{b}_y(k_y)|^2}.
\end{equation}
Both bands are two-fold degenerate due to chiral symmetry, {\it i.e.} $\{\mathcal{H}(\vec{k}),\sigma_3\otimes\tau_3\}=0$. The eigenstates of the two degenerate negative-energy bands can be expressed as tensor products of the form
\begin{align}
& \Psi_{1}(\vec{k})=\psi^x_{+}(k_x)\otimes\varphi^{1}_-(k_x,k_y),
\nonumber\\
& \Psi_{2}(\vec{k})=\psi^x_{-}(k_x)\otimes\varphi^{2}_-(k_x,k_y)
\,.\label{Psi12}
\end{align}
Here, $\psi^x_{\pm}(k_x)$ are the eigenstates of $h_x(k_x)$ with energies $\pm |\vec{b}_x(\vec{k})|$, respectively. Furthermore, $\varphi^{n}_-(\vec{k})$ ($n=1,2$) is the eigenstate with the negative energy of the two-by-two Hamiltonian
\begin{equation}
h_{xy}^n(\vec{k})=\vec{b}^n_{xy}(\vec{k})\cdot\vec{\tau}
\end{equation} whose characteristic vector is
\begin{equation}
\vec{b}^n_{xy}(\vec{k})=\left(b^1_{y}(k_y),b^2_{y}(k_y),(-1)^{n-1}|\vec{b}_{x}(k_x)|\right).
\end{equation}
For later discussion, we also define the polar angle of the characteristic vectors above as
\begin{align}
\theta_{xy}(\vec{k})=\arccos\left(|\vec{b}_x|/\sqrt{\vec{b}_{x}^2+\vec{b}_{y}^2}\right)\,.
\end{align}
In passing, we note that the generalized BBH model described by Eq.~(\ref{hBBH}) is identical to the ``separability-preserved" case proposed in Ref.~\cite{benalcazar2021chiral}.
For a 2D crystal, the bulk dipole moments in the directions $\mu=x,y$ are determined by the 2D Zak phases~\cite{RMP_Resta1994,multipole_science2017,multipole_PRB_2017} as
\begin{align} \label{eq:polarization_mu_definition}
p_\mu&=\frac{1}{2\pi}\int_{\text{BZ}}\,\Tr[\mathcal{A}_{\mu}(\vec{k})]\,\mathrm{d}^2\vec{k}\,\text{mod}\,1,
\end{align}
In the SM~\cite{SM} we show that, for the generalized BBH model, the dipole moments are equal to
\begin{align}
p_\mu&=\mathcal{N}_\mu\,\text{mod}\,1,
\end{align}
which are always trivial.
The integrals~(\ref{eq:polarization_mu_definition}) of the Berry connections $\mathcal{A}_\mu$ over the occupied bands are the 2D Zak Berry phases, and $\mathcal{N}_\mu$ are the winding numbers of the loops $k_\mu\to\vec{b}_\mu(k_\mu),k_\mu\in[0,2\pi]$.
An important part of our work is answering how to characterize the quadrupole moment of the 2D generalized quadrupole model with chiral symmetry. Using the definitions given above, we obtain an exact result for the Wilson loops along the $x$ and $y$ directions starting from the base point $\vec{k}=(k_x,k_y)$,
\begin{align}
\mathcal{W}_{x,\vec{k}}(k_y)&= \exp{i \int_{k_x}^{k_x+2\pi}d\phi_x(k'_x)\frac{s_0+\sin{\theta_{xy}(k'_x,k_y)}s_1}{2}},\\
\mathcal{W}_{y,\vec{k}}(k_x)&=\exp{i \int_{k_y}^{k_y+2\pi}d\phi_y(k'_y)\frac{s_0-\cos{\theta_{xy}(k_x,k'_y)s_3}}{2}}.
\end{align}
Here we define an additional set of Pauli matrices $\vec{s}=(s_1,s_2,s_3)$ where the two spinor components correspond to $\Psi_1(\vec{k})$ and $\Psi_2(\vec{k})$, respectively. We give the details of the derivation in the SM~\cite{SM}. The above expression allows us to solve the eigen equation
\begin{align}
\mathcal{W}_{\mu,\vec{k}}(k_{\bar{\mu}})\,v_j^\mu(\vec{k})=e^{i2\pi \nu^j_\mu(k_{\bar{\mu}})}\,v_j^\mu(\vec{k}),\label{wilson_loop}
\end{align}
where $j=\pm$ labels different solutions of this equation, $\bar{\mu}=y$ (${\bar \mu} = x$) when $\mu=x$ ($\mu = y$) and the quantities
\begin{align}
\nu^j_\mu(k_{\bar{\mu}})&=\int_0^{2\pi}\frac{dk_\mu}{{4\pi}}\frac{d\phi_\mu}{dk_\mu}\left[1+j\frac{|\vec{b}_{\bar{\mu}}(k_{\bar{\mu}})|}{\sqrt{\vec{b}_x^2(k_x)+\vec{b}^2_y(k_y)}}\right]\label{nu_u}
\end{align}
are called the Wannier centers. The eigenstates $v_j^\mu(\vec{k})$ with the corresponding Wannier center $\nu_\mu^j(k_{\bar{\mu}})$ generate the Wannier bands of the Wilson loop $\mathcal{W}_{\mu,\vec{k}}(k_{\bar{\mu}})$, not to be confused with the Wannier functions defined later in Sec.~\ref{sec:qinvarant}.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{Fig2.pdf}
\caption{(a) Discontinuous deformation of the Wannier gap $\Delta\nu_x$ when the winding number $\mathcal{N}_x$ is switched from 0 to by increasing the parameters $w_x$ from 0.99 to 1.01 in the model with fixed nearest-neighbor hoppings $u_x=u_y=1$ and $v_x=v_y=1.5$ as shown in Fig.~\ref{fig:2D_diagram1}. (b)-(c) correspond to the Wannier centers dependent on $k_y$ at the points $w_x=1 - 10^{-5},1 + 10^{-5}$, respectively.}\label{fig:transition}
\end{figure*}
The energy gap of the 1D insulator chain with chiral symmetry discussed in Sec.~\ref{sec:1D} closes when the winding number changes. But the generalized BBH model with $\pi$ flux will not necessarily have a gap closure. However, the Wannier gap defined as $\Delta\nu_\mu=\nu_\mu^+-\nu_\mu^-$ will close if the winding number $\mathcal{N}_{\bar{\mu}}$ changes because $\Delta\nu_\mu \propto|\vec{b}_{\bar{\mu}}|$ and the loop of $|\vec{b}_{\bar{\mu}}(k_y)|$ always touches the origin when the winding number changes. This change of $\mathcal{N}_{\bar{\mu}}$ also causes a discontinuous deformation of the curve $\Delta\nu_\mu(k_{\bar{\mu}})$ since the range of $\phi_\mu(k_\mu)$ suddenly changes to encompass the new winding number even when the Wannier gap $\Delta\nu_\mu(k_{\bar{\mu}})$ does not close. Since the transition from one topological class to another must cause a discontinuity in the Wannier bands, their topology can be described by the set of winding number $(\mathcal{N}_x,\mathcal{N}_y)$. These are defined in terms of two independent 2D Zak phases. Thus, the topological phase transition can be described by a $\mathbb{Z}\times \mathbb{Z}$ group. A transition in ${\cal N}_\mu$ is associated with a discontinuity in $\Delta\nu_\mu$, {\it i.e.} the Wannier gap in the same direction $\mu = x,y$. At the same time, the Wannier gap $\Delta\nu_{\bar \mu}$ in the orthogonal direction ${\bar \mu} = y, x$ closes. As the 1D winding numbers are topological invariants protected by chiral symmetry, we conclude that the topological phases of the Wannier bands are preserved by chiral symmetry \emph{even without spatial symmetry}.
In each direction, the Wannier bands are non-degenerate. This result corresponds to the classical case when two dipoles are oriented in the same direction $\mu$ and are separated in the orthogonal direction $\bar{\mu}$, but the total polarization in a unit cell is trivial ({\it i.e.}, it is an integer). Hence it is possible to define the quadrupole moment as in the standard BBH model through the polarizations within a sector of the Wannier bands~\cite{multipole_PRB_2017,multipole_science2017}. This polarization can be related to the 2D Zak-Berry phases of the Wannier basis, which is defined as
\begin{align}
w^\mu_j(\vec{k})=\big(\Psi_1(\vec{k}), \Psi_2(\vec{k})\big)\cdot v_j^\mu(\vec{k}),
\end{align}
where $\mu$, $j$ and $v_j^\mu(\vec{k})$ are defined after Eq.~(\ref{wilson_loop}) and in Eq.~(\ref{nu_u}). The polarizations can now be calculated from the formula given in Eq.~(\ref{eq:polarization_mu_definition}), with the trace of the Berry connection $\textrm{Tr}[\mathcal{A}_{\mu}(\vec{k})]$ replaced by $iw^{\bar \mu}_j{}^\dagger(\vec{k}) \partial_{k_\mu} w^{\bar \mu}_j(\vec{k})$.
For the generalized BBH model, we can choose $w^x_\pm(\vec{k})=\big(\Psi_1(\vec{k}) \pm \Psi_2(\vec{k})\big)/\sqrt{2}$ and $w^y_{\pm}(\vec{k})=\Psi_{1,2}(\vec{k})$ (see details in the SM~\cite{SM}). With these Wannier bases, we find that the Wannier-sector polarizations are linked to the winding numbers of the 1D chains in the $x$- and $y$-directions:
\begin{align}
p_{\mu}^{j}&=\frac{\mathcal{N}_\mu}{2}\,\text{mod}\,1.\label{sec_p}
\end{align}
We then obtain an exact expression for the off-diagonal quadrupole moment,
\begin{align}
q_{xy}=\sum_{j=\pm}p_x^{j}p_y^{j}=\frac{\mathcal{N}_x\mathcal{N}_y}{2}\,\text{mod}\,1.\label{qxy_eq}
\end{align}
The above results, especially Eq.~(\ref{sec_p}), clearly show that the Wannier-sector dipole moments $(p^{\pm}_x,p^{\pm}_y)$ can take the values in a $\mathbb{Z} _2\times\mathbb{Z} _2$ set. In turn, the quadrupole moment is quantized as $q_{xy}=0$ or $1/2$. Note that the quantization of $q_{xy}$ only requires the protection of chiral symmetry and does not need any additional spatial symmetry. This conclusion is consistent with the results from Ref.~\cite{2020DisEQI} where the authors use the real-space operator to understand the quadrupole moment and show that spatial symmetry is not necessary for the quantization of the quadrupole moment, even in a disordered system.
As an example of the generalized 2D quadrupole model with chiral symmetry beyond the standard BBH model, we introduce a long-range hopping term in the $x$ direction, as shown in Fig.~\ref{fig:2D_diagram1}(a). The corresponding characteristic vectors are
\begin{align}
\vec{b}_x(k_x)=(&u_x+v_x\cos{k_x}+w_x\cos{(2k_x)},\nonumber\\&v_x\sin{k_x}+w_x\sin{(2k_x)},0),\\ \vec{b}_y(k_y)=(&u_y+v_y\cos{k_y},v_y\sin{k_y},0),
\end{align}
where $u_\mu$, $v_\mu$ are the nearest-neighbor hopping terms, and $w_x$ is the long-range hopping along $x$ direction. This may seem unrealistic since the long-range hopping is usually small and can be neglected in real materials. A possible way to realize such a model is to couple two BBH layers in AB stacking (or a so-called ``stacking fault''~\cite{Partial2019}) pattern and regard the two layers as one. We show a schematic diagram for the two-layer stacking model in Fig.~\ref{fig:2D_diagram1}(b). Apart from the chiral symmetry discussed above, this model is reflection symmetric along the $x$ and $y$ axes.
When $w_x=0$, the model simplifies to the standard BBH model of Refs.~\cite{multipole_PRB_2017,multipole_science2017}. A non-trivial quadrupole moment $q_{xy}=1/2$ exists when $|u_\mu|<|v_\mu|$. With the formulas we have derived, we can characterize the Wannier-sector polarization by the winding number in the $x$-direction, $\mathcal{N}_x$. We find that~\cite{SM} when $|u_x+w_x|<|v_x|$, $\mathcal{N}_x=1$, while for $|u_x+w_x|>|v_x|$, $\mathcal{N}_x=\text{sign}[(w_x+u_x)(w_x-u_x)]+1$. The corresponding $\mathbb{Z}_2$-quantized Wannier-sector polarization $p_x^{\pm}$ is $1/2$ if $|u_x+w_x|<|v_x|$ and becomes $0$ when $|u_x+w_x|>|v_x|$. To be more concrete, we study the case $u_x=u_y=1$, $v_x=v_y=3/2$. As we let $w_x$ increase from 0 to 2, the winding number $\mathcal{N}_x$ switches from $1$ to $0$ at the point $w_x=1/2$ and later switches from $0$ to $2$ at the point $w_x=1$. Note that the Wannier-sector polarization remains trivial for $w_x>1/2$, even when $w_x$ exceeds $1$. As stated above, we expect a discontinuous behaviour of $\Delta\nu_x(k_y)$ around the critical point $w_x=1$ where the winding number $\mathcal{N}_x$ changes from 0 ($1/2<w_x<1$) to 2 ($w_x>1$). This is analyzed in more detail in Fig.~\ref{fig:transition}. In Fig.~\ref{fig:transition}(a) we show $\Delta\nu_x$ as a function of $w_x$ at $k_y=0,\pi$ to highlight the sharp change of the gap when $w_x$ increases from 0.99 to 1.01. Also, we show the corresponding Wannier centers $\nu_x^\pm(k_y)$ with parameters $w_x=1 - 10^{-5}$ and $w_x= 1 + 10^{-5}$ in Fig.~\ref{fig:transition}(b)-(c), respectively. Clearly, these figures provide substantial evidence that when $\mathcal{N}_{\mu}$ ($\mu=x,y$) changes, the Wannier gap $\Delta_{\mu}(k_{\bar{\mu}})$ ($\bar{\mu}=y,x$) jumps suddenly. These results are consistent with our previous analysis of the Wannier band topology.
\section{Corner charges and the bulk quadrupole invariant}
\label{sec:qinvarant}
We now consider a finite square lattice of the generalize BBH model with $N_x\times N_y$ unit cells and four sub-lattice sites $|\alpha_x,\beta_y\rangle$ ($\alpha_x,\beta_y=A,B$) per unit cell. For simplicity, we set $N_x$ and $N_y$ to be odd numbers and the lattice constant to be 1. The center of unit cells are located at $(x,y) = (m_x,m_y)$, where $m_x\in[-(N_x-1)/2,(N_x-1)/2]$ and $m_y\in[-(N_y-1)/2,(N_y-1)/2]$.
The Hamiltonian Eq.~(\ref{hBBH}) can be realised in real space as the lattice Hamiltonian $H=H_x\otimes I_{N_y}\tau_z+I_{N_x}\sigma_0\otimes H_y$, where $H_\mu$ is the corresponding lattice realization of $h_\mu$. For nontrivial topological phases, this model can exhibit zero-energy corner states at half filling (see Fig.~\ref{fig:2D_diagram2}).
In fact, in these non-trivial phases, all the corner states with zero energy are tensor products of the zero modes of $H_x$ and $H_y$.
Thus, the total number of corner zero-modes is $4\mathcal{N}_x\mathcal{N}_y$, and is exactly determined by the winding numbers of $H_x$ and $H_y$~\cite{SecondChiral2019,Shin2019,Shin2021}. As an example, we show the 48 eigenvalues near the zero energy of the generalized model in Fig.~\ref{fig:2D_diagram2}(a) for $u_x=u_y=1$, $v_x=v_y=3/2$ and $w_x=2$.
We find that the quadrupole moment $q_{xy}$ we obtained in Eq.~(\ref{qxy_eq}) as a $\mathbb{Z}_2 $ invariant does not correspond to the number of the corner states and the localized charge density in this case. We show the relative charge density $\Delta\rho$ distribution of the lattice with half-filled eigenstates in Fig.~\ref{fig:2D_diagram2}(b), where we filled the corner states with a negative eigenvalue ($-1$) of the chirality $\sigma_3\otimes\tau_3$. These states can be selected by adding a term $\delta \sigma_{3} \tau_{3}$ in the Hamiltonian and take the limit $\delta \rightarrow 0^-$, {\it i.e.} an infinitesimal negative number \cite{multipole_PRB_2017,multipole_science2017}. Here the relative charge density $\Delta\rho$ is defined as the difference between the density $\rho(\vec{r})$ and the average density $\bar{\rho}$ (which is 2 since every unit cell has two filled states). Note that we set the electron charge $\text{e}=1$ throughout our whole paper.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{Fig3.pdf}
\caption{(a): The energy spectrum of the long-range model with $41\times41$ unit cells and parameters $u_x=u_y=1$, $v_x=v_y=1.5$ and $w_x=2$ (see Fig. \ref{fig:2D_diagram1} (a) ). Here we show the minimum 48 energy values and see that 8 zero modes are generated. (b): The relative charge density distribution $\Delta\rho(\vec{r})=\rho(\vec{r})-\bar{\rho}$ for the half-filled square system with corner zero modes with eigenvalue $-1$ of the chiral operator. The localized charge can be observed around every kink. (c) Probability density distributions $\rho_W$ of the Wannier functions $W_{1\vec{0}}$ an $W_{2\vec{0}}$ of the central unit cell. The left one is for $W_{1\vec{0}}$ and the right one is for $W_{2\vec{0}}$.}\label{fig:2D_diagram2}
\end{figure}
So far, we have discussed the Wannier-band topology of the general quadrupole model by calculating the sector polarization based on the nested-Wilson loop formalism. In Ref.~\cite{benalcazar2021chiral}, a multipole winding operator is defined to determine the higher-order topological class for the generalized BBH model. In this paper, we would like to find a more straightforward definition of the bulk quadrupole moments for the same purpose. In analogy to the dipole moment, which can be simply calculated by taking the expectation of the position operator in the localized Wannier functions at the central unit cell, we find that the bulk quadrupole can be defined in an alternative way by evaluating the expectation value of the square position operator in the same functions~\cite{localizedWannier1997,Resta2011,RMPVanderbilt2012,EQM2020,EQM2021,Watanabe2019,Watanabe2020}. We introduce the alternative expression as follows
\begin{align}
\mathcal{N}_{\mu\nu}&=\int_{\text{all}} \mathrm{d}^2\vec{r} \,x_{\mu}x_{\nu}\, W_{n\vec{0}}^\ast(\vec{r}) W_{n\vec{0}}(\vec{r})\nonumber\\
&=\int_{\text{FBZ}} \frac{d^2\vec{k}}{(2\pi)^2}\sum_{n}^{N_\text{occ}}-\Psi^\dagger_{n}(\vec{k})\,\frac{\partial^2}{\partial k_\mu\partial k_\nu}\,\Psi_{n}(\vec{k}),\label{Nxy_k}
\end{align}
where the Wannier functions are defined as
\begin{equation}
W_{n\vec{R}}(\vec{r})=\sum_{k}e^{-i\vec{k}\cdot\vec{R}}\Psi_{n\vec{k}}(\vec{r})/(N_xN_y).
\end{equation}
The derivation of the above equation is given in SM~\cite{SM}. We call the Wannier functions $W_{1\vec{0}}(\vec{r})$ and $W_{2\vec{0}}(\vec{r})$ at the central unit cell with $\vec{R}=\vec{0}$ as the central Wannier functions over the bands $n=1,2$. These standard Wannier functions are not related to the Wannier bases we have constructed in Sec.~\ref{sec:general}. We show the density distributions of the Wannier functions $W_{1\vec{0}}(\vec{r})$ and $W_{2\vec{0}}(\vec{r})$ for the parameters $u_x=u_y=1$, $v_x=v_y=3/2$ and $w_x=2$ in Fig.~\ref{fig:2D_diagram2}(c). We clearly see these functions are localized as required.
Applying Eq.~(\ref{Nxy_k}) to the generalized BBH model, we obtain a quantized result for the off-diagonal element
\begin{align}
\mathcal{N}_{xy}&=\frac{\mathcal{N}_x\mathcal{N}_y}{2}.\label{eq:Nxy}
\end{align}
This winding number is protected by chiral symmetry and is unchanged under any continuous deformation of parameters that do not change the Wannier topology. Let us call this the quadrupole invariant to distinguish from the quadrupole moment $q_{xy}$. The result \eqref{eq:Nxy} is consistent with the quadrupole moment calculated through the Wannier-sector polarization of the BBH model~\cite{multipole_PRB_2017,multipole_science2017}.
We can then relate $\mathcal{N}_{xy}$ to $q_{xy}$ by
\begin{align}
q_{xy}=\mathcal{N}_{xy}\,\text{mod}\,1.
\end{align}
We know that the number of zero modes at every corner of a finite square lattice is $2\mathcal{N}_{xy}$ based on our previous analysis of the lattice Hamiltonian. The identification of $\mathcal{N}_x\mathcal{N}_y$ as a topological invariant for the generalized BBH model has been made before in Ref~\cite{SecondChiral2019}. What we have done here is to reveal the physical interpretation of that statement and its connection to the quadrupole moment. We also give expressions for the diagonal elements $\mathcal{N}_{xx}$ and $\mathcal{N}_{yy}$ of the quadrupole tensor in the SM~\cite{SM}. These are not quantized and will change continuously as the system's parameters are tweaked.
\section{Gauge issues of the quadrupole invariant}
The quadrupole invariant defined in Eq.~(\ref{Nxy_k}) is a gauge-dependent quantity. Fixing the gauge by choosing the eigenstates $\Psi_n(\vec{k})$ in Eq.~(\ref{Psi12}) yields the exact winding numbers for the 2D Zak phases \cite{SM}. However, a further $\vec{k}$-dependent gauge transformation can break the quantization of this quadrupole invariant, as was stated in Ref.~\cite{Watanabe2019}. The gauge-fixing applied must have a physical basis since it does capture the Wannier topology and boundary properties, as we have shown. We have no analytical insight into this issue at the moment, but the following discussion provides plausible arguments for the validity of our choice.
First of all, the bulk Wannier states of a full periodic system must be chosen to have a common gauge with the edge states of a ribbon in the periodic direction \cite{Quadrupole_Wannier2021}. This can be justified by the fact that in a ribbon system, the bulk-like extended states deep in the interior region are not modified in the periodic direction. To make the gauge of bulk and edge states consistent, we require the projection of the bulk central Wannier functions onto the edges to give the same probability density as the central Wannier functions of the edge states in a ribbon that is periodic in the $x$ or $y$ direction. We call this choice the ``edge-consistent gauge''. For the generalized BBH model, we use the gauge in Eq.~ (\ref{1d_gauge}) to describe the 1D edge states in the ribbons and the gauge in Eq.~(\ref{Psi12}) for the 2D bulk states. We have checked that this gauge choice is edge-consistent by numerical calculation, as shown in the SM~\cite{SM}. Since the chosen gauge of the edge states gives the correct number of the corner states discussed in Sec.~\ref{sec:qinvarant}, the consistent gauge of the bulk states should also be able to capture this.
Secondly, in several previous studies of different types of BBH models~\cite{multipole_PRB_2017,multipole_science2017,anomalousHOTs2018,type2_2020,Changan2020}, the fractional charge at the corners $Q_\text{c}$ has been constructed as the summation of the corresponding edge polarizations and the bulk quadrupole moment $Q_\text{c}=p^{\text{edge}}_x+p^{\text{edge}}_y-q_{xy}$. This formula, however, has been given without specifying a common gauge between the Wannier functions of the periodic and ribbon systems, as pointed out in Ref.~\cite{Quadrupole_Wannier2021}. In the SM \cite{SM}, we calculate $Q_\text{c}$ in the edge-consistent gauge we use. We find that
\begin{equation}
Q_\text{c}=p^{\text{edge}}_x=p^{\text{edge}}_y=q_{xy}=\frac{\mathcal{N}_{x}\mathcal{N}_{y}}{2}\,\text{mod}\,1,
\end{equation}
{\it i.e.} all quantities are half-quantized in the topological phase. This result agrees with the BBH one. In Ref.~\cite{multipole_PRB_2017,multipole_science2017,Changan2020,benalcazar2021chiral}, it is shown that
in that case $p^{\text{edge}}_x=p^{\text{edge}}_y=Q_c=q_{xy}=1/2$ in the topological phase.
Thus, it seems that gauge fixing has at least been implicitly assumed there.
Finally, we observe that the gauge of the 1D edge states can only be changed in one direction by a phase factor $U_\mu=\exp{[-i\Theta_\mu(k_\mu)]}$ ($\mu=x,y$). Hence, the corresponding gauge transformation of 2D bulk states must be the multiplication of these transformations, such that $U=U_xU_y$. By using the gauge-invariant quantum metric tensor, we can prove that the quadrupole invariant remains quantized and that its fractional part is unchanged under the action of (factorized) gauge transformations such is $U$ \cite{SM}.
\section{Conclusion}
In conclusion, we have provided exact expressions of the Wannier-sector polarizations and quadrupole moment of the generalized BBH model through the Wilson loop formalism and revealed the topological $\mathbb{Z} \times\mathbb{Z} $ class of the Wannier bands protected by chiral symmetry. Furthermore, we have provided an alternative expression of the quadrupole tensor and found a quadrupole invariant that is quantized if an edge-consistent gauge is assumed. This has enabled us to give a bulk interpretation of the quadrupole winding number proposed in a recent paper~\cite{benalcazar2021chiral}. Quantization of this quadrupole invariant in more general situations and how to resolve the gauge redundancy of the quadrupole invariant will be addressed in future studies, though we have provided arguments of plausibility for our choice. Since the quadrupole moment is related to the quantum metric tensor~\cite{RMP_Resta1994,geometry_metric2000,Resta2011}, which has attracted attention in recent studies~\cite{GMT2018,QuantumMetric2019,EQM2020,EQM2021}, the relation between the higher-order topology and the underlying quantum geometry also requires further investigation.
The BBH model has been intensely investigated in recent years in various experimental setups~\cite{photonic2019,photonic2020,phononic2018,Acoustic2020,microwave_circuit2018,circuit2018,circuit2019}. Therefore, similar experimental methods could be used to realize the generalized quadrupole model we have discussed, which presents richer second-order topological phases, by coupling two BBH layers in a stacking fault pattern with an inter-layer hopping term.
\acknowledgements
L.Y. acknowledges funding through the China Scholarship Council under Grant 201906230305; A.P. acknowledges support from the European Commission under the EU Horizon 2020 MSCA-RISE-2019 programme (project 873028 HYDROTRONICS) and of the Leverhulme Trust under the grant RPG-2019-363. N.R.W. is supported by STFC grant ST/P004423/1. L.Y. thanks Aleksandr Kazantsev, Feng Liu, and Zhiguo Lv for valuable discussions.
\section{Bulk dipole momentum in 1D crystal}\label{sec.1Ddipole}
In second quantization, a generic 1D Hamiltonian with periodic boundary conditions can be written as
\begin{align}
\hat{H}=\sum_{k} c_{k, \alpha}^{\dagger}\left[h_{k}\right]_{\alpha, \beta} c_{k, \beta},
\end{align}
where $c_{k,\alpha}^\dagger$ is the electron creation operator for a state with momentum $k$ and orbital label $\alpha$. The time-independent Schr\"odinger equation can be written as
\begin{align}
h(k)u_n(k)=E_n(k)u_n(k).
\end{align}
The dipole moment of a 1D crystal can be calculated through the Wilson loop, which can be constructed through the multiplication of the exponents of the non-Abelian Berry connection
\begin{align}
\mathcal{W}_{k}=\lim_{\Delta_k\to 0}F_{2\pi-\Delta_k}F_{2\pi-2\Delta_k}\ldots F_{k+\Delta_k}F_{k},\label{Wilson_loop}
\end{align}
where $[F_k]_{n_1n_2}=u^\dagger_{n_1}(k+\Delta_k)u_{n_2}(k)$. For small $\Delta k$, $F_k$ can be approximated as
\begin{align}
F_k&\approx e^{i \Delta_k\mathcal{A}(k)},\\
[\mathcal{A}(k)]_{n_1n_2}&=i u^\dagger_{n_1}(k)\partial_ku_{n_1}(k).
\end{align}
We then solve for the eigenstates of the Wilson loop,
\begin{align}
\mathcal{W}_{k}v_j(k)=\phi^j v_j(k)\quad,\forall\,k.
\end{align}
Since the Wilson loop is unitary, its eigenvalues can be expressed as a phase
\begin{align}
\phi^j=e^{i 2\pi\nu^j}.
\end{align}
In fact, $\nu^j$ is the Wannier center relative to the middle of the unit cells. In general, we call the eigenstates $v_j$ with its Wannier center $\nu^j$ as the Wannier bands.
The corresponding classical polarization is defined as
\begin{align}
p&=\sum_j(\nu^j\bmod 1)\nonumber\\
&=\frac{1}{2\pi i }\ln\det[\mathcal{W}_{k+2\pi\gets k}]\nonumber\\
&=\frac{1}{2\pi }\int_k^{k+2\pi}\Tr[\mathcal{A}(k)]\,\mathrm{d} k\bmod 1.
\end{align}
Here $\int_k^{k+2\pi}\Tr[\mathcal{A}_k]\mathrm{d} k$ is the Berry phase of all occupied bands. Note that the Berry phase may not be trivial when the polarization is trivial. The latter usually captures the number of topological edge states of the finite system, and the bulk-boundary correspondence \cite{defect_Jeffrey2010}.
We will use the following relation as a gauge-fixing condition for the eigenfunctions of the Wilson loop
\begin{align}
v_j(k+\Delta_k)&=e^{-i \Delta_k\nu^j}F_kv_j(k).\label{wannier_gauge}
\end{align}
Even if we guarantee the fixed gauge for the Wannier bands, there are still different choices of the sub-lattice degree of freedom in the occupied bands. In general, we can mix the eigenstates in the occupied subspace as follow
\begin{align}
|u'_{m}(k)\rangle=\sum_{n=1}^{N_{\text{occ}}}|u_n(k)\rangle[U_k]^{nm}.
\end{align}
This transformation makes the connection change as
\begin{align}
\mathcal{A}_k^{'}=iU^\dag_k\partial_kU_k-U^\dag_k\mathcal{A}_kU_k.
\end{align}
This gives a new polarization as
\begin{align}
p^{'}&=p+\frac{i }{2\pi}\text{Tr[ln} U_k]\bigg|_{k}^{k+2\pi}\nonumber\\
&=p+\frac{i }{2\pi}\text{ln[det} U_k]\bigg|_{k}^{k+2\pi}\nonumber\\
&=p+n.\label{1D_p_gauge}
\end{align}
In the above derivation, we apply the unitary property det$[U_k]=e^{-i\phi(k)}$ and $\phi(k+2\pi)=\phi(k)+2\pi n$ derived from periodic condition of $U_k$.
\section{Bulk Dipole moment and Zak-Berry phase in 2D crystal}
\label{sec:2D_dipole}
\subsection{ formalism}
In two dimensions, we can extend the Wilson loop $\mathcal{W}_{x,\vec{k}}$ that runs along $k_x$ axis as a loop at a fixed value of $k_y$ in the 2D FBZ. With this extension, we have
\begin{align}
\mathcal{W}_{x,\vec{k}}v_j^x(\vec{k})=\phi^j_xv_j^x(\vec{k}),
\end{align}
where
\begin{align}
\phi^j_{x}(k_y)=e^{i 2\pi\nu_x^j(k_y)}.
\end{align}
Then, the 1D polarization at fixed $k_y$ can be written as
\begin{align}
p_x(k_y)=\sum_{j=1}^{N_{\text{occ}}}\nu_x^j(k_y)\bmod 1=-\frac{i }{2\pi}\ln\,\det[\mathcal{W}_{x,\vec{k}}].
\end{align}
Thus the total polarization along $x$ is
\begin{align}
p_x&=\frac{\Delta k_y}{2\pi}\sum_{k_y}p_x(k_y)=\frac{1}{2\pi}\Omega^{2D}_y \bmod 1,\\
\Omega^{2D}_x&=\frac{1}{2\pi}\int_{\text{BZ}}\,\Tr[\mathcal{A}_{x}(\vec{k})]\,\mathrm{d}^2\vec{k}.
\end{align}
By the same method, the total polarization along $y$ is
\begin{align}
p_y&=\frac{\Delta k_x}{2\pi}\sum_{k_x}p_y(k_x)=\frac{1}{2\pi}\Omega^{2D}_y\bmod 1,\\
\Omega^{2D}_y&=\frac{1}{2\pi}\int_{\text{BZ}}\,\Tr[\mathcal{A}_{y}(\vec{k})]\,\mathrm{d}^2\vec{k}.
\end{align}
The functions $\Omega_\mu^{2D}$ ($\mu=x,y$) are called the 2D Zak-Berry phases\cite{RMP_Resta1994,2DSSH2017}.
\subsection{Polarization in the generalized BBH model}
We study a two-by-two Bloch Hamiltonian of the form
\begin{align}
h=\vec{b}\cdot\vec{\sigma},
\end{align}
where $\vec{\sigma}$ is the vector of Pauli matrices $\vec{\sigma}=(s_1,s_2,s_3)$. Using spherical coordinates, we express the characteristic vector $\vec b$ as $\vec{b}=|\vec{b}|(\sin{\theta}\cos{\phi},\sin{\theta}\sin{\phi},\cos{\theta})$. Then the two eigenstates with eigenenergies $\pm|\vec{b}|$ can be written as
\begin{align}
\psi_+=\left(\begin{array}{c}
{ e^{-i \phi}\cos{\frac{\theta}{2}}} \\
{ \sin{\frac{\theta}{2}}}
\end{array}\right),\quad\psi_-=\left(\begin{array}{c}
{ e^{-i \phi}\sin{\frac{\theta}{2}}} \\
{ -\cos{\frac{\theta}{2}}}
\end{array}\right).\label{TwoBand}
\end{align}
The Bloch Hamiltonian of the generalized BBH model as considered in the main text is
\begin{align}
\mathcal{H}(k_x,k_y)=h_x(k_x)\otimes\tau_3 +\sigma_0\otimes h_y(k_y),\label{hBBH}
\end{align}
where $h_x(k_x)=\vec{b}_x(k_x)\cdot\vec{\sigma}$, $h_y(k_y)=\vec{b}_y(k_y)\cdot\vec{\tau}$. The four sub-lattice sites in one unit cell are denoted as $\ket{\alpha_x\,\beta_y}$ ($\alpha,\beta=A,B$). The eigenenergies of $\mathcal{H}(\vec{k})$ are $E_\pm(\vec{k})=\pm\sqrt{|\vec{b}_x(k_x)|^2+|\vec{b}_y(k_x)|^2}$. Both the negative and positive energy bands are two-fold degenerate due to the chiral symmetry
\begin{align}
\{\mathcal{H}(\vec{k}),\sigma_3\otimes\tau_3\}=0.
\end{align}
For the two degenerate occupied bands the orthogonal eigenstates are
\begin{align}
\Psi_{n}(k_x,k_y)&=\psi^x_{\sigma(n)}(k_x)\otimes\varphi^{n}_-(k_x,k_y).
\end{align}
Here $n=1,2$ while $\sigma(n)=+,-$, $\psi^x_{\sigma(n)}(k_x)$ are the eigenstates of $h_x(k_x)$ with eigen-energies $\sigma(n) |\vec{b}_x(\vec{k})|$ and $\varphi^{n}_-(\vec{k})$ are the eigenstates with negative energies of the two-by-two Hamiltonians $h_{xy}^n(\vec{k})=\vec{b}^n_{xy}(\vec{k})\cdot\vec{\tau}$ whose characteristic vectors are $\vec{b}^n_{xy}(\vec{k})=(b^1_{y}(k_y),b^2_{y}(k_y),(-1)^{n-1}|\vec{b}_{x}(k_x)|)$, respectively. We then express the eigenstates $\psi^x_\pm(k_x)$ of $h_x(k_x)$ and $\varphi^{1/2}_\pm(k_x)$ of $h^{1/2}_{xy}(\vec{k})$ in spherical coordinates,
\begin{align}
\psi^x_\pm(k_x)&=\frac{1}{\sqrt{2}}\left(\begin{array}{c}
{ e^{-i \phi_x(k_x)}} \\
{ \pm 1}
\end{array}\right),\\
\varphi^1_-(k_x,k_y)&=\left(\begin{array}{c}
{ e^{-i \phi_y(k_y)}\sin{\frac{\theta_{xy}(\vec{k})}{2}}} \\
{ - \cos{\frac{\theta_{xy}(\vec{k})}{2}}}
\end{array}\right),\\
\varphi^2_-(k_x,k_y)&=\left(\begin{array}{c}
{ e^{-i \phi_y(k_y)}\cos{\frac{\theta_{xy}(\vec{k})}{2}}} \\
{ - \sin{\frac{\theta_{xy}(\vec{k})}{2}}}
\end{array}\right).
\end{align}
The angles in the above expressions are defined through the polar decomposition of the characteristic vectors $\vec{b}_x$ and $\vec{b}^{1}_{xy}=\vec{b}_y+(0,0,|\vec{b}_x|)$:
\begin{align}
\vec{b}_x(k_x)&=|\vec{b}_x(k_x)|\left(\cos{\phi_x(k_x)},\sin{\phi_x(k_x)},0\right),\\
\vec{b}_{y}(k_y)&=|\vec{b}_y(k_y)|\left(\cos{\phi_y(k_y)},\sin{\phi_y(k_y)},0\right),\\
\vec{b}^{1}_{xy}(k_x,k_y)&=(|\vec{b}_y(k_y)|\cos{\phi_y(k_y)},|\vec{b}_y(k_y)|\sin{\phi_y(k_y)},|\vec{b}_x(k_x)|)\nonumber\\
&=\sqrt{ \vec{b}_x^2+\vec{b}_y^2}\,\left(\sin{\theta_{xy}(\vec{k})}\cos{\phi_y(k_y)},\sin{\theta_{xy}(\vec{k})}\sin{\phi_y(k_y)},\cos{\theta_{xy}(\vec{k})}\right).
\end{align}
Note that $\phi_\mu$ only depends on $k_\mu$ while $\theta_{xy}(\vec{k})$ is determined by both $k_x$ and $k_y$. More precisely,
\begin{align}
\theta_{xy}(k_x,k_y)=\arccos{(|\vec{b}_x(k_x)|/\sqrt{ \vec{b}_x^2(k_x)+\vec{b}_y^2(k_y)})}.
\end{align}
With the gauge choice Eq.~(\ref{TwoBand}), the two eigenstates $\Psi_n(\vec{k})$ can be transformed into each other by
\begin{align}
\Psi_2(k_x,k_y)&=-e^{-i\phi_y}\sigma_3\otimes\tau_1\mathcal{K}\Psi_1(k_x,k_y).
\end{align}
To obtain the bulk dipole moment of the generalized quadrupole model with a $\pi$ flux threaded through, we should first calculate the 2D Zak phase.
The diagonal entries of the non-Abelian Berry connections can be written as
\begin{align}
[\mathcal{A}_{x}(\vec{k})]_{nn}&=i \Psi^\dagger_{n}(\vec{k})\partial_{k_x}\Psi_{n}(\vec{k})=i \psi_-^{x\dagger}(k_x)\partial_{k_x}\psi^x_-(k_x)+i \varphi_-^{n\dagger}(\vec{k})\partial_{k_x}\varphi^{n}_-(\vec{k}),\\
[\mathcal{A}_{y}(\vec{k})]_{nn}&=i \Psi_{n}^\dagger(\vec{k})\partial_{k_y}\Psi_{n}(\vec{k})=i \varphi_-^{n\dagger}(\vec{k})\partial_{k_y}\varphi^{n}_-(\vec{k}).
\end{align}
Note that the two terms $i \psi^{x\dagger}_\pm(k_x)\partial_{k_x}\psi^x_\pm(k_x)$ are actually the same under the gauge choice $\psi^x_+(k_x)=\sigma_3\psi^x_-(k_x)$, which shows the chiral symmetry. Applying
\begin{align}
\varphi^{2}_-(\vec{k})=-e^{-i\phi_y}\tau_1\mathcal{K}\varphi^{1}_-(\vec{k}),
\end{align}
we have
\begin{align}
\frac{\mathrm{d} \phi_y}{\mathrm{d} k_y}-i\varphi_-^{1\dagger}(\vec{k})\partial_{k_y}\varphi^1_-(\vec{k})=i \varphi_-^{2\dagger}(\vec{k})\partial_{k_y}\varphi^2_-(\vec{k}).
\end{align}
With the above formulas, we can obtain the 2D Zak-phase
\begin{align}
\Omega^{2D}_x&=\frac{1}{2\pi}\int_{\text{BZ}}\,\Tr[\mathcal{A}_{x}(\vec{k})]\,\mathrm{d}^2\vec{k}=\mathcal{N}_x,\\
\Omega^{2D}_y&=\frac{1}{2\pi}\int_{\text{BZ}}\,\Tr[\mathcal{A}_{y}(\vec{k})]\,\mathrm{d}^2\vec{k}=\mathcal{N}_y,
\end{align}
and thus the dipole moments along $x$ and $y$ directions can be expressed as
\begin{align}
p_x&=\frac{\Omega^{2D}_x}{2\pi}\bmod 1=\mathcal{N}_x\bmod 1=0,\\
p_y&=\frac{\Omega^{2D}_y}{2\pi}\bmod 1=\mathcal{N}_y\bmod 1=0.
\end{align}
\section{Wannier-sector Polarization and Bulk Quadrupole Moment In 2D}
\subsection{Nested Wilson loop formalism}
To be well-defined, the topological quadrupole defined in Ref.~\cite{multipole_PRB_2017,multipole_science2017} applies only for a system with zero bulk Chern number, a vanishing dipole moment, and a gap in the Wannier-center bands. The last requirement originates from a quadrupole being constructed from two separated dipoles. For the generalized BBH model, another crucial point is that the Wannier bands in the $x$ and $y$ direction are both gapped, which hints at a separation of the two dipoles along the direction perpendicular to their aligned direction and thus yields the non-vanishing quadrupole.
We denote the two Wannier bands $\nu_x^j(k_y)$ for the Wilson loop along $k_x$ line from the base point $\vec{k}$ as $\nu_x^-$ and $\nu_x^+$. Then we have the following projected-position eigenstates on the occupied bands with momentum $k_y$ as
\begin{align}
\ket{\Psi^j_{R_x,k_y}}&=\sqrt{\frac{\Delta k_x}{2\pi}}\sum_{n=1}^{N_\text{occ}}\sum_{k_x}[v_j^x(\vec{k})]_ne^{-i k_xR_x}|u_n(\vec{k})\rangle,
\end{align}
where
\begin{align}
&P_{\text{occ}}g_x P_{\text{occ}}\ket{\Psi^j_{R_x,k_y}}=e^{i \Delta_{k_x}(\nu^j(k_y)+R_x)}\ket{\Psi^j_{R_x,k_y}},\\
g_x&=e^{i \Delta_{k_x}\mathtt{x}}=\sum_{\vec{k}, \alpha} \ket{\vec{k}+\vec{\Delta}_{k_x},\alpha}\bra{\vec{k}, \alpha},\\
P_{\text{occ}}&=\sum_{n=1}^{N_{\text{occ}}}\ket{u_n(\vec{k})}\bra{u_n(\vec{k})},
\end{align}
where $N_{\text{occ}}$ is the number of the occupied bands. Note that we use the Dirac brackets to denote the state vectors and the symbols without brackets to denote the their matrices representations.
For every sector of the Wannier bands at $k_y$, there is polarization along the $y$ direction, contributing to the quadrupole moment. Thus, to define the quadrupole moment, we project the position operator $\hat{y}$ onto the Wannier sector $\omega_x^s$, $s=\pm$. The projected position operator $\hat{y}$ can be expressed as follows
\begin{align}
P_{\omega_x^s}g_yP_{\omega_x^s}=P_{\omega_x^s}\sum_{\vec{k}, \alpha} \ket{\vec{k}+\vec{\Delta}_{k_y},\alpha}\bra{\vec{k}, \alpha}P_{\omega_x^s},
\end{align}
where the projector for the sector below or above the Wannier gap is
\begin{align}
P_{\omega_x^s}&=\sum_{j\in\omega_x^s}\sum_{n_1,n_2=1}^{N_\text{occ}}\sum_{\vec{k}}\ket{u_{n_2}(\vec{k})}\bra{ u_{n_1}(\vec{k})}[v_j^{x}(\vec{k})]_{n_2}[v_j^{x}(\vec{k})]_{n_1}^\ast.\label{sector_projection}
\end{align}
To simplify the notation, we will now use a special set of basis vectors called the Wannier-band basis,
\begin{align}
\ket{w_j^x(\vec{k})}&=\sum_n^{N_\text{occ}} \ket{u_n(\vec{k}) }[v_j^x(\vec{k})]_n.
\end{align}
Therefore, the projected position operator reduces to
\begin{align}
P_{\omega_x^s}g_yP_{\omega_x^s}=\sum_{j_1,j_2\in\omega_x^s}\sum_{\vec{k}}\ket{w^{x}_{j_1}(\vec{k}+\vec{\Delta}_{k_y})}\bra{ w^{x}_{j_2}(\vec{k})}(w_{j_1}^{x\dagger}(\vec{k}+\vec{\Delta}_{k_y})w_{j_2}^{x}(\vec{k})).
\end{align}
The eigenvalues of the sector-position operator $P_{\nu_x^s}g_yP_{\nu_x^s}$ give us the Wannier centers for the selected Wannier bands, we can diagonalize it by calculating the Wilson loop as
\begin{align}
\mathcal{W}^{\omega_x^{s}}_{y,\vec{k}}=\lim_{\Delta_{k_y}\to 0}F^{\omega_x^{s}}_{y,\vec{k}+(N_y-1)\vec{\Delta}_{k_y}}F^{\omega_x^{s}}_{y,\vec{k}+(N_y-2)\vec{\Delta}_{k_y}}...F^{\omega_x^{s}}_{y,\vec{k}+\vec{\Delta}_{k_y}}F^{\omega_x^{s}}_{y,\vec{k}},
\end{align}
where $[F^{\omega_x^{s}}_{y,\vec{k}}]_{j_1j_2}= w_{j_1}^{x\dagger}(\vec{k}+\vec{\Delta}_{k_y})w_{j_2}^{x}(\vec{k})$, $j_1,j_2\in\omega_x^s$.
The repeated indices imply the summation over all Wannier bands of the selected sector in the above definition. Usually, this Wilson loop constructed from the Wannier-band basis is called the ``nested Wilson loop". Since the nested-Wilson loop is unitary, the eigenvalues of $\mathcal{W}^{\omega_x{s}}_{k_y}$ are again phases:
\begin{align}
\mathcal{W}^{\omega_x^s}_{y,\vec{k}} \,v^{y}_{\omega_x^s,m}(\vec{k})=e^{i 2\pi\nu_y^{\omega_x^s,m}(k_x)}\, v^{y}_{\omega_x^s,m}(\vec{k}).
\end{align}
Finally, we can introduce the Wannier-sector polarization at $k_x$ as the summation of the Wannier-sector centers $\nu_y^{x,j}(k_x)$,
\begin{align}
p^{s}_y(k_x)=\sum_{m}\nu_y^{\omega_x^s,m}(k_x)\bmod 1.
\end{align}
The total polarization in the Wannier sector is
\begin{align}
p^{s}_y=\frac{\Delta k_x}{2\pi}\sum_{k_x}p^{s}_y(k_x)=\frac{1}{(2\pi)^2}\int_{\text{BZ}}\,\Tr[\mathcal{A}^{xs}_{y}(\vec{k})]\,\mathrm{d}^2\vec{k}\bmod 1,
\end{align}
where $[\mathcal{A}^{xs}_{y}(\vec{k})]$ is the non-Abelian Berry connection for the Wannier bases of the Wannier sector and the trace $\Tr$ denotes the summation over all Wannier bands in the sector $\omega_x^s$. By the same derivation, we can have the sector polarization in $x$ direction as
\begin{align}
p^{s}_x=\frac{\Delta k_y}{2\pi}\sum_{k_y}p^{s}_x(k_y)=\frac{1}{(2\pi)^2}\int_{\text{BZ}}\,\Tr[\mathcal{A}^{ys}_{x}(\vec{k})]\,\mathrm{d}^2\vec{k}\bmod 1.
\end{align}
The definiton of the quadrupole moment is
\begin{align}
q_{xy}=\sum_{s}p^{s}_xp^{s}_y.
\end{align}
\subsection{Multipole moments of the generalized BBH model}
Here we give the exact calculation of the quadrupole moment of the 2D generalized BBH model. Applying $\Psi_1(k_x,k_y)=\psi^x_+(k_x)\otimes\varphi^1_-(k_x,k_y)$ and $\Psi_2(k_x,k_y)=\psi^x_-(k_x)\otimes\varphi^2_-(k_x,k_y)$ in the expression for the Berry connection, we have
\begin{align}
\mathcal{A}_x(\vec{k})&=\left(\begin{array}{cc}
{i \matrixel{\psi_-^x(k_x)}{\partial_{k_x}}{\psi_-^x(k_x)}+i \matrixel{\varphi_-^1(\vec{k})}{\partial_{k_x}}{\varphi_-^1(\vec{k})}}&{i \matrixel{\psi_+^x(k_x)}{\partial_{k_x}}{\psi_-^x(k_x)}\,\varphi^1_-(\vec{k})^\dagger\varphi_-^2(\vec{k})} \\
{i \matrixel{\psi_-^x(k_x)}{\partial_{k_x}}{\psi_+^x(k_x)}\,\varphi^2_-(\vec{k})^\dagger\varphi_-^1(\vec{k})}&{i \matrixel{\psi_+^x(k_x)}{\partial_{k_x}}{\psi_+^x(k_x)}+i \matrixel{\varphi_-^2(\vec{k})}{\partial_{k_x}}{\varphi_-^2(\vec{k})}}
\end{array}\right),\\
\mathcal{A}_y(\vec{k})&=\left(\begin{array}{cc}
{i \matrixel{\varphi_-^1(\vec{k})}{\partial_{k_y}}{\varphi_-^1(\vec{k})}}&{0} \\
{0}&{i \matrixel{\varphi_-^2(\vec{k})}{\partial_{k_y}}{\varphi_-^2(\vec{k})}}
\end{array}\right).
\end{align}
Next we express all the elements of the Berry connections in spherical coordinate. The derivation relies on the following relations
\begin{align}
i \matrixel{\psi_\pm^x(k_x)}{\partial_{k_x}}{\psi_\pm^x(k_x)}&=i \matrixel{\psi_\pm^x(k_x)}{\partial_{k_x}}{\psi_\mp^x(k_x)}=\frac{1}{2}\frac{\mathrm{d} \phi_x}{\mathrm{d} k_x},\\
i \matrixel{\varphi_-^1(\vec{k})}{\partial_{k_x}}{\varphi_-^1(\vec{k})}&=i \matrixel{\varphi_-^2(\vec{k})}{\partial_{k_x}}{\varphi_-^2(\vec{k})}=0,\\
i \matrixel{\varphi_-^1(\vec{k})}{\partial_{k_y}}{\varphi_-^1(\vec{k})}&=\frac{\mathrm{d} \phi_y}{\mathrm{d} k_y}\sin{^2\theta_{xy}},\\
i \matrixel{\varphi_-^2(\vec{k})}{\partial_{k_y}}{\varphi_-^2(\vec{k})}&=\frac{\mathrm{d} \phi_y}{\mathrm{d} k_y}\cos{^2\theta_{xy}},\\
\varphi_-^1(\vec{k})^\dagger\varphi_-^2(\vec{k})&=(\varphi_-^2(\vec{k})^\dagger\varphi_-^1(\vec{k}))^{\ast}=\sin{\theta_{xy}}\,.
\end{align}
Thus, we obtain the following concrete expression for the Berry connections:
\begin{align}
\mathcal{A}_{x}(\vec{k})&=i \matrixel{\psi^x_-(k_x)}{\partial_{k_x}}{\psi^x_-(k_x)} s_0+\frac{1}{2}\sin{\theta_{xy}(\vec{k})}\frac{\mathrm{d} \phi_x}{\mathrm{d} k_x}s_1\nonumber\\&=\frac{1}{2}\frac{\mathrm{d} \phi_x}{\mathrm{d} k_x} \left(s_0+\sin{\theta_{xy}(\vec{k})}s_1\right),\\
\mathcal{A}_{y}(\vec{k})&=i \matrixel{\varphi^1_-(\vec{k})}{\partial_{k_y}}{\varphi^1_-(\vec{k})} s_3=\frac{1}{2}\frac{\mathrm{d} \phi_y}{\mathrm{d} k_y}\left(s_0-\cos{\theta_{xy}}(\vec{k}) s_3\right).
\end{align}
By multiplying the exponent of these non-Abelian connections in path order, we determine the Wilson loops along the $x$ and $y$ axis as
\begin{align}
\mathcal{W}_{x,\vec{k}}(k_y)&= \exp{i \int_{k_x}^{k_x+2\pi}d\phi_x(k'_x)\frac{s_0+\sin{\theta_{xy}(k'_x,k_y)}s_1}{2}},\\
\mathcal{W}_{y,\vec{k}}(k_x)&=\exp{i \int_{k_y}^{k_y+2\pi}d\phi_y(k'_y)\frac{s_0-\cos{\theta_{xy}(k_x,k'_y)s_3}}{2}}.
\end{align}
The above exact expressions for the Wilson loops allows us to solve the Wannier centers and the corresponding Wannier bases directly. The Wilson loop along $y$ direction has been defined to be diagonal,
\begin{align}
\nu^\pm_y(k_x)=\int_0^{2\pi}\frac{\mathrm{d} k_y}{4\pi}\frac{\mathrm{d} \phi_y}{\mathrm{d} k_y}\left(1\pm\cos{\theta_{xy}(k_x,k_y)}\right).
\end{align}
For the Wilson loop $ \mathcal{W}_{x,\vec{k}}(k_y)$ along the $k_x$ axis through the point $\vec{k}$, we have
\begin{align}
\nu^\pm_x(k_y)&=\int_0^{2\pi}\frac{\mathrm{d} k_x}{4\pi}\frac{\mathrm{d} \phi_x}{\mathrm{d} k_x}(1\pm\sin{\theta_{xy}(k_x,k_y)}).
\end{align}
Then the corresponding Wannier gaps are
\begin{align}
\Delta\nu_x(k_y)&=\int_0^{2\pi}\frac{\mathrm{d} k_x}{2\pi}\frac{\mathrm{d} \phi_x}{\mathrm{d} k_x}\sin{\theta_{xy}(k_x,k_y)}=\int_0^{2\pi}\frac{\mathrm{d} k_x}{2\pi}\frac{\mathrm{d} \phi_x}{\mathrm{d} k_x}\frac{|\vec{b}_y(k_y)|}{\sqrt{\vec{b}_x^2(k_x)+\vec{b}^2_y(k_y)}},\\
\Delta\nu_y(k_x)&=\int_0^{2\pi}\frac{\mathrm{d} k_y}{2\pi}\frac{\mathrm{d} \phi_y}{\mathrm{d} k_y}\cos{\theta_{xy}(k_x,k_y)}=\int_0^{2\pi}\frac{\mathrm{d} k_y}{2\pi}\frac{\mathrm{d} \phi_y}{\mathrm{d} k_y}\frac{|\vec{b}_x(k_x)|}{\sqrt{\vec{b}_x^2(k_x)+\vec{b}^2_y(k_y)}}.
\end{align}
To calculate the Wannier-sector polarization, we have to construct the Wannier-band basis\cite{multipole_PRB_2017}:
\begin{align}
w_{j}^{\mu}(\boldsymbol{k})=\left(\Psi_{1}(\boldsymbol{k}), \Psi_{2}(\boldsymbol{k})\right) \cdot v_{j}^{\mu}(\boldsymbol{k}).
\end{align}
By solving the eigenstates of the Wannier loops in the energy representation, we obtain the following bases at the base point $\vec{k}_0$
\begin{align}
w^x_{j}(\vec{k}_0)&=\frac{1}{\sqrt{2}}\left(\Psi_1(\vec{k}_0)+j \Psi_2(\vec{k}_0)\right),\\
w^y_{1}(\vec{k}_0)&=\Psi_2(\vec{k}_0),\\
w^y_{2}(\vec{k}_0)&=\Psi_1(\vec{k}_0).
\end{align}
where $j=\pm$ correspond to the Wannier basis $w^\mu_j(\vec{k}_0)$ of the Wannier sectors $\omega_\mu^s$, $s=+,-$. To consistently define the full Wannier basis, we need to maintain the gauge we choose in Eq.~(\ref{wannier_gauge}) which adds a $\vec{k}$-dependent phase to the bases. Recalling Eq.~(\ref{sector_projection}) we see that these phases cancel in the projector, so we can alternatively choose a simply family of Wannier bases to define the sector polarization as
\begin{align}
w^x_{j}(\vec{k})&=\frac{1}{\sqrt{2}}\left(\Psi_1(\vec{k})+j \Psi_2(\vec{k})\right),\\
w^y_{1}(\vec{k})&=\Psi_2(\vec{k}),\\
w^y_{2}(\vec{k})&=\Psi_1(\vec{k}).
\end{align}
at any $\vec{k}$. We then use the above expressions to calculate Berry connection for the Wannier basis,
\begin{align}
[\mathcal{A}^{xs}_{y}(\vec{k})]_{jj}&=i \matrixel{ w_j^x(\vec{k})}{\partial_{k_y}}{w_j^x(\vec{k})}\nonumber\\
&=\frac{1}{2}\Tr[\mathcal{A}_y(\vec{k})]+\frac{1}{2}\frac{\mathrm{d} \phi_y}{\mathrm{d} k_y}-j\,\text{Re}(e^{i \phi_y}[\mathcal{A}_y(\vec{k})]_{21})\nonumber\\
&=\frac{1}{2}\frac{\mathrm{d} \phi_y}{\mathrm{d} k_y},\\
[\mathcal{A}^{ys}_{x}(\vec{k})]_{jj}&=i \matrixel{ w_j^y(\vec{k})}{\partial_{k_x}}{w_j^y(\vec{k})}\nonumber\\
&=[\mathcal{A}_x(\vec{k})]_{11}=[\mathcal{A}_x(\vec{k})]_{22}=\frac{1}{2}\frac{\mathrm{d} \phi_x}{\mathrm{d} k_x}.
\end{align}
We integrate this to obtain the Wannier-sector polarizations,
\begin{align}
p_x^{s}&=\frac{\mathcal{N}_x}{2}\bmod 1,\label{sec_px}\\
p_y^{s}&=\frac{\mathcal{N}_y}{2}\bmod 1\label{sec_py}.
\end{align}
Finally, the off-diagonal quadrupole moment as defined in Refs.~\cite{multipole_PRB_2017,multipole_science2017} is the multiplication of the two sector-polarizations
\begin{align}
q_{xy}=\sum_{s=\pm}p_x^{s}p_y^{s}=\frac{\mathcal{N}_x\mathcal{N}_y}{2}\bmod 1.
\end{align}
\section{Generalized BBH model with long-range hopping}
In the main text, we propose a generalized BBH model with long-range hopping, whose Bloch Hamiltonian is
\begin{align}
\mathcal{H}(k_x,k_y)&=\vec{b}_x(k_x)\cdot\vec{\sigma}\otimes\tau_3 +\sigma_0\otimes \vec{b}_y(k_y)\cdot\vec{\tau}\label{longH},\\
\vec{b}_x(k_x)&=(u_x+v_x\cos{k_x}+w_x\cos{(2k_x)},v_x\sin{k_x}+w_x\sin{(2k_x)},0)\label{vectorbx}\\
\vec{b}_y(k_y)&=(u_y+v_y\cos{k_y},v_y\sin{k_y},0)\label{vectorby}.
\end{align}
To build such a model with tunable long-range hopping, we propose a two-layer system with a stacking fault, as shown schematically in Fig.~\ref{twolayer}. We can regard the two layers as a single-layer generalized BBH model, described by the Hamiltonian shown in Eq.~(\ref{longH}). The long-range hopping in the generalized model is the original nearest-neighbor hopping in a single layer. The interlayer hopping takes the role of the nearest-neighbor hopping term.
\begin{figure}
\centering
\includegraphics[width=0.35\textwidth]{FigSM1.pdf}
\caption{Coupling of two BBH lattice layers with a stacking fault or AB stacking pattern. }\label{twolayer}
\end{figure}
We now derive a method to calculate the winding number $\mathcal{N}_x$ of the 1D two-band model described by the characteristic vector
\begin{align}
\vec{b}_x(k_x)&=(u_x+v_x\cos{k_x}+w_x\cos{(2k_x)},v_x\sin{k_x}+w_x\sin{(2k_x)},0),
\end{align}
where $w_x$ corresponds to the long-range hopping. In an explicit matrix form, the Hamiltonian is
\begin{align}
h_x(k_x)=\left(\begin{array}{cc}
{0}&{u_x+v_xe^{-i k_x}+w_xe^{-2i k_x}} \\
{u_x+v_xe^{i k_x}+w_xe^{2i k_x}}&{0}
\end{array}\right).
\end{align}
As there is no term proportional to $\sigma_3$ because of the chiral symmetry, the eigenstates for this Hamiltonian can be written as
\begin{align}
\psi^x_\pm(k_x)=\frac{1}{\sqrt{2}}\left(\begin{array}{c}
{e^{-i [\tilde{\phi}_x(k_x)+k_x]} }\\
{\pm1}
\end{array}\right)=\left(\begin{array}{cc}
{e^{-i k_x}}&{0} \\
{0}&{1}
\end{array}\right)\tilde{\psi}^x_\pm(k_x).
\end{align}
Here we define
\begin{align}
\tilde{\psi}^x_\pm(k_x)=\frac{1}{\sqrt{2}}\left(\begin{array}{c}
{e^{-i \tilde{\phi}_x(k_x)} }\\
{\pm1}
\end{array}\right),
\end{align}
It can be shown that $\tilde{\psi}^x_\pm(k_x)$ are the eigenstates of the following Hamiltonian
\begin{align}
\tilde{h}_x(k_x)=\left(\begin{array}{cc}
{0}&{u_xe^{i k_x}+v_x+w_xe^{-i k_x}} \\
{u_xe^{-i k_x}+v_x+w_xe^{i k_x}}&{0}
\end{array}\right)=\tilde{\vec{b}}_x(k_x)\cdot\vec{\sigma},
\end{align}
where the characteristic vector is
\begin{align}
\tilde{\vec{b}}_x(k_x)=(v_x+(u_x+w_x)\cos{k_x},(w_x-u_x)\sin{k_x},0).
\end{align}
We can see that the curve $\tilde{\mathcal{C}}:k_x\in[0,2\pi]\to \tilde{\vec{b}}_x(k_x)$ is actually an ellipse with semi-major axes along $k_x$ and $k_y$ as $|u_x+w_x|$ and $|w_x-u_x|$. Combine these properties and the winding direction determined by the signs in front of the cosine and sine in $\tilde{\vec{b}}_x(k_x)$, we can conclude the winding number of $\tilde{\mathcal{C}}$ is
\begin{align}
\tilde{\mathcal{N}}_x= \begin{cases}0, & |u_x+w_x|<|v_x| \\ \sign [(w_x+u_x)(w_x-u_x)], &|u_x+w_x|>|v_x|\end{cases}.\label{Nx_u}
\end{align}
This result has been shown previously, e.g., in Ref.~\cite{ExtendedSSH}. Applying the above results, we calculate the winding number for $\mathcal{C}:k_x\in[0,2\pi]\to \vec{b}_x(k_x)$ associated with the 1D Zak-Berry phase as
\begin{align}
\mathcal{N}_x=\frac{1}{2\pi}\int_0^{2\pi}\,\mathrm{d} k_x\,i \matrixel{\psi_-^x(k_x)}{\partial_{k_x}}{\psi^x_-(k_x)}=1+\frac{1}{2\pi}\int_0^{2\pi}\,\mathrm{d} k_x\,i \matrixel{\tilde{\psi}^x_-(k_x)}{\partial_{k_x}}{\tilde{\psi}^x_-(k_x)}=1+\tilde{\mathcal{N}}_x.
\end{align}
Finally, we obtain all possible values of $\mathcal{N}_x$ with the corresponding conditions as
\begin{align}
\mathcal{N}_x= \begin{cases}1, & |u_x+w_x|<|v_x| \\ \text{sign} [(w_x+u_x)(w_x-u_x)]+1 , &|u_x+w_x|>|v_x|\end{cases}.\label{Nx_u}
\end{align}
\section{Bulk Quadrupole moment}
We start from the quadrupole tensor definition in the Wannier representation
\begin{align}
\mathcal{N}_{\mu\nu}=\int_{\text{all}} \mathrm{d}^2\vec{r} \,x_{\mu}x_{\nu}\, W_{n\vec{0}}^\dagger(\vec{r}) W_{n\vec{0}}(\vec{r}).
\end{align}
Using Bloch functions to expand the localized Wannier functions, we have
\begin{align}
W_{n\vec{R}}(\vec{r})&=\frac{1}{N}\sum_{k}e^{i\vec{k}\cdot(\vec{r}-\vec{R})}u_{n\vec{k}}(\vec{r}),\\
N&=\frac{(2\pi)^2}{\Delta k_x \Delta k_y}
\end{align}
Now we use Bloch function to express the quadrupole tensor as
\begin{align}
&\mathcal{N}_{\mu\nu}=-\frac{1}{N^2}\int_{\text{all}} \mathrm{d}^2\vec{r} \sum_{k,k'}\sum_{n}\frac{\partial^2 e^{i(\vec{k}-\vec{k}')\cdot\vec{r}}}{\partial k_\mu\partial k_\nu}u_{nk'}^\dagger u_{nk}\nonumber\\
&=\frac{1}{N^2}\sum_{nkk'}\int_{\text{all}} \mathrm{d}^2\vec{r} \frac{\partial}{\partial k_\mu}\left[-\frac{\partial e^{i(\vec{k}-\vec{k}')\cdot\vec{r}}}{\partial k_\nu}u_{nk'}^\dagger u_{nk}\right]+\frac{\partial}{\partial k_\nu}\left[e^{i(\vec{k}-\vec{k}')\cdot\vec{r}}u_{nk'}^\dagger \frac{\partial }{\partial k_\mu}u_{nk}\right]-\left[e^{i(\vec{k}-\vec{k}')\cdot\vec{r}}u_{nk'}^\dagger\frac{\partial}{\partial k_\nu}\frac{\partial }{\partial k_\mu} u_{nk}\right]\nonumber\\
&=\frac{1}{N^2}\sum_{nkk'}\int_{\text{all}} \mathrm{d}^2\vec{r} \frac{\partial^2}{\partial k_\mu\partial k_\nu}\left[-e^{i(\vec{k}-\vec{k}')\cdot\vec{r}}u_{nk'}^\dagger u_{nk}\right]+\frac{\partial}{\partial k_\mu}\left[e^{i(\vec{k}-\vec{k}')\cdot\vec{r}}u_{nk'}^\dagger\frac{\partial u_{nk}}{\partial k_\nu} \right]+\frac{\partial}{\partial k_\nu}\left[e^{i(\vec{k}-\vec{k}')\cdot\vec{r}}u_{nk'}^\dagger \frac{\partial u_{nk}}{\partial k_\mu}\right]
\nonumber\\&\qquad\qquad
-e^{i(\vec{k}-\vec{k}')\cdot\vec{r}}u_{nk'}^\dagger\frac{\partial^2u_{nk}}{\partial k_\mu\partial k_\nu}\nonumber\\
&=\frac{1}{N}\sum_{nkk'}\int_{\text{cell}} \mathrm{d}^2\vec{r} \frac{\partial^2}{\partial k_\mu\partial k_\nu}\left[-\delta_{kk'}u_{nk'}^\dagger u_{nk}\right]+\frac{\partial}{\partial k_\mu}\left[\delta_{kk'}u_{nk'}^\dagger\frac{\partial }{\partial k_\nu} u_{nk}\right]+\frac{\partial}{\partial k_\nu}\left[\delta_{kk'}u_{nk'}^\dagger \frac{\partial }{\partial k_\mu}u_{nk}\right]-\left[\delta_{kk'}u_{nk'}^\dagger\frac{\partial}{\partial k_\nu}\frac{\partial }{\partial k_\mu} u_{nk}\right]\nonumber\\
&=\frac{1}{N}\sum_{nk}\int_{\text{cell}} \mathrm{d}^2\vec{r} \frac{\partial^2}{\partial k_\mu\partial k_\nu}\left[-u_{nk}^\dagger u_{nk}\right]+\frac{\partial}{\partial k_\mu}\left[u_{nk}^\dagger\frac{\partial }{\partial k_\nu} u_{nk}\right]+\frac{\partial}{\partial k_\nu}\left[u_{nk}^\dagger \frac{\partial }{\partial k_\mu}u_{nk}\right]-\left[u_{nk}^\dagger\frac{\partial}{\partial k_\nu}\frac{\partial }{\partial k_\mu} u_{nk}\right]\nonumber\\
&=\frac{1}{N}\sum_{k}\Tr\left[-i\frac{\partial}{\partial k_\nu}\mathcal{A}_\mu-i\frac{\partial}{\partial k_\mu}\mathcal{A}_\nu- u^\dagger_{nk}\frac{\partial^2}{\partial k_\mu\partial k_\nu}u_{nk}\right].\label{int_wannier}
\end{align}
In the above derivation, we have used the following formula
\begin{align}
\int_{\text{all}} \mathrm{d}^2\vec{r} e^{i\vec{k}\cdot\vec{r}}f(\vec{r})=\sum_{R}\int_{\text{cell}} \mathrm{d}^2\vec{r} e^{i\vec{k}\cdot(\vec{r}+\vec{R})}f(\vec{r})=\sum_{R}e^{i\vec{k}\cdot\vec{R}}\int_{\text{cell}} \mathrm{d}^2\vec{r} e^{i\vec{k}\cdot\vec{r}}f(\vec{r})=N\delta_{k,0}\int_{\text{cell}} \mathrm{d}^2\vec{r} f(\vec{r}),
\end{align}
where $f(\vec{r}+\vec{R})=f(\vec{r})$ and $\vec{k}$ is discretized as $k_\mu=2\pi l_\mu/(N_\mu a_\mu)$ ($l_\mu$ is an integer) to make the wave function satisfy the periodic boundary condition. As $\mathcal{A}_\mu$ are all real, the term $-i\frac{\partial}{\partial k_\nu}\mathcal{A}_\mu-i\frac{\partial}{\partial k_\mu}\mathcal{A}_\nu$ in the integral of Eq.~(\ref{int_wannier}) should vanish.
In the continuum limit and with the localization condition $\langle \vec{R}_i|\vec{R}_j\rangle=\delta_{ij}$, we obtain the following topological quadrupole tensor as
\begin{align}
\mathcal{N}_{\mu\nu}&=\int \frac{\mathrm{d} ^2\vec{k}}{(2\pi)^2}\Tr[-\matrixel{ \Psi_{n}(\vec{k})}{\frac{\partial^2}{\partial k_\mu\partial k_\nu}}{\Psi_{n}(\vec{k})}],\label{quadrupole_invariant}
\end{align}
where the trace denotes the summation over all occupied bands. For the main focus off-diagonal element $\mathcal{N}_{xy}$, we have
\begin{align}
\mathcal{N}_{xy}&=\int \frac{\mathrm{d} ^2\vec{k}}{(2\pi)^2}\Tr[-\matrixel{ \Psi_{n}(\vec{k})}{\frac{\partial^2}{\partial k_x\partial k_y}}{\Psi_{n}(\vec{k})}]\nonumber\\
&=\int \frac{\mathrm{d} ^2\vec{k}}{(2\pi)^2}(i\matrixel{\psi^x_+ }{\partial_{k_x}}{\psi^x_+})( i\matrixel{\varphi_-^1 }{\partial_{k_y}}{\varphi_-^1 })+(i\matrixel{\psi^x_-}{\partial_{k_x}}{\psi^x_-})( i\matrixel{\varphi_-^2 }{\partial_{k_y}}{\varphi_-^2 })\nonumber\\
&=\int \frac{\mathrm{d} ^2\vec{k}}{(2\pi)^2}\frac{1}{2}\frac{\mathrm{d} \phi_x}{\mathrm{d} k_x}\frac{\mathrm{d} \phi_y}{\mathrm{d} k_y}\nonumber\\
&=\frac{\mathcal{N}_x\mathcal{N}_y}{2}.
\end{align}
We also calculate the two diagonal elements as
\begin{align}
\mathcal{N}_{xx}&=\int \frac{\mathrm{d} ^2\vec{k}}{(2\pi)^2} \Tr [-\matrixel{ \Psi_{n}(\vec{k})}{\frac{\partial^2}{\partial k_x^2}}{\Psi_{n}(\vec{k})}]\nonumber\\
&=\int \frac{\mathrm{d} ^2\vec{k}}{(2\pi)^2} \Tr [\matrixel{ \frac{\partial}{\partial k_x} \Psi_{n}(\vec{k})}{\frac{\partial}{\partial k_x}}{\Psi_{n}(\vec{k})}]\nonumber\\
&=\int \frac{\mathrm{d} ^2\vec{k}}{(2\pi)^2} \Tr [\matrixel{ \frac{\partial}{\partial k_x} \psi_{n}^x}{\frac{\partial}{\partial k_x}}{\psi_{n}^x}]+\sum_{m=1,2}\matrixel{ \frac{\partial}{\partial k_x} \varphi_{-}^m}{\frac{\partial}{\partial k_x}}{\varphi_{-}^m}\nonumber\\
&=\int \frac{\mathrm{d} ^2\vec{k}}{(2\pi)^2}[(\frac{\mathrm{d} \phi_x}{\mathrm{d} k_x})^2+\frac{1}{2}(\frac{\partial \theta_{xy}}{\partial k_x})^2],
\end{align}
\begin{align}
\mathcal{N}_{yy}&=\int \frac{\mathrm{d} ^2\vec{k}}{(2\pi)^2} \Tr[-\matrixel{ \Psi_{n}(\vec{k})}{\frac{\partial^2}{\partial k_y^2}}{\Psi_{n}(\vec{k})}]\nonumber\\
&=\int \frac{\mathrm{d} ^2\vec{k}}{(2\pi)^2} \Tr[\matrixel{ \frac{\partial}{\partial k_y} \Psi_{n}(\vec{k})}{\frac{\partial}{\partial k_y}}{\Psi_{n}(\vec{k})}]\nonumber\\
&=\int \frac{\mathrm{d} ^2\vec{k}}{(2\pi)^2} \sum_{m=1,2}\matrixel{ \frac{\partial}{\partial k_y} \varphi_{-}^m}{\frac{\partial}{\partial k_y}}{\varphi_{-}^m}\nonumber\\
&=\int \frac{\mathrm{d} ^2\vec{k}}{(2\pi)^2}[(\frac{\mathrm{d} \phi_y}{\mathrm{d} k_y})^2+\frac{1}{2}(\frac{\partial \theta_{xy}}{\partial k_y})^2],
\end{align}
We can see that the off-diagonal element $\mathcal{N}_{xy}$ is topologically invariant under continuous parameter deformation of the $x$- and $y$-direction constitutive chains without closing the Wannier gaps.
\section{ Edge polarizations and edge-consistent gauge}
In the Refs.~\cite{multipole_PRB_2017,multipole_PRB_2017,Changan2020}, it has been shown through numerical calculation of the hybrid Wannier functions in the BBH model that the fractional edge polarizations parallel to the edge are localized at the boundary in ribbon systems. Thus, the dipole moments of the edge states should capture all the edge polarizations when the thermodynamic limit is taken. We derive the exact edge polarizations for the generalized BBH model around the boundary using this approximation.
First, we consider a ribbon along the $x$-axis described by the generalized BBH model, where the occupied $m_x-$th edge states along the up- or down-edge (labeled by U/D) are
\begin{align}
\Psi^\text{U/D}_{m_x}(k_x)=\psi^x_{-\sigma_{m_x}}(k_x)\otimes\varphi_{ \text{U/D}}^{\sigma_{m_x}},\quad {m_x}=1,..,\mathcal{N}_y,\label{edgex}
\end{align}
where $h_x(k_x)\psi^x_{-\sigma_{m_x}}(k_x)=-\sigma_{m_x}|\vec{b}_x(k_x)|\,\psi^x_{-\sigma_{m_x}}(k_x)$, $h_y\varphi^{\sigma_{m_x}}_{\text{U/D}}=0$ and $\tau_3\varphi^{\sigma_{m_x}}_{\text{U/D}}=\sigma_{m_x}\varphi^{\sigma_{m_x}}_\text{U/D}$ ($\sigma_{m_x}=\pm1$). $\varphi^{\sigma_{m_x}}_{\text{U/D}}$ and $\sigma_{m_x}$ labels the chirality. If we apply the gauge in Eq.~(\ref{TwoBand}) for $\psi^x_{-\sigma_{m_x}}(k_x)$, then we obtain the edge polarization along $x$-axis as
\begin{align}
p^\text{edge}_x&=\sum_{m_x=1}^{\mathcal{N}_y}\frac{1}{2\pi }\int_0^{2\pi}dk_x\,i\matrixel{\Psi^\text{U/D}_{m_x}(k_x)}{\partial_{k_x}}{\Psi^\text{U/D}_{m_x}(k_x)}\bmod 1\nonumber\\
&=\frac{\mathcal{N}_x}{2}\mathcal{N}_y \bmod 1.
\end{align}
Similarly, the occupied $m_y-$th edge states for a ribbon along the $y$-axis are
\begin{align}
\Psi^\text{L/R}_{m_y}(k_x)=\varphi^{\tau_{m_y}}_{\text{L/R}}\otimes\psi_{-}^{y}(k_y),\quad {m_y}=1,..,\mathcal{N}_x,\label{edgey}
\end{align}
where $h_y(k_y)\psi^y_{-}(k_y)=-|\vec{b}_y(k_y)|\,\psi^y_{-}(k_y)$, $h_x\varphi^{\tau_{m_y}}_{\text{L/R}}=0$ and $\sigma_3\varphi^{\tau_{m_y}}_{\text{L/R}}=\tau_{m_y}\varphi^{\tau_{m_y}}_\text{L/R}$ ($\tau_{m_y}=\pm1$). L/R denotes the edge where the zero modes $\varphi^{\tau_{m_y}}_{\text{L/R}}$ localize in the $x$-direction and $\tau_{m_y}$ gives the chirality of it. The corresponding edge polarization parallel to the $y$-axis is
\begin{align}
p^\text{edge}_y&=\sum_{m_y=1}^{\mathcal{N}_x}\frac{1}{2\pi }\int_0^{2\pi}dk_y\,i\matrixel{\Psi^\text{L/R}_{m_y}(k_y)}{\partial_{k_y}}{\Psi^\text{L/R}_{m_y}(k_y)}\bmod 1\nonumber\\
&=\frac{\mathcal{N}_y}{2}\mathcal{N}_x \bmod 1.
\end{align}
These results are consistent with the typical BBH model with $p^\text{edge}_x=p^\text{edge}_y=1/2$ in the topological phase~\cite{multipole_PRB_2017,multipole_science2017}.
In the main text, we propose a specific gauge for the bulk and edge Wannier functions. To address the edge-consistent properties of this gauge, we study the bulk and edge Wannier functions of the central unit cell
\begin{align}
|W^{\text{bulk}}_{n\vec{0}}\rangle&=(N_xN_y)^{-1}\sum_{\vec{R}\vec{k}\alpha_x\beta_y}e^{i\vec{k}\cdot\vec{R}}[\Psi_{n}(\vec{k})]_{\alpha_x,\beta_y}|\vec{R},\alpha_x,\beta_y\rangle,\\
|W^{\text{U/D}}_{m_x\vec{0}}\rangle&=N_x^{-1}\sum_{R_xk_x\alpha_x}e^{ik_xR_x}[\psi^x_{-\sigma_{m_x}}(k_x)]_{\alpha_x}|R_x,\alpha_x\rangle,\\
|W^{\text{L/R}}_{m_y\vec{0}}\rangle&=N_y^{-1}\sum_{R_yk_y\beta_y}e^{ik_yR_y}[\psi_{-}^{y}(k_y)]_{\beta_y}|R_y,\beta_y\rangle,
\end{align}
with their corresponding densities
\begin{align}
\rho^\text{bulk}_n(\vec{R},\alpha_x,\beta_y)&=|\langle\vec{R},\alpha_x,\beta_y|W^{\text{bulk}}_{n\vec{0}}\rangle|^2,\\
\rho^\text{U/D}_{m_x}(R_x,\alpha_x)&=|\langle R_x,\alpha_x|W^{\text{U/D}}_{m_x\vec{0}}\rangle|^2,\\
\rho^\text{L/R}_{m_y}(R_y,\beta_y)&=|\langle R_y,\beta_y|W^{\text{L/R}}_{m_y\vec{0}}\rangle|^2.
\end{align}
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{FigSM2.pdf}\\
\caption{The reduced probability density of the bulk and edge Wannier functions at the corresponding center unit cells of the model with long-range hopping described in Eq.~(\ref{longH}). The parameters used are $u_x=u_y=1$, $v_x=v_y=3/2$ and $w_x=2$, which are chosen to generate the topological phase. We can see the 2D bulk Wannier functionss and the 1D edge states have a common density distribution which shows the gauge we choose for the bulk is edge-consistent.}\label{edge_gauge}
\end{figure}
For an edge-consistent gauge, the reduced density of the center Wannier functions is required to satisfy as we presume in the discussion section of the main text \begin{align}
\frac{1}{N_\text{occ}}\sum_{n=1}^{N_\text{occ}}\sum_{R_y,\beta_y}\rho^\text{bulk}_n(\vec{R},\alpha_x,\beta_y)=\frac{1}{\mathcal{N}_y}\sum_{m_x=1}^{\mathcal{N}_y}\rho^\text{U/D}_{m_x}(R_x,\alpha_x)=\rho_x(R_x,\alpha_x),\\
\frac{1}{N_\text{occ}}\sum_{n=1}^{N_\text{occ}}\sum_{R_x,\alpha_x}\rho^\text{bulk}_n(\vec{R},\alpha_x,\beta_y)=\frac{1}{\mathcal{N}_x}\sum_{m_y=1}^{\mathcal{N}_x}\rho^\text{L/R}_{m_y}(R_y,\beta_y)=\rho_y(R_y,\beta_y).
\end{align}
To check whether the gauge we use in the main text for the generalized BBH model satisfies the above conditions, we use the numerical calculation for the Wannier functionss of the system described in Eq.~(\ref{longH}) with $u_x=u_y=1$, $v_x=v_y=3/2$ and $w_x=2$ and show the consistency of the reduced density distribution between the edge and bulk in Fig.~\ref{edge_gauge}.
\section{Gauge transformation of the Quadrupole invariant}
The quadrupole invariant we defined in Eq. (\ref{quadrupole_invariant}) can be expressed by the quantum geometry metric as
\begin{align}
\mathcal{N}_{\mu\nu}&=\int_{\text{FBZ}} \frac{\mathrm{d} ^2\vec{k}}{(2\pi)^2}g_{\mu\nu}(\vec{k})+\text{Tr}[\mathcal{A}_\mu\mathcal{A}_\nu]\nonumber\\
g_{\mu\nu}(\vec{k})&=\operatorname{Re} \sum_{n=1}^{N_{\text{occ}}} (\partial_{k_{\mu}} \Psi_{n }(\vec{k}))^\dagger \partial_{k_{\nu}} \Psi_{m }(\vec{k})-\sum_{n=1}^{N_{\text{occ}}} \sum_{m=1}^{N_{\text{occ}}} (\partial_{k_{\mu}} \Psi_{n }(\vec{k}))^\dagger \Psi_{m }(\vec{k})\,\Psi_{m }^\dagger (\vec{k}) \partial_{k_{\nu}} \Psi_{n}( \vec{k}).
\end{align}
Note that $\int_{\text{FBZ}} \mathrm{d} ^2\vec{k}\,g_{\mu\nu}(\vec{k})/(2\pi)^2=0$ is gauge invariant\cite{Resta2011,geometry_metric2000,Watanabe2019}. Applying the solutions of $\{\Psi_{1}(\vec{k}),\Psi_{2}(\vec{k})\}$, we find the off-diagonal quadrupole moment as
\begin{align}
\mathcal{N}_{xy}&=\int_{\text{FBZ}} \frac{\mathrm{d} ^2\vec{k}}{(2\pi)^2}(g_{xy}(\vec{k})+\text{Tr}[\mathcal{A}_x\mathcal{A}_y])\nonumber\\
&=\int_{\text{FBZ}} \frac{\mathrm{d} ^2\vec{k}}{(2\pi)^2}\text{Tr}[\mathcal{A}_x\mathcal{A}_y].
\end{align}
Under the following $U(2)$ gauge transformation
\begin{align}
\left(\begin{array}{cc} \Psi_1(\vec{k})&\Psi_2(\vec{k})\end{array}\right)\rightarrow\left(\begin{array}{cc} \Psi_1(\vec{k})&\Psi_2(\vec{k})\end{array}\right)U,
\end{align}
the non-Abelian Berry connections $\mathcal{A}_\mu$ transform as
\begin{align}
\mathcal{A}_\mu\to U^\dag \mathcal{A}_\mu U+U^\dag i\partial_{k_\mu} U
\end{align}
and the corresponding transformation for the quadrupole moment is
\begin{align}
\mathcal{N}_{xy}\to\mathcal{N}_{xy}+\frac{1}{(2 \pi)^{2}} \int_{\mathrm{FBZ}} d \vec{k} \,\text{Tr}\{\mathcal{A}_x (-iU\partial_{k_y} U^\dag)+\mathcal{A}_y (-iU\partial_{k_x} U^\dag)+(\partial_{k_x}U^\dag)(\partial_{k_y}U)\}.\label{Nxy_gauge}
\end{align}
Since the edge states are all quasi-1D states and should have common gauge for different bands in Eqs. (\ref{edgex}) and (\ref{edgey}), the physical gauge transformation for edge in $\mu-$direction can only be a phase factor
\begin{align}
U_\mu(k_\mu)=e^{-i\Theta_\mu(k_\mu)}.
\end{align}
Following the derivation in Eq. (\ref{1D_p_gauge}), we can obtain the corresponding transformation for the edge polarizations are
\begin{align}
p^\text{edge}_\mu&\rightarrow\frac{\mathcal{N}_\mu}{2}\mathcal{N}_{\tilde{\mu}}+Z_\mu\mathcal{N}_{\tilde{\mu}}\bmod 1=p^\text{edge}_\mu,\\
Z_\mu&=\frac{\Theta_\mu(2\pi)-\Theta_\mu(0)}{2\pi},
\end{align}
where $\bar{\mu}=y,x$ when $\mu=x,y$.
With the relation between the corner charge $Q_c$, edge polarizations and the bulk quadrupole, we presume that the physical gauge choice for the 2D bulk states should be restricted as
\begin{align}
U(\vec{k})=U_x(k_x)U_y(k_y)=e^{-i(\Theta_x(k_x)+\Theta_y(k_y))}.
\end{align}
Applying this transfomration to Eq. (\ref{Nxy_gauge}), we have
\begin{align}
\mathcal{N}_{xy}&\to\mathcal{N}_{xy}+\mathcal{N}_{x}Z_y+\mathcal{N}_{y}Z_x+2Z_xZ_y,
\end{align}
where $Z_\mu$ are integers for periodic $U(\vec{k})$. Note that the above transformation preserves the fractional quantization of $\mathcal{N}_{xy}$.
Similarly, the Wannier-sector polarizations and the quadrupole moment change as follow with gauge transformation $U$:
\begin{align}
p_x^{s}&\rightarrow \frac{\mathcal{N}_x}{2}+Z_x\bmod 1\\
p_y^{s}&\rightarrow \frac{\mathcal{N}_y}{2}+Z_y\bmod 1\\
q_{xy}&\rightarrow 2(\frac{\mathcal{N}_x}{2}+Z_x)(\frac{\mathcal{N}_y}{2}+Z_y)\bmod 1=\mathcal{N}_{xy}+\mathcal{N}_{x}Z_y+\mathcal{N}_{y}Z_x+2Z_xZ_y\bmod 1.
\end{align}
Here we can see that the corresponding gauge transformation of $q_{xy}$ and $\mathcal{N}_{xy}$ does not break their relation $q_{xy}=\mathcal{N}_{xy}\bmod 1$, which hints the separable gauge transformation is physical.
|
\section{Introduction}
\firstsection{..}
Visualizations serve a critical role in public discourse~\cite{claes2015, knudesen2018}, information dissemination~\cite{jones2017data, dork2013critical, zinovyev2010data}, civic engagement~\cite{airinei2010data, bohman2015data}, and decision making~\cite{kokina2017role, murphy2013data}.
The COVID-19 pandemic has emphasized the importance of data visualization literacy (DVL)---the ability to constructively reason with data---as experts and news outlets frequently use visualizations to increase public awareness, communicate public health decision-making, and drive action within different communities~\cite{shneiderman-covid}. The proliferation of visualizations from the pandemic has highlighted how the general public's limited DVL in tandem with poor visualization design (e.g. misleading visualizations~\cite{informatics7030035, doan2021misrepresenting}) is causing a rapid spread of both accurate and inaccurate information \cite{diseases2020covid, 2021-viral-visualizations}.
\vspace{-2pt}
We investigate how physically crafting visualizations (i.e., making and assembling with craft materials) through play and curiosity may serve to make visualization more comfortable and familiar to young children.
Despite significant attention to statistical literacy and, increasingly, general data literacy in elementary curricula \cite{lee2022data}, we still have limited knowledge of how to best foster DVL at early ages. Most work in DVL for children relies on a more constrained, formal educational context (i.e., a teacher-led approach). In addition, we lack formalized methods (i.e., visualization literacy tests for children) to evaluate the efficacy of these approaches. Broadly constraining DVL education to formal environments, children are restricted in where they learn and the datasets they work with (e.g., predefined datasets that children may not generally be interested in). Educational research highlights that allowing children to make artifacts driven by play and their own curiosity (known as a \textit{constructionist approach}) serves as a more approachable entry point for children to better grasp complex topics (e.g., mathematics~\cite{sarama2004building, sarama2004technology, verdine2014deconstructing}, artificial intelligence \cite{kahn2021constructionism}, science \cite{wickham2016constructionism}, and programming \cite{alimisis2009constructionism}).
Many aspects of DVL develop over prolonged study and experiences, such as the ability to critically reason with and design visualizations. To ground future pedagogical developments, our work explores the constructionist approach to provide preliminary steps towards familiarizing children with visualization.
We investigate \textit{\textit{Data is Yours}} as a design probe to explore how children can accumulate playful learning experiences and personalize visualizations made from everyday materials (e.g., paper, cardboard, mirrors) (\autoref{fig:teaser}). \textit{Data is Yours}\ as a toolkit stems from the constructionist \cite{papert2020mindstorms} and embodied cognition theories \cite{Wilson2002, anderson2012eroding} and lets children lead the learning/creation process. By crafting and playing with physical materials, children can give physical form to data. These physical actions can enable children to consciously reason about a visualization's visual encoding (e.g., colors and marks) freely and without the constraint of finding a ``right'' answer to any guiding task.
\vspace{-1.5pt}
The toolkit's workflow is illustrated in \autoref{fig:workflow}: children assemble the kit, collect data, create a paper template, and then construct a chart panel to represent their data.
Though past methods for fostering children's DVL show promise \cite{alper2017visualization, dasgupta2017scratch, bishop2020construct}, these tools typically rely on exclusively digital solutions. Digital tools may come off as ``black boxes'' to young children and limit their embodied experiences with visualizations \cite{resnick2000beyond}.
In contrast, physical representations enable children to engage in embodied learning and can serve as a more approachable introduction to visualization. Thus, we focus on enabling children to craft physicalizations, where the act of making scaffolds their understanding of different visualization components.
Our approach maintains the benefits of past digital approaches by using computer vision to dynamically render, update, and save digital versions of the physical charts.
Using \textit{Data is Yours}, our objectives are to investigate (1) how children might respond to physical crafting experiences in the context of working with visualizations and (2) how to expand DVL learning to informal learning spaces (i.e., learning that takes place outside of a structured classroom~\cite{oblinger2005leading}).
The toolkit is a product of a three-phase design process. We first conducted a formative evaluation using a preliminary prototype with educators and children to better understand current teaching practices. We then refined the prototype based on observations and feedback. Lastly, to investigate our two objectives, we deployed the toolkit in a series of in-person workshops with children ($n = 5$; aged 6--11) and conducted interviews with educators ($n = 5$).
Our observations reveal that the toolkit helped children creatively engage and interact with visualizations. Children with prior knowledge of data visualization reported our toolkit serving as an authoring tool that they envision using in their daily lives, while children with little to no experience found the toolkit as an engaging introduction to data visualization. Educators spoke of the benefits of \textit{Data is Yours}, emphasizing how it can be easily integrated into existing activities and spaces (e.g., museums, classrooms, homes). From these insights, we provide a future vision outlining how to expand the ways children can creatively engage with the data and visualizations found in their everyday lives.
\section{Background and Related Work}
Our work draws on practices both in education and in visualization. We explore prior approaches and investigate the potential of using constructionist practices to cultivate children's DVL.
\subsection{Tools to Improve Children's Data Visualization Literacy}
\label{sec:toolkit}
\vspace{-2pt}
Focusing on core DVL components\cite{lee2017vlat, boy2014principled, borner2019framework,Chevalier2018visualization}, prior tools for children's DVL help children improve their abilities to interpret and extract information from visually represented data. These tools primarily differ based on their types of activities (i.e., guided vs. child-driven) and intended context (i.e., formal vs. informal instruction).
The majority of existing DVL tools are actively \textit{guided} by an educator and used within a \textit{formal} learning context (i.e., classrooms). For example, C'est La Vis~\cite{alper2017visualization} teaches young children about pictographs and bar charts through concreteness fading---a pedagogical method where new concepts are first presented with concrete examples before progressively abstracting them---via animations. Similarly, Construct-A-Vis~\cite{bishop2020construct} helps children create free-form visualizations within a classroom setting. Both tools consider how teachers can configure datasets and visual encodings for children to work with. While this approach enables teachers to provide guidance and feedback, it largely limits children's engagement with data. Visualizations and data exist in various contexts beyond the classroom (e.g., news, games, books, infomercials, and advertisements). Yet this approach constrains children with which datasets they work with and where they learn.
Other approaches are \textit{child-driven} and used within \textit{informal} contexts (i.e., learning that takes place outside of a structured classroom~\cite{oblinger2005leading}). Gamification approaches, such as Huynh et al.'s role-playing game~\cite{huynh2020rpg} and Diagram Safari~\cite{gabler2019safari}, foster curiosity and experimentation by allowing children to learn through experiential play~\cite{oblinger2004next}. While both games let children pace themselves, they are limited to predefined datasets. In contrast, Scratch Community Blocks~\cite{dasgupta2017scratch} and data sculptures~\cite{bhargava2017data} help children work with datasets that the children think are important. This educational approach stems from the constructionist theory, which advocates for student-centered learning by having children build artifacts that are personally relevant and meaningful~\cite{papert2020mindstorms} (see \autoref{sec:constructionism}).
Like data sculpture~\cite{bhargava2017data} and Scratch Community Blocks \cite{dasgupta2017scratch}, we follow the constructionist approach. However, we investigate this approach from a do-it-yourself (DIY) perspective, drawing on constructive visualizations \cite{huron2014constructiveviz}. Manually constructing visualizations has helped adults explore~\cite{huron2014constructiveviz, huron2014constructingtokens}, learn~\cite{perin2021students, vandemoere2010physical}, reflect~\cite{gourlet2017cairn, thudt2018reflect}, and play~\cite{hopkins2020craft} with data. Children may similarly benefit from this approach. By constructing an artifact, children have an ``object-to-think-with'' to externalize their mental model, observe when their designs succeed and fail, and discuss their insights with others~\cite{papert1993children}. Research also shows that while modern technology provides rich functionalities, it may come off as ``black boxes'' to young children as its inner mechanisms are often hidden and thus poorly understood~\cite{resnick2000beyond}. To circumvent this challenge, like data sculpture~\cite{bhargava2017data}, we rely on low-cost physical materials that children can manipulate to create interactive physicalizations and learn from their embodied experiences.
Working with physical materials leverages the embodied cognition theory, which states that motor action and cognition are highly interrelated, and hence, mutually dependent upon each other~\cite{Wilson2002, anderson2012eroding}.
However, working with physical materials can be challenging and tedious when needing to update datasets and perform other data interactions (e.g., filter, select, encode). As a result, physicalizations traditionally limit children to represent mainly static datasets \cite{Jonathan2020teaching, Verhaert2021datablokken}. Our approach overcomes this limitation and differs from data sculptures in that we still leverage computational capabilities where children can dynamically render, update, and save their visualizations (see \autoref{sec:digital}).
\subsection{Using Toolkits to Construct Visualizations}
\label{sec:constructionism}
Papert postulated that making and designing are critical to learning: they position the learner as an active agent in the creation process, rather than as a passive recipient of materials designed for the learner~\cite{papert2020mindstorms}. D'Ignazio \& Bhargava~\cite{d2016databasic, d2018creative} recognized the benefits of the constructionist model when teaching visualizations to novices. They focused on hands-on learning with creative activities (e.g., analyzing music lyrics) to introduce new data concepts.
However, their work explicitly targets adults.
Research with children requires a different set of considerations \cite{punch2002children, druin2002role}, making D'Ignazio \& Bhargava's activities difficult to apply to children. Alternative considerations are needed to create meaningful artifacts for children's diverse interests and experiences.
Toolkits can help achieve the twin goals of meeting children's idiosyncratic needs and constructing visualizations in a creative, playful manner. Learning technologies illustrate that creativity support toolkits can support a wide range of explorations powered by children's interests~\cite{resnick2005design}. The success of these toolkits is driven by design principles, such as being approachable and supporting exploration, where children can construct diverse artifacts even with low-cost, easily available materials (e.g., Cricket~\cite{resnick2000beyond}, Makey Makey~\cite{silver2012makey}). Visualization also highlights the benefits of toolkits, such as the InfoVis Toolkit~\cite{fekete2004toolkit}, prefuse~\cite{heer2005prefuse}, and D3~\cite{bostockd3}. They can simplify creating complex visualizations using methods that facilitate critical thinking and ownership \cite{mendez2017bottom}.
Creative use of visual toolkits, such as danceON~\cite{payne2020dance}, shows how creative activities can engage novices (e.g., young women of color) in technical computational and data concepts while still providing personal and culturally relevant learning experiences. \textit{Data is Yours}\ has similar objectives. We explore how to best support children to playfully and creatively construct visualizations while broadening their engagement with data through physical construction.
We investigate the potential of physically constructing visualizations for DVL as it can help children become more comfortable and familiar with visualizations.
By learning from embodied experiences (see \autoref{sec:toolkit}), children may understand the different components of a variety of visualizations.
In turn, this approach may help children read and interpret existing visualizations. However, toolkits to help children author visualizations driven by their interests are limited and sparse (\autoref{sec:toolkit}). We take inspiration from both education and visualization to create a toolkit using everyday objects (e.g., cardboard, paper, mirrors) and paper templates to empower children to create, customize, and personalize their own visualizations. Grounded by constructionist and embodied cognition theories, we use \textit{Data is Yours}\ to explore how to leverage child-driven activities to author interactive visualizations and expand DVL learning to informal learning spaces. Our approach offers a new perspective on how to best cultivate children's DVL through creativity and play.
\section{Formative Study}
We developed an initial prototype and used it to conduct a formative study with children and educators. The prototype was designed to reach a broad audience using prevalent technology (e.g., smartphones). The objectives of the formative study were to (1) understand the methods and challenges associated with teaching data visualizations in elementary school and (2) collect feedback on the prototype to inform the development of a design probe for an exploratory study.
\vspace{-2pt}
\subsection{Formative prototype}
Inspired by the various tangible interactions printed fiducial markers enable~\cite{zheng2020paper, electronics10050604}, our preliminary prototype explored these markers as control points for children to map, read, and update paper-based visualizations~\cite{bae2021idc}. Each fiducial marker provides information based on its identity, position, and orientation (\autoref{fig:markers}).
Our initial setup included a single USB camera mounted above the paper graph. The camera was connected to a laptop which presented the digital visualizations.
However, to enable participation from a broader set of backgrounds, we designed a new prototype that used a smartphone. The smartphone acted both as a camera to track changes in markers as well as an interactive screen to show a digital representation of the physical visualization. Relative to personal computers, smartphones are considered to be more widely accessible hardware---in 2019, 95\% of U.S children had internet access at home, but 6\% (primarily low-income minorities) accessed the internet only via smartphones~\cite{nces2021}.
The prototype consisted of five vertical sliders that were mapped to five bars of a bar chart (\autoref{fig:formative}a).
Fiducial markers were placed behind each slider, and the smartphone was mounted at an angle that can view the fiducial markers via two mirrors enclosed in the cardboard casing (\autoref{fig:markers}a).
To further enhance authoring capabilities, we introduced paper templates that children could draw or write custom chart titles and labels. The templates would first be scanned by the phone camera and then applied to the digital chart (\autoref{fig:formative}b).
\vspace{-2pt}
\subsection{Method}
\label{sec:formative_methods}
After designing a phone-based prototype, we conducted a formative study with children and educators, which was approved by CU Boulder IRB.
All study materials are available in the supplemental materials\footnote{\href{https://osf.io/7u93z/}{https://osf.io/7u93z/}}.
\textbf{Children}. We interviewed two children (ages 10 \& 11; 1 female, 1 male). We first conducted a pre-interview using images of visualizations from popular media to determine their familiarity with visualizations and asked how the children currently create visualizations. We then used a think-aloud protocol asking participants to work through two activities for approximately 30 minutes. For each activity, we asked what data observations children noticed from the charts they created. In the end, we conducted a semi-structured interview to understand the children's experience in creating their own visualizations and learn how the prototype could be improved. We took notes and captured images and video recordings for analysis.
The first activity taught the children how to use the prototype. Children first sorted different cards with fruit images. Then the children entered the number of each fruit type to create a bar chart using the prototype's sliders. In the second activity, children generated their own data from a game. Children were given toy disks to slide toward a target (see \autoref{fig:workflow}b). The closer the disk was to the center, the more points children were rewarded. The game was played under five different conditions. Four of the conditions were predefined by the research team (left hand, right hand, closed eyes, five steps back), and one was child-defined (i.e., the child created their own game condition). Children created a bar chart of the points they scored per condition using the prototype and the paper templates.
\textbf{Educators}. To complement our formative study with the children, we interviewed 5 educators (3 female, 2 male) who work in a variety of settings: educational research organizations, formal classrooms, after-school programs, and online teaching. The objectives with the educators were to understand current practices, the challenges they faced when teaching data visualization, and how the prototype might help address those challenges. We conducted a semi-structured interview and an artifact walk-through where we shared the same prototype the children used and demonstrated the two aforementioned activities. We took notes and captured images and video recordings for analysis.
\begin{figure}[tb]
\centering
\captionsetup{farskip=0pt
\includegraphics[width=\columnwidth]{figures/formative_2.pdf}
\caption{(a) Formative prototype. (b) A child from formative study is moving cardboard sliders to author a bar chart. His paper template consists a mix of words and icons and is reflected on the smartphone display (indicated by dotted outlines). See \autoref{sec:flow2} for more details.}
\vspace{-0.2cm}
\label{fig:formative}
\end{figure}
\vspace{-2pt}
\subsection{Key Findings and Design Implications}
We identified three themes from the formative study: a need for approachability, engagement, and scaffolding and embodied learning.
\textbf{Approachability}.
Prior studies and our conversations with educators note that children even of the same general age range are likely to have a variety of experiences with data, drawn from both their personal experiences and educational backgrounds. This variety means that there is no standard method to teach children about data visualization~\cite{lee2022data}.
Though it is difficult to determine how to make visualization approachable for all children, we noted the promise of everyday materials.
Educators commented that because the prototype was made from materials that children are familiar with (e.g., paper), it can readily foster experimentation (i.e., if a child tried something and did not like the results or made a mistake, they could simply start over). We observed this during the study where children freely engaged with the prototype without fear of breaking something.
\textbf{Engagement}. Educators and children alike expressed a need to center the child’s learning around their interests. This is particularly pertinent as young children have short attention spans, yet for them to learn, they need to be driven by frequent, long-term engagement. Educators address this challenge by empowering students with autonomy and increasing their sense of ownership (e.g., letting children choose their own project topics). We saw this reflected in the formative investigation where both children found the first activity with the predefined fruit dataset as ``cheesy'' and one they would not be interested in doing again. Educators also suggested letting children construct parts of the prototype themselves to increase their investment in the learning activities.
From these comments, the final design probe expanded children’s ability to author their own visualizations in two ways. First, we designed the toolkit such that children can freely author visualizations using data of their choice~\cite{pasek2015learners}.
Second, we designed the probe for easy assembly and provided more options to personalize each chart (e.g., choosing colors and the number of data points) to increase their sense of ownership in the activities.
\textbf{Scaffolding \& Embodied Learning}. Educators highlighted how students struggle when initially designing visualizations (e.g., determining what information should go on which axis, setting up appropriate scales), remembering visualizations components (e.g., titles, axes labels), and interpreting visual representations (e.g., reading values between y-axis tick marks). These remarks generally aligned with our observations that children \textit{recognized} common visualizations (bar, line, pie) but sometimes struggled to articulate what they were conveying.
Educators also noted that though the prototype's input method of connecting physical movements to abstract changes in the bar's height supports embodied learning (see \autoref{sec:toolkit}), more explicit associations between the physical form and visualization would be useful.
Based on these comments, we designed the probe to tightly integrate contextual scaffolding into the physical artifact. We changed the design from a pre-assembled tangible interface to a DIY toolkit where children would physically construct the box and charts to create physicalizations. The chart panels and paper templates of the design probe were designed to be a complete visual representation on their own by containing all necessary information (e.g., visual encoding, title, axes) while still corresponding to the digital representation of the visualization.
Scaffolding is built into the design of the physical charts that children would reflect on while constructing (e.g., choosing the number of sliders as a proxy for defining the number of data points; \autoref{sec:physical-components}).
\vspace{-2pt}
\subsection{Design Goals}
\label{sec:design-goals}
We derived three core principles from our formative investigation and prior literature: the design probe should (1) be approachable to children; (2) provide children an engaging experience when authoring visualizations
and (3) support scaffolding and embodied learning.
We further divide these principles into six design goals to understand the usability, inclusiveness, and widespread applicability of the design probe in our exploratory study.\vspace{7pt}
\vspace{-1pt}
\begin{compactdesc}
\item[\textnormal{\textit{Approachability}}]
\item[\textbf{G1.}] \textbf{Foster experimentation:} Use low-cost materials to encourage children to experiment and iterate on designs without having to worry about losing or damaging materials.
\item[\textbf{G2.}] \textbf{Replicable:} Create a design with materials that can be replicated at home or informal learning spaces (e.g., after-school programs, libraries) to expand data engagement.\vspace{7pt}
\vspace{-1pt}
\item[\textnormal{\textit{Engagement}}]
\item[\textbf{G3.}] \textbf{Ease of use:} Create a low barrier of entry for children to build the toolkit and explore a wide range of datasets.
\item[\textbf{G4.}] \textbf{Support creative expression}: Enable children to create their own designs when authoring visualizations.\vspace{7pt}
\vspace{-1pt}
\item[\textnormal{\textit{Scaffolding \& Embodied Learning}}]
\item[\textbf{G5.}] \textbf{Craft visualizations:} Provide a contextualized, embodied experience to help children understand different visualization components.
\item[\textbf{G6.}] \textbf{Physical interactions:} Focus on making kit mechanisms transparent, connect physical manipulations to data operations, and help children understand why different features are necessary.\vspace{5pt}
\end{compactdesc}
\vspace{-1pt}
\section{Design Probe: Constructing Visualizations with Everyday Materials}
To address these goals, we created \textit{Data is Yours}: a design probe to study how constructionist practices can engage children to create visualizations in a playful, creative manner. It is a toolkit made out of everyday materials (e.g., paper, cardboard) and uses fiducial markers to help children author visualizations. \autoref{fig:workflow} illustrates the workflow of using the toolkit. After constructing the kit (G6), children mount a smartphone on top of the box and place a chart panel with a paper template onto the acrylic surface. Then children manipulate the physicalization to reflect their collected data and are able to see a corresponding visualization on the smartphone display. The toolkit is open source\footnote{\href{https://www.instructables.com/Data-Is-Yours-Toolkit-Assembly-Instructions/}{https://www.instructables.com/Data-Is-Yours-Toolkit-Assembly-Instructions/}} to enable children, parents, and educators to adopt and fabricate the toolkit on their own (G2).
\subsection{Physical Components}
\label{sec:physical-components}
The physical components of the toolkit include interchangeable chart panels, paper templates, and a cardboard-based box. These components are made from readily accessible materials that can be easily replicated (G2, G3).
Children first decide the number of data points required, assemble the chart panel, and then create a paper template to label the chart. The combination of the chart panel and the paper template results in a complete physicalization (\autoref{fig:chart-panels}).
\vspace{-2pt}
\subsubsection{Chart Panels}
The design probe supports three chart types (bar, line, pie) (\autoref{fig:chart-panels}) (G5). These charts were chosen based on their common use in media and our insights from the formative study. Each chart panel consists of a cardboard base, paper, and 3D printed sliders. To support scaffolding and embodied learning, children use these physical pieces to assemble a visualization. Children can draw connections of the existing idioms behind common visualizations through their physical interactions as opposed to being presented with a predefined chart. Note that though \autoref{fig:chart-panels} shows five data points for each chart, charts can be customized so that children can remove and add data points accordingly. A paper-based alternative to the 3D printed components is provided in our Instructable documentation.
\textbf{Bar chart.} The bar chart panel uses two cardboard panels, which are held together by 3D printed snap presses. Strips of colored paper are inserted between the cardboard and the handle base, acting as the physical ``bars'' of the bar chart. To support creative expression and help children think about visual attributes, children can choose different color strips of paper while authoring a bar chart (G4). For example, children can reason about visual semantics when five bars each have a unique color as opposed to the same color. Fiducial markers are attached to the back of the 3D printed snaps to track the height of the bars while children are physically manipulating them.
\textbf{Line chart}. The line chart panel uses the same cardboard body and 3D printed handles as the bar chart panel. This approach allows children to think about the relationship between a bar and line chart (i.e., how a bar chart can be transformed into a line chart with a temporal emphasis) through hands-on assembly (G5). Instead of wide paper strips, the line chart uses narrow paper strips that are placed between the cardboard and the 3D handle. These paper strips have pre-punched folding lines that shrink and stretch as the 3D handles move either closer or further away from each other. The paper strips are intentionally designed to help children observe and experiment with the strips as they move the 3D printed handles (G1).
\textbf{Pie chart.} The pie chart panel requires a cardboard base and a transparent plastic sheet. Children assemble the “slices” of a pie chart using cardstock circles. Each cardstock circle corresponds to a section of the pie chart and has a unique fiducial marker at the back. These cardstock circles are spirally inserted into the slit of the transparent plastic sheet and pinned by a 3D printed clip. Children can then rotate and manipulate the pie chart around the 3D printed clip. Children can choose pie sections of any color and arrange them in order of their preference.
\begin{figure}[tb]
\centering
\captionsetup{farskip=0pt
\includegraphics[width=\columnwidth]{figures/panel.pdf}
\caption{(a) Frontside of each chart panel (bar, line, pie). Blank paper templates layered on top. (b) Backside of each chart panel, illustrating where fiducial markers are placed when facing the camera (\autoref{fig:markers}). }
\vspace{-0.2cm}
\label{fig:chart-panels}
\end{figure}
\subsubsection{Paper Templates}
\label{sec:paper-templates}
Each chart panel requires a paper template, which is a layer of paper that covers the top of the chart panel (\autoref{fig:chart-panels}a). The paper templates are made from regular printing paper (G2), and children can be readily provided with multiple copies of the same paper templates to foster experimentation (G1). The paper templates' design subtly scaffolds children about visualization design by having them fill in the slots for different chart components (e.g., titles, axes labels, y-axis scale). The blank templates also enable children to creatively customize the charts (e.g., by drawing, writing) (G4).
The paper template design for the bar chart and line chart is the same. It consists of a title block, five x-axis label blocks, and tear-off rectangles. The left-most rectangle folds back so children can label the y-axis scale. The tear-off rectangles are to encourage children to think about the data they are representing (G5). Before creating the physicalization and rendering the visualization, children must decide how many rectangles to tear off depending on the number of their data points. Children can tear off up to four rectangles, and those that are not torn will cover the sliders of the respective chart panel. The low cost of paper allows children to experiment and try again if they make the mistake of ripping off too many rectangles (G1).
The paper template for the pie chart similarly includes a title block but features a color legend and a cut-out circle. The color legend helps children consider what the different colors represent within the pie chart. Children can also take notes on their pie chart in the white space of the template. For unused colors, we provided white stickers for children to cover the corresponding legend blocks.
\subsubsection{Box}
The main support structure of the toolkit is a folded cardboard box (\autoref{fig:teaser}c). Four other cardboard components are required to create the final structure. These five components can be quickly assembled and disassembled by simply inserting and unfastening tabs without needing glue or other adhesive materials. The box was designed so that the assembly does not require fine motor skills, making it more generally approachable and appropriate for children (G3). In addition, the cardboard material allows children to decorate the box, promoting creative expression while providing a sense of ownership (G4).
Inside the assembled box, children place two mirrors to capture the fiducial markers with the phone camera (\autoref{fig:markers}a). Afterward, children insert an acrylic plate from either side of the box. The acrylic panel acts as a support for the interchangeable chart panels. Then the chart panels and paper templates are directly placed on top of the acrylic plate.
\subsection{Digital Components}
\label{sec:digital}
\textit{Data is Yours}\ supports a web app where children can dynamically render, update, and save their visualizations. The app consists of three user flows (\autoref{fig:app}) that children enter after constructing the toolkit and creating their physicalizations.
The first flow teaches children how to interact with the chart panels to author a visualization, the second flow enables children to scan and save their own chart, and the third flow allows children to revisit previously saved charts. All three flows start via their respective buttons on the home screen.
\subsubsection{Flow 1: Interacting with the Chart Panels.} Flow 1 provides a tutorial of the system's digital functions and represents the first experience children have when connecting the physicalization to the digital visualization. After constructing and placing the physical chart panel on top of the acrylic panel, children can explore the basic features of their respective digital charts.
Children can move the physical components and see how the digital representations change as a result. They can also change the color of the digital chart by tapping on the visualization components, such as an individual bar, to match the chosen colors represented in the physical panel.
Children can also ``pause'' the chart in its current position and lift the phone to share their rendered visualizations. The pause functionality prevents the digital chart values from incorrectly changing as the phone camera's field of view is moved around. Given Flow 1's purpose as a tutorial, the data types and scales in Flow 1 are the same as those in the formative study: visualizing different types and quantities of fruit with a predefined dataset.
\begin{figure}[tb]
\centering
\captionsetup{farskip=0pt
\includegraphics[width=\linewidth]{figures/app_flow.pdf}
\caption{Screenshots of web app. (a) Home page (b) Bar chart tutorial in Flow 1 (c) Menu in Flow 2 (d) Scanning instructions for paper templates (e) Example of a scanned template (f) Saved visualizations in Flow 3.}
\vspace{-0.2cm}
\label{fig:app}
\end{figure}
\subsubsection{Flow 2: Authoring Visualizations}
\label{sec:flow2}
Once children have created the physical panels, collected relevant data of their choice, and prepared the corresponding paper templates, they can enter Flow 2 to author interactive visualizations. Flow 2 requires two steps: scanning and authoring.
\textit{Scanning and defining the chart.} Children are guided on how to scan their paper template through a series of GIFs and instructional text displayed on the screen. To scan their paper templates, children remove the physical panel and place the paper template face-down on the acrylic panel. The app displays a ``Scan'' button to initiate the process. Once the scanning is complete (i.e., taking a picture), children are instructed to flip the paper template upright, reposition the physical panel, and place the template on top of the chart panel to complete the physical chart. For only the line and bar chart, children are prompted to enter information corresponding to the x- and y-axes of the charts, such as the number of data points and the maximum value of their collected dataset.
\textit{Authoring and interacting with the chart.} To author visualizations, children follow the same general process as Flow 1. The predefined dataset is replaced with information from the paper template that the children scanned into the chart. Information from the templates, such as the hand-written chart title and x-axis labels, are displayed at the corresponding locations on the screen by taking cropped images. The scale of the y-axis is calculated based on the input from the previous step. Children can then move the physical sliders to represent the data values they have collected.
This information is translated onto the digital visualization.
In addition to the ``Pause'' button, Flow 2 also allows children to save a chart. Clicking the ``Save'' button results in a small visual indicator that the chart in its current state has been saved.
\subsubsection{Flow 3: Revisiting Saved Charts} Once children have saved their charts, they can view them within the web app by selecting the ``Saved Charts'' button from the home screen. They can browse through a list of chart images in order of the day and time at which they were saved. These images are non-interactive and meant to serve as ``snapshots'' of data collected in time.
\subsection{System Architecture \& Implementation}
Once the toolkit is fully constructed, we have two mirrors inside which direct the phone's rear-facing camera at the chart panels (\autoref{fig:markers}a). This configuration enables a web-based application to detect the markers on the interface and render the charts. This approach allows us to make a variety of interfaces that function with just paper. The front-end is implemented with a combination of HTML5, CSS, JavaScript, and a Javascript port~\cite{js-port} of the ArUco Computer Vision Marker library~\cite{garrido2016generation, romero2018speeded}. The library offers up to 1000 unique markers for tracking a variety of values (e.g., position, rotation).
\section{Exploratory Study}
To investigate how constructionist practices can foster children's comfort and familiarity with data visualization, we conducted an exploratory study using the design probe in a series of workshops with children (\autoref{sec:kids_eval}) and interviews with educators (\autoref{sec:teachers_eval}). Workshops can elicit rich qualitative insights for early stages of applied visualization research~\cite{kerzner2019workshops}. Using guidelines by Kerzner et al.~\cite{kerzner2019workshops}, Huron et al.~\cite{huron2016using, huron2017workshop}, and Terre des hommes~\cite{terre2020focus}, we designed our workshop around the constraints of working with children in a limited timeframe while still evaluating the research objectives of the toolkit.
Note: the workshop is a means of investigating the constructionist approach where we primarily collected qualitative data rather than quantitatively testing children’s learning outcomes. Current DVL tests are meant for adults and their efficacy for children has not been fully researched\footnote{Existing standardized tools\cite{boy2014principled, lee2017vlat, borner2019framework} target adults. E.g., two components of VLAT focus on determining range and extremum for various visualizations. Such concepts are generally not taught till children are older (11--14 y/o).}.
All activities with the children and educators were approved by the CU Boulder IRB. Study materials are available in supplemental materials. A comprehensive overview of the participants can be found in \autoref{tab:participants}.
\begin{table}[tb]
\small
\begin{tabular}{
>{\centering}
p{2cm}
>{\raggedright}
p{1.5cm}p{4cm}
}
\toprule
\textbf{Participant} & \textbf{Role} & \textbf{Age / Occupation}\\\midrule
P1 &
Child &
6 years old\\
P2 &
Child &
6 years old\\
P3 &
Child &
7 years old\\
P4 &
Child &
8 years old\\
P5 &
Child &
11 years old\\
E1 &
Educator &
After-school program\\
E2 &
Educator &
After-school program\\
E3 &
Educator &
After-school program\\
E4 &
Educator &
Educational Researcher\\
E5 &
Educator &
Elementary School Librarian\\
\bottomrule
\end{tabular}
\vspace{-0.5em}
\caption{Summary of participants' backgrounds in the exploratory study.}
\label{tab:participants}
\vspace{-0.1cm}
\end{table}
\subsection{Workshops with Children}
\label{sec:kids_eval}
\textbf{Participants}. We conducted two two-hour workshops with children ($n = 5$) between 6--11 years old to understand how the constructionist approach may differently impact different age groups. This age range aligns with the development stage where children are able to formulate representational thought through symbolic thinking (i.e., words and pictures) \cite{goswami2002blackwell}.
The first workshop was a group workshop at an after-school program with four children (2 male, 2 female; aged 6--8 years old), and the second workshop involved one child (1 female; aged 11 years) at a public museum.
We recruited the participants by sharing virtual flyers and emails about the workshop at informal learning spaces (e.g., public museums, local makerspaces, and after-school program venues).
\textbf{Procedure}. The workshop was divided into four stages, where the first stage consisted of icebreaker activities, the second and third stages focused on constructing and using the toolkit, and the last stage was for concluding and reflection. \autoref{fig:workflow} illustrates the general workflow of the workshop. Before engaging in the activities, the research team first provided an overview of the workshop to the parents and guardians of the children, gained consent from parents, and collected assent from the children.
We took photographs and collected video-audio recordings as the children were constructing and interacting with the toolkit.
\textbf{Stage 1. Introduction and Icebreaker Activities.} The workshop facilitators introduced themselves to the children. Children were asked about their perception of data and visualization in general. Afterward, they were shown a series of common visualizations (bar, line, pie) and asked if they recognized the charts. Then children engaged in an open-ended discussion about what they noticed in each of the images to gauge their familiarity with these visualizations more concretely.
\textbf{Stage 2. Toolkit Construction.} In the second stage, children were introduced to \textit{Data is Yours}. Workshop facilitators first provided a finished toolkit for children to play with and encouraged children to inspect the sliders and mirrors. Children were instructed on how to use the \textit{Data is Yours}\ by moving the sliders. At this time, children played with the toolkit using the predefined fruit dataset used in the formative study, observing changes displayed on the smartphone. Afterward, children moved on to building their own \textit{Data is Yours}. Each child was given a cardboard fold-out template and a packet containing the various components required for assembly. Facilitators guided the children through the construction process by demonstrating the physical operations required for assembly. After constructing the kit, children were provided craft materials (e.g., pens, pipe cleaners, stickers) to decorate the outer body of the toolkit.
\textbf{Stage 3. Using the Toolkit}. In the third stage, children used their toolkit to author visualizations. Working under the constraints of a guided and timed workshop, children were asked to generate and collect data using the same game from the formative study for three rounds (c.f., \autoref{sec:formative_methods}). In Round 1, children played the game under five different conditions to author a bar chart. Two of the conditions were predefined by the research team (throwing disks using the left hand and right hand) and the remaining three conditions were defined by the children. In Round 2, from the five conditions, children were asked to repeat one game condition of their choice four more times. This activity was intended to help children generate data to create a line chart. In Round 3, children calculated the total score for all five conditions and visualized each condition’s contribution (i.e., proportion) to the total score for the pie chart.
After finishing the game of disk darts, children were shown how to use the paper templates to translate their collected data into chart axes and other chart elements (e.g., title, legend). They then scanned the templates and authored the different visualizations (i.e., bar, line, pie) using the toolkit. Afterward, children analyzed the resulting graphs and compared their results to others.
\textbf{Stage 4. Conclusion and Reflection.} In the fourth stage, we quickly recapped the activities the children engaged with. Children were asked semi-structured questions to gain their overall impression of the toolkit and data visualization in general.
\vspace{-2pt}
\subsection{One-on-One Interview with Educators}
\label{sec:teachers_eval}
\textbf{Participants.} We recruited five educators (3 female, 2 male) from a local school district and via our network of education researchers. All educators had experience with teaching visualization to young children in the context of either mathematics or science curricula.
Our aim with the educators was to gain feedback on how \textit{Data is Yours}\ might be integrated into informal learning spaces. Following the guidelines by Kerzner et al.~\cite{kerzner2019workshops}, we recruited educators with diverse experiences.
\textbf{Procedure.} Each participant provided a verbal introduction of their background at the start of the interview. The facilitators then performed a demo of the toolkit, illustrating how to build the toolkit and showcasing the different features (e.g., scanning with paper templates, switching panels to create different visualizations). After the demo, educators partook in a concurrent think-aloud protocol where they were encouraged to build and interact with the toolkit in an open-ended session for approximately 20 minutes. Educators then engaged in a semi-structured interview where they answered additional survey questions about their impressions, ideas, and concerns about the toolkit. Educators were audio-recorded and photographed working with our toolkit.
\section{Results}
We qualitatively analyzed materials (i.e., notes, images, videos) from our exploratory study.
We conducted a content analysis on these materials using the design goals that emerged from the formative study as \textit{a priori} categories~\cite{weber1990basic}. Two researchers performed this categorization independently, after which a final set of observations corresponding to each design goal was agreed upon (\autoref{sec:design-goals}). Findings were extracted from these grouped observations. For the interviews with the educators, we summarized their impressions, feedback, and concerns also through the lens of our design goals and research objectives.
\vspace{-2pt}
\subsection{Workshops with Children}
Children provided positive impressions of the workshop and enjoyed engaging with \textit{Data is Yours}, despite having greatly varying baseline knowledge of data visualizations (as assessed by the ice-breaker discussions).
In the beginning, the younger children (P1-P4) were not able to assign names to the visualizations, explaining they did not frequently interact with or create visualizations. Yet, they found the workshop particularly fun, with P4 expressing, ``\textit{I really loved it!}'' P3 strongly expressed wanting a version of the toolkit that would work with his iPad, so that he could use it anytime without relying on his parents' smartphones. In contrast, P5, the oldest child of the two workshops, was able to identify, interpret, and explain the function of all charts and found the workshop with the toolkit interesting, but only a ``little fun.''
\vspace{-3pt}
\subsubsection{Approachability}
\textbf{G1: Foster Experimentation.}
Initially, P1-P4 were hesitant to perform certain physical operations, such as folding the flaps of the cardboard base, and even caused minor tears in the cardboard. We then explained to the children that these issues are normal and were, in fact, the reason why we used cardboard and everyday materials: mistakes were a welcome part of the process. After our explanation, P1-P4 began to more confidently interact with the cardboard base and intentionally experiment with different assembly configurations of the components, such as having the phone camera cutout be on the right instead of the left (\autoref{fig:workshop-pics}a). Similarly, with the paper templates, three of the four younger children (P1, P2, P4) initially ripped the template in half when trying to remove the tear-off rectangles. After providing more copies of the template, they were able to successfully author a template. These examples speak to the children's willingness to try again despite the challenges they faced and to experiment with the toolkit in unexpected ways to learn more about the content and function it provided.
\textbf{G2: Replicable.} P1-P4 mentioned they envisioned themselves attempting to recreate the toolkit at home with support from their family. Though children were given pre-packaged kits to construct their own toolkit during the workshop, they provided suggestions on how to further improve the replicability of \textit{Data is Yours}. P2 and P3 suggested that the overall toolkit should be bigger so it can support larger screen interfaces like a tablet. P3's motivation for this request stemmed from wanting increased agency by using a device he had access to at home. P5, in contrast, felt more comfortable with the idea of constructing a toolkit by herself and specifically envisioned how the toolkit could be used by children of her age at her school's library. She explained how the school librarian would often make children do activities related to data visualization (e.g., create a pie chart to track genres of books read), and that the toolkit might make for an interesting way to author multiple charts related to that information. \textit{Data is Yours}'s design showcases the breadth of how different children envision the toolkit being reused in different contexts.
\vspace{-2pt}
\subsubsection{Engagement}
\textbf{G3: Ease of Use.}
In both workshops, we provided verbal instructions on how to construct the box and also demonstrated relevant actions in real-time.
The components of the toolkit were designed so that it would be easy for children to assemble the box by themselves.
This was true for P5, but the younger participants all required more time and some hands-on assistance to complete the process. The younger children found it difficult to securely insert some cardboard flaps into the pre-cut slots especially when the pieces were not already connected. They also needed instructions while arranging fiducial markers on the back of the panels in the right order. P5 took 15 minutes to complete the box assembly, while P1-P4 needed around 35 minutes.
Despite this variance, \textit{Data is Yours}\ successfully enabled all children to transform their data into digital charts. All participants were able to successfully create visualizations even from the workshop's first activity: translate sorted stacks of fruit images into a digital bar chart using the toolkit slider. The interface also encouraged the participants to construct the kit by themselves and promoted collaboration. In the group workshop, despite no instructions to work together, P1-P3 started to naturally collaborate to help each other create and provide comments on each other's bar chart (\autoref{fig:workshop-pics}b).
\textbf{G4: Support creative expression.}
After constructing the toolkit, children personalized their box using a range of craft materials including sketch pens, colored paper, and stickers. Children particularly enjoyed this step, spending significant time (20 minutes on average) deciding which materials to use and how to design their box. The younger children began by writing their names on the box, adding stickers, coloring the sides, and drawing faces on the box to personify it. P5, who had the greatest familiarity with visualizations, focused on many functional additions, such as textual instructions for where to place the mobile phone, a handle to carry the box made using pipe cleaners that were inserted into the flutes of the cardboard, and two small knobs to keep the paper template in place (\autoref{fig:workshop-pics}d).
Children's creative expression was not limited to solely the box but also included the paper templates and paper strips for the bar charts. With the paper templates, children came up with a range of titles for their visualizations, such as ``Frisbee,'' ``Throw-it Game,'' ``[P3's] Game,'' and ``Targeting.'' Children also exhibited creative variations in how they labeled their x-axes: three children expressed the x-axis labels with full words (\autoref{fig:workshop-pics}c), one used a combination of letters and numbers, and one drew icons. With the paper strips, children used three types of color schemes: all of the same colors (P3--to use their favorite color), all different colors (P1, P2, P5--to make use of all available options), or two alternating colors (P4--to imitate the facilitator's toolkit). Even the simple decisions of choosing paper strips, deciding titles, and naming x-axis labels highlight how children were excited to showcase their creative expressions.
\vspace{-3pt}
\subsubsection{Scaffolding \& Embodied Learning}
\textbf{G5: Craft visualizations.}
Although children faced some issues during the construction process, most children were able to create, analyze and draw meaning from their visualizations. Across both workshops, children spent about 30 minutes building and interacting with the chart panels. Towards the end of the workshop, P2 and P5 inferred from their charts that their performance with one hand was better than the other due to better control of their dominant hand.
P5 noticed a slight upward trend in her line chart as she performed better over time as she played the game. She elaborated that the resemblance of the physical line chart to its digital form drew her attention to the line's shape. Similarly, P1-P4 were also able to draw parallels between the physical bar chart and the digital representation. While playing with the bar chart control panels (moving sliders up and down), they noted how the height of the paper strips proportionally matched the heights of the digital bars. The ability to physically craft the visualization and the manner in which the digital representation mirrored the movement of the 3D printed sliders helped reinforce the various visualization components that they interacted with.
\begin{figure}[tb]
\centering
\captionsetup{farskip=0pt
\includegraphics[width=\linewidth]{figures/workshop.pdf}
\caption{(a) P3 assembling the cardboard base (b) P1--P3 collaborating (c) P1 interacting with her authored bar chart (d) P5's decorated toolkit.}
\vspace{-0.2cm}
\label{fig:workshop-pics}
\end{figure}
\textbf{G6: Physical Interactions.} After interacting with the facilitators' toolkit, children immediately raised questions about and tried to explain the box's mechanism. P3 hypothesized the mirrors were necessary for the camera to see fiducial markers ``...to provide some kind of coding that's connect[ing] the markers and phone.'' P3 continued to question the toolkit's mechanism, asking ``\textit{So basically you build the data [visualization] with [the toolkit], and then you print [the visualization] out?}'' This curiosity and fascination motivated the younger children to take the lead in the learning and creation process. For example, P1-P4 expressed great excitement in building their own \textit{Data is Yours}\ and chose to construct the toolkit before playing the game of disk darts (\autoref{fig:workflow}b) when offered a choice. Though the construction process took longer for the younger children, they were engaged throughout the entire process. They continuously questioned how certain components and features were related to data visualizations. For example, P4 asked, ``\textit{Why are there these slots when we aren't using them?}'' when he noticed the two slots at the base of the box that were meant to let in light to the bottom mirror. Similarly, P2 questioned why there were multiple colors for the paper strips and why it mattered for visualization.
While the toolkit fostered their curiosity, children also noted improvements in the toolkit's construction process. P5 wished that the toolkit box would be more stable while she was decorating her box. Additionally, while constructing visualizations, she noticed that the sliders did not smoothly move for all panels (bar, line, pie). Consequently, for the bar and line charts, she spent approximately three minutes rapidly moving the sliders up and down the channels in an attempt to manually smooth the cardboard structure. P1-P4 faced a bottleneck during the panel assembly process when they tried to affix the two parts of the sliders to the chart panel. The grooves within some of the sliders were worn down after a few cycles of opening and closing, and the younger children requested assistance during this step.
\begin{figure*}[t]
\centering
\captionsetup{farskip=0pt
\includegraphics[width=\linewidth]{figures/scenarios_1.pdf}
\caption{Sketches of incorporating \textit{Data is Yours}\ into various scenarios to expand ways children engage with data.}
\vspace{-0.1cm}
\label{fig:future-vision}
\end{figure*}
\vspace{-2pt}
\subsection{Interview with Educators}
\vspace{-2pt}
Educators were overall appreciative of \textit{Data is Yours}. E1, E2, and E5 specifically appreciated the DIY aspect of the toolkit. E1's and E2's after-school programs emphasize making and physical activities, and they noted the toolkit was a clever solution to introduce children to data visualizations in line with those pedagogical goals. All three educators also felt that the ability to personalize both the overall appearance of the box and the charts can help children better identify with the activities and take more ownership of the results. As a researcher for early childhood education, E4 emphasized the importance of physical manipulation for young children's education, pointing out that even adults use physical operations to help them think more clearly (e.g., using their hands to count). Based on these comments, E4 noted \textit{Data is Yours}\ could help students have another mental metaphor to work with when reasoning about abstract concepts in data.
However, both E3 and E5 commented on the toolkit's limitations concerning the mobile phone. While E5 commented that the toolkit is a powerful introduction for younger children (i.e., 6-8 years old), ``\textit{the [current] toolkit is going to [reach] a ceiling where [children] have already reached to highest capability they can go [with the phone].}'' Similarly, E3 questioned the choice of a mobile phone. She felt smartphones provide a useful introduction to the common visualizations but will be quickly limited in their ability to create complex charts later on. She suggested a similar toolkit scheme could incorporate a tablet or laptop in the future.
We assume that these concerns arise from the perceived limitation of the compact physical form in addition to the mobile phone's capability relative to laptops when visualizing data. Both educators found it difficult to imagine how the toolkit can incorporate more than five data points in its current form while retaining sufficiently large marks for children to work with.
For future work, E3--E5 stressed how \textit{Data is Yours}\ could be integrated into existing curriculum (e.g., connecting with schools or after-school programs). They commented the strength of \textit{Data is Yours}\ stems from how easily it can be integrated into existing activities. E1 and E4 suggested integrating \textit{Data is Yours}\ into informal learning spaces that are likely to already have their own instructional content relating to visualizations (e.g., science museums).
E1 explained that in her past experience as a middle school teacher, she frequently asked her students to record their daily homework progress using stickers in a notebook, similar to the process of creating and saving charts in Flow 3. E1 also expressed the challenges she experienced when students misunderstood or were unwilling to create visualizations for their science lab reports. She stated a toolkit like \textit{Data is Yours}\ might have helped her students with those two activities more easily, while also resulting in engaging, informative visualizations throughout the school year.
\section{Discussion}
We used \textit{Data is Yours}\ as a design probe to understand two core research questions: (1) how might children respond to physical crafting experiences in the context of working with visualizations and (2) how can we expand DVL learning to informal learning spaces?
We found that \textit{Data is Yours}\ helped younger children learn more about visualization components, while older children who were more familiar with data visualizations found \textit{Data is Yours}\ to be more of an engaging authoring tool (\autoref{sec:data-is-yours}).
Educators immediately saw the benefits of our approach and how it could be expanded and integrated with other activities in informal learning spaces
(\autoref{sec:research-agenda}).
\vspace{-2pt}
\subsection{Data is Yours}
\label{sec:data-is-yours}
Integrating physical crafting into a visualization authoring context was useful for children but that utility differed based on the children's age and prior knowledge of data visualization.
Older children, who are likely to be more familiar with visualizations, viewed the toolkit primarily as a platform for creating their own visualizations with more agency. At the end of the workshop, P5, the oldest child, clearly stated she did not learn anything new about data visualization from the workshop. However, she noted how \textit{Data is Yours}\ was a more engaging method of authoring charts that she could see herself using in her daily life. Although she had used spreadsheet programs like Google Sheets to visualize her reading habits as pie charts, she described the limitations of the ``black box'' nature of purely digital tools: ``\textit{[With Google sheets] you just type in a bunch of numbers and [the visualization] appears, but with \textit{Data is Yours}\ I can actually create charts with my friends and talk about it.}'' P5 also spoke positively of how she could use \textit{Data is Yours}\ and the paper templates to author charts using any dataset (e.g., exercise time, how long she spends coding). P5's statements support the idea that the physical and interactive nature of the \textit{Data is Yours}\ results in a more engaging and empowering process compared to exclusively digital visualization software.
\vspace{-2pt}
\subsection{Limitations \& Future Work}
Our exploratory study does not aim to provide any direct pedagogical evidence of our approach improving children's DVL. Rather the qualitative observations we collected provide preliminary evidence that the children enjoyed interacting with data physically, can familiarize themselves with foundational concepts of visualization (e.g., reason about visual semantics, incorporate chart elements, set up appropriate scales) through curiosity and play, and were empowered to creatively engage with data.
We currently lack more formal assessment methods that are appropriate for children, such as revised visualization literacy assessment tests, that would allow for a more formal DVL measurement.
Though \textit{Data is Yours}\ is designed to help children author visualizations using data of their choice, the tool's expressiveness was not fully explored within the constraints of a two-hour workshop.
For example, the importance of working with personally meaningful datasets varied based on age. P1-P4 did not mind interacting with the predefined fruit datasets. P5 expressed a similar sentiment as the two formative children, who were similarly aged, indicating that the fruit datasets felt childish and the self-generated game datasets were more enjoyable.
Additionally, a majority of the time in the workshop was dedicated to assembling \textit{Data is Yours}. While further observations of how children engaged with their visualizations would have provided more in-depth insights, we found that assembling \textit{Data is Yours}\ helped children accumulate playful learning experiences that sparked an interest in and new comfort with visualization.
Extensive DVL encompasses complex knowledge and reasoning processes that go beyond the two-hour activities described here.
Thus, further studies will be required to see how children will use \textit{Data is Yours}\ to author data visualizations in an unconstrained and longitudinal setting.
Assessing children's DVL frequently focuses on reading, creating, and interpreting visualizations (\autoref{sec:toolkit}). However, DVL also encompasses critical engagement with visualizations, such as understanding when charts are misleading~\cite{camba2022deception, laurer2020}. Our toolkit currently requires facilitator feedback to let children know about misleading aspects of their visualizations. While we did not observe such errors, they are possible given the open-ended authoring functionality (e.g., using misleading axis values). To develop critical design literacy, future solutions could incorporate visual linting and similar verification strategies~\cite{hopkins2020visualint}.
\vspace{-2pt}
\subsection{A Research Agenda for Children's DVL}
\label{sec:research-agenda}
With any learning technology for children, research stresses the importance of three principles: ``low floor'' (how easy is the technology for children to learn and use?); ``high ceiling'' (does the technology advance children in making more complex projects?); and ``wide walls'' (how inclusive is the technology in supporting all children's interests?)~\cite{resnick2005design}.
Our observations from the workshops indicate that \textit{Data is Yours}\ successfully accomplished a \textit{low floor}. Young, inexperienced children were able to craft and interpret basic visualizations using the system. But our approach requires more consideration as to how to create a \textit{high ceiling}. E3 and E5 expressed concerns about the toolkit's current dependency on a smartphone. However, as children become older (e.g., teenagers), we anticipate they will not need the scaffolding from the physical panels to learn and interact with visualizations. Rather, research shows various ways to interact with data that are particularly suited for mobile devices~\cite{brehmer2019, brehmer2021interacting}. E3's and E5's comments indicate that future iterations of the toolkit could benefit from further considerations on how to better leverage state-of-the-art visualization techniques to engage older children.
Expanding \textit{Data is Yours}\ to additional contexts would also increase the variety and complexity of charts children can create.
Children and educators suggested several ways the toolkit can support \textit{wide walls}, including integrating the toolkit into activities that children naturally enjoy (\autoref{fig:future-vision}).
For instance, children can have a more sustained engagement with data at home by working with their family to recreate visualizations from news media or progressively tracking changes around them over time (e.g., tracking children's height). These activities, in turn, can seamlessly introduce children to different types of data (e.g., qualitative, geographic, categorical, temporal).
\textit{Data is Yours}\ can also be incorporated into group activities that already take place in informal learning spaces (e.g., science museums). The constructionist approach shows how produced artifacts can be ``shown, discussed, probed, and admired''~\cite{papert1993children}, serving as a locus for collaborative discussions. These actions align with our observations from the group workshop where children collaborated with each other and provided feedback on their peers' visualizations without being prompted to.
We believe expanding \textit{Data is Yours}\ into daily contexts and collaborative settings can help children be more critical of the data found around them.
\section{Conclusion}
We explored how a constructionist approach can foster children's data visualization literacy (DVL) using \textit{Data is Yours}.
Our observations from a series of workshops with children ($n=5$; 6--11 years old) and interviews with educators ($n=5$) reveal that physical crafting with visualizations benefits children in different ways depending on their past experiences with data. This exploratory study confirms the need to further investigate how to expand the ways children can interact with data and visualizations.
We reflect on findings and current research gaps to identify new challenges in children's DVL research.
We hope that our approach can inspire new ways for children to creatively and playfully interact with the visualizations and data that are woven into the fabrics of everyday life.
|
\section{Introduction}
The basic uncoupled space-time fractional diffusion equation
\begin{equation}\label{basicSTEq}
\partial^\alpha_t u+ (-\Delta)^s u= 0,\qquad 0< \alpha< 1,\,\,
0< s< 1
\end{equation}
models anomalous diffusion. Here the Caputo derivative $\partial^\alpha_t$
(time-fractional derivative) describes particle trapping phenomena; The fractional Laplacian $(-\Delta)^s$ (space-fractional derivative)
describes long particle jumps. See \cite{meerschaert2002governing} for more background information on (\ref{basicSTEq}). See \cite{chen2012space} for a probabilistic interpretation for (\ref{basicSTEq}).
In this paper, we consider uncoupled space-time fractional operators involving time-dependent coefficients which generalize $\partial^\alpha_t+ (-\Delta)^s$
and formulate the corresponding inverse problems.
\subsection{Main results}
We will first study the initial exterior problem
\begin{equation}\label{Cafracspacetime}
\left\{
\begin{aligned}
\partial^\alpha_t u+ \mathcal{R}^s_{A(t)} u + q(t)u&= 0
\quad \,\,\,\Omega\times (0, T)\\
u&= g\quad \,\,\,\Omega_e\times (0, T)\\
u&= 0\quad \,\,\,\Omega\times \{0\}
\end{aligned}
\right.
\end{equation}
where $\Omega$ is a bounded domain and $\Omega_e:= \mathbb{R}^n\setminus \Omega$.
Here the time-dependent fractional operator
$\mathcal{R}^s_{A(t)}$ is formally defined by
\begin{equation}\label{RsAt}
\mathcal{R}^s_{A(t)} u(x):= \lim_{\epsilon\to 0^+}
\int_{\mathbb{R}^n\setminus B_\epsilon(x)}(u(x)- R_{A(t)}(x, y)u(y))K(x,y)\,dy
\end{equation}
where
\begin{equation}\label{Kxy}
K(x, y):= c_{n, s}/|x-y|^{n+2s},
\end{equation}
$A(\cdot, t)$ is a time-dependent real vector-valued magnetic potential and
\begin{equation}\label{cost}
R_{A(t)}(x, y):= \cos((x-y)\cdot A(\frac{x+y}{2}, t)).
\end{equation}
The operator $\mathcal{R}^s_{A(t)}$ coincides with the fractional Laplacian $(-\Delta)^s$ when $A\equiv 0$.
Under appropriate regularity and support assumptions on magnetic potential $A$, electric potential $q$ and the exterior data $g$,
(\ref{Cafracspacetime}) is well-posed, so we can define the associated Dirichlet-to-Neumann map by
\begin{equation}\label{tDN}
\Lambda_{A, q}g:= \mathcal{R}^s_A u_g|_{\Omega_e\times (0, T)}.
\end{equation}
Our goal here is to determine both $A$ and $q$ from the exterior partial measurements of $\Lambda_{A, q}$.
The following theorem is the first main result in this paper, which can be viewed as a space-time fractional analogue of Theorem 1.1 in \cite{li2021fractional}.
\begin{thm}
Suppose $\Omega\subset B_r(0)$ for some constant $r> 0$, $\mathrm{supp}\,A_j(t)\subset\Omega$ for $t\in [0, T]$, $A_j\in C([0, T]; L^\infty(\mathbb{R}^n))$,
$q_j\in C([0, T]; L^\infty(\Omega))$,
$W_j$ are open sets s.t. $W_j\cap B_{3r}(0)= \emptyset$
($j= 1, 2$). Let
$$W^{(1, 2)}= \{\frac{x+ y}{2}: x\in W_1, y\in W_2\}.$$
Also assume $W^{(1, 2)}\setminus \Omega\neq \emptyset$. If
\begin{equation}\label{IDN}
\Lambda_{A_1, q_1}g|_{W_2\times (0, T)}
= \Lambda_{A_2, q_2}g|_{W_2\times (0, T)}
\end{equation}
for any $g\in C^\infty_c(W_1\times (0, T))$, then $A_1(t)= \pm A_2(t)$ and $q_1= q_2$ in $\Omega\times (0, T)$.
\end{thm}
We remark that the seemingly unnatural assumptions on $W_j$ in the statement are necessary (see the remark after Theorem 1.1 in \cite{li2021fractional}). However, if we replace $\mathcal{R}^s_{A(t)}$ by $(-\Delta)^s$ in (\ref{Cafracspacetime}), only interested in the determination of $q$, then the assumptions on $W_j$ in the statement can be simply replaced by $W_j\subset \Omega_e$ (see Subsection 3.2 for details).
We will next study the semilinear problem
\begin{equation}\label{Semifracspacetime}
\left\{
\begin{aligned}
\partial^\alpha_t u+ (-\Delta)^s u + a(x, t, u)&= 0
\quad \,\,\,\Omega\times (0, T)\\
u&= g\quad \,\,\,\Omega_e\times (0, T)\\
u&= 0\quad \,\,\,\Omega\times \{0\}
\end{aligned}
\right.
\end{equation}
where the nonlinearity satisfies
\begin{equation}\label{fracpower}
a(x, t, z)= \sum^m_{k=1}a_k(x, t)|z|^{b_k}z,
\end{equation}
$a_k\geq 0$ are smooth in $\bar{\Omega}\times [0, T]$, $b_1= 0$ and the powers $0< b_2<\cdots< b_m$ are not necessarily integers.
We will see that the Dirichlet-to-Neumann map
\begin{equation}\label{semiDN}
\Lambda_a: g\to (-\Delta)^su_g|_{\Omega_e\times (0, T)}
\end{equation}
is well-defined at least for $g\in C^\infty_c(\Omega_e\times (0, T))$.
Our goal here is to determine the nonlinearity $a$ from the exterior partial measurements of $\Lambda_a$.
The following theorem is the second main result in this paper, which can be viewed as a space-time fractional analogue of Theorem 1.1 in \cite{li2021inversediff}.
\begin{thm}
Let $W_1, W_2\subset \Omega_e$ be nonempty and open. If
\begin{equation}\label{semiIDN}
\Lambda_{a^{(1)}} g|_{W_2\times (0, T)}= \Lambda_{a^{(2)}} g|_{W_2\times (0, T)},
\qquad g\in C^\infty_c(W_1\times (0, T)),
\end{equation}
then $a^{(1)}_k= a^{(2)}_k$ in $\Omega\times (0, T)$, $k= 1,2,\cdots,m$.
\end{thm}
\subsection{Connection with earlier literature}
So far there have been many contributions in the study of inverse problems for time-fractional but space-local operators. See, for instance, \cite{kian2018global} for a work in this direction where the authors determined various time-independent smooth
coefficients appearing in the equation from the knowledge of the associated Dirichlet-to-Neumann map based on the inverse spectral theory.
The study of (Calder\'on type) inverse problems for space-fractional operators dates back to \cite{ghosh2020calderon}. See \cite{ghosh2017calderon, ghosh2020uniqueness, ghosh2021calder, covi2022higher, li2021determining, li2022elastic} for further results in this direction.
The proof of unique determination results in all these works heavily relies on exploiting the unique continuation property of the fractional Laplacian, which makes inverse problems for space-fractional operators often more manageable than their space-local counterparts. We adopt this framework here as well.
Our approaches will be mainly based on the ones introduced in \cite{li2021fractional, li2021inversediff}.
To the best knowledge of the author, there are very few existing rigorous works on inverse problems for space-time fractional operators. Calder\'on type problems for the coupled space-time fractional operator $(\partial_t-\Delta)^s$
has been studied in \cite{lai2020calderon, banerjee2022calder}. This fractional evolutionary operator actually behaves more like a fractional elliptic operator, which distinguishes it from our
uncoupled space-time fractional operators. Another work in this direction can be found in \cite{helin2020inverse} where the authors determined the Riemannian metric appearing in the uncoupled space-time fractional equation up to an isometry from the knowledge of the associated source-to-solution map in the setting of closed manifolds (also see \cite{feizmohammadi2021fractional, quan2022calder} for such geometric inverse problems for other fractional operators).
We also mention that our first inverse problem involving a magnetic potential can be viewed as a space-time fractional analogue of the classical magnetic Calder\'on problem studied in \cite{nakamura1995global, ferreira2007determining, krupchyk2014uniqueness}. See \cite{covi2020inverse, lai2021inverse} for the study of Calder\'on type problems for a different kind of fractional magnetic operators.
\subsection{Organization}
The rest of this paper is organized in the following way. In Section 2, we will summarize the background knowledge. In Section 3, we will first show the well-posedness of (\ref{Cafracspacetime}); Then we will derive the integral identity for the Dirichlet-to-Neumann maps and the Runge approximation property
of the space-time fractional operator to prove Theorem 1.1. In Section 4, we will first prove an a priori $L^\infty$ estimate and the well-posedness of (\ref{Semifracspacetime}); Then we will derive a linearization result to prove Theorem 1.2.
\medskip
\noindent \textbf{Acknowledgements.} The author would like to thank Professor Gunther Uhlmann for helpful discussions.
\section{Preliminaries}
Throughout this paper we use the following notations.
\begin{itemize}
\item Fix the space dimension $n\geq 2$.
\item Fix the fractional powers $0< \alpha< 1$ and $0< s< 1$.
\item $\Omega$ denotes a bounded domain with smooth boundary and
$\Omega_e:= \mathbb{R}^n\setminus\bar{\Omega}$.
\item $B_r(0)$ denotes the open ball centered at the origin with radius $r> 0$
in $\mathbb{R}^n$.
\item $c, C, C', C_1,\cdots$ denote positive constants.
\item $X$ denotes a Banach space and $X^*$ denotes the continuous dual space of $X$.
\item $\langle f, u\rangle$ denotes the standard $L^2$ distributional pairing
between $f$ and $u$ if $u$ (resp., $f$) is a (spatial) $n$-variable function (resp., functional).
\item For functions $f(t), g(t)$ ($t> 0$), we use the standard convolution notation
$$(f* g)(t)= \int^t_0 f(\tau)g(t-\tau)\,d\tau.$$
\end{itemize}
\subsection{Function spaces}
Throughout this paper we refer all function spaces to real-valued function spaces.
We use $H^r$ to denote the standard Sobolev space $W^{r,2}$.
We use $C^\alpha$ to denote the standard H\"older space $C^{0, \alpha}$.
Let $U$ be an open set in $\mathbb{R}^n$. Let $F$ be a closed set in $\mathbb{R}^n$. Then
$$H^r(U):= \{u|_U: u\in H^r(\mathbb{R}^n)\},\qquad
H^r_F(\mathbb{R}^n):=
\{u\in H^r(\mathbb{R}^n): \mathrm{supp}\,u\subset F\},$$
$$\tilde{H}^r(U):=
\mathrm{the\,\,closure\,\,of}\,\, C^\infty_c(U)\,\,\mathrm{in}\,\, H^r(\mathbb{R}^n).$$
For $r\in\mathbb{R}$, we have the natural identifications
$$H^{-r}(\mathbb{R}^n)= H^r(\mathbb{R}^n)^*,\quad \tilde{H}^r(\Omega)= H^r_{\bar{\Omega}}(\mathbb{R}^n),\quad
H^{-r}(\Omega)= \tilde{H}^r(\Omega)^*.$$
We use $C([0, T]; X)$ (resp., $C^\alpha([0, T]; X)$)
to denote the space consisting of the corresponding Banach space-valued continuous (resp., $C^\alpha$-continuous) functions on $[0, T]$.
$L^2(0, T; X)$ (resp., $H^1(0, T; X)$) denotes the space consisting of the corresponding Banach space-valued $L^2$-functions (resp., $H^1$-functions).
\subsection{Riemann-Liouville and Caputo derivatives}
Throughout this paper we use the notation
$$\phi_\alpha(t):= \frac{t^{\alpha-1}}{\Gamma(\alpha)},\qquad t>0$$
where $\Gamma$ is the standard Gamma function. It is straightforward to verify that
\begin{equation}\label{alpha1}
\phi_\alpha * \phi_{1-\alpha}= 1,\qquad t>0.
\end{equation}
The left Riemann-Liouville fractional integral of order $\alpha$ is defined by
$$I^\alpha_{0,t}u:= \phi_\alpha* u$$
and the left Riemann-Liouville fractional derivative of order $\alpha$ is defined by
$$D^\alpha_{0,t}u:= \partial_t(I^{1-\alpha}_{0,t}u)
= \partial_t(\phi_{1-\alpha}* u).$$
Correspondingly, the right Riemann-Liouville fractional integral of order $\alpha$ is defined by
$$I^\alpha_{t,T}u:= \frac{1}{\Gamma(\alpha)}\int^T_t (\tau- t)^{\alpha-1}u(\tau)\,d\tau$$
and the right Riemann-Liouville fractional derivative of order $\alpha$ is defined by
$$D^\alpha_{t,T}u:= -\partial_t(I^{1-\alpha}_{t,T}u).$$
The (left) Caputo fractional derivative of order $\alpha$ is defined by
$$\partial^\alpha_t u(t):= D^\alpha_{0,t}(u(t)- u(0)).$$
It is straightforward to verify that $\partial^\alpha_t u= I^{1-\alpha}_{0,t}(\partial_t u)$ if $u$ is sufficiently regular.
For (scalar-valued) functions $f(t), g(t)$, we have the following integration by parts formula (see, for instance, (3.6) in \cite{Warma2019approx})
\begin{equation}\label{fracibp}
\int^T_0 g\partial^\alpha_t f= \int^T_0 f D^\alpha_{t,T}g +
(f I^{1-\alpha}_{t,T}g)|^{t= T}_{t= 0}.
\end{equation}
Here the value of $I^{1-\alpha}_{t,T}g$ at $t=T$ should be interpreted in the limit sense. We also have the following inequality
(see, for instance, Lemma 2.1 in \cite{vergara2017stability})
\begin{equation}\label{Hconvex}
\partial^\alpha_t(H(f(t))\leq H'(f(t))\partial^\alpha_t f(t)
\end{equation}
where $H\in C^1(\mathbb{R})$ is convex. We note that both (\ref{fracibp})
and (\ref{Hconvex}) can be generalized for Banach space-valued $f(t), g(t)$
if we replace the pointwise product by the distributional pairing.
Let $L$ be a sectorial operator (i.e. a closed, densely defined linear operator which generates an analytic semigroup) in $X$. For sufficiently regular $f$, we can take Laplace transform to show that
\begin{equation}\label{SPformula}
u(t)= S_{\alpha}(t)u_0+ \int^t_0 P_{\alpha}(t-\tau)f(\tau)\,d\tau
\end{equation}
satisfies the continuity at $t=0$, and this formula gives the solution of
\begin{equation}\label{CdeL}
\partial^\alpha_t u= Lu+ f,\qquad u(0)= u_0;
\end{equation}
We can also take Laplace transform to show that
\begin{equation}\label{PPformula}
u(t)= P_{\alpha}(t)h_0+ \int^t_0 P_{\alpha}(t-\tau)f(\tau)\,d\tau
\end{equation}
satisfies that $I^\alpha_{0,t}u$ is continuous at $t=0$, and this formula gives the solution of
\begin{equation}\label{RLdeL}
D^\alpha_{0,t} u= Lu+ f,\qquad I^\alpha_{0,t}u(0)= h_0.
\end{equation}
Here $\{S_{\alpha}(t)\}$, $\{P_{\alpha}(t)\}$ are the resolvent families associated with $L$, which can be expressed by integrals involving the semigroup $\{S(t)\}$
generated by $L$ and the Wright type function $\Phi_\alpha(t)$. Since we are not going to use the explicit expressions of $S_{\alpha}(t), P_{\alpha}(t)$, we refer interested readers to (2.1.9) in \cite{gal2020fractional} for details.
\subsection{Space-fractional operators}
It is well-known that the fractional Laplacian $(-\Delta)^s$ can be defined by
\begin{equation}\label{fracLa}
(-\Delta)^s u(x):= c_{n, s}\lim_{\epsilon\to 0^+}
\int_{\mathbb{R}^n\setminus B_\epsilon(x)}\frac{u(x)-u(y)}{|x-y|^{n+2s}}\,dy
\end{equation}
and for each $r\in \mathbb{R}$, we have
$$(-\Delta)^s: H^r(\mathbb{R}^n)\to H^{r-2s}(\mathbb{R}^n).$$
We will see that many properties of the fractional Laplacian are preserved under the perturbation by the magnetic potential $A$.
Throughout this paper we assume
$$A\in C([0, T]; L^\infty(\mathbb{R}^n)),\qquad\mathrm{supp}\,A(t)\subset \Omega$$
for each $t\in [0,T]$ and
$q\in C([0, T]; L^\infty(\Omega))$.
It has been shown that (see (1) and (6) in \cite{li2021fractional}) for $u, v\in H^s(\mathbb{R}^n)$,
\begin{equation}\label{biRsAt}
\langle \mathcal{R}^s_{A(t)}u, v\rangle:=
2\int_{\mathbb{R}^n}\int_{\mathbb{R}^n}(u(x)- R_{A(t)}(x, y)u(y))v(x)K(x,y)\,dxdy,
\end{equation}
and $\mathcal{R}^s_{A(t)}$ is symmetric:
\begin{equation}\label{RsAsym}
\langle \mathcal{R}^s_{A(t)}u, v\rangle
= \langle \mathcal{R}^s_{A(t)}v, u\rangle.
\end{equation}
We define the bilinear form $B_t$
associated with $A, q$ by
\begin{equation}\label{tdbf}
B_t[u, v]:= \langle \mathcal{R}^s_{A(t)}u, v\rangle+ \int_\Omega q(t)uv,\qquad
t\in [0, T].
\end{equation}
\begin{lemma}
$B_t$ is bounded:
\begin{equation}\label{Bbdd}
|B_t[u,v]|\leq C_0||u||_{H^s}||v||_{H^s},\qquad u, v\in H^s(\mathbb{R}^n).
\end{equation}
$B_t$ is coercive:
\begin{equation}\label{Bcoer}
B_t[u, u]\geq c_1||u||^2_{H^s}- c_2||u||^2_{L^2},\qquad
u\in \tilde{H}^s(\Omega).
\end{equation}
Here $C_0, c_1, c_2$ do not depend on $t$.
\end{lemma}
The estimates above have been proved in Section 2 in \cite{li2021fractional}.
\begin{prop}
Suppose $\Omega\subset B_r(0)$ for some $r> 0$, $W$ is a nonempty open set s.t. $$W\cap B_{3r}(0)= \emptyset.$$
(i)\, For $g\in C^\infty_c(W\times (0, T))$, we have
$$\mathcal{R}^s_A g|_{\Omega\times (0, T)}
= (-\Delta)^s g|_{\Omega\times (0, T)};$$
(ii) (Proposition 2.4 in \cite{li2021fractional}) If
$$u\in \tilde{H}^s(\Omega),\qquad \mathcal{R}^s_{A(t)}u|_W= 0,$$
then $u= 0$ in $\mathbb{R}^n$.
\end{prop}
It is straightforward to verify (i) based on the support assumptions on $A$ and $g$. (ii) is based on the following unique continuation property of $(-\Delta)^s$ (see Theorem 1.2 in \cite{ghosh2020calderon}).
\begin{prop}
Let $u\in H^r(\mathbb{R}^n)$ for some $r\in \mathbb{R}$. Let $W\subset \mathbb{R}^n$ be nonempty and open. If $$(-\Delta)^su= u= 0\quad\text{in}\,\,W,$$
then $u= 0$ in $\mathbb{R}^n$.
\end{prop}
\section{Linear problem}
\subsection{Forward problem}
We will study (\ref{Cafracspacetime}) as well as the related problem
\begin{equation}\label{RLfracspacetime}
\left\{
\begin{aligned}
D^\alpha_{0,t} u+ \mathcal{R}^s_{A(t)} u + q(t)u&= 0
\quad \,\,\,\Omega\times (0, T)\\
u&= g\quad \,\,\,\Omega_e\times (0, T)\\
I^\alpha_{0,t}u&= 0\quad \,\,\,\Omega\times \{0\}
\end{aligned}
\right.
\end{equation}
for $g\in C^\infty_c(\Omega_e\times (0, T))$.
By using the substitution $u= w+g$,
(\ref{Cafracspacetime}) can be converted into the initial value problem
\begin{equation}\label{Caivp}
\left\{
\begin{aligned}
\partial^\alpha_t w+ \mathcal{R}^s_{A(t)} w + q(t)w&= f
\quad \,\,\,\Omega\times (0, T)\\
w&= 0\quad \,\,\,\Omega\times \{0\}
\end{aligned}
\right.
\end{equation}
and (\ref{RLfracspacetime}) can be converted into the (integral) initial value problem
\begin{equation}\label{RLivp}
\left\{
\begin{aligned}
D^\alpha_{0,t} w+ \mathcal{R}^s_{A(t)} w + q(t)w&= f
\quad \,\,\,\Omega\times (0, T)\\
I^\alpha_{0,t}w&= 0\quad \,\,\,\Omega\times \{0\}
\end{aligned}
\right.
\end{equation}
Based on Lemma 2.1 in Subsection 2.3 and Theorem 3.1 in \cite{zacher2009weak}, we have the following well-posedness result.
\begin{prop}
Suppose $f\in L^2(0, T; H^{-s}(\Omega))$.
Then (\ref{RLivp}) has a unique solution satisfying
$$w\in L^2(0, T; \tilde{H}^s(\Omega)),\qquad
I^\alpha_{0,t}w\in H^1(0, T; H^{-s}(\Omega)).$$
This actually implies
$I^\alpha_{0,t}w\in C([0,T]; L^2(\Omega))$ (see Theorem 2.1 in \cite{zacher2009weak}).
\end{prop}
We note that
$$0\leq 1- R_{A(t)}(x,y)=
2\sin^2(\frac{1}{2}(x-y)\cdot A(\frac{x+y}{2}, t))\leq
C_A\min\{1, |x-y|^2\}$$
and (based on (\ref{fracLa}) and (\ref{RsAt})) we have
$$|((-\Delta)^s- \mathcal{R}^s_{A(t)}) u(x)|\leq
\int (1- R_{A(t)}(x, y))K(x, y)|u(y)|dy.$$
Since we have
$$\int (1- R_{A(t)}(x, y))K(x, y)dy\leq C,\qquad x\in \mathbb{R}^n,$$
by the generalized Young’s inequality (see, for instance, Proposition 0.10 in \cite{folland1995introduction}), we get
\begin{equation}\label{L2bdd}
||((-\Delta)^s- \mathcal{R}^s_{A(t)})u||_{L^2(\mathbb{R}^n)}\leq
C||u||_{L^2(\mathbb{R}^n)}.
\end{equation}
Now we know that the operator ${R}^s_{A(t)} + q(t)$
can be written as
$$(-\Delta)^s+ F(t,\cdot)$$
where the linear operator
$$F(t,\cdot):=
\mathcal{R}^s_{A(t)}- (-\Delta)^s+ q(t)$$
is $L^2$ bounded (uniformly in $t$).
We also note that for sufficiently regular $f$, both (\ref{RLivp}) and (\ref{Caivp}) can be written as the integral equation (see (\ref{SPformula}) and (\ref{PPformula}))
$$w(t)= \int^t_0 P_{\alpha}(t-\tau)(f(\tau)- F(\tau, w(\tau))\tau\,d\tau.$$
Here $\{P_\alpha(t)\}$ is the resolvent family associated with
the sectorial operator
$$L: w\to -(-\Delta)^s w|_\Omega,\qquad D(L)= \{w\in \tilde{H}^s(\Omega):
(-\Delta)^s w|_\Omega\in L^2(\Omega)\}.$$
Based on Theorem 1.1 in \cite{kojima2021existence}, we have the following well-posedness result.
\begin{prop}
Let $f\in C^\alpha([0, T]; L^2(\Omega))$
Then (\ref{Caivp}) (or (\ref{RLivp})) has a unique solution
$$w\in C([0,T]; L^2(\Omega)).$$
\end{prop}
We remark that the well-posedness result above actually holds true for the general Lipschitz type nonlinear map $F(t,\cdot)$ and the regularity assumption on $f$ can be weakened. We refer readers to \cite{kojima2021existence} for more details.
In the rest of this section, we will assume $\Omega\subset B_r(0)$ for some constant $r> 0$ and $W$ is a nonempty open set in $\mathbb{R}^n$ s.t. $W\cap B_{3r}(0)= \emptyset$ unless otherwise stated.
We note that for $g\in C^\infty_c(W\times (0, T))$, by Proposition 2.2 (i)
we have
$$f:= -\mathcal{R}^s_A g|_{\Omega\times (0, T)}
= -(-\Delta)^s g|_{\Omega\times (0, T)},$$
which is smooth over $\bar{\Omega}\times [0,T]$.
Based on Proposition 3.1 and 3.2, we have the following well-posedness result
\begin{prop}
Let $g\in C^\infty_c(W\times (0, T))$
Then (\ref{Cafracspacetime})
(or (\ref{RLfracspacetime})) has a unique solution $u$ with $w:= u-g$ satisfying
$$w\in L^2(0, T; \tilde{H}^s(\Omega))\cap C([0,T]; L^2(\Omega)),\qquad
I^\alpha_{0,t}w\in H^1(0, T; H^{-s}(\Omega))\cap C([0,T]; L^2(\Omega)).$$
\end{prop}
We denote the solution operator $g\to u_g$ associated with (\ref{Cafracspacetime}) by $P_{A, q}$. We note that we have the same
well-posedness result for the dual problem
\begin{equation}\label{dualRLfracspacetime}
\left\{
\begin{aligned}
D^\alpha_{t,T} u+ \mathcal{R}^s_{A(t)} u + q(t)u&= 0
\quad \,\,\,\Omega\times (0, T)\\
u&= g\quad \,\,\,\Omega_e\times (0, T)\\
I^\alpha_{t,T}u&= 0\quad \,\,\,\Omega\times \{T\}
\end{aligned}
\right.
\end{equation}
and we denote the associated solution operator by $P^*_{A, q}$.
\subsection{Inverse problem}
\subsubsection{DN map}
Recall that we defined the Dirichlet-to-Neumann map $\Lambda_{A, q}$ associated with (\ref{Cafracspacetime}) in (\ref{tDN}). We also define
the Dirichlet-to-Neumann map $\Lambda^*_{A, q}$ associated with
(\ref{dualRLfracspacetime}) by
\begin{equation}\label{dualtDN}
\Lambda^*_{A, q}g:= \mathcal{R}^s_A (P^*_{A, q}g)|_{\Omega_e\times (0, T)}.
\end{equation}
The well-posed results ensure that $\Lambda_{A, q}$ and $\Lambda^*_{A, q}$
are well-defined at least for
$g\in C^\infty_c(W\times (0, T))$.
Let $g\in C^\infty_c(W_1\times (0, T))$ and
$h\in C^\infty_c(W_2\times (0, T))$.
By the definition of $B_t$, $P_{A, q}$ and $\Lambda_{A, q}$ we have
$$\int^T_{0}\langle \Lambda_{A, q}g(t), h(t)\rangle\,dt
= \int^T_{0}\langle \mathcal{R}^s_{A(t)}u(t), \tilde{h}(t)\rangle\,dt
- \int^T_{0}\langle \mathcal{R}^s_{A(t)}u(t), \tilde{h}(t)- h(t)\rangle\,dt$$
$$= \int^T_{0}\langle \mathcal{R}^s_{A(t)}u(t), \tilde{h}(t)\rangle\,dt+
\int^T_{0}\langle \partial^\alpha_t u(t)+ q(t)u(t), \tilde{h}(t)- h(t)\rangle\,dt$$
\begin{equation}\label{DNint1}
= \int^T_{0}B_t[u(t), \tilde{h}(t)]\,dt
+ \int^T_{0}\langle\partial^\alpha_t u(t), \tilde{h}(t)\rangle_\Omega\,dt\
\end{equation}
for any $\tilde{h}$ satisfying
$\tilde{h}- h\in L^2(0, T; \tilde{H}^s(\Omega))$.
Here $u:= P_{A, q}g$, $w:= u- g$ and
$$\langle\partial_t u(t), \tilde{h}(t)\rangle_\Omega=
\langle\partial_t w(t), \tilde{h}(t)- h(t)\rangle.$$
Similarly we have
\begin{equation}\label{DNint2}
\int^T_{0}\langle \Lambda^*_{A, q}h(t), g(t)\rangle\,dt
= \int^T_{0}B_t[u^*(t), \tilde{g}(t)]\,dt
+ \int^T_{0}\langle D^\alpha_{t,T}u^*(t), \tilde{g}(t)\rangle_\Omega\,dt
\end{equation}
where $u^*:= P^*_{A, q}h$ for any $\tilde{g}$ satisfying
$\tilde{g}- g\in L^2(0, T; \tilde{H}^s(\Omega))$.
\begin{prop}
For $g\in C^\infty_c(W_1\times (0, T))$ and
$h\in C^\infty_c(W_2\times (0, T))$, we have
\begin{equation}\label{dualsym}
\int^T_{0}\langle \Lambda_{A, q}g(t), h(t)\rangle\,dt
= \int^T_{0}\langle \Lambda^*_{A, q}h(t), g(t)\rangle\,dt.
\end{equation}
\end{prop}
\begin{proof}
Let $\tilde{h}= u^*$ in (\ref{DNint1}) and
let $\tilde{g}= u$ in (\ref{DNint2}).
Since $u(0)= I^\alpha_{t,T}u^*(T)= 0$, we have
$$\int^T_{0}\langle \Lambda_{A, q}g(t), h(t)\rangle\,dt
- \int^T_{0}\langle \Lambda^*_{A, q}h(t), g(t)\rangle\,dt$$
$$= \int^T_{0}\langle\partial^\alpha_t u(t), u^*(t)\rangle_\Omega
- \langle D^\alpha_{t,T}u^*(t), u(t)\rangle_\Omega\,dt
=\langle u(t), I^\alpha_{t,T}u^*(t)\rangle_\Omega|^{t= T}_{t= 0}= 0$$
based on the symmetry of $B_t$ and the integration by parts formula (\ref{fracibp}).
\end{proof}
Now we are ready to derive the integral identities for Dirichlet-to-Neumann maps.
For $g_j\in C^\infty_c(W_j\times (0, T))$ ($j= 1, 2$),
let $u_1= P_{A_1, q_1}(g_1)$
and $u^*_2= P^*_{A_2, q_2}(g_2)$,\\
i.e. $u_1$ is the solution of
\begin{equation}
\left\{
\begin{aligned}
\partial_t^\alpha u+ \mathcal{R}^s_{A_1(t)} u+ q_1(t)u &= 0\quad \,\,\,\text{in}\,\,\Omega\times (0, T)\\
u&= g_1 \quad \text{in}\,\,\Omega_e\times (0, T)\\
u&= 0\quad \,\,\,\text{in}\,\,\Omega\times \{0\}\\
\end{aligned}
\right.
\end{equation}
and $u^*_2$ is the solution of
\begin{equation}
\left\{
\begin{aligned}
D^\alpha_{t,T} u+ \mathcal{R}^s_{A_2(t)} u+ q_2(t)u &= 0\quad \,\,\,\text{in}\,\,\Omega\times (0, T)\\
u&= g_2 \quad \text{in}\,\,\Omega_e\times (0, T)\\
I^\alpha_{t,T}u&= 0\quad \,\,\,\text{in}\,\,\Omega\times \{T\}.\\
\end{aligned}
\right.
\end{equation}
Then by (\ref{DNint1}), (\ref{DNint2}), (\ref{fracibp}) and Proposition 3.4 we have
$$\int^T_{0}\langle \Lambda_{A_1, q_1}g_1(t), g_2(t)\rangle-
\langle\Lambda_{A_2, q_2}g_1(t), g_2(t)\rangle\,dt$$
$$= \int^T_{0}\langle \Lambda_{A_1, q_1}g_1(t), g_2(t)\rangle\,dt
-\int^T_{0}\langle \Lambda^*_{A_2, q_2}g_2(t), g_1(t)\rangle\,dt$$
$$= \int^T_{0}B^{(1)}_t[u_1(t), u^*_2(t)]
+ \langle\partial^\alpha_t u_1(t), u^*_2(t)\rangle_\Omega\,dt
- \int^T_{0}B^{(2)}_t[u^*_2(t), u_1(t)]
+ \langle D^\alpha_{t,T}u^*_2(t), u_1(t)\rangle_\Omega\,dt$$
$$=\int^T_{0}B^{(1)}_t[u_1(t), u^*_2(t)]\,dt
- \int^T_{0}B^{(2)}_t[u_1(t), u^*_2(t)]\,dt$$
\begin{equation}\label{intdif}
= \int^T_{0}\int_{\mathbb{R}^n}\int_{\mathbb{R}^n}
G(x, y, t)u_1(y, t)u^*_2(x, t)
-\int^T_{0}\int_\Omega(q_2- q_1)u_1u^*_2
\end{equation}
where
$$G(x, y, t):= 2(R_{A_2(t)}(x, y)- R_{A_1(t)}(x, y))K(x, y).$$
\subsubsection{Runge approximation}
\begin{prop}
Suppose $\Omega\subset B_r(0)$ for some constant $r> 0$ and $W$ is a nonempty open set in $\mathbb{R}^n$ s.t. $W\cap B_{3r}(0)= \emptyset$. Then
$$S:= \{P_{A, q}g|_{\Omega\times (0, T)}:
g\in C^\infty_c(W\times (0, T))\},$$
$$S^*:= \{P^*_{A, q}g|_{\Omega\times (0, T)}:
g\in C^\infty_c(W\times (0, T))\}$$
are dense in $L^2(\Omega\times (0, T))$.
\end{prop}
\begin{proof}
By the Hahn-Banach Theorem, it suffices to show that:
If $f\in L^2(\Omega\times (0, T))$ and $\int^T_{0}\int_\Omega wf= 0$ for all $w\in S$, then $f= 0$ in $\Omega\times (0, T)$.
In fact, for an arbitrary given $f\in L^2(\Omega\times (0, T))$, by (the dual version of) Proposition 3.1 we know that the solution of
\begin{equation}
\left\{
\begin{aligned}
D^\alpha_{t,T} v+ \mathcal{R}^s_{A(t)} v+ q(t)v &= f\quad \text{in}\,\,\Omega\times (0, T)\\
I^\alpha_{t,T}v&= 0\quad \text{in}\,\,\Omega\times \{T\}.\\
\end{aligned}
\right.
\end{equation}
satisfies
$$v\in L^2(0, T; \tilde{H}^s(\Omega)),\qquad
I^\alpha_{t,T}v\in H^1(0, T; H^{-s}(\Omega))\cap C([0,T]; L^2(\Omega)).$$
For $g\in C^\infty_c(W\times (0, T))$,
write $u_g:= P_{A, q}g$, then we have
$$\int^T_{0}\int_\Omega u_g f= \int^T_{0}\langle D^\alpha_{t,T}v(t)+
\mathcal{R}^s_{A(t)} v(t)+ q(t)v(t), u_g(t)- g(t)\rangle\,dt$$
$$= \int_{0}^T\langle\partial^\alpha_t u_g(t), v(t)\rangle
+ B_t[u_g(t), v(t)]\,dt
- \int_{0}^T\langle\mathcal{R}^s_{A(t)}g(t), v(t)\rangle\,dt$$
\begin{equation}\label{RAPid}
= -\int_{0}^T\langle\mathcal{R}^s_{A(t)}v(t), g(t)\rangle\,dt.
\end{equation}
The first equality holds since $u_g- g\in L^2(0, T; \tilde{H}^s(\Omega))$;
The second equality holds since $u_g(0)= I^\alpha_{t,T}v(T)= 0$ ensures
$$\int_{0}^T\langle D^\alpha_{t,T}v(t), u_g(t)\rangle=
\int_{0}^T\langle\partial^\alpha_t u_g(t), v(t)\rangle;$$
The last equality holds since $v\in L^2(0, T; \tilde{H}^s(\Omega))$ and $u_g$ is the solution of (\ref{Cafracspacetime}).
Hence, if $\int^T_{0}\int_\Omega wf= 0$ for all $w\in S$, then
(\ref{RAPid}) yields
$$\int_{0}^T\langle\mathcal{R}^s_{A(t)}v(t), g(t)\rangle\,dt= 0,\qquad
g\in C^\infty_c(W\times (0, T))$$
so for each $t$ we have
$$v(t)\in \tilde{H}^s(\Omega),
\qquad \mathcal{R}^s_{A(t)}v(t)|_W= 0,$$
which implies $v(t)= 0$ in $\mathbb{R}^n$ for each $t$ by Proposition 2.2 (ii) and thus $f= 0$ in $\Omega\times (0, T)$.
Similarly we can show $S^*$ is dense in $L^2(\Omega\times (0, T))$.
\end{proof}
We remark that a similar argument has been used to prove the approximate controllability result for $\partial^\alpha_t+ (-\Delta)^s$ (see Theorem 2.6 in \cite{Warma2019approx}).
\subsubsection{Proof of Theorem 1.1}
We note that time-fractional derivatives do not appear in the integral identity
(\ref{intdif}) so we can use the same argument as the one in \cite{li2021fractional} to prove Theorem 1.1 based on the Runge approximation property (Proposition 3.5). To avoid redundancy, we will skip inessential steps in the proof. We refer readers to Subsection 4.3 in \cite{li2021fractional} for full details.
\begin{proof}
We write $u_1= P_{A_1, q_1}(g_1)$ and $u^*_2= P^*_{A_2, q_2}(g_2)$
for $g_j\in C^\infty_c(W_j\times (0, T))$.
Based on (\ref{intdif}) and the assumptions stated in Theorem 1.1, we have
\begin{equation}\label{Gid}
\int^T_{0}\int_\Omega\int_\Omega G(x, y, t)u_1(y, t)u^*_2(x, t)
=\int^T_{0}\int_\Omega(q_2- q_1)u_1u^*_2.
\end{equation}
\textbf{Determine $A$}: We fix open sets $\Omega_j\subset \Omega$ s.t.
$\Omega_1\cap \Omega_2= \emptyset$.
We fix $\phi_j\in C^\infty_c(\Omega_j)$ and the constants
$a, b\in (0, T)$. We define
$$\tilde{\phi}_j(x, t):= 1_{[a, b]}(t)\phi_j(x).$$
By Proposition 3.5, we can choose
$g_1\in C^\infty_c(W_1\times (0, T))$ (resp. $g_2\in C^\infty_c(W_2\times (0, T))$) s.t. $u_1$ (resp. $u^*_2$) approximates
$\tilde{\phi}_1$ (resp. $\tilde{\phi}_2$) arbitrarily in $L^2(\Omega\times (0, T))$.
The key observation is that $\phi_1\phi_2= 0$ while $\phi_1\otimes\phi_2\neq 0$ in general. This enables us to take the limit on both sides of (\ref{Gid}) to get
$$\int^b_a\int_{\Omega_1}\int_{\Omega_2} G(x, y, t)
\phi_1(y)\phi_2(x)\,dxdydt= 0.$$
Since the choices of $\phi_1, \phi_2$ and $[a, b]$ are arbitrary, we can conclude that
$G(x, y, t)= 0$ for $x, y \in\Omega$ whenever $x\neq y$. Thus we know that
\begin{equation}\label{RA12}
R_{A_1(t)}(x, y)= R_{A_2(t)}(x, y),\quad x, y\in\Omega
\end{equation}
for each $t$. Then we can show that $A_1(t)= \pm A_2(t)$.
\textbf{Determine $q$}: Now (\ref{Gid}) becomes
$$\int^T_{0}\int_\Omega(q_2- q_1)u_1u^*_2= 0.$$
For $f\in L^2(\Omega\times (0, T))$,
again by Proposition 3.5, we can choose
$g_1\in C^\infty_c(W_1\times (0, T))$ (resp. $g_2\in C^\infty_c(W_2\times (0, T))$) s.t. $u_1$ (resp. $u^*_2$) approximates
$f$ (resp. constant function $1$) arbitrarily in $L^2(\Omega\times (0, T))$.
Then we take the limit to get
$$\int^T_{0}\int_\Omega(q_1-q_2)f= 0.$$
We conclude that $q_1= q_2$ since the choice of $f$ is arbitrary.
\end{proof}
We remark that if we are only interested in the problem
\begin{equation}\label{Laspacetime}
\left\{
\begin{aligned}
\partial^\alpha_t u+ (-\Delta)^s u + q(t)u&= 0
\quad \,\,\,\Omega\times (0, T)\\
u&= g\quad \,\,\,\Omega_e\times (0, T)\\
u&= 0\quad \,\,\,\Omega\times \{0\}
\end{aligned}
\right.
\end{equation}
instead of the more general (\ref{Cafracspacetime}), then the inverse problem will be reduced to the determination of $q$ from partial knowledge of $\Lambda_q$ (i.e. $\Lambda_{0, q}$), and in this case the assumptions on $W_j$ in the statement of Theorem 1.1 are not necessary for the unique determination.
In fact, based on the unique continuation property of fractional Laplacian (Proposition 2.3) instead of Proposition 2.2, we can use the same argument as the one in the proof of Proposition 3.5 to prove the following Runge approximation property.
\begin{prop}
Let $W\subset \Omega_e$ be nonempty and open. Then
$$S:= \{P_q g|_{\Omega\times (0, T)}:
g\in C^\infty_c(W\times (0, T))\},\qquad
S^*:= \{P^*_q g|_{\Omega\times (0, T)}:
g\in C^\infty_c(W\times (0, T))\}$$
are dense in $L^2(\Omega\times (0, T))$. (Here $P_q:= P_{0, q}$ and
$P^*_q:= P^*_{0, q}$.)
\end{prop}
Then we can use the argument in the second half of the proof of Theorem 1.1 to prove the following unique determination theorem.
\begin{prop}
Let $W_1, W_2\subset \Omega_e$ be nonempty and open. If
$$\Lambda_{q_1} g|_{W_2\times (0, T)}= \Lambda_{q_2} g|_{W_2\times (0, T)},
\qquad g\in C^\infty_c(W_1\times (0, T)),$$
then $q_1= q_2$ in $\Omega\times (0, T)$.
\end{prop}
\section{Semilinear problem}
\subsection{Forward problem}
Now we turn to the semilinear problem (\ref{Semifracspacetime}).
We first prove the following a priori $L^\infty$ estimate.
\begin{prop}
Let $g\in C^\infty_c(\Omega_e\times (0, T))$.
Suppose $u\in C([0, T]; L^2(\mathbb{R}^n))$ is a solution of
\begin{equation}\label{fgfracpara}
\left\{
\begin{aligned}
\partial^\alpha_t u+ (-\Delta)^s u + a(x, t, u)&= f
\quad \,\,\,\Omega\times (0, T)\\
u&= g\quad \,\,\,\Omega_e\times (0, T)\\
u&= 0\quad \,\,\,\Omega\times \{0\}.
\end{aligned}
\right.
\end{equation}
Then we have
$$||u||_{L^\infty}\leq \frac{T^\alpha}{\Gamma(\alpha+1)}||f||_{L^\infty(\Omega\times (0, T))}
+ ||g||_{L^\infty(\Omega_e\times (0, T))}.$$
\end{prop}
\begin{proof}
We fix $\varphi\in C^\infty_c(\mathbb{R}^n)$ s.t. $0\leq \varphi\leq 1$
and $\varphi= 1$ on $\bar\Omega \cup \mathrm{supp}_x g$. We define
$$\tilde{\varphi}(x, t):= (||f||_{L^\infty(\Omega\times (0,T))}
\frac{t^\alpha}{\Gamma(\alpha+1)}
+ ||g||_{L^\infty(\Omega_e\times (0, T))})\varphi(x).$$
Clearly $(-\Delta)^s\varphi\geq 0$ in $\Omega$ from the pointwise definition
(\ref{fracLa}) of $(-\Delta)^s$. Also note that
$$\partial^\alpha_t(t^\alpha)= \Gamma(\alpha+1),\qquad
\partial^\alpha_t(c)= 0$$
so we have
$$\partial^\alpha_t\tilde{\varphi}+(-\Delta)^s\tilde{\varphi}+ a(x, t, \tilde{\varphi})\geq ||f||_{L^\infty(\Omega\times (0, T))}$$
in $\Omega\times (0, T)$. Now we consider $\tilde{u}:= u- \tilde{\varphi}$.
Note that $\tilde{u}\leq 0$ in
$\Omega_e\times (0, T)$, $\tilde{u}\leq 0$ at $t= 0$ and
\begin{equation}\label{tildeest}
\partial^\alpha_t\tilde{u}+(-\Delta)^s\tilde{u}
+ a(x, t, u)- a(x, t, \tilde{\varphi})\leq 0
\end{equation}
in $\Omega\times (0, T)$. We write $\tilde{u}= \tilde{u}^+ - \tilde{u}^-$
where $\tilde{u}^{\pm}= \max\{\pm \tilde{u}, 0\}$. Then $\tilde{u}^+= 0$ in $\Omega_e\times (0, T)$ and $\tilde{u}^+= 0$ at $t=0$.
If we choose $H(y)= \frac{1}{2}(y^+)^2$ in (\ref{Hconvex}), then we have
\begin{equation}\label{tildeest1}
\frac{1}{2}\partial^\alpha_t(||\tilde{u}^+(t)||^2_{L^2(\Omega)})
\leq \langle \partial^\alpha_t\tilde{u}(t), \tilde{u}^+(t)\rangle.
\end{equation}
Note that by the definition of the Caputo derivative, we have
$$\frac{1}{2}\partial^\alpha_t(||\tilde{u}^+(t)||^2_{L^2(\Omega)})
= \frac{1}{2}\partial_t(\phi_{1-\alpha} *(||\tilde{u}^+(t)||^2_{L^2(\Omega)}- ||\tilde{u}^+(0)||^2_{L^2(\Omega)}))
= \frac{1}{2}\partial_t(\phi_{1-\alpha} *(||\tilde{u}^+( t)||^2_{L^2(\Omega)})$$
since $\tilde{u}^+= 0$ at $t=0$. Also note that
$u- g\in C([0, T]; L^2(\Omega))$ implies
$$\phi_{1-\alpha}* ||\tilde{u}^+(t)||^2_{L^2(\Omega)}= 0$$
at $t= 0$, so by (\ref{alpha1}) we have
$$\phi_{\alpha}* \partial_t(\phi_{1-\alpha}* ||\tilde{u}^+(t)||^2_{L^2(\Omega)})= \partial_t(\phi_{\alpha}* \phi_{1-\alpha}
*||\tilde{u}^+(t)||^2_{L^2(\Omega)})$$
\begin{equation}\label{tildeid1}
= \partial_t(1* ||\tilde{u}^+(t)||^2_{L^2(\Omega)})= ||\tilde{u}^+
(t)||^2_{L^2(\Omega)}.
\end{equation}
Now we let both sides of (\ref{tildeest}) act on $\tilde{u}^+$. Based on (\ref{tildeest1}) we can apply the convolution operation $\phi_{\alpha}\,*$ and
(\ref{tildeid1}) to get
$$\frac{1}{2}||\tilde{u}^+(t)||^2_{L^2(\Omega)}+ \phi_{\alpha}* \langle (-\Delta)^s\tilde{u}(t), \tilde{u}^+(t)\rangle$$
$$+ \phi_{\alpha}*\langle a(\cdot, t, u(\cdot, t))- a(\cdot, t, \tilde{\varphi}(\cdot, t)), \tilde{u}^+(\cdot, t)\rangle\leq 0$$
since $\phi_{\alpha}$ is a positive function.
Note that $a(x, t, u)- a(x, t, \tilde{\varphi})$ has the same sign as $\tilde{u}$ and
\begin{equation}\label{PlusfracLa}
\langle (-\Delta)^s\tilde{u}(t), \tilde{u}^+(t)\rangle\geq 0
\end{equation}
Hence we have $||\tilde{u}^+(t)||^2_{L^2(\Omega)}\leq 0$ and thus the only possibility is $\tilde{u}^+= 0$, i.e. $u\leq \tilde{\varphi}$ in $\Omega\times (0, T)$.
Also note that $a(x, t, \tilde{\varphi})+ a(x, t, u)$ has the same sign as $\tilde{\varphi}+ u$ so similarly we can consider $\tilde{u}:= -u-\tilde{\varphi}$ and show $-\tilde{\varphi}\leq u$ in $\Omega\times (0, T)$. Hence we have
$|u|\leq \tilde{\varphi}$ in $\Omega\times (0, T)$.
\end{proof}
We remark that we have
$$\langle (-\Delta)^s\tilde{u}(t), \tilde{u}^+(t)\rangle=
\int_{\mathbb{R}^n}\int_{\mathbb{R}^n}\frac{(\tilde{u}(x,t)- \tilde{u}(y,t))(\tilde{u}^+(x,t)- \tilde{u}^+(y,t))}{|x-y|^{n+2s}}dxdy.$$
Since $(\tilde{u}^+(x,t)- \tilde{u}^+(y,t))(\tilde{u}^-(x,t)- \tilde{u}^-(y,t))\leq 0$, we have
$$\langle (-\Delta)^s\tilde{u}(t), \tilde{u}^+(t)\rangle
\geq \int_{\mathbb{R}^n}\int_{\mathbb{R}^n}\frac{(\tilde{u}^+(x,t)- \tilde{u}^+(y,t))^2}{|x-y|^{n+2s}}dxdy\geq 0.$$
Hence (\ref{PlusfracLa}) holds true.
To study the well-posedness of (\ref{Semifracspacetime}),
we again use the substitutions $u= w+g$ and
$f= -(-\Delta)^s g|_{\Omega\times (0, T)}$ so (\ref{Semifracspacetime})
can be converted into the initial value problem
\begin{equation}\label{Capsemiivp}
\left\{
\begin{aligned}
\partial^\alpha_t w+ (-\Delta)^s w + a(x, t, w)&= f
\quad \,\,\,\Omega\times (0, T)\\
w&= 0\quad \,\,\,\Omega\times \{0\}.
\end{aligned}
\right.
\end{equation}
Our assumptions on $g$ and $a$ ensure that $f-a$ satisfies the conditions
(F1), (F2) (and thus (F3), (F4)) on Page 66 and the condition (F5) on Page 81 in \cite{gal2020fractional}. Hence we can apply Theorem 3.2.2, Corollary 3.2.3 and (3.2.32) in \cite{gal2020fractional}
to obtain the following well-posedness result.
\begin{prop}
There exists $0< T_{max}\leq \infty$ s.t. for $0< T < T_{max}$, (\ref{Semifracspacetime}) has a unique solution
$$w\in C([0, T]; L^\infty(\Omega)).$$ Furthermore, We have either (i) $T_{max}= \infty$ or (ii) $0< T< \infty$ and $lim_{t\to T_{max}}||u(\cdot, t)||_{L^\infty}= \infty$.
\end{prop}
We remark that the solution $w$ actually belongs to certain H\"older type function spaces. We refer readers to Subsection 3.2 in \cite{gal2020fractional} for more details.
By the previous $L^\infty$
estimate, we have $T_{max}= \infty$ for $g\in C^\infty_c(\Omega_e\times (0, T))$ under our assumptions on the nonlinearity $a$. Hence we do not need any smallness assumptions on the constant $T$.
Now we prove a linearization result.
For $g\in C^\infty_c(\Omega_e\times (0, T))$, we use
$u_g$ to denote the solution of the linear problem
\begin{equation}\label{linearfracspacetime}
\left\{
\begin{aligned}
\partial^\alpha_t u+ (-\Delta)^s u + a_1(x, t)u&= 0
\quad \,\,\,\Omega\times (0, T)\\
u&= g\quad \,\,\,\Omega_e\times (0, T)\\
u&= 0\quad \,\,\,\Omega\times \{0\}
\end{aligned}
\right.
\end{equation}
and we use $u_{\lambda, g}$ to denote the solution of the semilinear problem
\begin{equation}\label{lambdafrac}
\left\{
\begin{aligned}
\partial^\alpha_t u+ (-\Delta)^s u + a(x, t, u)&= 0\quad \,\,\,\Omega\times (0, T)\\
u&= \lambda g\,\,\,\,\, \Omega_e\times (0, T)\\
u&= 0\quad \,\,\,\Omega\times \{0\}
\end{aligned}
\right.
\end{equation}
in the rest of this section.
\begin{prop}
Let $w_{\lambda, g}:= u_g- \frac{u_{\lambda, g}}{\lambda}$. Then $\lim_{\lambda\to 0}w_{\lambda, g}= 0$
in $C([0,T]; L^\infty(\Omega))$.
\end{prop}
\begin{proof}
Note that $w_{\lambda, g}= 0$ in $\Omega_e\times (0, T)$ and
$$\partial_t w_{\lambda, g}+ (-\Delta)^s w_{\lambda, g}
+ a_1(x, t)w_{\lambda, g}= \frac{1}{\lambda}
\sum^m_{k=2}a_k(x, t)|u_{\lambda, g}|^{b_k}u_{\lambda, g}$$
in $\Omega\times (0, T)$. By the $L^\infty$
estimate (Proposition 4.1), we have
$$||w_{\lambda, g}||_{L^\infty}\leq \frac{T^\alpha}{\lambda\Gamma(\alpha+ 1)}
||\sum^m_{k=2}a_k(x, t)|u_{\lambda, g}|^{b_k}u_{\lambda, g}||_{L^\infty(\Omega\times (0, T))}$$
and $||u_{\lambda, g}||_{L^\infty}\leq
\lambda||g||_{L^\infty(\Omega_e\times (0, T))}$ so
$$||w_{\lambda, g}||_{L^\infty}
\leq \frac{T^\alpha}{\Gamma(\alpha+ 1)}\sum^m_{k=2}\lambda^{b_k}||a_k(x,t)||_{L^\infty}||g||^{b_k+1}_{L^\infty(\Omega_e\times (0, T))},$$
which implies $||w_{\lambda, g}||_{L^\infty}\to 0$ as $\lambda\to 0$.
\end{proof}
\subsection{Inverse problem}
We will use the same argument as the one in \cite{li2021fractional} to prove Theorem 1.2 based on the unique continuation property of the fractional Laplacian (Proposition 2.3), the Runge approximation property (Proposition 3.6), the $L^\infty$ estimate (Proposition 4.1) and the linearization result (Proposition 4.3).
We remark that this kind of regimes work for solving Calder\'on type inverse problems for a broad class of fractional evolutionary operators involving
the fractional Laplacian (see Theorem 2.18 in \cite{kowinverse} for more details).
To avoid redundancy, we will skip inessential steps in the proof. We refer readers to Subsection 4.2 in \cite{li2021inversediff} for full details.
\begin{proof}
Based on Proposition 2.3, the assumption (\ref{semiIDN})
implies that
$$u^{(1)}_{\lambda, g}= u^{(2)}_{\lambda, g}=: u_{\lambda, g}$$
in $\mathbb{R}^n\times (0, T)$.
Hence we have
\begin{equation}\label{aRid}
(a^{(1)}_1- a^{(2)}_1)u_{\lambda, g}
= R^{(2)}_1(x, t, u_{\lambda, g})
- R^{(1)}_1(x, t, u_{\lambda, g})
\end{equation}
in $\Omega\times (0, T)$ where
$$R^{(i)}_j(x, t, z):= \sum^m_{k=j+1}a^{(i)}_k(x, t)|z|^{b_k}z.$$
Now we note that
$$||a^{(1)}_1- a^{(2)}_1||_{L^2(\Omega\times (0, T))}$$
$$\leq ||a^{(1)}_1- a^{(2)}_1||_{L^\infty}||1- \frac{u_{\lambda, g}}{\lambda}||_{L^2(\Omega\times (0, T))}$$
\begin{equation}\label{2inftyest}
+ \frac{1}{\lambda}||(a^{(1)}_1- a^{(2)}_1)
u_{\lambda, g}||_{L^2(\Omega\times (0, T))}.
\end{equation}
For given $\delta> 0$, by Proposition 3.6 we can choose
$g\in C^\infty_c(W_1\times (0, T))$ s.t.
$$||1- u_g||_{L^2(\Omega)\times (0, T)}\leq \delta$$
and for this chosen $g$, we have
\begin{equation}\label{Rungelinear}
||1- \frac{u_{\lambda, g}}{\lambda}||_{L^2(\Omega\times (0, T))}\leq 2\delta
\end{equation}
for small $\lambda$ by Proposition 4.3. Since Proposition 4.1 implies that
$$||u_{\lambda, g}||_{L^\infty}\leq
\lambda||g||_{L^\infty(\Omega_e\times (0, T))},$$
by (\ref{aRid}) we have
$$\frac{1}{\lambda}||(a^{(1)}_1(\cdot,t)- a^{(2)}_1(\cdot,t))
u_{\lambda, g}(\cdot, t)||_{L^2(\Omega)}$$
$$\leq C'(\sum^m_{k=2}\lambda^{b_k}(||a^{(1)}_k||_{L^\infty}
+||a^{(2)}_k||_{L^\infty})||g||^{b_k+1}_{L^\infty})\to 0$$
as $\lambda\to 0$. Then by (\ref{2inftyest}) and (\ref{Rungelinear}) we get
$$||a^{(1)}_1- a^{(2)}_1||_{L^2(\Omega\times (0, T))}\leq
2\delta ||a^{(1)}_1- a^{(2)}_1||_{L^\infty(\Omega\times (0, T))}.$$
Now we conclude that $a^{(1)}_1= a^{(2)}_1$ since $\delta$ is arbitrary.
Iteratively, once we have shown $a^{(1)}_j= a^{(2)}_j$ ($1\leq j\leq m'-1$), we have
$$(a^{(1)}_{m'}(x, t)- a^{(2)}_{m'}(x, t))|u_{\lambda, g}|^{b_{m'}}
u_{\lambda, g}
= R^{(2)}_{m'}(x, t, u_{\lambda, g})
- R^{(1)}_{m'}(x, t, u_{\lambda, g})$$
in $\Omega\times (0, T)$.
Then we can repeat the procedure above to conclude that $a^{(1)}_{m'}= a^{(2)}_{m'}$.
\end{proof}
\bibliographystyle{plain}
{\small |
\section{Introduction}
LaSb$_2$ is a member of the light rare-earth diantimonides {\it R}Sb$_ 2$ ({\it R} = La-Nd, Sm) family in which the members all crystallize in the orthorhombic SmSb$_2$ structure. \cite{wan67a,hul79a} Single crystals of the members of the family were grown decades ago and basic physical properties were reported. \cite{can91a,bud98a} At that time \cite{bud98a} charge density wave (CDW) - like features in resistivity were observed (below 300 K) in PrSb$_2$ and NdSb$_2$, but not in other members of the family. LaSb$_2$ on the other hand is one of the first {\it stoichiometric} compounds where large, linear magnetoresistance was observed. {\cite{bud98a,you03a} Superconductivity with $T_c \approx 0.4 $~K was detected in LaSb$_2$, \cite{hul77a} and details of the superconducting state were studied at ambient and high pressure in Refs. \onlinecite{gal13a,guo11a}
The possibility of CDW in LaSb$_2$ was discussed over the years. Quantum oscillations data and band structure calculations \cite{goo04a} were considered consistent with a conjecture of existence of CDW transitions, on the other hand, optical conductivity measurements \cite{dit11a} were interpreted as ruling out the formation of a CDW phase above 20~K. Finally, signatures of CDW were discovered and studied in the La$_{1-x}$Ce$_x$Sb$_2$ substitutional series \cite{luc15a}, including a clear CDW-consistent feature in resistivity at $\sim 355$~K for pure LaSb$_2$. The observed suppression of CDW temperature with Ce - substitution (and the associated unit cell volume decrease) suggests the possibility of fair to moderate pressure sensitivity of this transition in pure LaSb$_2$.
This potential pressure sensitivity is experimentally addressed in this following by measurements of electrical resistivity of LaSb$_2$ in zero and applied magnetic field under pressure up to 12.9 kbar. We find that LaSb$_2$ is remarkably pressure sensitive with $T_{CDW}$ being suppressed to below 150 K by $\sim 4$~kbar and very likely fully suppressed by $\sim 6-7$~kbar. In addition we find that the resistive anomaly associated with $T_{CDW}$ changes dramatically when $T_{CDW}$ drops below the solidification temperature of the pressure media, suggesting that LaSb$_2$ is sensitive to even small deviations from purely hydrostatic pressure conditions. Finally we infer that for some pressure less than $\sim 2$~kbar there is a significant change in the degree of hysteresis in $T_{CDW}$ (changing from $\sim 2$ K at ambient pressure to more than 50 K above 2 kbar) and in the amount of Fermi surface that is gapped by the CDW transition (changing from roughly 10\% at ambient pressure to $\sim 50$~\% above 2 kbar).
\section{Experimental details}
Large single crystals of LaSb$_2$ were grown out of antimony flux.\cite{bud98a,can92a} Elemental La (99.99+ Ames Laboratory) and Sb (99.999+ Alfa Aesar) were combined in the ratio of La$_5$Sb$_{95}$ and placed into the bottom part of a 2 ml Canfield Crucible set (CCS). \cite{can16a,lsp} The CCS was sealed into an amorphous silica tube with silica wool on top of the CCS to act as a cushion during the decanting process. The sealed ampule was heated over 3 hours to 1100$^{\circ}$C, held at 1100$^{\circ}$C for 5 hours and then cooled to 1000$^{\circ}$C over 1 hour. After sitting at 1000$^{\circ}$C for 5 hours, the ampoule was cooled 675$^{\circ}$C over 99 hours. Upon reaching$^{\circ}$675 C, the ampule was removed from the furnace and decanted in a lab centrifuge. \cite{can20a} After cooling to room temperature, the ampule was opened, revealing large, sometimes crucible limited crystals of LaSb$_2$. The crystals grow as soft plates with the $c$-axis perpendicular to the plates. The samples are metallic-micaceous, and the layers is malleable and readily deformed. Due to the high malleability of LaSb$_2$, a reasonable quality powder x-ray diffraction and even Laue diffraction are difficult to perform and were not attempted here. However, examination of the La - Sb binary phase diagram \cite{mas92a} as well as x-ray diffraction data \cite{luc15a} on crystals grown in a similar way show unambiguously that the resulting crystals are indeed orthorhombic LaSb$_2$.
Standard, linear four-probe ac resistance at ambient pressure and under pressure was measured on bar - shaped samples in a $I || ab$ geometry. The 0.25 $\mu m$ diameter Pt wires were spot-welded to the sample and then the contacts were covered with Epo-Tek H20E silver-filled epoxy for better mechanical stability. The contact resistance values were $\sim 1 \Omega$ or lower. The frequency used was 17 Hz, typical current values were 3-5 mA. Magnetoresistance was measured in a transverse configuration, $H || c$, $I || ab$. The measurements were performed using the ACT option of a Quantum Design Physical Property Measurement System (PPMS) instrument. Pressure was generated in a hybrid, BeCu / NiCrAl piston - cylinder pressure cell (modified version of the one used in Ref. \onlinecite{bud86a}). A 40 : 60 mixture of light mineral oil and n-pentane was used as a pressure-transmitting medium. This medium solidifies at room temperature in the pressure range of 30 - 40 kbar, \cite{bud86a,kim11a,tor15a} which is above the maximum pressure in this work. Elemental Pb was used as a low temperature pressure gauge.\cite{eil81a} The measurements were performed both on increase and decrease of pressure. Given that the pressure inside of a piston cylinder cell changes with temperature, we used the results of Ref. \onlinecite{xia20a} to (approximately) incorporate this difference in the cited pressure values. In the following we will use $P_{LT}$ for the low temperature pressure values obtained using Pb and $P$ for recalculated pressure values at higher temperatures that are used in the $P - T$ and related plots shown in Figs. \ref{F4} and \ref{F5} below. It should be noted we only present finite pressure data that has $P_{LT} > 0$}.
\section{Results}
Resistivity measurements up to 380 K at ambient pressure are shown in Fig. \ref{F1}. In agreement with Ref. \onlinecite{luc15a}, a clear feature in resistivity resembling CDW transition is observed at $\sim 350$~K. This transition has small but resolvable hysteresis of $\sim 2$~K.
An example of the resistivity data taken under pressure ($P_{LT} = 0.1$~kbar is presented in Fig. \ref{F2}. Several features are noteworthy: (i) CDW transition is shifted to lower temperatures in comparison to ambient pressure; (ii) the hysteresis is increased significantly, from $\sim 2$~K at ambient pessure to about 50 K; (iii) the feature in resistivity associated with the CDW transition became significantly more pronounced, changing from a $\sim 10$~\% increase in resistance at ambient pressure to a nearly 100\% increase at 0.1 kbar; (iv) there is clear, low temperature magnetoresistance with a local minimum in $\rho(T)$ appearing near 40 K (such local minimum has been observed in resistivity of a number of semi-metals in magnetic field \cite{mun12a,jon17a} and is understood within the Kohler's rule physics \cite{pip89a}); and (v) 140 kOe magnetic field does not affect the CDW transition temperature (Fig. \ref{F2}(b), inset).
This last observation allows us to shorten the measurements time by evaluating the CDW transition hysteresis under pressure from comparison of $H = 0$ data on cooling and $H = 140$~kOe on warming. A subset of such data is shown in Fig. \ref{F3}. The resistivity at 300 K decreases slowly and monotonically under pressure. For $0.1~\text{kbar} \leq P_{LT} \leq 3.6$~kbar the feature associated with the CDW transition shifts to lower temperatures and gradually decreases in its amplitude. The $T_{CDW}$ continues to decrease for $4.5~\text{kbar} \leq P_{LT} \leq 5.5$~kbar, but the feature broadens and becomes barely resolved. Finally, no features in resistivity are observed in the $8.6~\text{kbar} \leq P_{LT} \leq 12.9$~kbar pressure range, so by or before 8.6 kbar the CDW transition is completely suppressed. The suppression of CDW transition is reflected in the base temperature magnetoresistance as well, even if in a subtle way.
\section{Discussion and Summary}
The temperature-dependent resistivity data taken at different pressures allows for the construction of a $P - T$ phase diagram for LaSb$_2$ presented in Fig. \ref{F5}. The symbols correspond to CDW transition temperatures at different pressures as determined from the minima in the derivatives $d \rho / dT$ of the electrical transport data taken on cooling and on warming, as illustrated in the insets of Figs. \ref{F1} and \ref{F2}. The open and closed symbols on the T-P phase diagram are associated with the abrupt decrease (and broadening) of the CDW-like feature in resistivity for the $4.5~\text{kbar} \leq P_{LT} \leq 5.5$~kbar datasets (Fig. \ref{F3}). Although it would be tempting to conject the presence of a transition in the 3-4 kbar pressure range, a more mundane explanation seems to be plausible. The CDW-like transition has unambiguous hysteresis pointing to its first order. It is likely that it is accompanied by (anisotropic) changes in the lattice parameters and volume. If this is hypothesis is correct, then we might expect that the experimental signature of the transition crucially depends on the level of hydrostaticity of the pressure transmitting medium (PTM) (for example, above and below of the medium freezing line, CaFe$_2$As$_2$ under pressure \cite{tor08a,yuw09a,can09a} is one of the recognized examples of such sensitivity). Indeed, the position of the freezing line of the PTM used in our measurements \cite{tor15a} (Fig. \ref{F5}) is in agreement with the pressure-temperature range where the abrupt change of the CDW-like signature in resistivity is observed. Clearly, measurements in He gas / liquid pressure medium would be required to address the issue of hydrostaticity.
To further explore the effects of applied pressure on LaSb$_2$, in figure \ref{F4} we present the pressure dependence of the 1.8 K resistivity and the relative change in the resistivity associated with CDW transition , $\Delta \rho_{CDW} / \rho_{CDW}$, which can serve as a simple caliper of how much of the Fermi surface is gapped as a result of the CDW transition. \cite{tor07a}. In addition, we plot the pressure dependence of the exponent $n$ in the fit of the zero field, low temperature ($ 1.8~\text{K} \leq T \leq 15~\text{K}$ ) resistivity to a power law, $\rho = \rho_0 + AT^n$, as well in Fig. \ref{F4}(c). For each of the plots we adopt the convention we used in figure \ref{F5} and show the data associated with finite $T_{CDW} < T_{solidification}$ as open symbols. Note that if $T_{CDW} = 0$~ K, then there is no signature of a CDW transition and there is no broadening of (nonexistent) features.
The base temperature resistivity (Fig. \ref{F4}(a)) decreases with increasing rapidity for pressure up to $\sim 5-6$~kbar, then the behavior changes abruptly and becomes only slightly pressure-dependent at higher pressures. This behavior is not unexpected, since below the pressure at which CDW is driven to $T = 0$~K an additional contribution from suppression of the resistive increase associated with the Fermi surface gapping due to CDW plays an important role. Similar behavior was observed e.g. in LaAu$_{0.970}$Sb$_2$ under pressure. \cite{xia20b}
The behavior of $\Delta \rho_{CDW} / \rho_{CDW}$ parameter is more complex. At ambient pressure it is rather small, suggesting, within a very simple model, that $\lesssim 10 \%$ of the Fermi surface is gapped. On pressure increase to $\sim 2$~kbar (at $T_{CDW}$) $\Delta \rho_{CDW} / \rho_{CDW}$ jumps to $\sim 1$, that implies gapping of roughly half of the Fermi surface. Then, in 5-7 kbar range, $\Delta \rho_{CDW} / \rho_{CDW}$ abruptly decreases, whereas at 8.6 kbar and above this feature is not observed any more.
The exponent $n$ ranges from $\sim 3.3$ to $\sim 3.7$ as pressure is increased, initially varying only weakly as $T_{CDW}$ is weakly suppressed. As $T_{CDW}$ is suppressed more rapidly, $n$ increases from 3.3 to 3.7. Once $T_{CDW}$ is suppressed to zero, $n(P)$ flattens.
It is noteworthy that whereas the low temperature resistivity ($\rho_{1.8~\text{K}}$, $n$) appears to be not very sensitive to $T_{CDW}(P)$ crossing $T_{solidification}(P)$ line, the changes in $\Delta \rho_{CDW} / \rho_{CDW}$ on this crossing are abrupt and unambiguous. This is actually not too surprising if the effects of non-hydrostaticity are thought to be a broadening of the resistive feature of the transition rather than a significant change in the low temperature state. If $T_{CDW}$ is well above base temperature or the $T \leq 15$~K range of power law fitting, then the affects of this broadening will be long past, at higher temperatures. On the other hand, the$\Delta \rho_{CDW} / \rho_{CDW}$ is inherently associated with the $T_{CDW}$ temperature region.
As mentioned above, the ambient pressure data (Fig. \ref{F1}) and the data for $0.1~\text{kbar} \leq P_{LT} \leq 3.6$~kbar (Fig.\ref{F3}) show very clear, CDW - like feature in resistivity, however comparing those (i) the hysteresis between data taken on cooling and on warming is notably different; (ii) the size of the CDW - like anomaly is significantly different. We believe that neither some difference in the in-plane direction of the current, nor possible temperature lag due to the thermal mass of the cell (rates of $\sim 0.3$~K/min were used) could explain these drastic differences. (Note, that at ambient pressure similar relative size of the anomaly was observed in Ref. \onlinecite{luc15a}, and no abrupt change of hysteresis under pressure in the same or similar cell and measurements protocol was reported in Ref. \onlinecite{bud20a}). This difference between ambient pressure and small applied pressure is fully consistent with the sensitivity of the CDW feature to hydrostatic / non-hydrostatic conditions discussed above. We speculate that (i) LaSb$_2$ mignt have very low pressure, significantly non-linear compressibility (not too surprising for van der Waals - like structure) that leads to rapid change of nesting, or (ii) there might be a dramatic pressure-induced change of electronic structure near the Fermi level (possible caused by some subtle structural transformation) at low pressure that causes these striking difference in the degree of nesting and thermal hysteresis. This possible transition is marked by the crosshatched area in the $P - T$ phase diagram (Fig. \ref{F5}). The complex (calculated) Fermi surface of LaSb$_2$ \cite{goo04a} is not inconsistent with the possibility of significant change of nesting under moderate pressure.
Regardless of possible complications caused by PTM freezing, and possible change of electronic structure in 1 kbar range, the data show that the CDW transition in LaSb$_2$ is very fragile, so that 6-7 kbar pressure is enough to suppress $T_{CDW}$ to $T = 0$~K, this corresponds to overall pressure derivative of $\sim - 50$~K/kbar, that is more than an order of magnitude larger in the absolute value than that for LaAgSb$_2$ \cite{tor07a} ($\approx - 4.3$~K/kbar). This might be due some specific details of the band structure of these materials, as well as to possible differences in elastic modulae.
A clear change in the base temperature magnetoresistance (MR) (figure \ref{F6}) can be seen when pressures exceed the 6-7 kbar range. For $P_{LT}$ = 8.6, 9.5 and 12.9 kbar a very clear, low field negative curvature / or shoulder develops around $H \sim 25$~kOe. For lower pressures the MR has a positive curvature over the whole applied field range. These data are consistent with a change in the band structure associated with the loss of the lower pressure CDW state. Given that MR is traditionally plotted as $\Delta \rho/\rho_0$ and given that $\rho_0$ changes rapidly and by very large amounts (due the changing nature of the CDW gapping) the lower pressure variation of the MR is non-monotonic. This said, the functional change in the magnetic field - dependent MR for $P_{LT} > 7$~ kbar is clear and unambiguous.
Our data on CDW suppression in LaSb$_2$ could be compared with the evolution of superconductivity in LaSb$_2$ under pressure in this material. \cite{guo11a} The $T_c$ reportedly increases under pressure, the superconducting dome of "3D superconductivity" has a maximum at about 4-5 kbar and then $T_c$ decreases with further pressure increase, but at a measurably slower rate. Qualitatively, such behavior appear to be consistent with a simple picture of coexistence and competition between CDW and superconductivity. \cite{bil76a,bal79a}
To summarize, the CDW state in LaSb$_2$ was found to be quickly (by 6-7 kbar) supressed under pressure with the overall pressure derivative of $\sim - 50$~K/kbar . We hypothesize that (i) the CDW-like transition has a structural component that causes enhanced sensitivity of CDW transition to the level of pressure hydrostaticity, and (ii) that there is either a very low pressure, non-linear compressibility that leads to rapid change of nesting or band structure very close to pressure induced change that dramatically changes degree of nesting of the CDW. All of these data and observations point out that LaSb$_2$ is a long overlooked, highly pressure sensitive, low dimensional material.
\begin{acknowledgments}
We would like to thank Hermann Suderow for useful communications. Work at the Ames National Laboratory was supported by the U.S. Department of Energy, Office of Science, Basic Energy Sciences, Materials Sciences and Engineering Division. The Ames National Laboratory is operated for the U.S. Department of Energy by Iowa State University under contract No. DE-AC02-07CH11358. PCC and SLB wish to acknowledge Hidalgo Alonso Quijano for inspiring this line of inquiry.
\end{acknowledgments}
|
\section{Introduction}
Ultra-hot Jupiters (UHJs) are gas giants with orbital periods of hours to days and typical day-side temperatures of $2500~\mathrm{K}$ or more. They form a continuum with the broader population of hot Jupiters, of which they constitute the high temperature end \citep{Baxter2020, Mansfield2021}. As a result of their temperatures, which push towards stellar values, they have distinct atmospheric properties that make them ideal laboratories of physics and chemistry. Indeed, their day-sides are expected to be cloud-free and close to chemical equilibrium \citep{Kitzmann2018, Lothringer2018, Parmentier2018, Helling2019_w18}. Clouds may still form in their night-side, but they do not completely sequester heavy elements, which are present in their transmission (e.g., \citealt{Hoeijmakers2018_k9}) and emission spectrum (e.g., \citealt{Pino2020}). As a result, UHJs currently constitute the only class of planets for which it is possible to measure elemental abundances with different volatility, including rock forming elements. This makes them very promising benchmarks for planet formation theories \citep{Lothringer2021}.
Observational campaigns with a variety of instruments from both ground and space have so far mainly focused on the day-side (through emission spectroscopy) and terminator regions (through transmission spectroscopy) of UHJs, because they yield the best signal-to-noise ratio (S/N). These methods have been successful in probing the local properties of their atmospheres. However, because of their short orbital periods, UHJs are expected to be tidally locked. Therefore, the global properties of UHJ atmospheres may deviate from the local ones, and it is necessary to account for three-dimensional (3D) effects (e.g., \citealt{Feng2016}). Observationally, this has so far been best achieved through phase curves.
Phase curves record the flux emitted from the planet throughout the orbit. In photometry, a relative calibration to the stellar flux registered during the secondary eclipse permits the measurement of the phase-resolved effective temperature of the planet, including its day-to-night contrast, and the longitude of its atmospheric hot spot. When observed through a low-resolution spectrograph (e.g., HST WFC3), the resulting spectrophotometric phase curve can resolve these properties in altitude, and additionally reveal the phase-dependent chemistry \citep{Stevenson2014, MikalEvans2022}.
Tens of hot and ultra-hot Jupiters have been observed through multi-wavelength photometric phase curves and a handful of them with spectrophotometric phase curves. These observations have revealed that the climate of UHJs may differ from that of their cooler counter-parts. Relatively small hot-spot offsets and small day-to-night contrasts (albeit observed with some scatter) indicate that heat transport likely occurs in the lack of strong equatorial jets, but is still efficient \citep{ParmentierCrossfield2018}. These observations contrast with trends observed for gas giants colder than $2500~\mathrm{K}$ \citep{Zhang2018, Wong2020}, which have weaker hot spot offsets but, unlike their hotter counter-parts, display larger day-to-night contrasts with increasing temperature, indicating a reduced heat transport efficiency. Theoretical studies have identified two key additional ingredients that emerge in UHJs, and are necessary to explain their different climate: thermal recombination of atomic hydrogen to $\mathrm{H}_2$ in their night-side, which leads to an increase in the efficiency of heat transport \citep{Bell2018}, and atmospheric drag, which affects their atmospheric circulation patterns. Indeed, the interaction of the significant planetary ionospheres, produced by the high temperatures, and of magnetic fields, has the capability to dampen waves that would otherwise lead to the formation of an equatorial jet \citep{Perna2010, Beltz2022a}. This would lead to smaller hot spot offsets, and potentially to a transition to day-to-night atmospheric flow \citep{Tan2019}. While the theory was moderately successful in explaining the overall properties of climate of UHJs, results obtained on individual UHJs still blurry this picture (e.g., \citealt{Addison2021}). Reconciling optical (e.g., TESS) and infrared (e.g., Spitzer) phase curves proves particularly challenging in some cases \citep{Wong2020, VonEssen2020}.
Very high-resolution spectroscopy (HRS; $R\gtrsim100000$) offers a unique observational angle to tackle the open questions in this field. Indeed, by individually resolving spectral lines, it is able to directly measure the Doppler-shift of spectral lines, and thus probe the strength of winds moving masses of gas around the atmosphere of target planets. \cite{Ehrenreich2020} showcased the power of this technique by measuring phase-resolved transmission spectroscopy of UHJ WASP-76b. They reveal a complex Doppler shift pattern, sensitive to temperature structure, rotation, dynamics, and night-side condensation \citep{Wardenier2021, Savel2022}. However, HRS transmission spectra generally probe above the planet photosphere (e.g., \citealt{Hoeijmakers2019}), and can only access a limited longitudinal region of the planetary atmosphere \citep{Wardenier2022}.
A complementary approach is to observe parts of a planetary phase curve with a high-resolution spectrograph. When observed through a high-resolution spectrograph, features in the planet continuum are lost, but the phase-resolved Doppler shift and Doppler broadening of emission/reflection lines uniquely constrains the orbit of the planet, and its rotation and dynamics \citep{CollierCameron1999, Kawahara2012, Brogi2013, Snellen2014}. This is supported by dedicated theoretical studies that, applying radiative transfer to General Circulation Models (GCMs), show that the combination of inhomogeneous specific intensity, rotation, and winds imprints kilometre-per-second level Doppler shifts and significant distortions in the line shapes and strengths \citep{Zhang2017, Beltz2022}. However, two challenges have so far hindered the use of this technique. First, the signal observed in this configuration is smaller compared to transmission spectroscopy, which has led most authors to stack spectra across planet phases to increase the significance of detections. Unfortunately, this operation washes out the phase-dependency of the line profiles, which is crucial to probe planetary climate. Second, HRS observations require complex data reduction methods, which makes the retrieval of atmospheric properties a non-trivial operation.
Both challenges can now be surpassed. The application of likelihood-based frameworks of recent development \citep{Brogi2019, Gibson2020, Pino2020} allows us to reliably determine atmospheric properties and their error bars starting from HRS emission spectra. Exploiting this key technical novelty, \cite{Beltz2021} presented the first indication that the HRS phase curve of HD209458b observed with CRIRES has the sensitivity to probe 3D effects in the planetary atmosphere. In addition, thanks to their relatively large planet-to-star flux ratio, UHJs offer the opportunity to observe HRS phase curves at unprecedented S/N. Indeed, \cite{Herman2022} and \cite{vanSluijs2022} present the first parameterized studies of phase-dependent \ion{Fe}{i} and CO line strengths targeting a UHJ, WASP-33b, for which \cite{Cont2021} had already suggested that three-dimensional effects are required to explain the phase-stacked optical HRS emission spectrum of TiO and \ion{Fe}{i}. In addition, \cite{Borsa2022} found evidence for a different chemistry and thermal profile in pre- and post-eclipse phases of UHJ KELT-20b.
In this paper, we further extend these previous studies to include, for the first time, a simultaneous parameterized analysis of phase-dependent line intensity and Doppler shift of the hottest UHJ: KELT-9b ($T_\mathrm{eq} = 3900~\mathrm{K}$; \citealt{Borsa2019}). In section \ref{sec:methods}, we present our data-reduction pipeline and custom model suite, designed to be able to capture in a parameterized way the phase-dependence of both \ion{Fe}{i} line intensity and Doppler shift. In section \ref{sec:results}, we present results from our retrievals, with a focus on phase-resolved, 3D properties of the atmosphere of KELT-9b. In section \ref{sec:discussion}, we discuss our results in the context of photometric phase-curve observations of KELT-9b and in the context of other observations and theory of UHJ climate with low and high-resolution spectroscopy; we additionally discuss the prospects of HRS phase curve studies with current and upcoming instrumentation, including unprecedented possibilities to study the orbital and atmospheric dynamics of UHJs. We present our conclusions in section \ref{sec:conclusions}.
\section{Methods}
\label{sec:methods}
The key methodological novelty introduced by this work, is that we extract information on the phase dependence of planetary atmospheric properties only by comparing planetary spectra at difference phases, without relying on external input such as a known systemic velocity, or the overall stellar flux level. In this section, we present all the crucial aspects to successfully apply this new approach to HRS phase curves.
\subsection{Ephemeris of KELT-9b}
\label{sec: Ephemeris of KELT-9b}
Knowing a reliable orbital ephemeris for KELT-9b is essential for our analysis, given the crucial impact it has for associating an accurate orbital phase to each spectrum. Unfortunately, even a quick look at recent transit photometry shows that the prediction by the only ephemeris available until recently (from \citealt{Gaudi2017}) disagrees with the observations by more than ten minutes at present epoch. In order to compute an updated orbital period $P$ and reference transit time $T_0$, we extract the TESS short-cadence PDCSAP light curves of KELT-9 from Sectors 14, 15 and 41 (2019 to 2021) and perform a JKTEBOP \citep{Southworth2008} transit fit to them, assuming a linear ephemeris and computing the error bars through a residual-permutation technique. The final result that we adopt for the subsequent analysis is:
\begin{equation}
\left \{\begin{array}{l}
P = 1.4811188 \pm 0.0000003 \textrm{ d} \\
T_0 = 2\,459\,006.3289 \pm 0.0001 \textrm{ BJD}_\textrm{TDB}
\end{array}\right .\textrm{ ,}
\end{equation}
where the chosen reference frame and time standard is the Barycentric Julian Day in Barycentric Dynamical Time, following the prescription by \citet{Eastman2010}. It is worth noting that our new estimate of $P$ is perfectly consistent with those recently published by \citet{PaiAsnodkar2022} and \citet{Ivshina2022}.
\subsection{Observations and data reduction}
\label{sec:methods:observations_and_data_reduction}
We focus our analysis on neutral iron lines, which offer the highest S/N in KELT-9b thanks to their large cross-section, and have well known line positions and strengths. This is crucial to identify astrophysical variations in line strengths and positions with accuracy. In the following, we describe a custom pipeline that we developed to extract the planet spectrum, model iron emission lines including simplified pseudo-3D effects, and how we compare models to data in a statistical framework.
\subsubsection{Night selection}
Our goal for this study was to obtain the most complete phase coverage between quadratures (orbital phases $0.25 < \varphi < 0.75$), while at the same time minimizing the size of the dataset to keep it computationally feasible to run Markov-Chain Monte Carlos on the data. We note that the secondary eclipse (SE) of KELT-9b occurs between orbital phases $0.44<\varphi<0.56$, and the spectra falling within this interval were excluded from the analysis.
Overall, we combined three nights of proprietary HARPS-N (\citealt{Cosentino2012}; $R=115000$, spectral coverage $390$--$690~\mathrm{nm}$) data\footnote{Observed in GIARPS (GIANO-B + HARPS-N) mode \citep{Claudi2017}; but we only use the HARPS-N data in this work.}
with two nights of public CARMENES (\citealt{Quirrenbach2016, Quirrenbach2018}; $R\sim94600$, spectral coverage $520$--$960~\mathrm{nm}$ in the optical arm) data\footnote{http://caha.sdc.cab.inta-csic.es/calto/jsp/searchform.jsp}.
We define our naming convention for the observing nights and list relevant parameters in Table~\ref{table:observing_nights}. We visualize the orbital phases covered
in Figure~\ref{fig:orbital_coverage}. We only used the optical arm of CARMENES, where we expect most of the \ion{Fe}{i} signal to arise.
\begin{figure}
\resizebox{\hsize}{!}
{
\includegraphics[trim={0cm 1cm 2cm 0}]{orbit_with_phases.pdf}}
\caption{Schematic representation of the orbit of KELT-9~b to scale, with the orbital phases covered in this study colour-coded by observing night (Blue: HARPS-N N1; Orange: HARPS-N N2; Purple: HARPS-N N3; Red: CARMENES N1; Green: CARMENES N2). The exoplanet orbits counterclockwise (curved arrow), and the observer is positioned at the bottom of the figure.}
\label{fig:orbital_coverage}
\end{figure}
\begin{table*}
\caption{Details of the five observing nights included in this study.}
\label{table:observing_nights}
\centering
\begin{tabular}{cccccccc}
\hline \hline
Name & Date & $N_\mathrm{spectra}$ & Orbital phase & Airmass & $t_\mathrm{exp}$ & S/N & Instrument \\
\hline
HARPS-N N1 & 2018-07-22 & 84 & 0.233-0.457 & 1.02--1.63 & 300~s$^\star$ & 97-112-126 & HARPS-N \\
HARPS-N N2 & 2019-11-16 & 42 & 0.598-0.703 & 1.05--1.63 & 300~s & 55-63-89 & HARPS-N \\
CARMENES N1 & 2019-05-28 & 72 & 0.587-0.715 & 1.00--1.80 & 172~s & 65-96-109 & CARMENES \\
CARMENES N2 & 2018-09-02 & 44 & 0.557-0.741 & 1.01--1.80 & 300~s & 78-113-141 & CARMENES \\
HARPS-N N3 & 2021-09-05 & 78 & 0.577-0.772 & 1.02--2.09 & 300~s & 71-103-114 & HARPS-N \\
\hline
\end{tabular}
\tablefoot{Columns list the naming convention followed in this paper, the observing date, the number of observed spectra, the orbital phase range covered, the airmass range, the exposure time, typical S/N of the observations and the instrument. The quoted range in S/N is the 10-50-90 percentile of the median S/N calculated at the central pixel of each order.\\
$^\star$ In \cite{Pino2020} we reported an incorrect exposure time of 180 s. The correct exposure time is 300 s.}
\end{table*}
The HARPS-N nights were obtained within a Large Program (PI: Micela) awarded to the GAPS Collaboration \citep{Poretti2016}. This program already yielded results on atmospheric characterization (e.g., \citealt{Borsa2019, Pino2020, Guilluy2020, Rainer2021, Scandariato2021}) and is described in \cite{Guilluy2022}. We initially included the HARPS-N night published in \citet{Pino2020}, covering phases pre-SE, and a second HARPS-N night post-SE observed at lower S/N. To compensate for the latter, we added the two best archival nights of CARMENES post-SE. While performing the analysis, we observed a third HARPS-N night, also post-SE, which we integrated into the analysis to test for consistency between different instrument and observing dates.
The two CARMENES nights included spectra at very low S/N and very high airmass. Four spectra at visibly low S/N between phase 0.55 and 0.60 were removed from the night of 2018 September 2. To do so, a threshold of 35\% of the maximum median S/N calculated across all wavelengths was applied. Furthermore, we excluded from the analysis spectra taken at airmass greater than 1.8. With this choice, we minimized the size of the confidence intervals on the retrieved planetary orbital velocity. However, the exact choice of the airmass threshold does not significantly affect the combined analysis. Finally, we remove the 10 reddest orders of CARMENES from the analysis due to very low S/N.
At our unusually high level of precision obtained in the planet's radial velocity, it is important to account for the fact that the timestamps stored in the headers of both CARMENES and HARPS-N spectra contain BJD dates computed in the UTC frame rather than in the TDB frame. Indeed, between 2017 and 2021 included, the difference BJD UTC -- BJD TDB is fixed and amounts to 69.2 s, which is non-negligible for us. All the dates in the file headers are correctly weighted by the photometric barycentre of the exposure, i.e. they account for the fact that varying flux during the exposure can effectively offset the point where half of the flux is collected from the middle of the exposure.
All in all, our full set of data consists of 320 spectra covering orbital phases $\varphi$ = 0.233-0.772 excluding the secondary eclipse as mentioned above. While the reliance on just one HARPS-N night pre-SE might seem unbalanced, the exceptionally high quality of the pre-SE HARPS-N night results in similar confidence intervals from the pre-SE and the post-SE analysed separately.
\subsubsection{Data reduction}
\label{Sec: data reduction}
We design a pipeline that we can uniformly apply to HARPS-N and CARMENES. Both HARPS-N (we used the e2ds spectra) and CARMENES spectra are provided with a wavelength solution in the rest frame of the observer, in air for HARPS-N and in vacuum for CARMENES. We shift the HARPS-N wavelength solution to vacuum to be able to compare them with models. After deblazing and color-correcting the spectra, we thus proceed to shift them into the barycentric reference frame of the Solar System by applying the barycentric correction included in the respective file headers. We followed \citep{Pino2020} for this part of the analysis.
We then apply a telluric and stellar lines correction adapted from \citet{Giacobbe2021} and originally designed to clean near-infrared data from the more severe effects of telluric and instrumental systematic effects. It consists in the application of a Principal Component Analysis (PCA) algorithm to blindly identify ``state vectors'' in common between spectral channels and use them to describe the time-evolution of telluric lines, and remove them. We work in the barycentric rest-frame, where stellar lines are stationary and telluric lines are quasi-stationary. In this way, we slightly privilege correction of stellar lines over telluric lines, although practically the algorithm removes both components. This choice is driven by our focus on neutral iron lines that are typically present both in the stellar and planetary spectrum, but not in the Earth's atmosphere. Since in this work there are some differences with the algorithm in \citet{Giacobbe2021}, we fully describe our implementation of the PCA. We note that in exoplanet literature sometimes PCA is performed by only using the Singular Value Decomposition (SVD) algorithm, while in our case SVD is just one of the three main steps of the analysis. While conceptually very similar, the full PCA algorithm described here and the SVD-only algorithm differ in the way the wavelength channels in the data are weighted.
We describe the data as a cube with three axes: order number, time (or phase), and wavelength. One cube is stored for each night of observations. Due to the fixed layout of the echellogram for both CARMENES and HARPS-N, the dimensions along the order axis is always 61 for CARMENES and 69 for HARPS-N. The corresponding number of pixels along the wavelength dimension is 4096 for both spectrographs.
The PCA algorithm in this analysis is entirely written in Python and based on {\tt Numpy} methods. The code is run on each night and each order separately, recognizing not only that the atmosphere behaves differently on each night, but also that different orders are affected by different time-correlated sources, both astrophysical and instrumental. The input matrix for the algorithm is always a two-dimensional array, with time on the vertical axis (along a column) and wavelength on the horizontal axis (along a row).
The PCA algorithm, applied independently to each spectral order and each observing night, is outlined as follows:
\begin{enumerate}
\item Initial masking: non-finite flux values ({\tt NaN}), as well as values below 2\% of the median flux level (low-S/N) of the spectrum are flagged and a mask is created to keep track of such pixels. Spectral channels (data columns) with more than 1 invalid pixel are entirely masked.
\item Standardization: each column (each spectral channel) has its mean subtracted and is divided by its standard deviation. At the end of the process, each column has thus mean zero and unit standard deviation. This step makes sure that the PCA weights all wavelength bins equally while looking for common modes in the next step.
\item Singular Value Decomposition (SVD): this is computed via {\tt numpy.linalg.svd()} on the standardized matrix (step 2), with the option {\tt full\_matrices = False}. SVD decomposes the array in a set of orthogonal eigenvectors and corresponding eigenvalues. In our case, there are as many eigenvectors(values) as the number of spectral channels, i.e. the dimension of the wavelength axis. In PCA jargon, this choice corresponds to running the PCA in the ``time domain'', i.e. with time (or orbital phase) as an independent variable. Note that we only feed the SVD algorithm the columns of data that were not masked at step 1.
\item Component selection: SVD ranks eigenvectors according to their contribution to the data variance. Therefore, only the first few eigenvectors are needed to describe most of the flux variations in telluric lines. In our case, by visual inspection, we determined that 2 eigenvectors (also called ``components'') are sufficient to suppress telluric lines below the level of the noise for all the spectral orders, and we use those in the next step.
\item Multilinear regression: the eigenvalues obtained via SVD do not correctly describe the observed data, because they have been computed on the standardized flux array (step 2) rather than on the measured flux array (step 1). Therefore, we run a multilinear regression (MLR) between the set of two eigenvectors $\vec{v}_1(t), \vec{v}_2(t)$ determined at step 4 and the matrix stored at step 1. For each spectral channel $i$ in the matrix, the MLR calculates two eigenvalues $c_{1i}, c_{2i}$ plus an offset $c_{0i}$.
\item For each spectral channel $i$, we calculate the linear combination
$l_i(t) = c_{0i}+c_{1i}\vec{v}_1(t) + c_{2i}\vec{v}_2(t)$ and divide the corresponding column of the matrix at point 1 through it, obtaining a residual matrix.
\item High-pass filtering: columns in the residual matrix (step 6) that deviate by more than 3 times the standard deviation of the whole matrix are masked. Such mask is then merged via boolean {\tt OR} with the mask at step 1, and stored for future use during model reprocessing (Section~\ref{sec:modelling}). Low-order variations in each spectrum, i.e. along the wavelength axis, are fitted with a second order polynomial (excluding the masked pixels) and divided out.
\item The mean of each spectrum obtained in step 7 is subtracted out, which is necessary to compute the log-likelihood function. We note that the mean is calculated by excluding the masked pixels; these are then manually reset to zero so that they do not contribute to the log-likelihood function.
\end{enumerate}
The 5 resulting data cubes (one per observing night) are then compared to models.
\subsection{Models}\label{sec:modelling}
We produce a set of nested models to test the presence of 3D effects in the atmosphere of KELT-9b, generalizing the best-fit 1D model by \cite{Pino2020}. This is calculated employing a custom line-by-line radiative transfer code which solves the radiative transfer equation in its integral form using a ``linear in optical depth approximation'' for the source function \citep{Toon1989}. It assumes the temperature-pressure profile calculated by \cite{Lothringer2018}. Volume mixing ratios are calculated under the assumption of equilibrium chemistry using \texttt{FastChem} version 2 \citep{Stock2018}, with stellar (solar in KELT-9's case) composition. Following \cite{Pino2020} we only include lines from neutral iron, whose opacities we compute starting from the VALD3 database \citep{Piskunov1995, Ryabchikova1997, Kupka1999, Kupka2000, Ryabchikova2015}\footnote{Additional references specific to the used \ion{Fe}{i} line list: \cite{K14, BKK, BK, BPM, BWL, FMW}.}, and the bound-free ${\mathrm{H}^-}$ opacity from \cite{John1988}. We convolve the model with two kernels: the former is a Gaussian kernel with FWHM matching the velocity resolution of the HARPS-N/CARMENES spectrographs; the latter is a rotational kernel calculated assuming a rigidly rotating planet with equatorial velocity of 6.64~$\mathrm{km~s^{-1}}$. This value is obtained by assuming synchronous rotation and the planet radius from \cite{Gaudi2017}. We refer to \cite{Pino2020} for other details about the radiative transfer code.
Our one-dimensional, fiducial model 1C is built using the best-fit spectrum by \cite{Pino2020} as a template, and is described by three parameters: $K_\mathrm{p}$, $v_\mathrm{sys}$ and $\mathrm{S}$. The parameter $\mathrm{S}$ is a multiplicative scaling factor that accounts for a mismatch between the intensity of model lines and lines in observations \citep{Brogi2019}. In the lack of any additional contribution to Doppler shift (e.g., atmospheric winds), the parameters $K_\mathrm{p}$ and $v_\mathrm{sys}$ are physically interpreted as the maximum Keplerian orbital velocity of the planet (assuming a circular orbit in the case of model 1C) and the systemic velocity of the system, i.e. the constant component of the radial velocity of the system’s centre of mass relative to the solar system barycentre, which may also include an instrument-dependent radial velocity offset \citep{Deeg2018}. We keep nomenclature consistent with the literature, but caution the reader that additional velocity contributions present in the data will be captured by this model, changing the physical interpretation of these parameters (e.g., see Sec. \ref{sec: discussion_climate}). The same caveat applies to the following models.
Model 1C contains limited information about the phase-dependence of the planet spectrum. Indeed, it has a constant shape across phases, and has a varying Doppler shift with planetary phase that is forced to follow a Keplerian, circular orbit around the system centre of mass. This model has been used by \cite{Pino2020}, who show that it reproduces well the average \ion{Fe}{i} emission line of the planet in HARPS-N N1. We generalize this model by relaxing two separate assumptions: (1) the rest frame from which the signal arises is revolving around the star KELT-9 with a circular motion; (2) the flux emitted from the photosphere of the planet is uniform with longitude. Table \ref{table:models adopted} summarizes the models and their parameters, and in the following we describe them more in detail.
We design an eccentric orbit model 1E for KELT-9b using the parameters $h=\sqrt{e}\sin\omega$ and $k=\sqrt{e}\cos\omega$. $e$ is the eccentricity, while $\omega$ is the argument of periastron. This parameterization performs comparable or faster than alternatives in an MCMC, and, in addition, it is free of the burden of dealing with angular parameters. Finally, it naturally sets a uniform prior in $e$ and $\omega$ that is regarded as the most sensible choice for hot Jupiters \citep{Eastman2013}. This model is described by five parameters: $K_\mathrm{p}$, $v_\mathrm{sys}$ and $\mathrm{S}$ that are the same as for the circular orbit model 1C, and $h$ and $k$ are in addition.
Finally, we employ two simple modelling approaches to capture the possible variation in intensity of the iron lines as a function of longitude in the planet atmosphere. In both approaches, we neglect latitudinal variations of atmospheric properties, thus properties measured at a given longitude should be seen as a latitudinal average. In the first approach, the two four-scale factor models 4C (circular orbit) and 4E (eccentric orbit) divide the orbital range in 4 parts ($0.25 < \varphi < 0.35$, $0.35 < \varphi < 0.45$, $0.55 < \varphi < 0.65$, $0.65 < \varphi < 0.75$). We assign a different scale factor ($S_1$, $S_2$, $S_3$, $S_4$) to each of these phase ranges, while the orbital parameters ($K_\mathrm{p}$, $v_\mathrm{sys}$, and additionally $h$ and $k$ in the case of an eccentric orbit) are in common. With this choice, our models are able to capture an inhomogeneous intensity of \ion{Fe}{i} lines. For instance, if the \ion{Fe}{i} signal mostly comes from the planet day-side, $S_1$ and $S_4$ should be smaller than $S_2$ and $S_3$, since a larger portion of the night-side contributes to determine them. The shortcoming of this modelling approach is that it mixes information coming from different longitudinal slices of the planet atmosphere in a suboptimal way. Indeed, a given longitudinal slice of the planet atmosphere could be observable in multiple phase ranges. The second modelling approach that we consider in this work overcomes this limitation. We adopt three models based on reflection off a Lambert sphere \citep{CollierCameron2002}. While this is likely unrealistic, we argue that these models have the desirable feature that line intensity has a maximum (whose position is a parameter model Loff), and then decreases symmetrically when moving further away in longitude. This would, for instance, reproduce a shallower thermal gradient profile, or a decrease of neutral iron abundance while moving towards the night-side. In this model, we substitute the constant multiplicative scale factor $S$ with the phase function $g(\varphi)$:
\begin{align}
\label{Eq.: Lambert_scale_factor}
&g(\varphi) = \frac{S}{\pi} + \frac{S_\mathrm{Lambert} \cdot \left[ \sin\alpha + \left( \pi - \alpha\right)\cdot \cos\alpha \right]}{\pi}\ ,\\
&\cos\alpha = -\cos\left(2\pi\varphi - 2\pi\varphi_0\right)\ ,
\end{align}
where we have assumed a perfectly edge-on orbit and $\varphi$ varies between 0 (inferior conjunction) and 1. In model Lbase, the free parameters are $K_\mathrm{p}$, $v_\mathrm{sys}$, $S$ (constant scale factor contributing emission independent of planetary phase), $S_\mathrm{Lambert}$ (scale factor of additional phase-dependent \ion{Fe}{i} line emission), and $h$ and $k$, and we fix $\varphi_0 = 0$ (offset of maximum scale factor from sub-stellar point). We additionally test model L where we also suppress the phase-independent emission ($S=0$, $\varphi=0$).
All our models rely on several simplifying assumptions, and adopt a parameterized approach that is agnostic of the underlying physics. In future work, we will consider more physically motivated, parameterized models able to partially account for the interplay between stellar irradiation, atmospheric dynamics and structure, and thus predict variations in thermal profile and chemistry as a function of longitude (e.g., \citealt{DobbsDixon2022}).
\begin{table*}
\caption{Model description, parameters and priors. See text for detailed explanation of parameters.}
\label{table:models adopted}
\centering
\begin{tabular}{cclc}
\hline \hline
Model name & Parameters & Short description & Priors\\
\hline
\textbf{1C} & $K_\mathrm{p}$ & Circular orbit, & [100, 300] km s$^{-1}$ \\
(3 parameters) & $v_\mathrm{sys}$ & single scale factor & [$-70$, 20] km s$^{-1}$ \\
& $S$ & (as \citealt{Pino2020}) & [0.01, 100] \\
\hline
\textbf{4C} & $K_\mathrm{p}$ & Circular orbit, & [100, 300] km s$^{-1}$ \\
(6 parameters) & $v_\mathrm{sys}$ & four scale factors & [$-70$, 20] km s$^{-1}$ \\
& $S_1$, $S_2$, $S_3$, $S_4$ & & [0.01, 100] \\
\hline
\textbf{1E} & Same as 1C + & Eccentric orbit, & \\
(5 parameters) & $h=\sqrt{e}\sin\omega$, $k=\sqrt{e}\cos\omega$ & single scale factor & [$-1, 1$] \\
\hline
\textbf{4E} & Same as 4C + & Eccentric orbit, & \\
(8 parameters) & $h=\sqrt{e}\sin\omega$, $k=\sqrt{e}\cos\omega$ & four scale factor & [$-1, 1$] \\
\hline
\textbf{L} & $K_\mathrm{p}$ & Lambert sphere, & [100, 300] km s$^{-1}$ \\
(5 parameters) & $v_\mathrm{sys}$ & eccentric orbit & [$-70$, 20] km s$^{-1}$\\
& $S_\mathrm{lambert} / \pi$ & & [0, 10] \\
& $h=\sqrt{e}\sin\omega$, $k=\sqrt{e}\cos\omega$ & & [$-1, 1$] \\
\hline
\textbf{Loff} & Same as L + & Lambert sphere, & \\
(6 parameters) & $\varphi_0$ & eccentric orbit, & [$-0.15,0.15$]\\
& & offset from substellar point & \\
\hline
\textbf{Lbase} & Same as L + & Lambert sphere, & \\
(6 parameters) & $S/\pi$ & eccentric orbit, & [0, 10]\\
& & with baseline constant emission & \\
& & & \\
\end{tabular}
\end{table*}
\subsection{Cross-correlation-likelihood mapping}
\cite{Pino2020} used two different methods to stack the signal from thousands of iron lines, yielding the first detection of iron in the emission spectrum of KELT-9b: (1) a weighted mask method, and (2) a cross-correlation-likelihood mapping scheme by \cite{Brogi2019}. In this paper, we employ the cross-correlation-likelihood mapping method, and provide our rationale for this choice in Appendix \ref{sec_appendix: mask_vs_mapping}. We stress that each method has strengths and weaknesses, and the choice should be driven by the science goals.
The mapping is achieved through the use of the log-likelihood function
\begin{equation}\label{eq:log-L}
\log L = -\frac{N}{2}\log(s_f^2+s_g^2-2R)\,,
\end{equation}
where $N$ is the number of valid (i.e. unmasked) spectral channels on each spectrum and each order, $s_f^2$ and $s_g^2$ are the data and model variances, respectively, and $R$ is the cross-covariance between model and data. While cross-correlation does not appear explicitly in this formula, its numerator ($R$) and denominator $(s_fs_g)$ are present in the argument of the natural logarithm. We partition Eq. (\ref{eq:log-L}) by calculating it on each order and frame. This correctly weights for noise variations, directly estimated from the sample variance, across orders and frames. This methodology has been shown to produce unbiased and accurate results \citep{Brogi2019, Gibson2022}.
As pointed out since the early application of the log-likelihood framework \citep{Brogi2019,Pino2020,Gibson2020}, the model spectrum cannot be compared to the data directly, i.e. by just scaling and shifting it. In order to avoid biases in the retrieved parameters, it is important to reproduce on the model any possible alteration of the planet signal induced by the data reduction process. This is achieved by running a parallel data analysis chain (steps 1-6 in Section~\ref{Sec: data reduction}) on a data cube where the model is injected on top of the observed data at a reduced amplitude of 1\%\footnote{We tested that optimizing this additional scaling factor between 10\% and 0.001\% of the model flux has no impact on the result}, so that the injection is not altering the analysis, in particular the ranking of the SVD components at step 3. Each injected spectrum $F_\mathrm{inj}$ is obtained from the observed spectrum $F_\mathrm{obs}$ via\footnote{Except for a constant offset of 1, subtracting or dividing through the PCA-processed data at this stage differs by a factor of order $(F_\mathrm{pl}/F_\mathrm{\star})^2$, that is a difference of the order of $10^{-8}$ in absolute flux (or $10^{-4}$ in relative flux). Here we opt for subtracting to avoid possible issues with division by values close to zero.}
\begin{equation}
F_\mathrm{inj} = F_\mathrm{obs} \left( 1 + 0.01\frac{F_\mathrm{pl}(\vec\theta)}{F_\star}\right),
\end{equation}
where $\vec\theta$ is the state vector containing all the model parameters at the current MCMC step, and the model is expressed in planet/star units by normalizing through a stellar black body at the effective temperature of KELT-9.
After step 6 of the analysis, the observed dataset at the same stage is subtracted out, so that the resulting data cube represents the (noiseless) processed model. The latter is then multiplied by 100 to restore its original amplitude, mean-subtracted as in step 8, and used to calculate the log-likelihood values through Eq.~\ref{eq:log-L} after masking (i.e. setting to zero) the same pixels as for the observed data. We skip the application of a high-pass filter (step 7) on the model. While early tests indicate no impact on the final results, we believe that it is conceptually different to apply the filter on observed data (dominated by instrumental and photon noise) than on the noiseless model (dominated by signal variations).
The difference between the model injected at step 1 and the model processed as above is shown in Fig.\ref{fig:processing}, and it shows how even by visual inspection there are evident alterations in the depth and shape of the planetary lines that need to be accounted for. As expected, the effect is phase-dependent and particularly severe close to quadrature, where the planet trail is nearly parallel to telluric and stellar lines. In Appendix \ref{sec_appendix: injection}, we show that our reprocessing methodology is capable of recovering a phase-dependent signal accurately at our level of precision.
\begin{figure*}
\resizebox{\hsize}{!}
{
\includegraphics[trim={0 0 3cm 0}]{injection.pdf}
\includegraphics[width=0.42\textwidth]{injection_signal_loss.pdf}
}
\caption{Effects of telluric removal on the signal from KELT-9 b. {\sl Left:} the best-fit \ion{Fe}{i} emission model is Doppler-shifted based on the orbital parameters of KELT-9 b (panel a), injected on top of the 1st night of HARPS-N data, and passed through the analysis described in Section~\ref{Sec: data reduction}. The signal is recovered with evident alterations, as shown in the reference frame of the observer (panel b) and in the planet's rest frame (panel c). {\sl Top-right:} recovered amplitude of a strong injected Fe line as a function of orbital phase (solid) versus its amplitude at injection (dashed). {\sl Bottom-right:} percent loss of the Fe signal due to the telluric-removal analysis. Nearly 100\% of the signal is lost around quadrature.}
\label{fig:processing}
\end{figure*}
Currently, model reprocessing is the bottleneck of our retrieval code. In order to optimise it, we follow the practice of \citet{Gibson2022} and reuse for model reprocessing the same eigenvectors calculated by the SVD algorithm on the observed spectra (step 3). We also reuse the same mask for both analysis and reprocessing. With these choices, we speed up the calculation by a factor of about 2.
\section{Results}
\label{sec:results}
\subsection{Confirmation of iron emission lines from the day-side of KELT-9b}
We run an MCMC using model 1C, the same that was adopted by \cite{Pino2020}, on each individual night. We detect neutral iron in each of the unpublished HARPS-N N2 and N3, and in each of the two CARMENES N1 and N2.
This confirms the results by \cite{Pino2020} and \cite{Kasper2021}, based on pre-eclipse observations, and shows that neutral iron is present also at post-eclipse phases. Our best single-night detection remains HARPS-N N1, which featured better observing conditions and twice the exposures compared to HARPS-N N2. In HARPS-N N1, we measure a $K_\mathrm{p}$ which is incompatible at $2\sigma$ with \cite{Pino2020}, despite using the same data. This discrepancy is due to the different ephemeris adopted in this work (see Sec. \ref{sec: Ephemeris of KELT-9b}). We directly verified that our new PCA pipeline, when applied using the same ephemeris adopted by \cite{Pino2020}, reproduces their best-fit planetary orbital velocities. HARPS-N N3 features a similar total S/N compared to HARPS-N N1, but provides looser constraints on $v_\mathrm{sys}$ and $K_\mathrm{p}$, and a marginally lower scale factor. Furthermore, CARMENES provides looser confidence intervals on the parameters compared to HARPS-N. For comparison, the precision reached on all parameters in the best of the two CARMENES nights (CARMENES N1) is comparable to the precision reached in the worst HARPS-N night (HARPS-N N2).
It is not immediately clear why HARPS-N outperforms CARMENES in our analysis, and multiple factors could be at play. First, the S/N estimate for each CARMENES and HARPS-N dataset is provided by the respective pipelines, and are thus not necessarily comparable. A proper assessment of the S/N would require us to homogeneously reduce the raw frames of both instruments, which is out of the scope of this paper. Second, the amount of signal intrinsically carried in different wavelength ranges could differ, because of the difference in strength and number of \ion{Fe}{i} lines, and in planet-to-star flux ratio. In Appendix \ref{sec_appendix: intrisic_information_content} we show that bluer wavelengths (probed by HARPS-N) seem to carry more information about \ion{Fe}{i} compared to redder wavelengths (probed by CARMENES) in KELT-9b. Finally, telluric correction is more severe at the redder wavelengths probed by CARMENES. We can not exclude that, if present, the cumulative effect of slight telluric residuals left behind by our analysis, despite being buried in the photon noise, would more likely negatively impact the CARMENES observations.
\begin{figure*}
\resizebox{\hsize}{!}
{\includegraphics{Compare_nights_all_paramters.pdf}}
\caption{Posterior distributions of parameters of model 1C fit with an MCMC to HARPS-N N1 (blue), HARPS-N N2 (orange), CARMENES N1 (red), CARMENES N2 (green), HARPS-N N3 (purple), and all 5 nights combined (black). The lower panels show the normalized posteriors (higher peak for better constrained parameters), and the marked improvement obtained when combining all data sets. The upper panels show the $1\sigma$ (dark color) and $2\sigma$ (light colour) confidence intervals obtained from the corresponding posteriors in the lower panels.
}\label{fig:night_comparison}
\end{figure*}
Most of the individual posteriors are compatible with each other at $1\sigma$ (see Fig. \ref{fig:night_comparison} and Appendix \ref{sec_appendix: individual nights}). The only exceptions are: the posteriors for the scaling factor in HARPS-N N2 (orange posterior) and HARPS-N N3 (purple posterior), marginally incompatible with each other; the posteriors for $K_\mathrm{p}$ in HARPS-N N1 (blue posterior), which is only compatible with HARPS-N N2 (orange posterior). Since the discrepancy in the posteriors for the scaling factor between HARPS-N N2 and HARPS-N N3 (orange and purple posterior) is little more than $2\sigma$, we do not consider it significant. We thus refrain from interpreting it as an astrophysical signal, also in light of the significant impact of the PCA on the strengths of \ion{Fe}{i} lines (see Fig. \ref{fig:processing}) that needs to be very accurately captured by our injection method for a correct interpretation. We do note, however, that HARPS-N N3 was observed at least two years after the other epochs, so that any variability in the planet atmosphere on year-long timescales would be visible. On the other hand, the discrepancy in $K_\mathrm{p}$ is more likely real. It concerns the only pre-eclipse night that we observed, and could be indicative of a pre-eclipse / post-eclipse asymmetry. We exclude that this is caused by the nodal precession of KELT-9b \citep{Stephan2022}, which would only change the inclination of the orbit by about $1^\circ$ in three years. Its effect would only have a relative impact on $K_\mathrm{p}$ at the $10^{-4}$ level. We will come back to our interpretation of the variation of $K_\mathrm{p}$ in the next section and in section \ref{sec: discussion_climate}.
We also perform a combined analysis of the nights.
The combined analysis leads to a marked improvement in the precision of each parameter (the relative corner plot is displayed in Fig. \ref{fig:MCMC_eccentric_versus_circular}). Our best fit parameters are $v_\mathrm{sys} = -16.5^{+0.4}_{-0.4}~\mathrm{km~s^{-1}}$, $K_\mathrm{p} = 241.8^{+0.5}_{-0.5}~\mathrm{km~s^{-1}}$, and $\log S = 0.29^{+0.03}_{-0.03}$ (see Table \ref{table:retrieval results}). As \cite{Pino2020}, we find that our model under-predicts the Fe I line intensity by a factor of about 2. This could be due to an underestimation of the temperature of the upper atmosphere (e.g., \citealt{Fossati2020, Fossati2021}), an underestimation of the \ion{Fe}{i} volume mixing ratio, an overestimation of the stellar flux, the presence of other atomic and molecular species in the spectrum that partially mask neutral iron spectral features \citep{Kasper2021}, or a combination of these. Understanding the source of this discrepancy is out of the scope of this paper, and our conclusions do not hinge on reproducing exactly the Fe I line depth with our one-dimensional model template.
\subsection{Neutral iron lines trace a displacement from a circular orbit for KELT-9b}
\label{sec: results, eccentric_orbit}
Encouraged by the sub-kilometre-per-second precision reached on 5 combined nights with model 1C, we perform a fit using an eccentric orbit (model 1E, see Table \ref{table:models adopted}). We obtain bound posteriors for all parameters of model 1E. We convert the posteriors in $h$ and $k$ to posteriors in $e$ and $\omega$ by converting every MCMC sample of $h$ and $k$ using $e = h^2 + k^2$ and $\omega = \arctan (h/k)$, and building the corresponding probability distribution. We show the posteriors in Fig. \ref{fig:MCMC_eccentric_versus_circular}, and our best fit parameters are: $v_\mathrm{sys} = -15.9^{+0.35}_{-0.4}~\mathrm{km~s^{-1}}$, $K_\mathrm{p} = 239.9^{+0.8}_{-1}~\mathrm{km~s^{-1}}$, $\log S = 0.31^{+0.02}_{-0.03}$, $e = 0.016^{+0.003}_{-0.003}$, and $\omega = 150^{+13\,\circ}_{-11}$ (see Table \ref{table:retrieval results}).
This result is surprising at face value, and in apparent contrast with \cite{Wong2020}. Indeed, they report a tight upper limit to the eccentricity of the planet of $e<0.007$ at $2\sigma$ by using TESS photometry, and additionally find that the addition of parameters $e$ and $\omega$ is disfavoured by BIC and AIC in their analysis, indicating a likely circular orbit, as is expected for KELT-9b. In sections \ref{sec: discussion_eccentricity} and \ref{sec: discussion_climate} we reconcile our result with \cite{Wong2020} by interpreting our measured eccentric as apparent, and actually due to atmospheric dynamics (winds). For now, we just remark that model 1E captures a significant deviation from a circular orbit traced by \ion{Fe}{i} lines, independently of the source of the anomaly.
\begin{figure*}
\resizebox{\hsize}{!}
{\includegraphics[width=\hsize]{corner_eccentric_versus_circular.pdf}}
\caption{Posterior for the MCMC fit of models 1C (circular orbit, black histograms) and 1E (eccentric orbit, blue histograms). The $1\sigma$ and $2\sigma$ confidence intervals are indicated with darker and lighter shades, respectively.}
\label{fig:MCMC_eccentric_versus_circular}
\end{figure*}
Since model 1E has additional parameters compared to model 1C, we perform model comparison to determine whether the additional complexity is justified by the data, using several methods: likelihood ratio test, Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC)\footnote{These values should be computed in the maximum likelihood, which we have approximated with the median of the posterior. With a flat prior and perfectly symmetrical posteriors, this is a correct assumption.}. Since model 1C and 1E are nested, we apply the likelihood ratio test (as described in Appendix D in \citealt{Pino2020}) with 2 fixed parameters and obtain a $5\sigma$ preference for the eccentric model. We also obtain the apparently contrasting results that the BIC test strongly favours the circular orbit solution ($\mathrm{BIC_{1E}} - \mathrm{BIC_{1C}}= 7.9$), while AIC strongly favours the eccentric orbit solution ($\mathrm{AIC_{1E}} - \mathrm{AIC_{1C}}= -24.5$, which corresponds to a $4.5\sigma$ preference\footnote{Calculated from the relative likelihood of model 1C $\exp (\mathrm{AIC_{1E}} - \mathrm{AIC_{1C}})/2$ converted to $\sigma$ level using a two-tailed normal distribution test} for model 1E). In reality, this is not surprising: the BIC test penalizes higher complexity models much more compared to the AIC test, and this difference is more marked for larger data-sets. We fit simultaneously $n_\mathrm{tot}$ = 71978272 pixels. Thus, for a change of 2 in the number of parameters, the penalty term in the BIC test is $2\log(n_\mathrm{tot}) = 36.2$,
compared to 4 for the AIC test. In other words, the BIC test is much more demanding in terms of quality of fit for higher complexity models.
We now compare the posterior distributions for models 1C and 1E (see Fig. \ref{fig:MCMC_eccentric_versus_circular}). Eccentricity and argument of periastron are partially degenerate with $K_\mathrm{p}$. As a result, the marginalized distribution for $K_\mathrm{p}$ is slightly broader in model 1E.
All common parameters between models 1C and 1E are compatible at $2\sigma$. The eccentricity that we measure has a significance of more than $5\sigma$ level, which, in combination with the AIC test, suggests that the higher complexity model introduced by model 1E over 1C is justified.
We developed a new method to display the radial velocity displacement from the best-fit planet's rest frame ($\Delta v_\mathrm{rest}$) as a function of planet phase (planet trail) in the context of the likelihood framework by \cite{Brogi2019}. Conceptually, we can build the planet trail for a given solution as the conditional probability\footnote{As opposed to marginal probability: in this case we are interested in the value of $\Delta v_\mathrm{rest}$ for the given best fit solution, rather than trying to identify the most likely values of $\Delta v_\mathrm{rest}$ independently of the other parameters.} of $\Delta v_\mathrm{rest}$ at every planetary phase, given the best fit values for the model considered. In this representation, the confidence interval obtained at every phase is completely independent of all other phases, and confidence intervals of adjacent phases are not related. Practically, we divide the covered phase range in 0.02-wide phase bins, and for each phase bin we calculate the conditional likelihood as:
\begin{equation}
\label{Eq:planet_trail}
\mathcal{L}_\mathrm{trail}\left( \Delta v_\mathrm{rest}; \varphi \right) = \prod_\mathrm{exposure\,1, \varphi}^{\mathrm{exposure\,n}, \varphi} \mathcal{L}\left( K_\mathrm{p\,BF}, v_\mathrm{sys\, BF} + \Delta v_\mathrm{rest}, S_\mathrm{BF}; \varphi \right),
\end{equation}
where each phase bin is sampled by n exposures (n could differ in every phase bin), the subscript BF indicates the best fit value for a parameter, and $h$ and $k$ appear as additional parameters evaluated in the best fit position in the eccentric orbit case. The assumptions behind Eq. \ref{Eq:planet_trail} are that there is no atmospheric variability, and that each phase bin is small enough that we can neglect the planet motion within it. Finally, in every bin, we calculate the $1\sigma$ and $2\sigma$ deviations from the maximum likelihood in $\Delta v_\mathrm{rest}$ using Wilk's theorem. Figure \ref{fig:trail} displays the resulting trail in the best fit planetary rest frame using a box-plot \citep{Hyndman1996} for models 1C and 1E. Multi-peaked solutions likely correspond to phase bins where the best-fit solution is not very well constrained, so that $1\sigma$ and $2\sigma$ deviations capture additional spurious peaks. In the eccentric orbit case, the trail appears as a vertical line (except a few noisy phase bins), which is the expectation in the case that we are in the rest frame where the lines are originated. On the opposite, in the circular orbit case, the trail appears slanted. This indicates that a circular orbit is not fully capable of capturing the morphology of the planet trail or, in other words, that the rest frame from which the neutral iron signal is generated deviates from a circular orbit.
Concluding, our results markedly favour model 1E over model 1C, indicating the presence of a significant radial velocity anomaly detected for \ion{Fe}{i} lines. As we already mentioned, and further address in sections \ref{sec: discussion_eccentricity} and \ref{sec: discussion_climate}, such radial velocity anomaly is unlikely due to the orbital motion of the planet, and is more likely the result of atmospheric dynamics in KELT-9b.
\begin{figure*}
\resizebox{\hsize}{!}
{\includegraphics{Trail.pdf}}
\caption{Time-resolved confidence intervals as a function of the planet rest-frame velocity obtained by binning (i.e. co-adding) the individual likelihood functions by 0.02 in orbital phase. Black represents $1\sigma$ deviations from the best-fit rest-frame velocity in each phase bin and gray represents $2\sigma$ deviations. The left and centre panels show results for the 1C (circular) and 1E (eccentric) model, respectively, when fixing all the model parameters at their best-fit value, i.e. the median of the posterior distribution. Horizontal black dashed lines indicate phases corresponding to ingress and egress, and a vertical red dashed line indicates the best fit planet rest-frame. The right panel represents the total S/N achieved within each phase bin, with HARPS-N (lighter pink) and CARMENES (darker purple). \label{fig:trail}}
\end{figure*}
\subsection{Symmetric intensity of iron lines around secondary eclipse}
\label{sec: results_symmetry}
We then search a variation in the intensity of the observed neutral iron lines while the planet orbits around its host. First, we assume a circular orbit and fit a separate scale factor for 4 mutually exclusive ranges in phase (model 4C). Figure \ref{fig: single_versus_four_scale_factors}, left panel, shows the resulting posteriors for the scale factors. We show the full corner plots in Appendix \ref{sec_appendix:phase_dependent_intensity_corner_plots} (Fig. \ref{fig:4C_vs_1C}), and report best fit parameters and their errors in Table \ref{table:retrieval results}. The MCMC finds well constrained posteriors for all parameters, indicating that our data contains sufficient information to perform this kind of study. The posteriors of $K_\mathrm{p}$ and $v_\mathrm{sys}$ are in good agreement with those for model 1C. We attribute this to the fact that the scale factor (hence, neutral iron line intensity) and the orbital parameters (hence, neutral iron line Doppler shift) are not correlated at our level of precision.
\begin{figure*}
\resizebox{\hsize}{!}
{\includegraphics{Compare_scale_factors_4C_and_4E_final.pdf}}
\caption{Posterior distribution for the scale factors of models 1C and 4C (left panel), 1E and 4E (middle panel). The right panel shows the phase ranges corresponding to each of the 5 parameters for each panel, color-coded. \label{fig: single_versus_four_scale_factors}}
\end{figure*}
We additionally repeated the fit using model 4E, thus allowing for an eccentric orbit. We obtained bound posteriors for all parameters, and compare the resulting scale factors in the right panel of Fig. \ref{fig: single_versus_four_scale_factors}. We also show full corner plots in the Fig. \ref{fig:4E_vs_1E}, and report best fit parameters and their errors in Table \ref{table:retrieval results}. We observe that the scale factors are uncorrelated with all other parameters at our level of precision. As a result, the posteriors of the phase-resolved scale factors are very similar to those of model 4C, and the posteriors of orbital parameters are close to those of model 1E. We conclude that, at our level of precision, the Doppler-shift and line shape of the exoplanet atmosphere spectra can be modelled separately without any loss of information.
Finally, we run MCMC chains using Lambert sphere models (L, Loff, Lbase). We achieve convergence on all parameters in these models, and we report the full posteriors in Appendix (Fig. \ref{fig:Slam_vs_Soff}; Fig. \ref{fig:Slam_vs_Sbase}), and best fit parameters with their errors in Table \ref{table:retrieval results}. The orbital properties ($K_\mathrm{p}$, $v_\mathrm{sys}$, $e$, $\omega$) are in excellent agreement across the three models, confirming that Doppler and line intensity information are uncorrelated, and confirming the preference for an eccentric orbit.
We compare results from these models and results from model 1E and 4E in Fig. \ref{fig:collier_cameron_models}. A first clear result is that the peak of the emission is tightly constrained at the substellar point. This is supported by the posterior of the parameter $\varphi_0$ in model Loff, whose value is $\varphi = 0.00\pm0.01$,
which translates to an angular displacement from the substellar point of $0\pm5^\circ$. Secondly, model Lbase presents a degeneracy between $S_\mathrm{Lambert}$ and $S$. In other words, it is possible to explain the data both with an intensity profile which decreases moving towards the nightside (akin to model L, where contribution from the anti-stellar point is forced to 0), but also with a constant intensity profile independent of phase. Solutions with no $S_\mathrm{Lambert}$ contribution are also allowed, while the model supports at $>2\sigma$ significance the presence of a phase-independent contribution to emission. When comparing the intensity profiles (see Fig. \ref{fig:collier_cameron_models}), it is clear that both models L and Lbase agree in the range actually probed by observations (phases $0.25 < \varphi < 0.45$ and $0.55 < \varphi < 0.75$). In addition, all Lambert sphere models are in good agreement with the scale factors measured with model 4E in the probed planet phase range. We conclude that our data are not sufficient to support a strong evidence of a variation of the intensity of iron lines with longitude. This is further confirmed by the strong preference of model 1E over model L and Loff (e.g., $\mathrm{AIC_L - AIC_{1E} = 8.5}$, and same difference in BIC), and the marginal to strong preference over Lbase according to AIC and BIC ($\mathrm{AIC_{L_{base}} - AIC_{1E} = 2.4}$; $\mathrm{BIC_Lbase - BIC_{1E} = 18}$). We argue that Lbase is less disfavoured because it allows solutions with smaller to no dependence on orbital phase of the scale factor, further supporting our results.
On the other hand, we can confidently conclude that there is no detectable asymmetry (e.g. due to a hot-spot offset) in the intensity of lines in pre- and post-eclipse spectra. Indeed, model Loff is strongly disfavoured in terms of BIC and AIC compared to model L (by 19 and 3, respectively), and its parameter $\varphi_0$ presents a tight posterior around 0.
\begin{figure}
\centering
\includegraphics[width=\hsize]{scale_factor_versus_phase.png}
\caption{The upper panel shows 100 draws from the posterior for scale factors as a function of phase obtained by fitting models 1E (orange), 4E (blue), L (black), Loff (red) and Lbase (green) to the data. The phase coverage of the data is shown in the lower panel, in terms of the S/N of stellar spectra in each phase bin (HARPS-N: lighter pink; CARMENES: darker purple). Where multiple nights cover the same orbital phase bin their contributions are summed in quadrature.}
\label{fig:collier_cameron_models}
\end{figure}
\section{Discussion}
\label{sec:discussion}
\subsection{HRS phase curves as a new tool for exoplanet atmosphere characterization}
\label{sec:discussion_phase_curves}
Our results show that HRS emission spectroscopy observations of hot gas giants have the potential to reveal much more information when treated as a phase-curve rather than a single phase-collapsed emission spectrum, as was done in most literature works (e.g., \citealt{Pino2020, Yan2020,Nugroho2020, Borsa2022}). Conceptually, an HRS phase curve is similar to a classic low-resolution (LRS) phase curve observed from space (e.g., \citealt{Stevenson2016}). The key difference between LRS and HRS phase curves is the observable they record. For LRS phase curves, this is simply the flux of the planet-plus-star system as a function of planetary phase. On the other hand, HRS phase curves record two separate observable quantities.
The first observable quantity is the contrast of planet lines relative to the continuum generated by the star and planet as a function of phase (\citealt{Pino2020}; also noted and exploited by \citealt{Herman2022} and \citealt{vanSluijs2022}). Unlike in LRS phase curves, the spectral features of the planet continuum are lost because it is impossible to reconstruct the chromatic fibre losses, which are due to differential refraction from Earth atmosphere. Such losses introduce an additional, unknown wavelength and time dependence which need to be removed from the spectra in order to correct the stellar and telluric lines, and extract the planetary signal (e.g., Appendix A of \citealt{Pino2020}). This is typically performed with a normalization, which unfortunately also removes the exoplanet atmosphere continuum features. This may seem a strong limitation compared to LRS phase curves. However, when interpreted in an appropriate statistical framework \citep{Brogi2019}, HRS emission spectra still contain enough information to constrain the thermal profiles and the volume mixing ratios \citep{Pino2020, Kasper2021}, including their absolute values \citep{Line2021}. This extends to HRS phase curves. Indeed, both in this work and in \cite{Herman2022} and \cite{vanSluijs2022}, HRS phase curves are used to constrain the value of a scaling factor as a function of phase, which is a proxy for the steepness of the thermal gradient and volume mixing ratio of the species present in the cross-correlation template. In addition, compared to LRS phase curves, HRS phase curves probe higher up in the atmosphere, and, thanks to the simultaneous detection of lines of different strengths, a broader span of pressures.
The second observable quantity is the phase resolved Doppler-shift of planet lines. This is obtained by individually resolving the spectral lines (see section \ref{sec: results, eccentric_orbit}), and is only possible with HRS spectrographs. \cite{Herman2022} do not include this aspect in their analysis, and \cite{vanSluijs2022} indicate that at a resolution of 15000 this information is washed out. At the price of lower efficiency compared to lower resolution slit spectrographs, fibre-fed, $R\sim100000$ spectrographs open a new window into planetary climate and dynamics, which can not be obtained with any other technique (\citealt{Kawahara2012}; see sections \ref{sec: discussion_eccentricity}, \ref{sec: discussion_climate}).
Line contrasts and wind induced Doppler shifts are determined by the complex interplay between energy transport, chemistry and dynamics. They are thus intrinsically linked. An appropriate framework to interpret HRS phase curves needs to be able to reproduce both observable quantities simultaneously. This has already been shown by computing HRS spectra starting from GCMs of irradiated planets (e.g., \citealt{Zhang2017, Beltz2021, Beltz2022}). Unfortunately, such simulations are time-consuming, and it is currently not realistic to use them for data comparison over a large parameter space. However, our analysis demonstrates that, to first-order, the line contrasts (parameterized by the scale factor) and the positions (parameterized by $K_\mathrm{p}$, $v_\mathrm{sys}$, $h$, and $k$) are uncorrelated for KELT-9b. As a result, if Doppler shifts can be attributed to planetary climate, it is possible to build a first-order picture of winds (through line positions) and thermal/abundance structure (through line contrasts) even in the lack of a GCM analysis (see section \ref{sec: discussion_climate}).
\subsection{Sensitivity of HRS phase curves to slight eccentricities in UHJs}
\label{sec: discussion_eccentricity}
In section \ref{sec: results, eccentric_orbit} we have demonstrated that, with our new method based on the relative velocities of the planet probed by its atmospheric emission lines throughout the orbit, HRS phase curves have the potential to detect radial velocity deviations from a circular orbit of the order of a few kilometres-per-second. In section \ref{sec: results, eccentric_orbit}, we further demonstrated that, if due to a slightly eccentric orbit, deviations of this order would correspond to a solution of $e=0.016\pm0.003$ that we measured with high significance. In the specific case of KELT-9b, TESS photometry shows that this eccentricity is unlikely real (\citealt{Wong2020}, Sec. \ref{sec: results, eccentric_orbit}). However, it is interesting to note that, if KELT-9b had an eccentricity of this magnitude, we would have had the sensitivity to measure it with our new technique.
The challenge of measuring slight eccentricities of UHJs with traditional methods such as stellar radial velocities and transit photometry is manyfold. Indeed, many planets of this class orbit A-type or early F-type stars, which have fast rotational velocities (up to more than 100 $\mathrm{km s^{-1}}$). As a result, the stellar lines are significantly broadened, resulting in reduced stellar radial velocity precision. In addition, at the level of precision required, detailed modelling of second order effects, such as the radial velocity contamination by tidal bulges induced by the planet on the host star, need to be accounted for. Transit photometry offers an alternative avenue to measure slight eccentricities, but the interpretation of these measurements is not straightforward. For instance, an inhomogeneous day-side surface brightness of the planet could affect the retrieved mid-occultation timing \citep{Shporer2014}, whose accurate knowledge is necessary to measure a slight eccentricity with this method.
Our method is completely independent, and does not suffer from the same systematics as those based on stellar radial velocities and transit photometry. However, as we already discussed in Sec. \ref{sec: results, eccentric_orbit} and will discuss in more detail in Sec. \ref{sec: discussion_climate}, winds of moderate strength (a few kilometres-per-second) in UHJs can create signals of the same order of magnitude as those induced by a slightly eccentric orbit. Thus, while planetary radial velocities of UHJs could be used to measure or set tight upper limits to the potential slight eccentricities of UHJs, the interpretation of radial velocity anomalies in terms of orbital dynamics or climate could remain challenging.
It is out of the scope to explore in detail how to break such degeneracy. However, our work already shows that combination with independent measurements of eccentricities has the potential to discriminate between the two possibilities. In our case, \cite{Wong2020} provides a tight upper limit by combining information on the timing and duration of the primary and secondary eclipse of KELT-9b (measured by TESS). Indeed, these combined observables depend on the parameters $k$ and $h$, respectively \citep{Ragozzine2009}. Unfortunately, in many practical cases, only the timing difference can be measured to a useful precision, thus making it impossible to completely determine $e$ and $\omega$ \citep{Charbonneau2005}. Still, in the most favourable cases, this method has provided some of the best precision to date on eccentricities, pushing to measuring eccentricities smaller than 0.01 at $>5\sigma$ significance (e.g., \citealt{Nymeyer2011}). Extensive stellar radial velocities campaigns also demonstrated the capability to significantly measure eccentricities smaller than 0.01 in the best cases. For example, \citealt{Bonomo2017}, one of the largest homogeneous studies of eccentricities, significantly measured eccentricities smaller than 0.01 in 10 out of 231 planets. However, those cases all correspond to planets orbiting relatively slowly rotating host stars ($v\sin(i) < 20~\mathrm{km\ s^{-1}}$), and are the subject of multi-year radial velocity campaigns, and thus they still represent a minority.
In addition, different planetary wind patterns produce a different shape for the radial velocity anomaly (see Sec. \ref{sec: discussion_climate}), and so does an eccentric orbit. Through detailed modelling, and using a wide planetary phase coverage, it could thus be possible to distinguish between the radial velocity anomaly morphology induced by an eccentric orbit or winds. For instance, \cite{Kawahara2012} shows that the additional modulation induced by a slightly eccentric orbit over a circular one has half the period of the corresponding circular orbit, while planetary rotation, and, by extension, simple wind patterns, induce a deviation with a full orbital period (see also Sec. \ref{sec: discussion_climate}). Planetary radial velocities obtained towards quadrature and beyond (pushing into the night-side) are likely best suited to disentangle these solutions, which we will study in future work.
Finally, combining information from multiple species could prove the key to ultimately break this degeneracy. Eccentricity would produce the same radial velocity anomaly independently of the species or specific lines analysed. On the other hand, winds are expected to be altitude dependent. Lines of different oscillator strength probe different altitudes in the atmosphere, spanning several orders of magnitude in pressure with HRS (e.g., \citealt{Pino2020}). As a result, the radial velocity measured through different subsets of lines or atmospheric species that probe different altitudes would change depending on the wind pattern experienced by them.
Thanks to the advent of new generation spectrographs mounted on 8-metre class telescopes, with better throughput and broader spectral coverage (e.g., ESPRESSO@VLT, MAROON-X@Gemini-N), the same precision that we achieve on KELT-9b using \ion{Fe}{i} lines only will be reached on systems at least 2.5 magnitudes dimmer than KELT-9b (pushing up to $\mathrm{V}=9$). Despite the challenges outlined above, the prospect of a survey of eccentricities for this class of planets is tantalizing, as it would open a new window onto UHJ evolution history, and potentially of their internal structure \citep{Fortney2021}.
\subsection{The climate of KELT-9b}
\label{sec: discussion_climate}
As discussed in section \ref{sec: results, eccentric_orbit}, we consider unlikely that KELT-9b's orbit is actually eccentric. To reconcile our results with \cite{Wong2020}, we thus explore an alternative and more likely scenario in which the neutral iron signal comes from a localized region of the planet day-side atmosphere. Indeed, in this scenario, this region will show changes in net Doppler shift due to rotation and winds, and the neutral iron lines Doppler shift would track them.
We build a toy model to capture this effect. We define two sets of spherical polar coordinate systems to represent a spherical planet: $(\beta, \alpha)$ represent the latitude and longitude in a reference frame centred in the planet, with the observer located along the x-axis ($\alpha=0$); $(\theta, \phi)$ represent the latitude and longitude in a reference frame centred in the planet, co-rotating with the planet and with the sub-stellar point located along the x-axis ($\phi=0$). Since KELT-9b is transiting, we assume that the orbital inclination is $90^{\circ}$, which leads to the relations $\theta=\beta$ and $\alpha=\phi+\varphi$, where $\varphi$ is the planet phase. We assume a solid-body rotation for the planet, and an obliquity of $0^\circ$. We additionally allow for a zonal wind flow: we model both an eastward wind and a day-to-night wind, which are the wind circulation patterns largely predicted to be dominant by General Circulation Models of KELT-9b \citep{Komacek2017}. Then, the line-of-sight (LOS) component of the velocity of a mass of iron gas located at coordinates $(\beta, \alpha)$ is:
\begin{equation}
\label{Eq: vlos}
v_\mathrm{LOS}(\beta, \alpha) = -\left(R_\mathrm{p}\Omega+v_{\mathrm{wind}}\right)~\sin\alpha~\sin\beta\ ,
\end{equation}
where $R_\mathrm{p}$ is the radius of the planet and $\Omega$ is the angular velocity derived from the orbital period, and $v_{\mathrm{wind}}$ is always positive in the eastward jet case, but is negative (contrary to the sense of planet rotation) for $\phi < \pi$ in the day-to-night flow case. In addition, we need to take into account that not every surface element of the planet contributes to the net Doppler shift in the same way. Following \cite{Zhang2017}, we weight every contribution by the flux emitted in the direction of the observer. The projection of the normal to the local surface along the direction of observation results in a factor of $\mu = \cos \alpha \sin\beta$, which is thus the visibility weight. Finally, we weight every surface element with 1 if it is located in the day-side of the planet, and 0 if it is located in the night-side of the planet to account for the uneven brightness distribution.
We compare the predictions of this toy model with the planet trail obtained in the best fit circular orbit rest frame of the planet in Fig. \ref{fig: trail_and_winds}. In the case of pure rotation, we can intuitively understand the curve. Right after transit ($\varphi=0$), the day-side starts to be visible as it rotates towards the observer. As a result, the net Doppler-shift is negative and close in value to the rotational velocity. While more of the blue-shifted day-side is visible with increasing planet phase, the Doppler-shift becomes dominated by regions of the planet that have a smaller projected velocity component towards the observer due to geometry, and thus its magnitude decreases. After phase 0.25 part of the visible day-side is rotating away from the observer. At $\varphi=0.5$ the two parts cancel out exactly as the day-side is completely visible. The effect is symmetric for $\varphi>0.5$. The discontinuity at $\varphi=0$ is due to our assumption that the night-side is not contributing to the emission \citep{Beltz2022}. This reasoning is easily extended to the case of eastward wind, which simply adds to the planet rotation velocity, while the day-to-night flow case is slightly more complicated, but can be understood with similar reasoning.
The pure rotation curve (blue curve) is consistent with the trail at the $2\sigma$ level in the majority of the individual phase bins, both in pre-eclipse and post-eclipse. However, it does not seem to correctly capture the trends in neither phase range. In the pre-eclipse ($\varphi<0.5$) phase range, the trail appears to be steeper compared to the prediction of the pure rotation model. In other words, the rotation alone seems to be too slow to explain the pre-eclipse trail. The addition of zonal flows in the direction of rotation provides a possible solution. In this phase range, both an eastward (green curve) and a day-to-night (orange curve) flows are aligned with the planet rotation. We find that with the addition of an eastward wind model with $v_\mathrm{wind}=10~\mathrm{km~s^{-1}}$ the prediction matches the observed slope of the trail better compared to the pure rotation case. However, this model fails to reproduce the evidence for redshift observed right before eclipse, and it deviates from the trail in the post-eclipse range where the trail appears flatter. Both these trends are instead captured by a rotation plus day-to-night flow model with $v_\mathrm{wind}=5~\mathrm{km~s^{-1}}$. Indeed, in this configuration, the winds tend to redshift the signal close to secondary eclipse, and to counter-act the effect of rotation in post-eclipse phases, producing a nearly constant-with-phase trail.
\begin{figure*}
\sidecaption
\includegraphics[trim={0 1cm 0 0}, width=12cm]{trail_and_winds.pdf}
\caption{Planetary trail obtained in the best-fit circular orbit reference frame (same as left panel of Fig. \ref{fig:trail}, but zoomed on smaller radial velocity displacements from the best fit circular orbit rest frame), and comparison with predictions of net Doppler-shift in different cases: pure rotation (blue dashed curve), east-west wind of $10~\mathrm{km~s^{-1}}$ (green dashed curve), and day-to-night wind of $5~\mathrm{km~s^{-1}}$ (orange dot-dashed curve). The upper panel shows the contribution to Doppler shift from the portion of the planet disk seen at different phases due to the two wind geometries considered and planet rotation. Black portions of the disk correspond to the night-side, while red(blue) portions correspond to regions producing red(blue)shifted lines. Less weight should be given to phase bins close to quadrature ($\varphi < 0.35$; $\varphi>0.65$) because of the detrimental effect of our analysis on the signal-to-noise ratio in this phase range (see Fig. \ref{fig:processing}). \label{fig: trail_and_winds}}
\end{figure*}
Our results are broadly consistent with expectations from GCMs and \textit{Spitzer} phase curve observations of KELT-9b. Indeed, \cite{Mansfield2020} indicate a preference for strong drag GCMs to explain the day-side phase curve of this planet. In strong drag models, the drag timescale is short compared to the rotational timescale. As a result, the balance in the horizontal angular momentum is primarily between the drag force (e.g., dissipation due to the interaction between the planetary magnetic field and the ions in the atmosphere) and the day-night thermal forcing. This results in a day-to-night flow which is roughly symmetric around the substellar point \citep{Showman2011, Tan2019}. For KELT-9b, GCMs predict wind speeds of about $5~\mathrm{km~s^{-1}}$, a roughly symmetric thermal structure around the substellar point, and a day-to-night flow at the photospheric pressures probed by \textit{Spitzer} \citep{Tan2019, Mansfield2020}. This is in very good agreement with the results from our HRS phase curve, which shows symmetry in line intensities around the substellar point (Sec. \ref{sec: results_symmetry}), and shows a phase-dependent Doppler shift which is well reproduced by day-to-night wind flows with the speeds predicted by GCMs (this section). In addition, \cite{Wong2020} report a small eastward hot-spot offset of $5.2\pm0.9^{\circ}$, which is consistent with our upper limit to the scale factor asymmetry in the HRS phase curve. Indeed, in the presence of a hot-spot offset driven by an eastward super-rotating jet, we could expect the steepness of the thermal gradient and the iron ionization fraction to vary across the day-side in an asymmetric way, which would be picked by our retrievals as an asymmetry in the scale factor as a function of longitude \citep{vanSluijs2022}. However, \cite{Mansfield2020} report a larger hot-spot offset of $19\pm2^{\circ}$. If the vertical gradient in the thermal structure followed the same longitudinal pattern as the photospheric emission from \textit{Spitzer}, our retrieval should have the precision to detect it. It is anyway important to note that a coherent explanation of the \textit{TESS} and \textit{Spitzer} phase curves for this planet based on GCM models is still missing, and that neutral iron lines in this planet probe atmospheric pressures as low as $10^{-5}~\mathrm{bar}$ \citep{Pino2020}, which is significantly lower than the \textit{TESS} and \textit{Spitzer} photosphere ($10^{-1}$--$10^{-2}~\mathrm{bar}$), and well within the region where the double-gray assumption common to many GCMs does not apply \citep{Rauscher2012}. As a consequence, dedicated modelling effort to simultaneously fit LRS and HRS phase curves is necessary, and we leave it for future work.
\subsection{Implications for wind dynamics measured in transmission spectroscopy}
Day-to-night wind flow in KELT-9b was also reported by \cite{PaiAsnodkar2022} and \cite{BelloArufe2022} based on the Doppler shift of \ion{Fe}{ii} lines seen in transmission compared to the stellar systemic velocity. While \ion{Fe}{i} traces gas motions in the gravitationally bound part of the atmosphere of a planet, \ion{Fe}{ii} is likely tracing the dynamics of the evaporating exosphere \citep{Zhang2022}. Therefore, a full analysis combining high spectral resolution phase curves and transmission spectroscopy of UHJs has the potential to illuminate the physical processes leading to mass loss in this class of exoplanets. The challenge is that this requires the comparison between different geometries, and accurate modelling over a broad range of pressures, possibly including hydrodynamics and non-local thermodynamic equilibrium effects \citep{Huang2017, Hoeijmakers2019}. This is out of the scope of this paper, and we limit ourselves to a qualitative comparison, and to highlight potential pathways and caveats towards combining the two techniques.
Both \cite{PaiAsnodkar2022} and \cite{BelloArufe2022} report winds directed from day-side to night-side, with speeds between $5$ and $10~\mathrm{km~s^{-1}}$. At a qualitative level, both the wind speed and pattern that they report is consistent with our physical interpretation of the \ion{Fe}{i} Doppler shifts emerging from the HRS phase curve. This may suggest that the transition in dynamical regimes is not sharp across the pressures probed by \ion{Fe}{i} and \ion{Fe}{ii} in KELT-9b. At face value, this is surprising, since the two sets of observations likely bridge from the gravitationally bound part of the atmosphere to the exosphere. However, both our analysis and the analyses by \cite{PaiAsnodkar2022} and \cite{BelloArufe2022} only account for Doppler shift of lines, and yet, line shapes are also distorted differently by different wind patterns \citep{Seidel2020}. A more detailed analysis that accounts for this effect is thus necessary to properly assess whether a transition in the dynamical regime experienced by \ion{Fe}{i} and \ion{Fe}{ii} takes place.
Additionally, \cite{PaiAsnodkar2022} report potential variability in the measured wind speed at several $\mathrm{km~s^{-1}}$, although this does not reflect in strong brightness temperature variations for the planet \citep{Jones2022}. We leave the study of the effect of variability to future work, and only note that our observations likely probe deeper layers in the atmosphere, and that our results assume that no variability across our observed epochs is present.
It is also important to note that the measurement of the systemic velocity from the stellar lines of KELT-9 appears to be challenging.
This highlights a potential weakness of transmission spectra as probes of atmospheric dynamics \citep{BelloArufe2022}, since they must rely on the accuracy of the systemic velocity. On the other hand, our measurement only depends on the relative shift of planetary lines during the orbit, and it is therefore independent of the value of the systemic velocity. Our result thus constitutes an independent measurement of the likely presence of day-to-night winds at the several kilometre-per-second level, blowing at the altitudes probed by neutral atomic species in KELT-9b, which could indeed extend to the upper atmosphere probed by \ion{Fe}{ii} lines in transmission.
Looking forward, our technique could also be used to revisit results from transmission spectroscopy without using the systemic velocity measured through stellar lines, which is potentially inaccurate at the level of a few $\mathrm{km~s^{-1}}$ in this planet \citep{Pino2020}. For instance, our measured value of $v_\mathrm{sys}$ has a precision of about $0.5~\mathrm{km~s^{-1}}$, which could still be sufficient to compare with Doppler displacements measured in the transmission spectrum of the planet to reveal day-to-night winds of several $\mathrm{km~s^{-1}}$. However, our measurement is not guaranteed to be a completely accurate estimate of the systemic velocity itself, because its value could be affected by winds blowing neutral iron gas (as it happens in transmission). Still, $v_\mathrm{sys}$ values measured assuming a circular or an eccentric orbit (capturing winds in our interpretation) are compatible at $1\sigma$ (Table D.1). This provides some confidence that our systemic velocity is accurate at the level of $1~\mathrm{km~s^{-1}}$. It is also consistent with values measured on stellar lines by \cite{PaiAsnodkar2022}, \cite{Hoeijmakers2019}, and \cite{BelloArufe2022}, but seems more difficult to reconcile with the value measured by \cite{Borsa2019} and \cite{Gaudi2017}, from whose measurements it deviates by $3~\mathrm{km~s^{-1}}$ and $4~\mathrm{km~s^{-1}}$, respectively. Ultimately, while a more detailed analysis of the potential systematics and astrophysical effects impacting the measurement of $v_\mathrm{sys}$ with each technique is required, our analysis provides additional confidence to the literature estimates of wind speeds measured from transmission spectra in KELT-9b.
\subsection{KELT-9b in the context of ultra-hot Jupiters}
Our current understanding of the physical processes regulating the climate of hot and ultra-hot gas giants has mostly come from space-borne phase curve studies. One of the most striking results from these surveys is the relation between day-night temperature contrast, phase offset and level of stellar irradiation. Phase curve observations have found a systematic increase of day-night temperature contrast with increasing temperature up to about $2500~\mathrm{K}$, accompanied by a decrease in hot-spot offset, which is also a prediction of General Circulation Models (GCMs). More recently, it was found that hotter planets showcase a surprisingly low day-to-night contrast, which is interpreted as likely due to cooling of the day-side due to $\mathrm{H}_2$ dissociation, and warming of the night-side due to $\mathrm{H}_2$ recombination \citep{Bell2018, Tan2019}. It is less clear whether there is a trend in hot-spot offset: \cite{Zhang2018} suggest an increase of the offset for planets hotter than $3400~\mathrm{K}$, but this is based on a few data points and was challenged by the lack of a large offset in individual ultra-hot Jupiters (e.g., WASP-103b, \citealt{Kreidberg2018}).
KELT-9b constitutes the high end of the population of the most irradiated planets, and thus an ideal test-case for climate theories of ultra hot Jupiters. Our results (1) indicate that there is no strong asymmetry between east and west day-side thermal structure or volume mixing ratio of iron from KELT-9b, and (2) indicate a possible preference for day-to-night flow over eastward flow. Both results are consistent with predictions of GCMs, and partially agree with results from photometric phase curves of this planet (section \ref{sec: discussion_climate}). However, as noted in section \ref{sec: discussion_climate}, it is not trivial to directly compare results from LRS and HRS phase curves because they probe very different pressure regions in the atmosphere.
It is however interesting to compare our results on KELT-9b with recent work studying the HRS phase curve of WASP-33b through neutral iron \citep{Herman2022} and carbon-monoxide \citep{vanSluijs2022} lines. Both studies report a phase-dependence of the scale factor, with larger scale factors found at phases $\varphi>0.5$ (post-eclipse). Indeed, using a model similar to our Loff and Lbase models, \cite{Herman2022} found that the scale factor peaks at $22\pm12^\circ$, which is westwards of the sub-stellar point. They refrain from interpreting this result in terms of hot-spot offset, but \cite{vanSluijs2022} points out that this is consistent with GCM predictions: the thermal profile tends to get more isothermal inside the hot-spot, leading to a smaller scale factor. In other words, the westward offset in the peak scale factor of iron lines found by \cite{Herman2022}, and confirmed with CO by \cite{vanSluijs2022}, can be interpreted as indirect evidence for an eastward hot-spot offset. Furthermore, based on HARPS-N emission spectroscopy, \cite{Borsa2022} report an evidence of asymmetry in thermal structure between pre- and post-eclipse spectra of UHJ KELT-20b, although it is not statistically significant. They also find that \ion{Fe}{ii} and \ion{Cr}{i} are only observable in post-eclipse, and not in pre-eclipse, despite achieving a similar total S/N across both phase ranges. \cite{Johnson2022} were not able to confirm the detection of \ion{Fe}{ii} and \ion{Cr}{i} with LBT-PEPSI, however, they did not investigate in detail the reason for the discrepancy which could be due to differences in analysis techniques. Therefore, even if additional work is required to understand results on KELT-20b, observations of this UHJ further reinforce the expectation that climate can induce phase-dependent signals in HRS phase curves.
With its symmetric neutral iron scale factor within $10^\circ$ of the substellar point at $2\sigma$, KELT-9b shows qualitatively different behaviour. Taken together, these first results with HRS phase curves mark the opening of a new observational window to study the relation between climate and irradiation, which eventually could help to shed light on questions that have been opened by LRS phase curve population studies. Excitingly, we can expect a population study of HRS phase curves in the coming years thanks to new generation high-resolution spectrographs.
\section{Conclusions}
\label{sec:conclusions}
In this paper we have demonstrated for the first time that the optical HRS phase curve of UHJ KELT-9b, observed between quadratures ($0.25 <\varphi <0.75$) with 4-metre-class telescopes, has the necessary precision to measure two separate observable quantities: the phase-dependent \ion{Fe}{i} emission line strengths, and the phase-dependent \ion{Fe}{i} emission line Doppler shifts. This result sheds new light on the atmospheric circulation of KELT-9b. Our work has implications for the study of orbital dynamics, and of climate of UHJs, and allows us to draw a pathway towards optimal exploitation of HRS phase curves also in the context of complementary observations and in the JWST era.
In terms of studies of orbital dynamics of UHJs, we draw the following conclusions:
\begin{enumerate}
\item Thanks to the favourable planet-to-star flux ratio, and by exploiting state-of-the-art likelihood frameworks, 4-metre class telescopes are able to measure the orbital velocities of UHJs with sub-kilometre-per-second precision.
\item We detect a radial velocity anomaly of a few $\mathrm{km~s^{-1}}$ in KELT-9b. If attributed to a slightly eccentric orbit, this corresponds to $e = 0.016\pm0.003$ ($5\sigma$).
\item This demonstrates that HRS phase curves are potentially sensitive to eccentricities smaller than 0.02 in UHJs, although there are potential degeneracies with atmospheric winds.
\item To reconcile this result with the upper limit on eccentricity from \textit{TESS} transit timing and duration of KELT-9b by \cite{Wong2020}, we interpret this eccentricity as apparent, and due to atmospheric winds.
\end{enumerate}
In terms of climate of KELT-9b and UHJs, we draw the following conclusions:
\begin{enumerate}
\item KELT-9b has symmetric \ion{Fe}{i} line intensity around the sub-stellar point within $10^{\circ}$ ($2\sigma)$. The simplest explanation is the lack of a strong hot spot offset at the probed altitudes.
\item We detect a radial velocity anomaly in the \ion{Fe}{i} emission lines that we attribute to atmospheric dynamics (winds). By accounting for planet rotation and a day-to-night wind flow blowing at a few kilometre-per-second, we can reproduce the observed radial velocity anomaly.
\item Taken altogether, these observations favour predictions of GCMs with strong atmospheric drag for KELT-9b, confirming that the climate of UHJs is likely different from that of their cooler counterparts.
\end{enumerate}
This study demonstrates the potential role of HRS phase curves in the upcoming JWST era, and complementarity with other techniques:
\begin{enumerate}
\item HRS phase curves provide unique information that complements spectrophotometric and photometric phase curves. Even in the JWST era, this technique will remain relevant. In addition, HRS phase curves extend HRS phase-resolved transmission spectroscopy to a broader range of phases, and to the photosphere.
\item Current HRS optical spectrographs hosted on 8-metre class telescopes (e.g., ESPRESSO@VLT, MAROON-X@Gemini-N) could measure HRS phase curves of UHJs hosted around stars brighter than $V=9$ (at least 10) with a precision comparable to ours. In the short term future, we can expect a statistical survey of HRS optical phase curves, targeting neutral iron but also additional species that are spectrally active in this wavelength range (e.g., \ion{Fe}{ii}, TiO, \ion{Ti}{i}, \ion{Ti}{ii}, \ion{Ca}{ii}).
\end{enumerate}
Finally, we highlight lessons learned, and areas where additional work is necessary:
\begin{enumerate}
\item We recommend that future works on HRS phase curves consider both the phase-dependence of line intensities and shapes, and of Doppler-shifts, as they provide complementary information.
\item Dedicated work is necessary to fully explore the optimal way to combine HRS phase curves of exoplanets with stellar radial velocities and transit photometry to determine the orbital properties of targeted systems. However, our results already show that the combination of these methods opens unprecedented opportunities to simultaneously constrain the orbital and atmospheric properties of exoplanets.
\item The advantage of combining HRS phase curves, phase-resolved transmission spectroscopy at high spectral resolution and photometric and spectrophotometric phase curves of a planet is that they probe different altitudes and longitudes in UHJ atmospheres. This is also a challenge, as it requires a fully 3D GCM model with all key ingredients that are peculiar to UHJ atmospheres. Such a model is still not available.
\item Much information in HRS phase curves is found close to quadrature, where Doppler shifts, thermal structure and chemistry deviate the most from pure HRS emission spectroscopy. These are the phases where PCA-based and other classic HRS data reduction techniques most strongly alter and reduce the planet signal. We encourage future studies to design alternative methods to correct or mask telluric and stellar lines that better preserve the planetary signal.
\end{enumerate}
\begin{acknowledgements}
We thank the anonymous referee, whose feedback improved the quality and clarity of the manuscript. Based on observations or data obtained in the framework of GAPS. We made use of the Astropy Project tools and resources \citep{astropy:2013,astropy:2018, astropy:2022}. We acknowledge financial contribution from PRIN INAF 2019. Based on observations made with the Italian Telescopio Nazionale Galileo (TNG) operated by the Fundaci\'on Galileo Galilei (FGG) of the Istituto Nazionale di Astrofisica (INAF) at the Observatorio del Roque de los Muchachos (La Palma, Canary Islands, Spain). J.M.D acknowledges support from the European Research Council (ERC) European Union’s Horizon 2020 research and innovation programme (grant agreement no. 679633; Exo-Atmos) and the research programme VIDI New Frontiers in Exoplanetary Climatology with project number 614.001.601, which is (partly) financed by the Dutch Research Council (NWO).
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
Materials scientists and physicists have been intrigued by properties of the ternary transition metal pnictides with the chemical formula MT$_{4}$Pn$_{12}$ (M = alkali metal, alkaline-earth metal, lanthanide or light actinide element; T = Fe, Ru, or Os; Pn = P, As, or Sb), which crystallize in the filled skutterudite structure with space group $Im\bar{3}$ (No. 204)~\cite{Book1}. Examples include excellent thermoelectric performance ~\cite{Sales1996, Keppens1998}, and strongly correlated electron phenomena such as unconventional superconductivity ~\cite{EDBauer2002, Kotegawa2003, Adroja2005}, quadrupolar ordering ~\cite{Hachitani2006, Curnoe2002, Sugawara2002, Iwasa2002, Kohgi}, non-Fermi liquid behavior \cite{Takeda2000}, hybridization gap or Kondo phenomena~\cite{CeRu4Sb12_1,Takeda2000,Adroja2003,Adroja2007,Baumbach2008,Sanada2004} as well as unusual metal-insulator transitions ~\cite{Sekine}. From various factors that give rise to pletora of unuusal physical properties of MT$_{4}$Pn$_{12}$, the hybridization of localized $f$-electron states with conduction electron states is of prime importance. As shown for, e.g., the [Ru$_{4}$As$_{12}$] sublattice, its subtle interplay with the M cation results in multiband superconductivity, non-Fermi-liquid behavior, conventional superconductivity, and low-lying ferromagnetic order can be realized for La ~\cite{5}, Ce~\cite{6}, Pr ~\cite{7}, and Nd~\cite{8}, respectively. It is worth emphasizing that due to their cubic structure, MT$_{4}$Pn$_{12}$ compounds are unique in that they can be used to study multiband effects~\cite{Seyfarth,Jeitschko,Sato}. This opens an opportunity to compare and contrasts multiband effects in the superconductivity of anisotropic materials, such as iron pnictides~\cite{Tarantini,Ren}, heavy-fermion compounds ~\cite{Seyfarth, Kasahara,Hill,Kittaka}, and
topological insulators ~\cite{Hor,Zhang}.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{Fig1.pdf}
\caption{(Top) The body-centered cubic structure of the filled skutterudite LaRu$_4$As$_{12}$ with space group $Im\bar{3}$. (bottom) The low-temperature electrical resisitivity of LaRu$_4$As$_{12}$ showing a sharp superconducting transition at 10.4 K. Inset: Zero-field-cooled (ZFC) and field-cooled (FC) magnetization in presence of 30 mT applied field as a function of temperature for a collection of LaRu$_{4}$As$_{12}$ single crystals.}
\label{rt}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{Fig2.pdf}
\caption {(a)-(b) The asymmetry spectra of the TF-$\mu$SR data in the low time region, obtained in a 40 mT applied magnetic field at $T$ = 12.5 K (i.e. above $T_c$) and $T$ = 0.31 K (i.e. below $T_c$), respectively. (c)-(d) The maximum entropy plots at $T$ = 12.5 K and $T$ = 0.31 K, respectively. A zoomed-in section of the low-field region is shown in the inset.}
\label{tfmusr}
\end{figure}
\par
Most of the La-filled skutterudites exhibit BCS-type superconductivity with the superconducting critical temperature $T_c$ ranging from 0.4 K to 10.4 K~\cite{Jeitschko,Sato}. Among all completely filled compounds LaT$_4$Pn$_{12}$, the highest transition temperature accompanied by the highest upper critical field $H_{c2}$(0) = 10.2 T shows LaRu$_4$As$_{12}$~\cite{Namiki}. These parameters are substantially enhanced as compared to the sister compounds LaOs$_4$As$_{12}$~\cite{Shirotani} and PrRu$_4$As$_{12}$~\cite{Shirotani1997} with $T_c$ = 3.2 K and 2.3 K, respectively. This enhancement is even more intriguing, if one takes into account that LaRu$_4$As$_{12}$ does not show any distinct differences in both the conduction electron density of states at the Fermi level and the vibrational properties~\cite{Klotz}. Another remarkable feature of LaRu$_{4}$As$_{12}$ is several arguments for multiband superconductivity~\cite{Namiki,Klotz}. Specifically, the temperature dependence of the lower critical field $H_{c1}$ displays the sharp anomaly deep in the superconducting state, indicative of a rarely observed case of almost decoupled bands~\cite{Namiki}. Furthermore, a very recent observation of suppression of anharmonic phonons by artificial nonmagnetic defects is in line with a multiband superconductivity in the filled skutterudite LaRu$_{4}$As$_{12}$~\cite{Mizukami11}. Considering all of these factors, we are compelled to investigate this compound further.
\par
In this work, we investigated LaRu$_{4}$As$_{12}$ utilizing muon spin rotation and relaxation($\mu$SR), as well as inelastic X-ray scattering (IXS) measurements. $\mu$SR is a microscopic and very sensitive technique and provides direct information on the gap symmetry from the temperature dependence of superfluid density measured using transverse field (TF) $\mu$SR. Further zero-field (ZF) $\mu$SR is very sensitive to detect very small local magnetic field inside the (type II) superconducting state and hence can provide direct information on the broken time reversal symmetry. Inelastic X-ray scattering provides direct information on phonon dispersion and hence it is very important to investigate change in the phonon dispersion above and below the superconducting transition in LaRu$_{4}$As$_{12}$. These measurements are important to understand the mechanism of the superconductivity in LaRu$_{4}$As$_{12}$. Using TF-$\mu$SR measurement, we show clear evidence of multiband superconductivity in LaRu$_4$As$_{12}$. As a result, in addition to Fe-based and cuprates superconductors, LaRu$_4$As$_{12}$ could be an unique multiband superconductor~\cite{Bochenek}. The preservation of time-reversal symmetry is revealed in ZF-$\mu$SR study. The observed phonon dispersion, using IXS, shows good agreement with the DFT-calculated phonon dispersion.
\section{Experimental Details}
Single crystals of LaRu$_4$As$_{12}$ were grown by mineralization in a molten Cd:As flux using a procedure previously reported~\cite{Henkie}. The low-temperature electrical resistivity of a LaRu$_{4}$As$_{12}$ single crystal was investigated by a conventional four-point ac technique using a $^4$He cryostat (PPMS). Magnetic measurements for a collection of LaRu$_4$As$_{12}$ single crystals were performed down to 1.8 K at H = 30 mT utilizing a superconducting quantum interference device magnetometer (MPMS). The MUSR spectrometer at ISIS Pulsed Neutron and Muon Source, Rutherford Appleton Laboratory, UK, was used to measure the muon spin rotation and relaxation of LaRu$_4$As$_{12}$.
The powder sample (1 g) of LaRu$_4$As$_{12}$ obtained from very small single crystals was mounted on a silver holder using diluted GE varnish (99.999\%). This was placed on the dilution refrigerator's cold finger. Zero-field and transverse field measurements were taken at temperatures ranging from 0.31 K to 13.0 K. For ZF measurements, an active correction coil was employed to cancel any stray magnetic fields at the sample space to a level of 10$^{-4}$ mT.
The ZF-$\mu$SR measurement assists in identifying the spontaneous internal field associated with time-reversal symmetry breaking~\cite{Sonier}. The TF-$\mu$SR measurements were carried out in the presence of a 40 mT external magnetic field, which is much higher than the lower critical field ($\mu_0H_{c1}$ = 1.3 mT) and significantly lower than the upper critical field ($\mu_0H_{c2}$ = 10.3 T) of LaRu$_4$As$_{12}$. To avoid any flux trapping effects, we applied the transverse field much above $T_{c}$ and subsequently cooled the sample down to base temperature. The WiMDA software was used to evaluate the experimental data~\cite{Pratt2000}.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{Fig4.pdf}
\caption {(a) The temperature dependence of $\sigma_{sc}(T)$ data with fits with various gap models. The red solid line shows the fit to the two-gap $s$+$s$-wave model, while the blue line shows the fit to the single-gap isotropic $s$-wave model. (b) Temperature variations in the internal field confirm the appearance of Meissner state.}
\label{tfmusr2}
\end{figure}
\par
\begin{figure}[b]
\centering
\includegraphics[width=0.9\linewidth]{Fig5.pdf}
\caption{The ZF-$\mu$SR spectra for LaRu$_4$As$_{12}$ at 0.31 K (black) and 12.5 K (blue). As stated in the text, the solid line fits the data.}
\label{zfmusr}
\end{figure}
\par
High-resolution inelastic X-ray scattering (IXS) experiments were carried out at the beamline BL35XU of the SPring-8 synchrotron radiation facility in Harima, Japan~\cite{BL35XU-Baron-JPCS-2000}. Si(11 11 11) backscattering optics with 21.747 keV photon energy was chosen for the present work resulting in an energy resolution of 1.5 meV. The $Q$-resolution was set to (0.06 0.06 0.06) in reciprocal lattice units along [1 0 0] and [1 1 0] directions. Constant $Q$ scans were carried out in an extended range of reciprocal space at temperatures 2, 20, and 300~K. Few supplementary scans were performed at 6 and 14~K. i.e below and above $T_ c$. The sample temperature was controlled by a helium closed-cycle refrigerator. The quality of the single-crystal sample was checked by measuring the rocking curve widths on a series of Bragg reflections along with high-symmetry directions, say [1 0 0], [1 1 0], and [0 0 1]. These widths are less than 0.1 degrees and close to the resolution limit of the BL35XU spectrometer, $\sim$ 0.02 degrees. Phonon properties of LaRu$_4$As$_{12}$ were modelled by density functional theory (DFT) based lattice dynamics calculations. The first-principles calculations were performed with the Vienna ab-initio simulation package (VASP) utilizing projector augmented wave potentials and the generalized gradient approximation of Perdew-Burke-Ernzerhof (GGA-PBE) for the exchange-correlation terms~\cite{VASP-Kresse-1996,PAW-Kresse-1999,GGA-PBE-PRL-1996}. The electronic minimization, as well as the subsequent calculation of Hellmann-Feynman (HF) forces, were carried out with a 2x2x2 supercell and a Monkhorst-Pack k-mesh of 4x4x4~\cite{Monkhorst-PRB-1976}. The first-order Methfessel-Paxton method with a $\sigma=0.1$ for the band occupancies was applied and forces were relaxed to 10$^{-7}$. The relaxed structure resulted in a lattice parameter of a = 8.6074~\AA~ and the fractional coordinates of $y = 0.3497$, $z = 0.1507$ for the As position. The dynamical matrix and, thus, the phonon eigenvectors and eigenstates were derived from the HF forces by the direct method implemented in the software package PHONON~\cite{Phonon-Parlinski-1999}. The HF forces were computed for symmetry non-equivalent displacements of 0.03~\AA~ in high symmetry directions of the atoms. For the estimation of phonon intensities, the eigenvectors were scaled by the atom-specific squared electron number $Z^2$ and an averaged electronic form-factor. Within the dynamic range given by the experimental setup, the highest phonon intensities were derived for $Q$ numbers around the (600) Bragg reflection and the high symmetry directions [600]$\rightarrow$[611],[600]$\rightarrow$[700]$\rightarrow$[6.5 .5 .5].
\section{Results and discussion}
\subsection{Structure \& Physical Properties}
LaRu$_4$As$_{12}$ has a body-centered cubic structure with space group $Im\bar{3}$ (No. 204) which crystallizes in the CoAs$_3$-type skutterudite structure packed with La atoms. Green are La, yellow are Ru, and As are red atoms, as seen in Fig.~\ref{rt}(a). The electropositive element La, which lacks four-fold rotational symmetry, is at the centre of the large icosahedron cage formed by As atoms. Ru, a transition metal ion, creates a primitive cubic sublattice between the cages. The atomic positions are (0, 0, 0) for 1 La atom in the crystallographic position (2$a$), (1/4, 1/4, 1/4) for 4 Ru atoms in (8$c$), (0, $y$, $z$) for 12 As atoms in (24$g$), with $y$ = 0.3499 and $z$ = 0.1502. Fig.~\ref{rt} (b) displays the temperature variation of resistivity in the zero applied magnetic field in low temperature limit. The electrical resistivity reveals an on set of superconductivity at $T_c$ = 10.4 K. The low residual resistivity data revealed the good quality of the sample. The superconducting volume fraction in the LaRu$_4$As$_{12}$ sample was calculated using magnetization measurements. The inset of Fig.~\ref{rt}(b) shows the magnetization data for both zero field cooled (ZFC) and field cooled (FC) measurements. In the ZFC measurement, full diamagnetic shielding is observed below $T_c$ = 10.4 K. The upper critical field is estimated to be 10.3 T, exceeding the Pauli paramagnetic limit 18.4 $T_c$ = 18.9 T.
\begin{figure}[t]
\begin{center}
\includegraphics[width = \linewidth, height = \linewidth]{Fig6.pdf}
\caption{\textcolor{black}{Selected spectra of LaRu$_4$As$_{12}$ recorded along high-symmetry directions at the three different $T$ indicated in the figures. Lines with data correspond to standard Lorentzians fitted to the data (300~K in left panel and 20~K in right panel). Filled areas highlight the total fit functions.}}
\label{fig_IXS_spectra}
\end{center}
\end{figure}
\subsection{Superconducting gap structure: TF-$\mu$SR}
To further understand the multiband nature of superconductivity as suggested by the earlier experimental~\cite{Juraszek2020} and theoretical studies~\cite{Klotz}, we have performed TF-$\mu$SR measurements in LaRu$_4$As$_{12}$. The TF-$\mu$SR muon spin precession signals collected at $T$ = 12.5 K and 0.31 K in $H$ = 40 mT are shown in Figs.~\ref{tfmusr}(a)-(b). Due to the flux-line lattice state, which causes spatial inhomogeneity of the magnetic field distribution below $T_c$, the data in Figs.~\ref{tfmusr}(a)-(b) indicate a distinct contrast in the relaxation rate above and below $T_c$. The maximum entropy plots in Figs.~\ref{tfmusr}(c)-(d) show one peak at 12.5 K at the applied field. An extra peak at 0.31 K is observed on the lower side of the applied field in addition to the peak at 40 mT, confirming type-II superconductivity in this sample. Use of straight Fourier transformation to extract frequency spectra from transverse $\mu$SR data is not practicable, since at long times the data are noisy since the count rate is low, due to the shortness of the muon lifetime. On the other hand the maximum entropy method overcomes this problem and allows simultaneous analysis of spectra from multiple detectors with different phases to yield a single frequency spectrum \cite{Rainford}. The inset of Fig.~\ref{tfmusr}(d) shows a zoomed section of the low-field region of the maximum entropy plot, which reveals a broad peak near 30 mT associated with the vertex lattice formation.\\
\par
Due to the single peak of the internal field in the superconducting state, we have fitted the data using a Gaussian function having one component. To illustrate the characteristics of the vortex state, we have fitted the data using the expression~\cite{Bhattacharyya1, Bhattacharyyarev,ThCoC2,Panda},
\begin{equation}
A_\mathrm{TF}(t) = A_\mathrm{0}\cos(\gamma_\mathrm{\mu}H_0t+\Phi)\exp(-\frac{\sigma^{2}t^{2}}{2})+A_\mathrm{bg}\cos(\gamma_\mathrm{\mu}H_{bg}t+\Phi)
\end{equation}
where the first component is the sample contribution and the second is the background contribution. The initial asymmetries of the sample and silver holder contributions are described by $A_0$ and $A_{bg}$, respectively, where $A_{bg}$ does not undergo any depolarization; the internal fields of the sample and sample holder are described by $H_0$ and $H_{bg}$, respectively. The muon gyromagnetic ratio is $\gamma_\mathrm{\mu}$/2$\pi$ = 135.53 MHz/T; $\sigma$ is the Gaussian muon spin relaxation rate; $\Phi$ is the initial phase of the signal at the detectors. The values of $A_{0}$=0.66(1) and $A_{bg}$=0.34(1) were estimated by fitting 0.31 K data and which were kept fixed during the analysis of other temperature data. The superconducting depolarization rate can be determined by subtracting the nuclear contribution: $\sigma_\mathrm{sc} = \sqrt{\sigma^{2}-\sigma_\mathrm{n}^2}$, where $\sigma_{\mathrm{n}}$ = 0.047 $\mu$s$^{-1}$, is the normal state contribution, which is temperature independent as evident in spectra above $T_c$. Since $\sigma_\mathrm{sc}$ is connected to the magnetic penetration depth ($\lambda$) by $\sigma_\mathrm{sc}~\approx~1/\lambda^2$ for a triangular lattice~\cite{Chia,Sonier}, the temperature variation of $\sigma_\mathrm{sc}(T)$ can be used to measure the nature of the superconducting gap. Then, using the functional form given below, a single- and a two-gap $s$-wave model is fitted to the data:
\begin{align}
\label{MuonFit3}
\frac{\sigma_{sc}(T)}{\sigma_{sc}(0)} &= \frac{\lambda^{-2}(T)}{\lambda^{-2}(0)} \nonumber \\
&= w\frac{\lambda^{-2}(T,\Delta_{0,1})}{\lambda^{-2}(0,\Delta_{0,1})}+(1-w)\frac{\lambda^{-2}(T,\Delta_{0,2})}{\lambda^{-2}(0,\Delta_{0,2})}
\end{align}
where $\lambda(0)$ is the value of the penetration depth at T = 0 K, $\Delta_{0,i}$ is the value of the i$^{th}$ (i = 1 or 2) superconducting gap at T = 0 K and $w$ is the weighting factor of the first gap. Each term in Eq. (2) is evaluated using the standard expression within the local London approximation ($\lambda$ $>>$ $\xi$ ) ~\cite{Prozorov} as
\begin{align}
\label{MuonFit2}
\frac{\sigma_{\mathrm{sc}}\left(T\right)}{\sigma_{\mathrm{sc}}\left(0\right)} &= \frac{\lambda^{-2}\left(T,\Delta_{0,i}\right)}{\lambda^{-2}\left(0,\Delta_{0,i}\right)} \nonumber \\
&=1+\frac{1}{\pi}\int_{0}^{2\pi}\int_{\Delta\left(T,\phi\right)}^{\infty}\left(\frac{\partial f}{\partial E}\right) \frac{E\mathrm{d}E\mathrm{d}\phi}{\sqrt{E^{2}-\Delta_i^2\left(T,\phi_{}\right)}},
\end{align}
where $f= [1+\exp(E/k_\mathrm{B}T)]^{-1}$ is the Fermi function, the temperature and angle-dependent gap function is expected to follow the relationship: $\Delta_{i}(T,\phi) = \Delta_{0,i}\delta(T/T_\mathrm{C})\mathrm{g}(\phi)$. Here $\Delta_{0,i}$ denotes the maximum gap value at zero temperature and $\mathrm{g}(\phi)$ is the angular dependence of the gap, $\mathrm{g}(\phi)$ = 1 for an isotropic $s$-wave model and for $s+s$ wave gaps. Here $\phi$ is the azimuthal angle. The temperature depedence of the superconducting gap symmetry is expected to follow, $\delta(T/T_\mathrm{C}) = \tanh[1.82[1.018(T_\mathrm{C}/T-1)]^{0.51}]$\cite{Annett,Pang}. At any coupling strength, this gap function is accurate enough to explain the temperature dependency.\\~\\
\begin{table}
\caption{Fitted parameters obtained from the fit to the$\sigma_\mathrm{sc}$(T) data of LaRu$_{4}$As$_{12}$ using different gap models.}
\vspace{0.5cm}
\resizebox{\linewidth}{!}{
\begin {tabular}{|c|c|c|c|c| }
\hline
{Model} & {$\Delta_{i}(0)$(meV)} &{2$\Delta_{i}(0)$/k$_\mathrm{B}T_\mathrm{C}$} &w& $\chi^{2}$ \\ \hline
s-wave & 1.59 & 3.57& 1 & 3.1\\
s+s wave & 1.656, 0.064 & 3.73, 0.144 & 0.87& 1.3 \\
\hline
\end{tabular}}
\label{Tabel}
\end{table}
\par
\begin{figure*}[t]
\begin{center}
\includegraphics[width = \linewidth]{Fig7.pdf}
\caption{\textcolor{black}{Left, total and partial phonon densities of states of LaRu$_4$As$_{12}$ computed with DFT and normalized to 51 phonon modes in the total signal. Right, phonon energies derived from the Lorentzian fits. Symbol sizes correspond to the logarithm of the peak intensities, and error bars indicate the full width at half maximum of the peaks. Dotted lines with the data report the DFT computed phonon dispersion relation. Blue shaded signal indicates the DFT-computed phonon intensities. The energy scale of both, the Z$(\omega)$ and the phonon dispersion, has been scaled up by 7~\% for a better match with the experimental data.}
\label{fig_DFT_dispersion}}
\end{center}
\end{figure*}
\begin{figure}[b]
\begin{center}
\includegraphics[width = \linewidth]{Fig7bis.pdf}
\caption{\textcolor{black}{Generalized spectra S$(\omega)$ at $T=2, 20$ and $300$~K from top to bottom.
Left and right, data from scans for high symmetry mid-point and end-point scan in direction H$\rightarrow$P.
Lines with the data correspond to standard Lorentzians fitted to the data. Gray shaded area highlights the fitted total signal.}
\label{fig_generalized_spectra}}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width = \linewidth]{Fig8bis.pdf}
\caption{\textcolor{black}{Temperature dependence of peak energies derived through Lorentzian fits to the generalized S$(\omega)$ spectra.
Different $T$ are disclosed by the colors as blue (2~K), green (20~K), and red (300~K).
Characters indicate the different setups applied with
A ($\Gamma\rightarrow$ N, 600 $\rightarrow$ 6 .5 .5),
B ($\Gamma\rightarrow$ N, 6.5.5 $\rightarrow$ 611),
C ($\Gamma\rightarrow$ H, 600 $\rightarrow$ 6.500),
D ($\Gamma\rightarrow$ H, 6.500 $\rightarrow$ 700),
E (H $\rightarrow$ P, 6.5.5.5),
and F (H $\rightarrow$ P, 6.75.25.25)).
Results for E and F are highlighted in Fig.~\ref{fig_generalized_spectra} and for all other orientations in Fig.~8 of the Supplementary Information.
Boxes framed by dashed black lines highlight frequencies at which a turnover from mode hardening to softening upon cooling is particularly obvious.
Boxes framed by dotted red lines indicate frequencies at which a softening is observed below 300~K not in line with a standard quasi--harmonic response.
See text for details. The symbol size approximates a diameter of 1 meV. The corresponding reliability parameters of the fits are reported in Table I of the supplementary information section. They are of the order of 0.1 meV.}
\label{fig_IXS_energypoints}}
\end{center}
\end{figure}
\par
The symmetry of the superconducting gap was determined by fitting the data of $\lambda^{-2}(T)$ or $\sigma_{sc}(T)$ using (a) a single gap $s$-wave and (b) a multigap $s$+$s$-wave gap model, as presented in Fig.~\ref{tfmusr2}(a). Because it produces a high $\chi^2$ goodness of fit value (here $\chi^2=\sum{(\sigma_{obs}-\sigma_{cal})^2}/{\sigma_{error}^2(M-N)}$, the total number of data points is M, and the number of free parameters is N), the typical BCS single gap cannot characterise the $\lambda^{-2}(T)$ data. A recent study of LaRu$_4$As$_{12}$ by Juraszek {\it et al.}~\cite{{Juraszek2020}} confirms that the lower critical field behavior is well suited to multiband isotropic $s$-wave superconductivity. The multiband character of LaRu$_4$As$_{12}$ is further strengthened by the two-gap model employing the isotropic $s$+$s$-wave with $\chi^2$ = 1.3, compared to $\chi^2$ = 3.1 for single gap $s$-wave fit. The estimated larger gap is 2$\Delta_1(0)/k_BT_C$ = 3.73$\pm$0.2 in the $s$+$s$-wave model, which is close to the value of 3.53 driven from BCS theory, and the estimated smaller gap is 2$\Delta_2(0)/k_BT_C$ = 0.144$\pm$0.01 in the $s$+$s$-wave model. The smaller value of the second gap is observed in many Fe-based superconductors, for example RbCa$_{2}$Fe$_{4}$As$_{4}$F$_{2}$ $2\Delta_{1}/k_{B}T_{C}$ = 6.48 and $2\Delta_{2}/k_{B}T_{C}$= 0.7 $T_{C}$ = 29.2 K) [see Table-I of Ref.~\cite{Bhattacharyyarev}]. The different size of the gaps is arising from the Fermi surface topology. Multigap features are common in iron-based superconductors e.g. Ba$_{1-x}$K$_{x}$Fe$_{2}$As$_{2}$~\cite{Khasanov2009,Avci}, cuprate superconductors~\cite{Khasanov2007} also in Bi$_{4}$O$_{4}$S$_{3}$~\cite{Biswas}. We calculated the values of London penetration depth $\lambda_L(0)$= 240(4) nm for $s$+$s$ wave fit, carriers density $n_\mathrm{s} = 8.6(3) \times 10^{27}$ carriers m$^{-3}$, and effective mass of the quasiparticle $m^{*} = 1.749(2) m_\mathrm{e}$ for LaRu$_4$As$_{12}$ using the method described in Ref.~\cite{HfIrSi,CeIr3,Zr5Pt3}. As shown in Fig.~\ref{tfmusr2}(b), the internal field of the flux line lattice is temperature dependent and decreases with decreasing temperature, as in other superconductors, indicating diamagnetic shift. Because of the formation of a flux line lattice in the superconducting state, the total internal field inside the sample is reduced.
\subsection{Time reversal symmetry: ZF-$\mu$SR}
Here we present our ZF-$\mu$SR results for detecting the spontaneous internal field associated with time reversal symmetry breaking in the superconducting state.
As shown in Fig.~\ref{zfmusr}, we measured the time evolution of ZF asymmetry spectra in both the normal and superconducting states. The ZF-$\mu$SR spectra were fitted using a combination of Lorentzian and Gaussian Kubo-Toyabe relaxation functions~\cite{ThCoC2}:
\begin{equation}
A_\mathrm{ZF}(t) = A_\mathrm{2}G_\mathrm{KT}(t)e^{-\lambda_\mathrm{ZF}t}+A_\mathrm{bg}
\end{equation}
here,
\begin{equation}
G_\mathrm{KT}(t) = [\frac{1}{3}+\frac{2}{3}(1-\sigma_\mathrm{KT}^{2}t^{2})\exp({-\frac{\sigma_\mathrm{KT}^2t^2}{2}})]
\end{equation}
is the Gaussian Kubo-Toyabe function, the sample and silver holder asymmetry contributions are represented by $A_\mathrm{2}$ and $A_\mathrm{bg}$, respectively. As shown in Fig.~\ref{zfmusr}, both asymmetry spectra fall on top of each other, indicating that there is no change in relaxation rate in the superconducting state compared to the normal state. This denotes the lack of a spontaneous internal field. As a result, time reversal symmetry is preserved in the superconducting state. Furthermore, the fits to the ZF data yield, $\sigma_\mathrm{KT} = 0.11(1) ~\mu \mathrm{s}^{-1}$ and $\lambda_\mathrm{\mu} = 0.010(2) ~\mu \mathrm{s}^{-1}$ at 0.31~K and $\sigma_\mathrm{KT} = 0.105(3) ~\mu \mathrm{s}^{-1}$ and $\lambda_\mathrm{\mu} = 0.011(2) ~\mu \mathrm{s}^{-1}$ at 12.5~K.
\subsection{Inelastic X-Ray Scattering}
As outlined above, we have investigated the lattice dynamics of the multiband superconductor LaRu$_4$As$_{12}$ ($T_c$ =10.3 K) by IXS at 300 K, 20 K, and 2 K.
All spectra recorded at those $T$ are reported in Figs.~1--7 of the Supplementary Information.
For the highest intensities, the measurements were focused near the (6, 0, 0) Bragg position by performing longitudinal and transverse scans. Fig.~\ref{fig_IXS_spectra} shows some selected spectra recorded along high-symmetry directions at the sampled temperatures of 300 K, 20 K, and 2 K.
It can be seen that the overall behavior of the compound is rather harmonic. No strong phonon renormalization can be identified over the entire explored temperature range. This applies not only to the high-symmetry directions shown in Fig.~\ref{fig_IXS_spectra} but also to any other direction examined.
All spectra have been fitted after the temperature correction with the Bose occupation number.
Experimentally determined resolution functions have been convoluted with the fitted model.
The model was formed from a set of standard Lorentzian functions whose number was adjusted to the number of peaks identified. To compare data recorded at the three different temperatures the final fits were carried out to the Stokes-line data only. However, we confirmed that both side fits of the spectra at 300~K resulted in very similar parameters. Differences might be attributed to the strong elastic line in the prior data treatment, which has been suppressed in the final fits.
Fig.~\ref{fig_DFT_dispersion} reports all derived phonon energies along high-symmetry directions compared with our DFT computed dispersion relation and the total and partial phonon densities of states Z$(\omega)$. Note, that for best match, the energy scale of the DFT data has been scaled up by 7~\%. This underestimation of eigenfrequencies and overestimation of structural dimensions is a well-known feature of the GGA-PBE and has been reported before~\cite{Koza-JPSJ-2013, Tutuncu-PRB-2017}. The size of the energy symbols reflects the measured peak intensities on a logarithmic scale, and the peak widths are indicated as error bars. Vertical arrows indicate the ${\bf Q}$ points at which the spectra shown in Fig.~\ref{fig_IXS_spectra} have been recorded.
Improving the statistical significance of the spectra was necessary to derive a conclusion upon the temperature dependence of the phonon modes.
S$({\bf Q},\omega)$ spectra from different analyzers recorded at a fixed setup were corrected for zero shift of the elastic line and merged to a generalized S$(\omega)$ signal.
An example of the resulting data is presented in Fig.~\ref{fig_generalized_spectra}.
All spectra have been fitted with standard Lorentzians accounting for the resolution function by convolution as well as the Bose thermal occupation.
Fig.~\ref{fig_IXS_energypoints} reports the frequencies derived from those fits.
Different temperatures monitored at a single setup are disclosed by different colors.
Open symbols indicate less reliable frequencies which are strongly biased by the presents of the elastic line on one hand and a strong localized mode on the other.
Frequencies for which a switchover from hardening (going from 300~K to 20~K) to softening upon cooling below T$_C$ (going from 20~K to 2~K) is observed are highlighted by the dashed black line boxes, see Fig.~\ref{fig_IXS_energypoints}.
Some other frequencies showing a softening already below 300~K are framed by dotted red line boxes.
The mode hardening between 300 and 20~K can be associated with the lattice contraction and is comprehensible within the theory of quasi-harmonic crystals.
Clearly, the mode softening below T$_C$ does not follow this classical quasi-harmonic concept.
It rather points at a weak electron-phonon coupling effect when entering the superconducting state below ~10.5~K.
This observation is supported by the theoretically calculated $T_c$ = 11.56~K based on the electron-phonon coupling compared to the measured $T_c$= 10.45 K for LaRu$_4$As$_{12}$ ~\cite{Tutuncu2017}.
Note that the significance of the conclusion upon the detectability of the electron-phonon coupling in the present IXS data is qualified by the statistics of the results.
Further experiments are required to apprehend if specific vibrational eigenmodes are particularly involved in the electron-phonon interaction.
A participation of specific eigenmodes would result in a characteristic directional dependence of the phonon response and should be studied in future experiments.
\section{CONCLUSION}
In conclusion, we have determined the superconducting state of LaRu$_4$As$_{12}$ using TF-$\mu$SR and its lattice dynamics has been studied by inelastic X-ray scattering. The temperature dependence of magnetic penetration depth has been found to be consistent with a two-gap $s$+$s$-wave model of multiband superconductivity. The larger gap to $T_c$ ratio calculated within the $s$+$s$-wave scenario, 2$\Delta_1(0)/k_{\mathrm{B}}T_c$ = 3.73$\pm0.2$, is very close to the value of 3.53 for an ordinary BCS superconductor. The ZF-$\mu$SR spectra at 0.31 K and 12.5 K resemble each other very closely, indicative that the time-reversal symmetry is preserved in the superconducting state of LaRu$_4$As$_{12}$.
Furthermore, IXS analyses of the Brillouin region reveal phonon modes between 300 and 20 K that show a weak temperature dependence comprehensible as an effect of the crystal's volume changes.
Comparatively to 20 K, many of the 2 K modes show a weak softening of the phonon frequencies indicating that the electron-phonon interactions driving the
the superconductivity of LaRu$_4$As$_{12}$ exert a visible effect on the vibrational eigenstates.
This observation is also in agreement with the theoretically calculated $T_c$ = 11.56 K based on the electron-phonon coupling compared to the measured $T_c$= 10.45 K ~\cite{Tutuncu2017}. From a broader perspective our findings should be relevant for a wider class of multiband superconductors, including these with lower crystallographic symmetries such as iron-arsenide superconductors.\\
\section{Acknowledgements}
We are thankful to Prof. Tanmoy Das and Priyo Adhikari for their insightful discussion on superconducting gap structure. We are grateful to Prof. Zygmunt Henkie for providing single crystals of LaRu$_4$As$_{12}$. AB thanks the Science \& Engineering Research Board for the CRG Research Grant (CRG/2020/000698). The IXS experiment was performed under the approval of JASRI
(Proposal No. 2018B1493). DTA would like to acknowledge the funding from the Royal Society of London for Newton Advanced Fellowship between UK and China and for the International Exchange between UK and Japan and EPSRC-UK (Grant reference EP/W00562X/1).
|
\section{Introduction}
Functional testing in the semi-conductor context refers, broadly, to testing whether the device under test performs as designed, and detecting and possibly preparing repair of faulty behavior (e.g.~by activating spare parts of a chip). With increasing integration density, functional testing has become one of the most expensive and time-consuming parts of chip production. With the number of memory cells on a DRAM chip in the billions, it is clearly out of the question to test all possible states of the memory as the time needed would by far exceed the life-time of the Universe. Rather, test engineers focus on specific parts of the chip, such as address decoders, actual memory cells, sense-amps etc., and known error mechanisms. The latter typically include interaction of a cell with neighboring cells. E.g.~flipping one memory cell might flip a physically neighboring cell through various mechanisms. Many such error mechanisms can be tested in parallel by writing and reading out specific patterns into the memory, such as checkerboard patterns of 0s and 1s. \\
While there is no mass production of quantum-information-processing devices yet, it is expected that sooner or later similar considerations of quantum functional testing will be needed. Clearly, due to the essential necessity of quantum-coherent superpositions, quantum functional testing is an even more daunting task than classical functional testing.
Several methods have been proposed for characterizing quantum channels as well as possible. Most focus has been on ``quantum channel tomography'' \cite{chuang_prescription_1997,PhysRevA.63.020101,PhysRevLett.86.4195,PhysRevLett.90.193601,PhysRevA.64.012314,leonhardt_tomographic_1995,kosut_quantum_2008,christandl_reliable_2012,blume-kohout_robust_2012,rey-de-castro_time-resolved_2013,ohliger_efficient_2013}, with prominent methods such as randomized benchmarking
\cite{knill_randomized_2008,bendersky_selective_2009,cramer_efficient_2010,magesan_scalable_2011,magesan_efficient_2012,gaebler_randomized_2012,onorati_randomized_2018,cross_validating_2018,nielsen_gate_2021}, cross-entropy benchmarking \cite{boixo_characterizing_2018} (the latter was used in \cite{Arute2019} for showing ``quantum supremacy''), Bayesian techniques
\cite{huszar_adaptive_2012,huszar_adaptive_2012,ferrie_self-guided_2014,kueng_near-optimal_2015,granade_practical_2016}, compressed sensing \cite{gross_quantum_2010,flammia_quantum_2012}, matrix-product state representations \cite{baumgratz_scalable_2013,lanyon_efficient_2016}, multiplexed measurements \cite{titchener_scalable_2018}, or machine learning
\cite{Fiderer2021,quek_adaptive_2021,torlai_neural-network_2018}; see also the reviews and analyses in \cite{mohseni_quantum-process_2008,banaszek_focus_2013,schwemmer_systematic_2015}. Recently, for NISQ devices the fitting and extrapolation of error models \cite{Arute2019}, and the study of correlations of errors \cite{PRXQuantum.1.010305} have started to receive increased attention.
Functional testing takes a perspective that differs from quantum channel tomography. At the heart is a
decision problem: the device should be accepted or rejected (or
possibly qualified as repairable, with an additional output of what repair is needed), and since testing time is expensive, this decision must be reached as quickly as possible. In particular, the test engineer wants to waste as little time as possible on devices that do {\em not} function properly. In addition, it is desirable to provide a level of confidence that the decision is correct, and tests continue until such a level of confidence for a decision is reached, or testing is stopped if a specified maximum number of tests have been performed without reaching the required level of confidence. Naturally then, new tools from statistical analysis, notably the Chernoff bound \cite{Chernoff1952} and its quantum generalization, the quantum Chernoff bound \cite{Audenaert2007,Audenaert2008,Nussbaum2009}, become important. They provide asymptotic error rates for discriminating classical probability distributions, or quantum states \cite{pirandola_fundamental_2019}, respectively.
The world of classical functional testing divides into analogue and digital testing. In the former falls testing of e.g.~sense amplifiers, current drivers etc.~that can be characterized by a set of continous parameters, in the latter testing of Boolean functionality (including the identity function for memories). Quantum devices, even quantum memories, fall naturally and at least partly in the realm of analogue testing, due to the fact that quantum states vary continuously, in spite of the fact that in the context of quantum error correction errors can be ``quantized''. E.g.~a single-qubit memory cell should be able to store and retrieve any linear superposition of $\ket{0}$ and $\ket{1}$, parametrized by two angles of the Bloch vector, and a quantum channel is typically characterized by a set of continuous parameters. Fundamentally, the decision problem should therefore be based on quantum parameter estimation and has similarity with quantum metrology --- the difference being, however, that in the latter one strives at as high sensitivity as possible, whereas in quantum functional testing one strives at arriving at estimates with large confidence as quickly as possible. Crucial for this is the availability of a ``spec'', i.e.~a specification of the device provided by its producer (called ``the company'' in the following).
The specific quantum-mechanical context opens room for new principles for functional testing. In this paper we introduce and investigate three such principles: iteration of the channel, efficient decision criteria, and non-greedy Bayesian updates. In Sec.~\ref{sec.it} we discuss the relevance of iterating the channel, and in Sec.~\ref{ngu} we present the Bayesian context in which our protocols are studied. In Sec.~\ref{applications} we apply these principles to the problem of certifying quantum channels.
Ideas in the direction of a quantum-functional-testing approach have recently also been put forward under the label ``quantum system certification'' \cite{Kliesch2021}.
\section{Iterating a quantum channel}\label{sec.it}
\subsection{Intuition: coherent enhancement of errors \label{motiv}}
Any unitary single-qubit channel $\Phi:\mathbb C^2\to \mathbb C^2$ can be written as $\rho\mapsto \rho'=\Phi(\rho)=R_{\hat{\bm n}}(\theta)\rho R_{\hat{\bm n}}^\dagger(\theta)$ with
\begin{equation}
\label{eq:ch}
R_{\hat{\bm n}}(\theta)=e^{-i \frac{\theta}{2} \hat{\bm n}\cdot\bm\sigma}\,,
\end{equation}
where $\hat{\bm n}$ is a unit vector about which the Bloch vector of the state is rotated by an angle $\theta$, and $\bm\sigma$ is the vector of Pauli matrices. To see why iterations of a channel might be beneficial for speeding up the decision problem, consider a phase gate with $R_{{\bm e}_z}(\theta)$, i.e.~a rotation about the $z$-axis, and an initial state $\ket{+}$, with $\ket{\pm}=(\ket{0}\pm\ket{1})/\sqrt{2}$. After application of the channel, one measures ideally in a basis $\{R_{{\bm e}_z}(\theta)\ket{+},R_{{\bm e}_z}(\theta)\ket{-}\}$, which will lead to probabilities 1 and 0, respectively, to find $+$ and $-$. Now suppose that the gate actually enacts the (fixed) faulty unitary $R_{{\bm e}_z}(\theta')$ with $\theta'=\theta+\varepsilon$, where $\varepsilon$ is a small angle. The probability to find $+$ and $-$ in the same measurement are then, respectively, $\simeq 1-\varepsilon^2/4$ and $\simeq\varepsilon^2/4$. The chance to detect the faulty behavior can be enhanced by sampling the same channel $N_0$ times, in which case the probability to find a ``$-$'' increases to $\simeq N_0\varepsilon^2/4$. But a much more efficient strategy is to iterate the channel ${\cal N}$ times and only measure at the end in the correspondingly adapted basis $\{R_{{\bm e}_z}({\cal N}\theta)\ket{+},R_{{\bm e}_z}({\cal N}\theta)\ket{-}\}$, in which case the faulty channel will lead to a probability for the outcome $''-''$ given by $({\cal N}\varepsilon/2)^2$ (for ${\cal N}\varepsilon\ll 1$). For fair comparison, and supposing that the application of the channel dominates the total time required for the experiment rather than the measurement itself, we can sample the output after ${\cal N}$ iterations $N=N_0/{\cal N}$ times. The probability to find an outcome ``$-$'' will then be $\simeq (N_0/{\cal N})({\cal N}\varepsilon/2)^2=N_0 {\cal N} \varepsilon^2/4$, i.e.~the coherence of the channel can be exploited to gain a factor ${\cal N}$ in efficiency.
Iterations of channels have been proposed previously in the context of ``long sequence gate set tomography'' \cite{nielsen_gate_2021}. In the remainder of this section we study in more detail under what conditions iterations of the channel are beneficial: If $\varepsilon$ is not so small or is random with a certain probability distribution, or if decoherence effects play a role, then the advantage of iterating is not so clear. In addition, the answer might depend on the measurements available. A more rigorous assessment using a Chernoff bound is then in order; this will be developed in the next subsection. In Sec.~\ref{ngu} we shall propose a setup which automatically determines at each step of the protocol whether to iterate or measure directly.
\subsection{Quantum Chernoff bound}
\label{secchernoff}
The problem of {\em discriminating} two different quantum channels, such as an ideal channel from a faulty one (as opposed to fully characterizing one of them), can be cast in the framework of quantum hypothesis testing for quantum states (see e.g.~\cite{Audenaert2008}) by calculating the quantum Chernoff bound between the output states returned by the channels considered. Classically, the error probability $P_{\textrm{e}}$ in discriminating two probability distributions (by betting on the probability distribution that is more likely to generate some observed outcome) decreases exponentially with the number $N$ of draws from the distributions in the limit $N\to\infty$,
\begin{equation}
P^{\textrm{cl}}_{\textrm{e},N}\sim e^{-N \xi}\,.
\label{Pe}
\end{equation}
The asymptotic error exponent $\xi$ (called ``Chernoff bound'' in the following) was found by Chernoff in \cite{Chernoff1952} and later generalized to the quantum situation in \cite{Audenaert2007,Nussbaum2009}. In the quantum case we are given $N$ identical copies of a quantum state (or density matrix) that can only be one out of two possible states, $\rho$ or $\tau$, and are supposed to determine which quantum state has been provided. In a Bayesian setting, let $\pi_{0} >0$ and $ \pi_{1} >0$ be the prior probabilities with which $ \rho $ and $ \tau $ occur, respectively. In the absence of any a priori knowledge on $\tau$, we have $\pi_{0}=\pi_{1}=\frac12$. The minimal error probability based on an optimized POVM measurement can be shown \cite{Audenaert2008} to take the form
\begin{align}
\label{chernoff2}
&P_{\textrm{e},N}= \frac12\left(1-\Vert \pi_{1}\tau^{\otimes N}-\pi_{0}\rho^{\otimes N} \Vert_1\right)\,,
\end{align}
where $ \Vert O\Vert_1=\tr\vert O\vert $ is the trace norm. This error probability decreases again exponentially in the asymptotic limit $N\to\infty$, which allows one to define the \emph{quantum Chernoff bound} $ \xi_{\qcb} $,
\begin{align}
&\xi_{\qcb} =-\underset{N\rightarrow\infty}{\lim}{\frac{1}{N}}\log P_{\textrm{e},N}\,.
\end{align}
In \cite{Audenaert2008} an analytical formula for $ \xi_{\qcb} $ was found in terms of the two states $\rho$ and $\tau$ we wish to discriminate. It reads
\begin{equation}
\xi_{\qcb}(\rho,\tau) = -\log(\underset{0\leq s \leq 1}{\inf} \tr(\rho^{1-s}\tau^{s}))\,.
\label{chernoff}
\end{equation}
In the context of quantum functional testing, we set $ \tau $ as the output of the ideal channel and $ \rho $ as the output of the faulty channel, for given input state $\rho_{\textrm{in}}$. The quantum Chernoff bound $\xi_{\qcb}$ then gives the asymptotic rate of decrease of the minimal probability of not discriminating the faulty from the ideal channel upon measurement of the output. In order to maximize the rate at which the outputs are discriminated, we additionally maximize $ \xi_{\qcb} $ over all possible input states $\rho_{\textrm{in}}$. It turns out that it is sufficient to maximize over pure input states: indeed, denoting with $\rho_{\textrm{in}}=\sum_i p_i \ket{\psi_i}\!\bra{\psi_i}$ any decomposition of $\rho_{\textrm{in}}$ in pure states, one has:
\begin{align}\notag
&\xi_{\qcb} (\Phi_{\mathrm{ideal}}(\rho_{\textrm{in}}),\Phi_{\mathrm{faulty}}(\rho_{\textrm{in}}))\\
\notag
&\qquad\leqslant \sum_i p_i \xi_{\qcb} (\Phi_{\mathrm{ideal}}(\ket{\psi_i}\!\bra{\psi_i}),\Phi_{\mathrm{faulty}}(\ket{\psi_i}\!\bra{\psi_i}))\\
&\qquad\leqslant \sup_{i} \xi_{\qcb} (\Phi_{\mathrm{ideal}}(\ket{\psi_i}\!\bra{\psi_i}),\Phi_{\mathrm{faulty}}(\ket{\psi_i}\!\bra{\psi_i})),\label{eq:convexity}
\end{align}
where the first inequality holds by virtue of the joint convexity of $\xi_{\qcb}$
\cite{Calsamiglia2008,Audenaert2008,lieb_convex_1973} and linearity of quantum channels.
As a consequence of \eqref{eq:convexity} it is always better to consider pure input states rather than mixed ones. In what follows we will therefore, in the case of qubit channels, optimize over initial pure states parametrized by their Bloch vector $\mathbf{r}_\textrm{in}=(\sin\alpha\cos\beta,\sin\alpha\sin\beta,\cos\alpha)$.
\begin{figure}[!t]
\centering{\includegraphics[scale=0.6]{Fig1draft}}
\caption{Quantum Chernoff bound for the phase gate $R_{{\bm e}_z}(\pi/10)$ and initial state $\ket{+}$ as a function of the deterministic error $ \varepsilon $ for the channel applied once (blue dots, with $N=30$ samplings of the output), twice (orange squares, with $N=15$ samplings of the output) or three times (green rhombi, with $N=10$ samplings of the output) before measurement. We plot $\xi_{\qcb}$ divided by ${\cal N}$ to take into account the reduction of the number of samplings if the total test time is fixed and the time for applying the channel dominates over the measurement time.}
\label{pi10rot}
\end{figure}
\subsection{Phase gate}\label{phasegate}
As an example, let us consider the phase gate $R_{{\bm e}_z}(\theta)$ discussed in Sec.~\ref{motiv}. The phase implemented experimentally may be faulty and differ from the nominal one by some amount $\varepsilon$. Our aim is to evaluate the quantum Chernoff bound $\xi_{\qcb}$ for the distinguishability of the output $\tau=\Phi_\text{ideal}(\rho_{\textrm{in}})$ of the ideal channel and $\rho=\Phi_{\textrm{faulty}}(\rho_{\textrm{in}})$ of the faulty channel, with $\rho_{\textrm{in}}$ the input state.
After optimizing over all possible pure input states, the quantum Chernoff bound \eqref{chernoff} between the output states from the ideal channel and the faulty one gives an upper bound to the distinguishability of the two channels.
We shall study both the case where the error $\varepsilon$ on the rotation angle is unknown but fixed, and the case where it has some probability distribution $P(\varepsilon)$; in the latter situation the output of the faulty channel is a mixed state given by $\int P(\varepsilon)\Phi_{\theta+\varepsilon}(\rho_{\text{in}})\mathrm{d}\varepsilon$.
\subsubsection{Deterministic phase error}\label{sec.fixer}
As noted in Sec.~\ref{motiv}, if one performs measurements in the adapted basis $\{R_{{\bm e}_z}({\cal N}\theta)\ket{+},R_{{\bm e}_z}({\cal N}\theta)\ket{-}\}$, the faulty channel for small $\varepsilon$ leads to a probability for the outcome ``$-$'' given by $\simeq ({\cal N}\varepsilon/2)^2$, independent of $\theta$, implying an advantage $\propto {\cal N}$ for fixed total test time. Our numerical results for the quantum Chernoff bound corroborate this: In Fig.~\ref{pi10rot} we plot the quantum Chernoff bound \eqref{chernoff} for a small rotation angle $\theta=\pi/10 $ as function of the fixed unknown error $\varepsilon$. Clearly, for this value of $\theta$, the coherent enhancement of the distinguishability by iterating the channel occurs for all deterministic errors considered.
Note that in the above reasoning we disregard the fact that the rotation $R_{{\bm e}_z}({\cal N}\theta)$ of the measurement basis may suffer itself from errors, see \cite{nielsen_gate_2021}.
The situation changes when considering a fixed measurement basis, say, the basis $\{\ket{+},\ket{-}\}$, irrespective of ${\cal N}$, see Appendix \ref{classchernoff}.
The result for the classical Chernoff bound in this case is reported in Eq.~\eqref{defsmin}. Then, the optimal number of iterations depends on $\theta$, see Fig.~\ref{xy}. While for sufficiently small $\theta$ iterating the channel is always beneficial (within the three values ${\cal N}=1,2,3$ considered) as $\xi\sim{\cal N}^2$, this is no longer the case for larger values of $\theta$ (see discussion in the Appendix).
\begin{figure}
\includegraphics[scale=0.6]{Fig2draft}
\caption{Quantum Chernoff bound for the Z gate as a function of the angle $ \alpha $ parametrizing the input states. The error $ \varepsilon $ is uniformly distributed over the interval [0,1/5]. The blue curve corresponds to the gate applied once (${\cal N}=1$), the orange one (dashed) to the gate applied twice (${\cal N}=2$) and the green one (dotted) to the gate applied three times (${\cal N}=3$) before each measurement. Same number $N_0=300$ of samplings of the channel, and $N=N_0/{\cal N}$ measurements of the output.}
\label{Zgate}
\end{figure}
\subsubsection{Random phase error}
Let us now consider a random phase error, uniformly distributed over an interval $[0,w]$, and suppose that the rotation angle is $\theta=\pi$
(thus the channel implements a $Z$ gate). In Fig.~\ref{Zgate} we compare the quantum Chernoff bound for $w=0.2$ and different numbers $\mathcal{N}$ of iterations of the channel, as a function of the angle $\alpha$ charaterizing the input state via its Bloch vector $\mathbf{r}_\textrm{in}$.
We find that for narrow distributions of errors the quantum Chernoff bound is always greater if the channel is iterated, up to three times, rather than applied only once, whereas for wide distributions, $w\gtrsim 2$, iterations can be disadvantageous, as shown in Fig.~\ref{interval}. Qualitatively, this might be understood by the fact that the coherent enhancement of the error is lost for such strongly mixed states, whereas the number of samplings is reduced by a factor ${\cal N}$ compared to the non-iterated case.
\begin{figure}
\includegraphics[scale=0.6]{Fig3draft}
\caption{Quantum Chernoff bound for the Z gate as a function of the width $w$ of a uniform distribution of errors.
There are $N_0=30$ samplings of the channel, and $N=N_0/{\cal N}$ measurements of the output, with ${\cal N}=1$ (blue dots), ${\cal N}=2$ (orange squares), and ${\cal N}=3$ (green rhombi).
}
\label{interval}
\end{figure}
\subsection{Phase gate with dephasing}\label{pgd}
The phase gate of Sec.~\ref{phasegate} can be realized by having a qubit evolve for a controllable time $t$ under the Hamiltonian $H(\omega)=\frac{\omega}{2}\sigma_{z} $ \cite{Granade2012, Ferrie2013,Fiderer2021}, giving $\theta=\omega t$ for $\hbar=1$. If in addition the qubit suffers from a dephasing
process with characteristic time $T_2$, the coherences decay exponentially as $\exp(-\gamma t)$ with $\gamma=T_2^{-1}$. We now assume that the error $\varepsilon$ affects the value $\gamma$ at which the true device operates. Just as for $\theta$, one would like to find as quickly as possible whether the true value $\gamma_{\textrm{true}}$ of $\gamma$ coincides with the value $\gamma_\textrm{c}$ within a given interval. Due to the decoherence process, it is intuitively less clear
whether iterating the channel helps. In order to assess whether iterating helps, we derive an exact expression for the quantum Chernoff bound. After time $t$, an initial pure state parametrized by angles $\alpha$ and $\beta$ becomes
\begin{equation}
\label{rhodeph}
\rho_t(\gamma)=\begin{pmatrix}
\cos ^2\frac{\alpha }{2} & \frac{\sin \alpha}{2} e^{-\gamma t- i (\omega t+\beta )} \\
\frac{\sin \alpha}{2} e^{-\gamma t+i (\omega t+\beta )} & \sin ^2\frac{\alpha }{2}
\end{pmatrix}\,.
\end{equation}
If we want to distinguish channels with parameters $\gamma$ and $\gamma'$, the quantum Chernoff bound is obtained from Eq.~\eqref{chernoff} by evaluating the infimum of $T=\tr[\rho_t(\gamma)^{s}\rho_t(\gamma')^{1-s}]$ over $s$. We define the dimensionless parameters $\tau=\gamma t$, $\varepsilon=(\gamma'-\gamma)/\gamma$, $\tau'=\gamma' t=\tau(1+\varepsilon)$ and $\phi=\omega t + \beta$. It turns out that $T$ does not depend on the complex phase $\phi$, which can be therefore discarded. Moreover $T$ is symmetric about $\alpha=\pi/2$, $T(\alpha)=T(\pi-\alpha)$. The value $\alpha=\pi/2$ is thus a local extremum. Numerically, the minimum over $s$ and $\alpha$ is always achieved at $\alpha=\pi/2$, which corresponds to a qubit initially prepared in the pure state $\ket{+}=\frac{1}{\sqrt{2}}(\ket{0}+\ket{1})$. If we fix $\alpha=\pi/2$ for the initial state, the output of the channel reads
\begin{equation}
\label{rhodetau}
\rho(\tau)=\frac12\begin{pmatrix}
1 & e^{-\tau} \\
e^{-\tau}& 1
\end{pmatrix}\,.
\end{equation}
We now wish to distinguish channels with parameters $\tau$ and $\tau'$. Iterating the dephasing channel ${\cal N}$ times amounts to multiply $t$, and thus $\tau$ and $\tau'$, by a factor ${\cal N}$.
At fixed $\tau$ the quantum Chernoff bound is given by the infimum of $T(s)=\tr[\rho(\tau)^{s}\rho(\tau')^{1-s}]$ over $s$. Setting
\begin{equation}
x=\frac{1+e^{-\tau(1+\varepsilon)}}{2},\qquad y=\frac{1+e^{-\tau}}{2},
\label{defxy}
\end{equation}
$T$ simplifies to
\begin{equation}
\label{Tpisur2}
T=x^{1-s}y^s+(1-x)^{1-s}(1-y)^s\,.
\end{equation}
Remarkably, this expression exactly coincides with Eq.~\eqref{deff} giving the classical Chernoff bound in the context of a faulty rotation angle; only the definition of $x$, $y$ differs. The minimum of \eqref{Tpisur2} over $s$ thus yields a value $s_{\min}$ given by Eq.~\eqref{defsmin}, which provides an explicit analytic expression $\xi_{\qcb}=-\log T(s_{\min})$ for the quantum Chernoff bound of the dephasing channel. This quantity is plotted in Fig.~\ref{xydephasing}: as can be expected intuitively, the smaller the error $\varepsilon$, the larger the number of measurements needs to be to distinguish the faulty from the ideal channel.
\begin{figure}
\includegraphics[scale=0.9]{xydephasing.pdf}
\caption{Quantum Chernoff bound for the dephasing channel as
function of $\tau$ for (from bottom to top) $\varepsilon=0.1$ (solid), 0.2 (dashed), 0.3 (dotted). Note that varying ${\cal N}$ now amounts to varying $\tau$.
\label{xydephasing}}
\end{figure}
We can now address the question whether iterating the channel helps or not. For ${\cal N}$ iterations of the channel $\tau$ is changed to ${\cal N} \tau$, and a series expansion of $T(s_{\min})$ at small ${\varepsilon}$ gives
\begin{equation}
\label{xipetiteps}
\frac{\xi_{\qcb}}{{\cal N}}=\frac{{\cal N}\tau ^2 {\varepsilon} ^2}{8 \left(e^{2 {\cal N}\tau }-1\right)}+O\left({\varepsilon} ^3\right)
\end{equation}
for the quantum Chernoff bound in units of the number of iterations. At fixed $\varepsilon$ and $\tau$, \eqref{xipetiteps} is a decreasing function of ${\cal N}$.
This means that iterating the channel leads to a slower decrease of the error probability with the number of steps. This is in contrast with the case of the rotation channel of the previous section, where the QCB is systematically smaller for smaller ${\cal N}$ (see Fig.~\ref{Zgate}), and thus iterating the channel increases the success probability. A way of understanding this intuitively is to observe that following Eq.~\eqref{rhodetau}, $\rho(\tau)$ goes to the maximally mixed state as $\tau$ increases. Therefore, at large values of $\tau$ (which corresponds to larger ${\cal N}$) states $\rho(\tau)$ and $\rho(\tau+\varepsilon)$ both flow towards the same state and become less and less distinguishable.
\section{Bayesian experimental design} \label{ngu}
The previous section was based on the frequentist approach to parameter estimation. However, in the light of wanting to speed up the decision problem, it is natural to use Bayesian parameter estimation combined with optimal experimental design strategies: Instead of waiting for measurement results from a possibly large
number of measurements, each measurement can be used to update the knowledge about the unknown parameter. Any past measurement outcome can be used for determining which subsequent measurement promises the largest information gain, given the current knowledge about the parameter. In Sec.~\ref{sec.it} we discussed the advantage that can be expected from iterating the quantum channel; in the Bayesian approach, the option to iterate the channel without measurement can easily be included by considering the identity operator as a possible POVM. It is then the information already gathered that determines at each step if it is more
advantageous to continue iterating the channel or to do a measurement. This naturally leads to ``non-greedy'' Bayesian updates and experimental designs, where one tries to find an optimal sequence of measurements rather than just the next one. We briefly review the Bayesian methods and adaptive strategies, adapt them to the present context of quantum functional testing, and then study their performance.
\subsection{Bayesian methods and adaptive strategies}
\label{smc}
At the core of the Bayesian methods of parameter estimation is the Bayes rule for conditional probabilities $\prob(\mathbf{x}|\mathbf{D})=\prob(\mathbf{D}|\mathbf{x})\prob(\mathbf{x})/\prob(\mathbf{D})$, where $ \mathbf{x} $ is the vector of unknown parameters and $\mathbf{D}=\{d_1,d_2,...,d_n\} $ are the data gathered from $n$ experiments $ \mathbf{e}=\{e_1,e_2,...,e_n\} $. In the above formula, $\prob(\mathbf{x}) $ is called the \emph{prior} and it incorporates any a priori knowledge on the parameters, $ \prob(\mathbf{x}|\mathbf{D})$ is referred to as the \emph{posterior}, $ \prob(\mathbf{D}|\mathbf{x}) $ is the \emph{likelihood}, which in the present context is given by the Born rule, and $\prob(\mathbf{D})$ is a normalization factor. \\
An analytic calculation of the posterior probability distribution soon becomes impractical after few experiments/updates. Thus numerical methods, often involving Monte Carlo algorithms, have been implemented to overcome these difficulties. We consider here the so-called \emph{sequential} Monte Carlo algorithms (SMC) \cite{Doucet2000}, where data are processed sequentially, that is, the posterior at step $k$ is taken as the prior at step $k+1$.
In the "particle filtering" approach to SMC, the prior continuous probability distribution $ \prob(\mathbf{x}) $, corresponding to a case where parameter space is continuous, is replaced with a discrete approximation that can be written as $\sum_i w_i\delta(\mathbf{x}-\mathbf{x}_i)$, where the $\mathbf{x}_i$ are the locations of "particles" (which sample the space of parameters) and $ w_i $ the associated "weights" (probabilities). A particle filter $ \{w_i',\mathbf{x}_i'\} $ for the posterior probability distribution $\prob(\mathbf{x}|\mathbf{D})$ can be obtained by setting $\mathbf{x}_i'=\mathbf{x}_i$ and updating the weights from the Bayes rule as $w'_i=w_i \prob(\mathbf{D}|\mathbf{x_i})/\sum_j w_j\prob(\mathbf{D}|\mathbf{x_j})$. This update method suffers from numerical instability as data are collected, a problem known as impoverishment of the filter.
Resampling
new particle locations can solve this problem \cite{Doucet2009}. A resampling algorithm substitutes the impoverished filter with a new enriched one that still approximates the same probability distribution, only modifying the choice of particle locations $\mathbf{x}_i$.
Including new knowledge coming from the data through the update of the conditional posterior probability distribution suggests adaptive strategies to choose the next best experiment. A lot of effort has been put in Bayesian experimental design \cite{Chaloner1995}, which recasts the problem of choosing the best experiment into the theoretical problem of optimizing a \emph{utility function} $U$. At step $n$, the function $U$ depends on the possible future $m$ designs/experiments $ \mathbf{e'}=\{e_{n+1},e_{n+2},...,e_{n+m}\} $, the possible future data $\mathbf{D'}=\{d_{n+1},d_{n+2},...,d_{n+m}\} $ and the model parameters $ \mathbf{x} $. One may consider the optimal Bayesian design $ \mathbf{e'}^* $ among the ones in the design space $ \mathbf{E}$. If $U(\mathbf{e'},\mathbf{x},\mathbf{D'})$ is the utility function, then the optimal design $\mathbf{e'}^*$ is the one that maximises the expected utility
\cite{Ryan2016, Granade2012},
\begin{equation}
\mathbf{e'}^*=\argmax_{\mathbf{e'}\in \mathbf{E}} E[U(\mathbf{e'},\mathbf{x},\mathbf{D'})]\,,
\label{utility}
\end{equation}
where $ E[...] $ denotes the expectation value of the utility function over the possible outcomes of the experiments.
\subsection{Greedy vs non-greedy designs}\label{greedy}
If the above optimization is performed at each step, this gives a \emph{sequential} or \emph{adaptive} design. Many adaptive experimental designs involve optimizing only over a single future experiment and outcome $\{e_{n+1},d_{n+1}\}$; this case where $m=1$ is called the \emph{myopic} (or \emph{greedy}) approach. However, an update can occur in the optimization after one or more actions, i.e.~$ m\geq 1 $. Even classically, the case $m=1$ is in general not optimal, since one could look ahead to \textit{all} future observations and decisions that can be made at each step in the experiment. For quantum functional testing, this $m\geq 2$ approach appears even more attractive due to the possibility of coherent enhancement of errors discussed above. Optimizing the experimental design over several steps is
known as \emph{backward induction} \cite{Bernardo2009}. It is computationally very expensive due to the exponential proliferation of combinations, and hence
remains limited to relatively simple design problems. Here we use an intermediate strategy, which goes beyond the single step optimization, but is limited to a modest number of steps ahead ($ 1\leq m\leq 4 $).
Common choices of utility functions $U$ for parameter estimation problems are functions of the posterior covariance matrix or information-based utilities
\cite{Ryan2016, Granade2012}. Going from prior to posterior distribution should increase the available information. The information gain ($\ig$) \cite{Lindley1956} is defined via the Kullback-Leibler divergence (KLD) between the prior distribution $\prob(\mathbf{x})$ (omitting the dependence on earlier experiments $ \mathbf{e} $ and outcomes $ \mathbf{D} $) and the posterior distribution $\prob(\mathbf{x}|\mathbf{e'},\mathbf{D'})$ where we gained some knowledge $ \mathbf{D'} $ from the experiments $ \mathbf{e'}$. The information gain resulting from these additional experiments is
\begin{align}
\begin{aligned}
\ig(\mathbf{e'},\mathbf{D'})
&=\int_{\mathbf{X}}\prob(\mathbf{x}|\mathbf{e'},\mathbf{D'})\log\frac{ \prob(\mathbf{x}|\mathbf{e'},\mathbf{D'})}{\prob(\mathbf{x})} d\mathbf{x}\\
&\equiv E_{\mathbf{x}|\mathbf{e'},\mathbf{D'}}
\left[\log\frac{ \prob(\mathbf{x}|\mathbf{e'},\mathbf{D'})}{\prob(\mathbf{x})}\right]\,.
\label{kld}
\end{aligned}
\end{align}
In \eqref{kld} we take the integral over the sample space $\mathbf{X}$ of possible model parameters $\mathbf{x}$, and the index of $ E $ specifies the probability distribution to be integrated over for the calculation of the expectation value: in the present case, $E_{\mathbf{x}|\mathbf{e'},\mathbf{D'}}$ implies to use the probabilities $ \prob(\mathbf{x}|\mathbf{e'},\mathbf{D'}) $.
We now define the mutual information (MI) as the information gain averaged over all possible outcomes \cite{cover1999elements}. Suppose each of the $m$ experiments has a finite number $ l $ of possible outcomes. The MI at fixed $m$ is then given by taking the integral over the data $\mathbf{D'}$, conditional on the design $ \mathbf{e'} $, hence by the expected value
\begin{equation}
\textrm{MI}(\mathbf{e'})\equiv \overline{\ig}^{(m)}=\sum_k p_k \ig(\mathbf{e'},\mathbf{D'}_k)
\label{avIG}
\end{equation}
where $ k $ runs up to the total number $l^m$ of possible outcome sequences, $p_k $ is the probability of obtaining the $k$th sequence of outcomes, and $\ig(\mathbf{e'},\mathbf{D'}_k)$ is the information gain of the $k$th sequence. For a channel $\Phi_\mathbf{x}$ depending on parameter $\mathbf{x}$ the probability $p_k$ is given by the Born rule for the state $\Phi_\mathbf{x}(\rho_{\textrm{in}})$. Namely, if experiments $\mathbf{e'}=\{e_{n+i},1\leq i\leq m\}$ are realized by POVMs with operators $M_{i,j}$ (with outcomes labeled by $j$ for experiment $e_{n+i}$), then the $k$th sequence of outcomes $\mathbf{D'}_k=\{j_1,j_2,...,j_m\}$ is obtained with probability
\begin{equation}
\prob_{\mathbf{e'}}(\mathbf{D'}_k|\mathbf{x})=\tr (T_k^\dagger T_k \Phi_\mathbf{x}(\rho_{\textrm{in}})),\quad
T_k=M_{m,j_m}...M_{2,j_2}M_{1,j_1}
\end{equation}
\cite{Nielsen00},
and the probability $p_k$ then reads
\begin{equation}
p_k=\prob_{\mathbf{e'}}(\mathbf{D'}_k)=\int_{\mathbf{X}}\prob_{\mathbf{e'}}(\mathbf{D'}_k|\mathbf{x})\prob(\mathbf{x})d\mathbf{x}.
\end{equation}
Note that in this case the prior distribution $\prob(\mathbf{x})$ in Eq.~\eqref{kld} is not the one relative to the previous single step, but rather the one of $m$ steps back.
In Sec.~\ref{applications} we shall compare the efficiency of two different utility functions $U$, the negative variance of the posterior distribution (VAR) and the mutual information (MI), and see how they perform in the case of single parameter estimation.
\subsection{Experimental design and decision criteria}\label{design}
Our general framework is the following. We conceive a scenario in which a company provides us with a device which realizes a quantum channel depending on some parameter $\mathbf{x}$. The company claims that the nominal value of $\mathbf{x}$ is within the \emph{confidence region} $\mathbf{x}_\textrm{c}\pm\Delta\mathbf{x}$, where $ \Delta\mathbf{x} $ defines a small interval around each component of the vector $ \mathbf{x}_\textrm{c} $.
Our aim is to establish as quickly as possible whether the vector of true parameters $\bf{x}_{\textrm{true}}$ characterizing the channel lies in that confidence region. In the case of a single parameter $ x_\textrm{c} $ this region is simply an interval centered around $ x_\textrm{c} $; in the following we refer to this interval as ``spec'' (for ''specification by the company''). The functional test process consists of two stages; in the first one, we update the probability distribution of $\mathbf{x}$ following an adaptive strategy, and in the second one we use a \emph{decision criterion} to accept or reject the spec given by the company, and reach a conclusion.\\
We consider two different decision criteria. The first criterion (which we label MEAN) only relies on the position of the mean of the posterior distribution obtained, looking whether it falls inside or outside the spec. The second criterion compares the spec interval with a chosen \emph{credible region} \cite{Granade2012,Ferrie2014,Bernardo2005} for the posterior distribution (intuitively, the region where most of the probability lies). We took the \emph{highest posterior density} (HPD) region as the credible region, which in our case is defined as the region containing $ 95\% $ of the probability mass; the "HPD" criterion is defined by accepting the value $ x_\textrm{c} $ provided by the company if $ 95\% $ of the HPD credible region falls inside the spec, and rejecting it if $ 95\% $ falls outside.
In the remaining cases this criterion remains inconclusive, and we count these instances as failed tests, i.e.~they contribute to the estimated error probability. The overlap of this credible interval with the spec is central to the decision method used in \cite{Kruschke2018}.
\begin{figure*}
\includegraphics[scale=0.6]{batchup3bis}
\caption{Sketch of the certification protocol in the
experiment-limited scenario. In a batch update the state between
single updates is not reset to the initial input state. The obtained
data $d_1,\ldots,d_m$ are used to find the next set of $m$
measurements (including possibly the identity) that maximizes the
expected increase of information.}
\label{batch}
\end{figure*}
\section{Application to quantum channels}\label{applications}
We now apply this scheme to the problem of certifying quantum channels governed by a single parameter, focusing again on the phase gate and the dephasing channel. The best sequence of measurements of length $ m $ is the one which maximises the utility function $U$. We denote with $\mathcal{M}$ the set of measurements one can choose from; each measurement is performed on the output state given by the application of the channel on an input state. Besides the measurements in $\mathcal{M}$, we introduce the possibility of choosing an extra action, which consists in applying again the channel over the output state instead of measuring; this is implemented in practice by including the identity operator in the set $\mathcal{M}$. The choice of this action entails that the input state is not always reset to the initial one between single actions (measurement or identity channel). We assume that the certification process is experiment-limited, which means that the number of actions $N_0$ is fixed, while the number of updates of the posterior distribution depends on $ m $ as $N=N_0/m$. We sketch this certification protocol in Fig.~\ref{batch}.
In what follows, we analyze and compare the choice of $U$ (either MI or VAR) and the role of the decision criterion (MEAN or HPD).
\subsection{Phase gate}
\subsubsection{Setup}
\label{1stmodel}
Let us consider the phase gate $R_{{\bm e}_z}(\theta)$ defined in Sec.~\ref{sec.it}. We fix the initial input
state as $\rho_{\textrm{in}}=\vert + \rangle \langle + \vert$, whose Bloch vector points along the $ \hat{x} $ axis.
We consider an experiment which involves two-outcome measurements, given by two POVM elements, e.g.~projectors $\Pi_1^{(i)}=(\mathbb{I}+\sigma_i)/2$ (and $\Pi_2^{(i)}=\mathbb{I}-\Pi_1^{(i)}$), with $\sigma_i$ the Pauli matrices $ \sigma_x $ and $\sigma_y$.
The probability of obtaining one of the two outcomes is given by the Born rule $ p_{1,2}^{(i)}=\tr(\Pi_{1,2}^{(i)}\rho_{\textrm{out}}) $, where at the first update $\rho_{\textrm{out}}=\Phi(\rho_{\textrm{in}})$.
To simulate an experiment, we choose as initial prior a uniform distribution of $\theta$ over the $[-\pi,\pi]$ interval. We fix the true angle $\theta_{\textrm{true}}$, and the angle $\theta_\textrm{c}$ and a spec around it provided by the company as $[\theta_\textrm{c}-\Delta\theta_\textrm{c},\theta_\textrm{c}+\Delta\theta_\textrm{c}]$.
According to the central limit theorem (or to the Bernstein-von Mises theorem in the context of Bayesian inference) the posterior distribution
tends towards a normal distribution for $ N\gg 1 $ measurements, and the standard deviation of the posterior distribution decreases as $
\sigma=\sigma_{\rm prior}/\sqrt{N} $ for large $N$.
A possible choice for the number of measurements after which one can decide to stop the update process and apply the decision criterion is that $\sigma$ becomes comparable with the standard deviation $\sigma_{\rm spec}$ of a uniform distribution inside the spec. This corresponds to $ N_0=\sigma^2_{\rm prior}/\sigma^2_{\rm spec}$ measurements for $m=1$.
The set $\mathcal{M}$ one can choose from at each step includes the POVM elements defined above, as well as the identity matrix, corresponding to simply iterating the channel without measurement. To compare the efficiency of the different optimization methods (corresponding to different choices of $m$ and utility function $U$), we
calculate the probability of success, that is, the probability of reaching the correct conclusion, for the two decision criteria discussed in Sec.~\ref{design}.
In particular, since we fixed the total number of actions, we also consider as not successful for the "HPD" criterion the cases in which it remains inconclusive.
\subsubsection{Numerical results}
We consider sequences of measurements of length $m=1$ to 4, and both utility functions -- variance of posterior and mutual information -- discussed above. Our computations are performed with the Qinfer package \cite{Granade2017} which uses the SMC methods discussed in Section \ref{smc}. In this package, the credible region used in the HPD criterion is defined by ordering particles by decreasing weight and selecting the region including the highest-weight particles such that the desired credibility level ($ 95\% $) is reached. This corresponds to the HDP region defined above. In principle, there could be an issue with the connexity of the support (for instance if the distribution is bimodal). This has been discussed e.g.~in \cite{Ferrie2014b}. Here, in practice, posterior distributions always end up having Gaussian-like shapes, and thus the above construction of the credible region meets its purpose. We used a particle filter with 2000 particles, and we checked that for this choice the numerical values of the MI agree up to the second decimal digit with the analytical result obtained after one Bayesian update of the probability distribution.
In Fig.~\ref{newpanel} we show the results obtained for a fixed angle $ \theta_{\textrm{true}}=\pi/10 $ and several values of $ \theta_\textrm{c} $ with a fixed $\Delta\theta_\textrm{c}= \frac{\pi}{18}$. In order to reach a width of the posterior distribution of the order of the width of the spec we take $ N_0\approx 300 $. We find that the optimization over sequences of measurements of length $m\geq 2$ gives an advantage over the greedy approach $m=1$, both for the negative variance and for the mutual information. Namely, the final posterior distribution for $m\geq 2$ is narrower than the one obtained for $m=1$ with the same number of actions (measurements or iterations of the channel), despite a smaller number of Bayesian updates. After a few initial oscillations, the best sequence of actions always coincides with the one which contains $m-1$ times the identity and a single measurement ($\sigma_x$ or $\sigma_y$). By contrast, in the greedy approach, the identity measurement is never chosen, as it does not lead to any information gain by itself. \\
\begin{figure}
\includegraphics[scale=0.45]{newfig67.pdf}
\caption{(Upper panel) Success probabilities for correctly classifying the phase gate $R_{{\bm e}_z}(\theta)$ as in or out of spec concerning $\theta$
for different lengths of measurement sequences. Here we consider $\theta_{\textrm{true}}=\pi/10$ and $\Delta\theta_\textrm{c}=\pi/18$. Utility function: mutual information (MI); decision criterion: position of the estimated mean value (MEAN).
Probabilities are averaged over 2000 random particle filters with 2000 particles each.
The probabilities increase going from length 1 to 4. The minimum is reached when $ \theta_{\textrm{true}} \simeq\theta_\textrm{c}-\Delta\theta_\textrm{c} $. Statistical fluctuations of the estimates of $p$ as judged from the minimum, which should be at $p=0.5$, amount to $ \sim 0.05 $. (Lower panel) Success probabilities for measurement sequences of length 4 for all four combinations of the two utility functions MI and VAR, and decision criteria HPD and MEAN.
}
\label{newpanel}
\end{figure}In the upper panel of Fig.~\ref{newpanel}
we present the results for the MI as utility function and MEAN as decision method, comparing sequences of $m=1,2,3 $ and $ 4 $ measurements.
The probability of success decreases and reaches the minimum value when the true angle $ \theta_{\textrm{true}}$ almost coincides with one of the boundaries of the spec region (i.e.~when $ \theta_{\textrm{true}}\simeq \theta_\textrm{c}\pm\Delta\theta_\textrm{c} $). In that case the posterior distribution is expected to converge to a Gaussian centered on the boundary; the mean fluctuates around the boundary from one random realization to the other, yielding only a $50\%$ chance of reaching the correct conclusion from the results of the decision criterion based on the position of the mean. The error made by having only a limited number of realizations can be estimated by considering the fluctuations of the mean at that point, where the theoretical value of the success probability is exactly $0.5$. The difference between the largest and smallest values for different ${\cal N}$ is $ \sim 0.05 $, yielding an indication of how large statistical errors bars are for this point. We can see that the probability of success grows going from sequences of measurements of length 1 (greedy approach) to 4.
In the bottom panel of Fig.~\ref{newpanel} we compare the results obtained for sequences of length $m=4$ for all four combinations of the utility functions MI, VAR and final decision criteria MEAN, HPD. We see that MEAN always gives greater success probabilities than the HPD criterion, whereas the utility functions has little influence. We expect bad performance of the HPD criterion when $\theta_{\textrm{true}}$ is close to $\theta_\textrm{c}\pm \Delta_\textrm{c}$ as the final distribution has insufficient overlap with both the spec and the interval outside the spec, i.e.~the criterion remains inconclusive and the test is counted as unsuccessful. This is confirmed in Fig.~\ref{newpanel} for $\theta_{\textrm{true}}=\theta_\textrm{c}- \Delta_\textrm{c}$. On the other hand, the MEAN criteria leads to essentially random classification at the border of the spec, resulting in about 50\% success rate (but also 50\% of wrongly classified gates). Depending on the use case either strategy can be more advantageous, the first one being safer, the second leading to less waste of devices still in spec. \\
\subsection{Phase gate with dephasing}
\subsubsection{Setup}
We now consider the phase gate with dephasing, defined in Sec.~\ref{pgd}. Suppose that we want to estimate the decay parameter $\gamma$. The controllable quantities are the evolution time $t$, which we fix to some constant value (or to a multiple of it when the channel is iterated), and the measurements to perform after the evolution, which are chosen with an adaptive strategy as in Sec.~\ref{1stmodel}.
We again choose the experiment-limited scenario, where $N$ is fixed, and we investigate whether iterating the channel, represented by the time evolution, can bring an advantage in drawing the right conclusion about the parameter to estimate.
\subsubsection{Numerical results}
To run the certification protocol we chose a uniform prior distribution of $\gamma$ over the interval $[0,1]$. We allow for the same two-outcome measurement as in Sec.~\ref{1stmodel}, and fix again the number of actions to $N_0=300$.
We set a spec $\gamma\in[\gamma_\textrm{c}-\Delta\gamma_\textrm{c},\gamma_\textrm{c}+\Delta\gamma_\textrm{c}]$ with $\Delta\gamma_\textrm{c}=0.03$ provided by the company. We simulate an experiment with $\gamma_{\textrm{true}}=0.1$ and $t=5$ (for $k$ iterations of the channel we take $t=5k$). The adaptive strategy based on the two different utility functions never selects as optimal sequences of measurements the ones containing the identity, so never allowing for more than one application of the channel. We show the results obtained for the probability of success in Fig.~\ref{panel2}.
In the upper panel we show that the probability of success is larger for $m=2$ or $3$ for the mutual information strategy with the decision criterion based on the mean as long as the true value $\gamma_t$ is inside the spec, whereas for $\gamma_t$ outside the spec, $m=3$ turns out to perform worse than $m=1,2$. This is explained by the fact that the posterior probability distributions have not completely converged to their asymptotic form for $N_0=300$. We show in the Appendix that for $N_0=1200$ the order is as expected, with $m=2,3$ performing better than $m=1$.
In the bottom panel we show the difference between the probabilities of success for sequences of length 3, finding as in the case of the rotation angle (see Fig.~\ref{newpanel}) that both the utility functions perform better with the 'MEAN' as final decision criterion.
We showed in Fig.~\ref{xydephasing} that iterating the dephasing channel does not enhance the distinguishability between the ideal and faulty channels. Indeed, for $\gamma=0.1$ and $t=5$, the corresponding value $\tau$ is in a regime in Fig.~\ref{xydephasing} where $\xi_{\rm QCB}$ is not a decreasing function of $\tau$.
Fig.~\ref{panel2n01200} for $N_0=1200$ shows that iterations slightly enhance the probability of success for the correct identification of the channel (ideal or faulty). This might be seen to contradict the findings based on the quantum Chernoff bound. However, quantum Chernoff bound and success probabilities give answers to two distinct questions. If one wishes to distinguish channels, they should be as far apart as possible in the space of states, i.e.~${\cal N}$ should be as small as possible according to Eq.~\eqref{rhodetau}. However if the question is to detect as fast as possible whether a probability distribution is within or outside some spec interval, it is their distance in the space of the parameter we are probing that matters: in that case, for a fixed number $N_0$, larger ${\cal N}$ may correspond to a faster decrease in the width of the posterior distribution, thus to a higher distinguishability.
\begin{figure}
\includegraphics[scale=0.45]{newfig7_2.pdf}
\caption{(Upper panel) Success probabilities for correctly classifying the phase gate with dephasing as in or out of spec concerning $\gamma$ for different lengths of measurement sequences. Here we consider $\gamma_{\textrm{true}}=0.1$ and $\Delta\gamma_\textrm{c}=0.03$, so that for $\gamma_\textrm{c}<0.13$ $\gamma_{\textrm{true}}$ is within the spec and outside for $\gamma_\textrm{c}>0.13$. Utility function: mutual information (MI); decision criterion: position of the estimated mean value (MEAN). The probabilities are averaged over 2000 random particle filters with 2000 particles each and $N_0=300$.
The success probabilities improve only slightly from $m=1$ to $m=2$ and for $\gamma_\textrm{c}\lesssim \gamma_{\textrm{true}}+\Delta\gamma_\textrm{c}$. Then they coincide, showing no advantage in iterating the channel. Statistical fluctuations of the estimates of $p$ as judged from the minimum, which should be at $p=0.5$, amount to $ \sim 0.05 $. (Lower panel) Success probabilities for measurement sequences of length 3 for all combinations of the utility functions (MI and VAR) and decision criteria (HPD and MEAN) considered.
}
\label{panel2}
\end{figure}
\section{Conclusions}
To summarize, we have described how the perspective of quantum functional testing leads to a new twist of quantum parameter estimation, namely one where the rapid gain of information about a parameter and a decision problem to accept or reject a device is center stage, rather than ultimate sensitivity. We have introduced several ingredients that allow one to speed up the decision problem, most importantly the possibility of coherent amplification of errors by iteration of a quantum channel before measurement. We have analysed the interest of iterations first with Chernoff bounds, both classical and quantum, and found that for single qubit gates and deterministic (or narrowly distributed) errors in the rotation angles with negligible decoherence, iterations of the channel tend to be beneficial, where however attention has to be paid to choosing appropriately adapted measurements as well. We then introduced non-greedy Bayesian updates and experimental design, where an agent decides on the fly whether it is more advantageous to continue iterating the quantum channel or measuring, and what kind of measurement to perform, based on the previously accumulated data. This setup was investigated for two different utility functions and two decision criteria. The simulations showed that the estimate of rotation angles profits from iterating the quantum channel. In contrast, for the dephasing parameter the situation is more complex and depends on whether or not the true value of the parameter is inside or outside the spec. We also stressed the fact that assessing whether the parameter value is within a certain interval or not is distinct from the problem of distinguishing two quantum channels.
\section*{Acknowledgments}
This work is funded by the Ministry of Economic Affairs, Labour and Tourism Baden W\"urttemberg in the frame of the Competence Center Quantum Computing Baden-W\"urttemberg (project 'QORA').
|
\section{Introduction}
Recent progress in computer vision and natural language processing has enabled a wide range of possible applications to generative models. One of the most promising applications is text-guided image generation (text-to-image models). Solutions like DALL-E~2~\cite{ramesh2022} and Stable Diffusion~\cite{rombach2021highresolution} use the recent advances in joint image and text embedding learning (CLIP~\cite{pmlr-v139-radford21a}) and diffusion models~\cite{pmlr-v37-sohl-dickstein15} to produce photo-realistic and aesthetically-appealing images based on a textual description.
However, in order to ensure the high quality of generated images, these models need a proper \emph{prompt engineering} \cite{Liu_2022} to specify the exact result expected from the generative model. In particular, it became a common practice to add special phrases (\emph{keywords}) before or after the image description, such as ``trending on artstation,'' ``highly detailed,'' etc. Developing such prompts requires human intuition, and the resulting prompts often look arbitrary. Another problem is the lack of evaluation tools, so practically, it means that the user subjective judges the quality of a prompt by a single generation or on a single task. Also, there is currently no available analysis on how different keywords affect the final quality of generations and which ones allow to achieve the best images aesthetically.
In this work, we want to bridge this gap by proposing an approach for a large-scale human evaluation of prompt templates using crowd workers. We apply our method to find a set of keywords for Stable Diffusion that produces the most aesthetically appealing images. Our contributions can be summarized as follows:
\begin{itemize}\itemsep0em
\item We introduce a method for evaluating the quality of generations produced by different prompt templates.
\item We propose a set of keywords for Stable Diffusion and show that it improves the aesthetics of the images.
\item We release all the data and code that allow to reproduce our results and build solutions on top of them, such as finding even better keywords and finding them for other models.
\end{itemize}
\section{\label{sec:prompts}Prompts and How to Evaluate Them}
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{Example_img.pdf}
\caption{\label{fig:example_images}Comparison of the keyword sets. Left: no keywords vs. our approach. Right: 15 most popular keywords vs. our approach. Images are cherry-picked.}
\end{figure}
Consider a standard setup for generative models with text inputs. A model input a natural language text called \emph{prompt} and outputs a text completion in the case of the text-to-text generation or an image in the case of text-to-image generation. Since specifying the additional information increases the quality of the output images \cite{Liu_2022}, it is common to put specific keywords before and after the image description:
\begin{equation*}
\text{prompt} = [\text{keyword}_1,\ldots,\text{keyword}_{m-1}]\;[\text{description}]\;[\text{keyword}_m,\ldots,\text{keyword}_{n}].
\end{equation*}
Consider a real-world example when a user wants to generate an image of a cat using a text-to-image model.\footnote{\url{https://lexica.art/?q=a+cat&prompt=28f5c644-9310-4870-949b-38281328ffd0}} Instead of passing a straightforward prompt
\emph{a cat}, they use a specific prompt template, such as \emph{Highly detailed painting of a calico cat, cinematic lighting, dramatic atmosphere, by dustin nguyen, akihiko yoshida, greg tocchini, greg rutkowski, cliff chiang, 4k resolution, luminous grassy background}. In this example, the \textbf{description} is \emph{painting of a calico cat} and the \textbf{keywords} are \emph{highly detailed, cinematic lighting, dramatic atmosphere, by dustin nguyen, akihiko yoshida, greg tocchini, greg rutkowski, cliff chiang, 4k resolution, luminous grassy background}.
Since aesthetics are difficult to evaluate computationally, we propose a human-in-the-loop method for evaluating the keyword sets. Our method inputs a set of descriptions and a set of the keyword set candidates and outputs a list of keyword sets in the decreasing order of their aesthetic appeal to humans:
\begin{enumerate}\itemsep0em
\item For each pair of a description and a keyword set, generate the image.
\item For each image description, sample $nk\log_2(n)$ pairs of generated images~\cite{pmlr-v70-maystre17a}, where $n$ is the number keyword sets to compare and $k$ is the number of redundant comparisons.
\item Run a pairwise comparison crowdsourcing task in which the workers are provided with a description and a pair of images, and they have to select the best image without knowing the keyword set.
\item For each description, aggregate the pairwise comparisons using the Bradley-Terry algorithm \cite{Bradley_1952}, recovering a list of keyword sets ordered by their visual appeal to humans.
\item For each keyword set, compute the average rank in the lists recovered for the descriptions.
\end{enumerate}
As a result, we quantify the quality of a keyword set as its rank averaged per description.
\section{Iterative Estimation of the Best Keyword Set}
One of the advantages of our approach is that the keywords can be evaluated iteratively. Once we have compared a number of keyword sets, we can request a small additional number of comparisons to evaluate the new set of keywords. This allows us to apply discrete optimization algorithms, such as a genetic algorithm, to retrieve from a large pool of keywords the most influential keywords.
We pick a set of keyword sets for initialization, rank the keywords using the approach in Section~\ref{sec:prompts}, and use it as an initial population for the genetic algorithm. Then we repeat the following steps multiple times to obtain the best-performing keyword set:
\begin{enumerate}\itemsep0em
\item Obtain the next candidate keyword set using the genetic algorithm.
\item Sample $k\left((n+1)\log_2(n + 1) - n\log_2 n\right)$ pairs of images generated using keywords from the new candidate set and already evaluated keyword sets. We do this to sustain $kn\log_2 n$ comparisons in total.
\item Evaluate the quality of the obtained keyword set (Section~\ref{sec:prompts}).
\end{enumerate}
\section{Experiment}
We perform an empirical evaluation of the proposed prompt keyword optimization approach in a realistic scenario using the publicly available datasets.
\subsection{Setup}
To construct a set of possible keywords, we have parsed the Stable Diffusion Discord\footnote{\url{https://discord.com/invite/stablediffusion}} and took the 100 most popular keywords. For image descriptions, we decided to choose prompts from six categories: \emph{portraits}, \emph{landscapes}, \emph{buildings}, \emph{interiors}, \emph{animals}, and \emph{other}. We took twelve prompts for each category from Reddit and \url{https://lexica.art/} and manually filtered them to obtain only raw descriptions without any keywords.
We use a simple genetic algorithm to find the optimal prompt keyword set. The algorithm was initialized with two keyword sets: one is an empty set, and another set contained the 15 most popular keywords that we retrieved before. We limited the maximum number of output keywords by 15 as otherwise, the resulting prompts became too long.
In order to evaluate the keyword sets, we generate four images for each prompt constructed by appending comma-separated keywords to the image description in alphabetical order. Each image was generated with Stable Diffusion model \cite{rombach2021highresolution} with 50 diffusion steps and 7.5 classifier-free guidance scale using the DDIM scheduler~\cite{song2020denoising}. Then, we run crowdsourcing annotation on the Toloka crowdsourcing platform.\footnote{\url{https://toloka.ai/}} The crowd workers had to choose the most aesthetically-pleasing generated image in $3n \log_2 n$ pairs for each image description, where $n$ is the number of currently tried keyword sets.
After the annotation is completed, we run the Bradley-Terry~\cite{Bradley_1952} aggregation from the Crowd-Kit~\cite{HCOMP2021/CrowdKit} library for Python to obtain a ranked list of keyword sets for each image description. The final evaluation metric used in the genetic algorithm to produce the new candidate sets is the average rank of a keyword set (as described in Section~\ref{sec:prompts}). We use 60 image descriptions for optimization (10 from each category) and 12 for the validation of the optimization results.
Since crowdsourcing tasks require careful quality control and our task involved gathering subjective opinions of humans, we followed the synthetic golden task production strategy proposed for the IMDB-WIKI-SbS dataset~\cite{imdb-wiki-sbs}. We randomly added comparisons against the images produced by a simpler model, DALL-E Mini~\cite{Dayma_DALL_E_Mini_2021}. We assumed that DALL-E Mini images are less appealing than the ones generated by Stable Diffusion, and choosing them was a mistake. Hence, we suspended the workers who demonstrated an accuracy lower than 80\% on these synthetic golden tasks.
\subsection{Results}
\begin{table}[t]
\centering
\caption{\label{tab:results}Average rank of the baseline keywords and the ones found by the genetic algorithm. Rank is averaged over 60 prompts on train and over 12 prompts on validation (val); maximal rank is 56.}
\resizebox{\textwidth}{!}{\begin{tabular}{rrrrcrrrr}\toprule
\multicolumn{4}{c}{\textbf{Train}} & \phantom{abc} & \multicolumn{4}{c}{\textbf{Validation}} \\
\cmidrule{1-4} \cmidrule{6-9}
No Keywords & Top-15 & Best Train & Best Val & & No Keywords & Top-15 & Best Train & Best Val \\ \midrule
$3.5$ & $14.25$ & $\boldsymbol{43.60}$ & $39.32$ & & $5.42$ & $12.50$ & $38.00$ & $\boldsymbol{46.00}$ \\
\bottomrule
\end{tabular}}
\end{table}
We ran the optimization for 56 iterations on 60 image descriptions since we have a fixed annotation budget. To ensure that our method did not overfit, we ran the evaluation on another 12 descriptions (validation). According to the evaluation results in Table~\ref{tab:results}, we found that our algorithm was able to find a significantly better set of keywords than the 15 most popular ones (Top-15). Also, we see that any set of prompt keywords is significantly better than no keywords at all (No Keywords).
We see that most results hold on the validation set, too, but the metrics have more noise. Overall, the best set of keywords on the training set of 60 prompts is \emph{cinematic, colorful background, concept art, dramatic lighting, high detail, highly detailed, hyper realistic, intricate, intricate sharp details, octane render, smooth, studio lighting, trending on artstation}. An example of images generated with this keyword set is shown in Figure~\ref{fig:example_images}.
\subsection{Discussion}
We show that adding the prompt keywords significantly improves the quality of generated images. We also noticed that the most popular keywords do not result in the best-looking images. To estimate the importance of different keywords, we trained a random forest regressor~\cite{Breiman:01} on the sets of keywords and their metrics that is similar to W\&B Sweeps.\footnote{\url{https://docs.wandb.ai/guides/sweeps}} We found that the most important keywords, in reality, are different from the most widely used ones, such as ``trending on artstation.'' The most important keyword we found was ``colorful background.''
There are several limitations to our approach. We can not conclude that the found set of keywords is the best one since the genetic algorithm can easily fall into a local minimum. In our run, it tried only 56 keywords out of the 100 most popular ones. Also, our evaluation metrics are based on ranks, not absolute scores, so they are not sensitive enough to determine the convergence of the algorithm.
However, since we release all the comparisons, generated images, and code, it is possible for the community to improve on our results. For instance, one can run a genetic algorithm from a different initialization, for a larger number of iterations, or even with more sophisticated optimization methods. This can easily be done by comparing the new candidates with our images and adding these results to the dataset.
\section{Related Work}
The aesthetic quality evaluation is one of the developing topics in computer vision. There are several datasets and machine learning methods aiming at solving this problem~\cite{Sheng_2018,zhang2021image}. However, the available datasets contain human judgments on image aesthetics scaled from 1 to 5. Our experience shows that the pairwise comparisons that we used in this paper are a more robust approach as different humans perceive scales differently and subjectively. Also, they specify training a model to evaluate the aesthetics but not on the generative models. Large language models, such as GPT-3~\cite{GPT3}, have enabled a wide range of research tasks on prompt engineering~\cite{BetterFewShot,Liu2022PretrainPA,fantasticallyOrderedPrompts,CrossTaskGeneralization,HowContextAffects,PromptProgramming,RetrivePrompts}. Recent papers also discover the possibilities of prompt engineering for text-to-image models and confirm that prompts benefit from the added keywords~\cite{Liu_2022}. To the best of our knowledge, we are the first to apply it to find the best keywords.
\section{Conclusion}
We presented an approach for evaluating the aesthetic quality of images produced by text-to-image models with different prompt keywords. We applied this method to find the best keyword set for Stable Diffusion and showed that these keywords produce better results than the most popular keywords used by the community. Despite the fact that our work focuses on the evaluation of keywords for text-to-image models, it is not limited by this problem and can be applied for an arbitrary prompt template evaluation, for example, in the text-to-text setting. This is a direction for our future work. Last but not least, we would like to encourage the community to continue our experiment and find better keyword sets using our open-source code and data.\footnote{\url{https://anonymous.4open.science/r/sd-keyword-search-F7F2}}
\clearpage
\bibliographystyle{plain}
|
\section{Introduction}
\input{progress_report/1_definition}
\section{Related Work}
\input{progress_report/2_related_work}
\section{Background}
\input{progress_report/3_background}
\section{Method}
\input{progress_report/4_method}
\section{Experiment}
\input{progress_report/6_experiment}
\section{Conclusions}
\input{progress_report/7_conclusion}
{\bibliographystyle{ieee_fullname}
\subsection{Signal Fitting with Implicit Representations}
Implicit neural representations learn implicitly defined functions to model many different types of signals such as image, video, and audio processing in a continuous, memory-efficient way. The following equation is a general form of loss to train the implicit function $\Phi$ to represent the signal $f$ on $\Omega$:
\begin{equation}
\mathcal{L}=\int_{\Omega}
\Vert \Phi(\mathbf{x})
- f(\mathbf{x})
\Vert d\mathbf{x}.
\label{eq-inr}
\end{equation}
It turns out that sinusoidal representation networks (SIRENs) can effectively learn representing complex natural signals by using periodic activation functions such as sin activation functions~\cite{NEURIPS2020_53c04118}. SIRENs converge significantly faster than existing networks that use other activation functions as well as feature high image fidelity. It shows that reconstructed images from SIRENs have more realistic image properties so that their gradients and laplacians are close to the ground truth.
\subsection{Poisson Equation}
A supervised neural network is generally trained to provide an output that is close to the ground truth. Implicit neural representations also focus on learning implicit functions of given signals. However, reconstructed images from SIRENs that are trained to fit given images may not reflect complex patterns such as edges in a logo. It is because there is no supervision to learn exact edges in images and the network is optimized to fit average color-level of pixels.
The implicit neural representation can also be supervised solely by derivatives of a given signal~\cite{NEURIPS2020_53c04118}. That is, the model is never provided with the actual function values. The Poisson equation is the simplest elliptic partial differential equation (PDE) that requires to fit the first or higher-order derivatives of a given signal. In this problem, an unknown ground truth signal $f$ is estimated from discrete samples of either is gradients $\nabla f$ or Laplacian $\Delta f=\nabla \cdot \nabla f$ as
\begin{equation}
\mathcal{L}_{\textrm{grad.}}=\int_{\Omega}
\Vert \nabla_\mathbf{x}\Phi(\mathbf{x})
- \nabla_\mathbf{x} f(\mathbf{x})
\Vert d\mathbf{x},
\label{eq-poisson-grad}
\end{equation}
\begin{equation}
\mathcal{L}_{\textrm{lapl.}}=\int_{\Omega}
\Vert \Delta_\mathbf{x}\Phi(\mathbf{x})
- \Delta_\mathbf{x} f(\mathbf{x})
\Vert d\mathbf{x}.
\end{equation}
Solving the Poisson equation enables the reconstruction of images from their derivatives so that the implicit function represents edges of images much better. Actually, a way to fit gradients with SIREN for a RGB image has not been directly described in the original paper. We train SIRENs to minimize the difference between automatically computed gradients and true gradients calculated by sobel filters for each RGB channel.
Figure~\ref{SIREN comparison} shows one example of the results. SIREN trained to fit gradients can describe complex patterns such as the logo in the jacket and clearly reconstruct eyes of the astronaut. However, there is remaining intensity variations due to the ill-posedness of the problem.
\subsection{Sobel Filter}
In the SIREN paper, the Sobel filter was used to calculate the ground truth gradient. The Sobel filter is used in image processing and computer vision, particularly within edge detection algorithms where it creates an image emphasizing edges. Technically, it is a discrete differentiation operator, computing an approximation of the gradient of the image intensity function. The Sobel filter is based on convolving the image with a 3x3 filter in the horizontal and vertical direction.
\begin{equation}
\mathbf{G}_x=
\begin{bmatrix}
-1 & 0 & 1\\
-2 & 0 & 2\\
-1 & 0 & 1
\end{bmatrix}
,\quad
\mathbf{G}_y=
\begin{bmatrix}
-1 & -2 & -1\\
0 & 0 & 0\\
1 & 2 & 1
\end{bmatrix}
\end{equation}
$\mathbf{G}_x$ denotes the Sobel filter for the horizontal direction and $\mathbf{G}_y$ denotes the Sobel filter for the vertical direction. The gradient approximation that the Sobel filter produces is relatively crude, in particular for high-frequency variations in the image.
\begin{figure}[ht!]
\includegraphics[width=8cm, trim = 0cm -1cm -1cm 0cm, clip]{progress_report/sobel-problem.pdf}
\centering
\caption{Induced problem of the Sobel filter. (a) shows that the gradients computed by convolving the Sobel filter can be overestimated compared to the original gradient in the simple example. (b) shows that the reconstructed dog image from SIREN trained with these estimated gradients has extremely large differences at the edges of the image.}
\label{fig:sobel-problem}
\end{figure}
\subsection{Gradient Magnitude Adjustment}
Even though the Sobel filter is widely used as a computationally efficient way to approximate the gradients of images, it is problematic to feed the estimated gradients as the ground truth to the model. Figure~\ref{fig:sobel-problem} describes which problem can be induced from the Sobel filter. Consider the simple case where the gradient value is a constant $a$ in the horizontal direction. When we apply convolution with the Sobel filter, the estimated gradient is $8a$. That is, the Sobel filter overestimate the gradients. To fit the overestimated gradient values, the model sacrifice the correct pixel values such as the dog image in Figure~\ref{fig:sobel-problem} (b). Even a little contrast exists in the original input image, the difference of the values in the edge is overestimated and the reconstructed image becomes very unrealistic.
To address this issue, we propose to adjust the magnitude of gradients computed by filter for realistic image reconstruction. We define the magnitude of the filter as the total sum of the absolute values used in the filter. For instance, the Sobel filter, described in Figure~\ref{fig:sobel-problem}, has magnitude 8 which means that the estimated gradient values are 8 times larger than the actual gradient values. Therefore, we need to offset the gap between the estimated gradients and the actual gradients.
We consider two main components to feed proper gradient values to the model. The first one is the magnitude of the filter which is introduced previously, since it leads to overestimation. The second thing is the size of the image. Let $W^T,H^T$ be the range of the input coordinates and $W,H$ be the width and height of the image respectively. Note that convolving process assumes that the size of the interval between adjacent pixels is 1. However, SIREN uses the different range of input coordinates. As a result, the derivatives of the implicit function with respect to input may have different scale. Therefore, we need to multiply the gradients by $\frac{W}{W_T}, \frac{H}{H_T}$, after normalizing it by filter magnitude. Then, we can feed adjusted gradient values to the model for solving the Poisson Equation.
\begin{algorithm}
\setstretch{1.35}
\caption{Gradient with Magnitude Adjustment}\label{alg:gma}
\begin{algorithmic}[1]
\Procedure{GMA}{$I_{W\times H}$, ${W_T, H_T}$ , $F^x$, $F^y$}
\State $M_F^x \gets |F^x|$ \Comment{compute filter magnitude}
\State $M_F^y \gets |F^y|$
\State $F^x \gets F^x/M_F^x$ \Comment{normalize by filter magnitude}
\State $F^y \gets F^y/M_F^y$
\State $F^x \gets F^x \times \frac{W}{W_T}$ \Comment{normalize by image size}
\State $F^y \gets F^y \times \frac{H}{H_T}$
\State $grad_x \gets F^x \circledast I_{W\times H}$ \Comment{convolution}
\State $grad_y \gets F^y \circledast I_{W\times H}$
\State \textbf{return} $(grad_x, grad_y)$
\EndProcedure
\end{algorithmic}
\end{algorithm}
\begin{figure*}[ht]
\includegraphics[width=16cm, trim = 0cm 0cm 0cm 0cm, clip]{progress_report/recon-figure.png}
\centering
\caption{The implicit representation of the RGB images in skimage data with SIREN-pixel, SIREN-grad, EoREN~ without the channel tuning module and EoREN. }
\label{exp_constructed_rgb}
\end{figure*}
\subsection{EoREN}
We add one more additional module to help our model correctly represent pixel values while keeping complex signals by gradient adjustment. We propose Edge-oriented Representation Network (EoREN) that can represent complex patterns in the signal with avoiding the ill-posedness problem through the channel tuning module.
\begin{equation}
\Phi(\mathbf{x}) = (h \circ g) (\mathbf{x})
\end{equation}
EoREN~is composed of two functions $g$ and $h$. $g$ represents original implicit representation function of the signal and $h$ is an additional shift transformation for each dimension of outputs.
\begin{gather}
g(\mathbf{x}) = \mathbf{W}_n(g_{n-1} \circ g_{n-2} \circ \cdots \circ g_0)(\mathbf{x}) + \mathbf{b}_n \\
g_i: \mathbf{x}_i \mapsto \mathrm{Activation}(\mathbf{W}_i \mathbf{x}_i + \mathbf{b}_i)
\end{gather}
Here, $g_i : \mathbb{R}^{M_i} \rightarrow \mathbb{R}^{N_i}$ is the $i^{th}$ layer of the network with the weight matrix $\mathbf{W}_i \in \mathbb{R}^{N_i \times M_i}$ and the biases $\mathbf{b}_i \in \mathbb{R}^{N_i}$. The function $g$ is a MLP that acts as an implicit function of the signal. It can be SIREN with the use of sine activation functions.
The objective of our implicit function is not to optimize the values of the signal $f$, but to match the gradients of the signal. Our implicit function $g$ is trained to solve the Poisson equation for fitting true gradients $\nabla_\mathbf{x} f(\mathbf{x})$ for each coordinate $\mathbf{x}\in\Omega$. That is, it minimizes the loss described in Equation~\ref{eq-poisson-grad}.
\begin{equation}
\mathcal{L}_{\mathrm{grad.}} = \int_\Omega ||\nabla_\mathbf{x} g(\mathbf{x}) - \nabla_\mathbf{x} f(\mathbf{x})|| d\mathbf{x}
\end{equation}
We call this implicit function \textit{Edge-oriented module}. The Edge-oriented module can be easily supervised to learn complex shapes in a given images. However, since it is optimized only to fit $\nabla_\mathbf{x} f(\mathbf{x})$, $g$ may not follow the actual signal $f$. In the case of image fitting, $g$ may not represent true colors of the image $f$. Therefore, we need to calibrate the distribution of reconstructed images.
We propose an additional shift transformation layer, called \textit{channel tuning module}, to adjust differences between the output and the target signal. Figure~\ref{fig-eoren-archi} shows the overall architecture of our model. $\mathbf{\alpha}$ and $\mathbf{\beta}$ are vectors of trainable parameters.
\begin{equation}
h(\mathbf{x}) = \alpha * g(\mathbf{x})+\beta
\end{equation}
It is updated to transform the output of the Edge-oriented module so that it fit the true values of the signal $f$. The loss to optimize $\alpha$ and $\beta$ are as follows:
\begin{equation}
\tilde{\mathcal{L}} = \int_\Omega || f(\mathbf{x}) - \Phi(\mathbf{x})|| d\mathbf{x}
\label{eq-loss-offset}
\end{equation}
Note that we do not backpropagate the loss in Equation~\ref{eq-loss-offset} to optimize $g$. When we set $\alpha=\mathbf{1}$, even if we add the additional channel tuning module, the derivatives of EoREN~do not change due to the chain rule, i.e., $\nabla\Phi= \nabla g$. Theoretically, it is enough to set $\alpha=\mathbf{1}$ and only train $\beta$ parameter to fit correct pixel values. However, it has been empirically observed that training $\alpha$ parameter helps the model to fit the distribution of pixel values more precisely. Therefore, in our experiment, we decide to train $\alpha$ in the channel tuning module. Consequently, EoREN~can learn true color of the image without hindering the representability for gradients.
\subsection{Signal Reconstruction}
We first show the signal reconstruction of our model. We test our model with three color images in skimage data. We choose three images, the astronaut, the rocket and the coffee. The results of reconstruction are described in the Figure \ref{exp_constructed_rgb}. The first two columns in Figure~\ref{exp_constructed_rgb} are the images from SIREN with pixel/gradient fitting. Gradient fitting images successfully represent the edge parts of the image, while they fail to match the RGB values of the image. The results of SIREN with pixel fitting have plausible reconstructed RGB values, but fail to distinguish detailed edges on image, such as the patch in the astronaut image, towers in the rocket image and the background in the coffee image. The images reconstructed by our EoREN~are the 3th and 4th columns on Figure~\ref{exp_constructed_rgb}. EoREN without channel tuning module overestimate the edges, such as shadows in the coffee image and dark parts in the rocket image. Then, it results in generating high-contrast image compared to the original images. Meanwhile, EoREN with channel tuning module shows not only clear but also realistic images with similar color levels compared to the other gradient fitting models.
\begin{figure}[ht]
\includegraphics[width=8cm, trim = 0cm 0cm 0cm 0cm, clip]{progress_report/histogram.pdf}
\centering
\caption{The histogram of the pixel values and the gradients for the RGB astronaut image. }
\label{exp_hist}
\end{figure}
\begin{table*}[ht]
\centering
\caption{The result of image reconstruction for the baseline models and the EoREN.}
\begin{tabular}{l c c c c c c}
\toprule & \multicolumn{2}{p{4cm}}{Set14} & \multicolumn{2}{p{4cm}}{DIV2K} & \multicolumn{2}{p{4cm}}{MNIST} \\
\cmidrule(lr){2-3} \cmidrule(lr){4-5} \cmidrule(lr){6-7}
Measure & PSNR & SSIM & PSNR & SSIM & PSNR & SSIM \\
\midrule
SIREN+Pixel & \textbf{73.81} & \textbf{0.78} & \textbf{74.23} & \textbf{0.83} & 83.69 & 0.91 \\
SIREN+Grad & 56.46 & 0.33 & 55.45 & 0.31 & 54.54 & 0.74\\
EoREN & 66.01 & 0.46 & 64.36 & 0.41 & \textbf{93.96} & \textbf{0.97}\\
\bottomrule
\label{table_psnr_ssim}
\end{tabular}
\end{table*}
\subsection{Disentangled Property on Edge and Offset}
One advantage of EoREN~ is the separation of modules for edge and offset. We show the disentangled property in color and edge construction of our model by the reconstruction of the edge-oriented module and the channel-tuning module separately.
The distributions of the pixel values and the gradients for three models are shown in the Figure \ref{exp_hist}.
The SIREN with pixel fitting shows that most of the pixel values are matched. However, the model underestimates the gradients and results in the smooth image as shown in the Figure \ref{exp_constructed_rgb}. The SIREN with gradient fitting matches most of the gradients but fails to reconstruct pixel values. We find that two pixel distributions from the ground truth and the model output are similar in shape. Therefore, our proposed model EoREN~ linearly transforms the pixel values by the channel-tuning module and matches the pixel distributions.
\subsection{Benchmark Dataset}
Following the previous work in \cite{NEURIPS2020_53c04118}, we compare EoREN~ and SIREN on the benchmark dataset. We evaluate our model based on the PSNR and SSIM metric \cite{5596999} which is widely used metric for image reconstruction. The overall results are in the Table \ref{table_psnr_ssim}.
\subsubsection{Set14}
Set14 contains 14 images from simple to complex. The PSNR performance of EoREN, SIREN with pixel fitting, and SIREN with gradient fitting for each sample is in Figure \ref{exp_set14}. We observe that SIREN+pixel shows generally better PSNR scores than EoREN. However, there are some samples on which EoREN shows better PSNR than SIREN+pixel. We observe that EoREN achieves better performance when an image contains less complex edges. We shows four examples in Figure \ref{exp_set14_eg}. The images of Face and Flower in Set14 contains more edges than the images of Pepper and PPT3. This implies that EoREN~ reconstructs an image well when it contains more clear edges.
\begin{figure}[ht]
\includegraphics[width=8cm, trim = 0cm 0cm 0cm 3cm, clip]{progress_report/set14.pdf}
\centering
\caption{The PSNR performance of SIREN+pixel, SIREN+grad, and EoREN~ for Set14 dataset. EoREN~ outperforms SIREN+pixel for three images, coastguard, pepper, and ppt3. }
\label{exp_set14}
\end{figure}
\begin{figure}[ht]
\includegraphics[width=8cm, trim = 0cm 0cm 0cm 0cm, clip]{progress_report/set14_ex.pdf}
\centering
\caption{Reconstructed images by EoREN~ and SIREN with pixel fitting. the left panel contains the images which SIREN+pixel outperforms EoREN, whereas the right panel contains the images which EoREN~ outperforms SIREN+pixel.}
\label{exp_set14_eg}
\end{figure}
\subsubsection{DIV2K}
DIV2K contains high resolution realistic images and we trained three models for validation dataset of the high resolution images of DIV2K which constains 100 samples. The image has higher resolution than $256\times 256$ and training bigger size of an image requires larger model and more computational time. Therefore, we down-sample the image to $256\times 256$ size. The result in Table \ref{table_psnr_ssim} shows that EoREN~ has lower performance than SIREN+pixel. When we compare the performance of each samples, there is no case EoREN~ outperforms SIREN+pixel. We hypothesize that this is due to the fact that high resolution images contains more complex edge patterns than original $256\times 256$ image and down-sampling does not contain clear edge information.
\subsubsection{MNIST}
Finally, we tested three models with MNIST dataset. As the results of Table \ref{table_psnr_ssim} and the stacked box plot in Figure \ref{exp_mnist} show, EoREN~ significantly outperforms SIREN+pixel and SIREN+grad. We show an reconstruction example of digit six in Figure \ref{exp_6_digit}. When training SIREN with pixel fitting, there is a watermark on the reconstructed image, whereas, the reconstructed image of EoREN has no watermark. When the EoREN is trained without the channel-tuning module, it is equivalent to the SIREN+grad. As the performance of SIREN+grad suggests, the channel tuning module is necessary for the gradient fitting of image.
\begin{figure}[ht]
\includegraphics[width=8cm, trim = 0cm 0cm 0cm 1cm, clip]{progress_report/mnist_performance.pdf}
\centering
\caption{The performance of EoREN, SIREN with pixel fitting, and SIREN with gradient fitting for each digits. EoREN~ outperforms other models for all digits. We test for 10 samples each. }
\label{exp_mnist}
\end{figure}
\begin{figure}[ht]
\includegraphics[width=8cm, trim = 0cm 0cm 0cm 0cm, clip]{progress_report/6_digit.pdf}
\centering
\caption{Reconstruction sample of SIREN with pixel fitting and EoREN~ for digit six for 500 and 1000 epochs. }
\label{exp_6_digit}
\end{figure}
\subsection{Poisson Image Editing}
Two images $f_1$ and $f_2$ can be composited by the interpolation of them. The composite image $f_{1,2}$ can be trained using Possion image editing method \cite{perez2003poisson} which focuses on the edge of two images. Poisson image editing technique uses addition of the gradients of images and trained with single parametric function by
\begin{gather}
\mathcal{L}_{\mathrm{grad.}} = \int_{\Omega} || \nabla_\mathbf{x} \Phi(\mathbf{x}) - \nabla_\mathbf{x} f_{1,2}|| d\mathbf{x} \\
f_{1,2} = \lambda f_1 (\mathbf{x}) + (1-\lambda) f_2 (\mathbf{x})
\label{equation:possion}
\end{gather}
When model $\Phi$ is fully trained by Equation \ref{equation:possion}, we get two edge based composited image. The proportion of two images is controlled by the constant $\lambda \in [0,1]$. In the case of EoREN, there are already two hyperparameters $\gamma_1$ and $\gamma_2$ in a single image to control the training, the learning rate of the edge-oriented module and the offset-tuning module. Then, the composition of two images can be trained by two losses
\begin{gather}
\mathcal{L}_{1,2}^{grad} = \lambda \gamma_1 \mathcal{L}_{\mathrm{grad.}}^1 + (1-\lambda) \gamma_1 \mathcal{L}_{\mathrm{grad.}}^2 \\
\tilde{\mathcal{L}}_{1,2} = \lambda \gamma_2 \tilde{\mathcal{L}}^1 + (1-\lambda) \gamma_2 \tilde{\mathcal{L}}^2
\end{gather}
where $\mathcal{L}^i_{grad}$ is the loss of edge-oriented module of $f_i$ and $\tilde{\mathcal{L}}^i$ is the loss of offset-tuning module of $f_i$.
Therefore, there are total 3 number of independent hyperparameters to train composition. We show that by setting proper parameter, we can make the composition image with desired magnitude of edge and color. We test this Poisson image editing method for EoREN.
We show the result of the image composition in Figure \ref{exp_composition}. Two images have different degree of complexity.The first one is an image of a field with reeds, and the second one is an image of an eagle flying against the sky and sea. The first image contains more edges than the second image and therefore more difficult to train with gradients. The result shows that EoREN~ successfully combines two images when the portion of the second image is more than the first image. This implies that the channel tuning depends on the complexity of the edges.
\begin{figure}[ht]
\includegraphics[width=8cm, trim = 0cm 0cm 0cm 0cm, clip]{progress_report/composition.pdf}
\centering
\caption{The result of image composition for $\lambda=0.2, 0.6, 0.9$. EoREN~ reconstruct the composited image successfully when $\lambda$ is large indicating that EoREN~ prefers an image with simple patterns.}
\label{exp_composition}
\end{figure}
|
\section{Introduction}\label{Sec:Introduction}
Networks \cite{barabasi2013network,newman2018networks} abound and increasingly shape our world, ranging from infrastructural networks (transportation, telecommunication, power-grids, water, etc.) over social networks to brain and biological networks. In general, a network consists of a graph or underlying topology and a dynamic process that takes place on the network. Some examples of processes on a network are percolation \cite{karrer2014percolation} and epidemic spreading \cite{prasse2019viral,pastor2015epidemic}, that possess a phase transition \cite{sanchez2002nonequilibrium,stanley1971phase}. While most real-world processes on networks are non-linear, linearisation allows for hierarchical structuring of processes on the network \cite{jokic2020linear}.
The identification of communities and the corresponding hierarchical structure in real-world networks has been an active research topic for decades \cite{fortunato2010community}, although a single, precise definition of a community does not seem to exist \cite{budel2020detecting,newman2012communities}. In network science, a community is defined as a set of nodes that share links dominantly between themselves, while a minority of links is shared with other nodes in the network.
Newman proposed in \cite{newman2006modularity} a spectral clustering algorithm that reveals hierarchical structure of a network, by optimising modularity, a commonly used quality function of a graph partition. A heuristic, modularity-based two-step clustering algorithm, proposed by Blondel \textit{et al.} in \cite{blondel2008fast}, has proved to be computationally efficient and performed among the best in the comparative study conducted in \cite{lancichinetti2009community}. Recently, Peixoto proposed in \cite{peixoto2014hierarchical} a nested generative model, able to identify nested partitions at different resolutions, which thus overcomes an existing drawback of a majority of clustering algorithms, identifying small, but well-distinguished communities in a large network. Dannon \textit{et al.} concluded in their comparative study \cite{danon2005comparing} that those clustering algorithms performing the best tend to be less computationally efficient.
We refer to \cite{fortunato2010community} for a detailed review on existing clustering algorithms.
Our new idea here is the proposal of a linear clustering process (LCP) on a graph, where nodes move in one-dimensional space and tend to concentrate in groups that lead to network communities. Linear means "proportional to the graph", which is needed, because the aim is to cluster the graph and the process should only help and not distract from our main aim of clustering. A non-linear process depends intricately on the underlying graph that we want to cluster and may result in worse clustering! Our LCP leads to a new and non-trivial graph matrix $W$ in (\ref{def_martrix_W}) in Theorem \ref{clustering_law_additive_matrix_form}, whose spectral decomposition is at least as good as the best clustering result, based on the non-back tracking matrix \cite{krzakala2013spectral}. Moreover, the new graph matrix $W$ has a more "natural" relation to clustering than the non-back tracking matrix, that was not designed for clustering initially. Finally, our resulting LCP clustering algorithm seems surprisingly effective and can compete computationally with any other clustering algorithm, while achieving generally a better result!
In Section \ref{Sec:Network_Clustering}, we introduce notations for graph partitioning and briefly review basic theory on clustering such as modularity and the stochastic block model. We introduce the linear clustering process (LCP) on network in Section \ref{Sec_dynamic_clustering}, while the resulting community detection algorithm is described in Section \ref{Sec:Position_Vector_Analysis} and Section \ref{Sec_Iterative_Computation}. We compare the performance of our LCP algorithm with that of the non-back tracking matrix, Newman's and the Louvain algorithm and provide results in Section \ref{Sec_Results}, after which we conclude.
\section{Network or graph clustering}\label{Sec:Network_Clustering}
A graph $G\left( \mathcal{N},\mathcal{L}\right) $ consists of
a set $\mathcal{N}$ of $N=\left\vert \mathcal{N}\right\vert $ nodes and a set $\mathcal{L}$ of $L=\left\vert \mathcal{L}\right\vert $ links and is defined by the $N\times N$ adjacency matrix $A$, where $a_{ij}=1$ if node $i$ and node $j$ are connected by a link, otherwise $a_{ij}=0$. The $N\times1$ degree vector $d$ obeys $d=A\cdot u$, where the $N\times1$ all-one vector $u$ is composed of ones. The corresponding $N\times N$ degree diagonal matrix is denoted by $\Delta=\text{diag}\left( d\right) $.
The set of neighbours of node $i$ is denoted by $\mathcal{N}_{i}=\{k\mid
a_{ik}=1,k\in\mathcal{N}\}$ and the degree of node $i$ equals the cardinality
of that set, $d_{i}=\left\vert \mathcal{N}_{i}\right\vert $. The set of common
neighbours of node $i$ and node $j$ is $\mathcal{N}_{i}\cap\mathcal{N}_{j}$,
while the set of neighbours of node $i$ that do not belong to node $j$ is
$\mathcal{N}_{i}\setminus\mathcal{N}_{j}$. The degree of a node $i$ also
equals the sum of the number of common and different neighbours between nodes
$i$ and $j$
\begin{equation}
d_{i}=\left\vert \mathcal{N}_{i}\setminus\mathcal{N}_{j}\right\vert
+\left\vert \mathcal{N}_{i}\cap\mathcal{N}_{j}\right\vert
\label{eq_degree_in_set_notation}%
\end{equation}
The number of common neighbours between nodes $i$ and $j$ equals the $ij$-th
element of the squared adjacency matrix
\begin{equation}
\left\vert \mathcal{N}_{i}\cap\mathcal{N}_{j}\right\vert =\left(
A^{2}\right) _{ij} \label{common_neig_set_not}%
\end{equation}
because $(A^{k})_{ij}$ represents the number of walks with $k$ hops between
node $i$ and node $j$ (see \cite[p. 32]{van2010graph}). From
(\ref{eq_degree_in_set_notation}), (\ref{common_neig_set_not}) and
$d_{i}=(Au)_{i}=(A^{2})_{ii}$, we have
\[
\left\vert \mathcal{N}_{i}\setminus\mathcal{N}_{j}\right\vert =\left(
A^{2}\right) _{ii}-\left( A^{2}\right) _{ij}%
\]
and
\[
\left\vert \mathcal{N}_{i}\setminus\mathcal{N}_{j}\right\vert +\left\vert
\mathcal{N}_{j}\setminus\mathcal{N}_{i}\right\vert =\left( A^{2}\right)
_{ii}+\left( A^{2}\right) _{jj}-2\left( A^{2}\right) _{ij}%
\]
The latter expression is analogous to the effective resistance $\omega_{ij}$
between node $i$ and node $j$,
\[
\omega_{ij}=Q_{ii}^{\dag}+Q_{jj}^{\dag}-2Q_{ij}^{\dag}
\]
in terms of the pseudoinverse $Q_{ii}^{\dag}$ of the Laplacian matrix $Q=$
$\Delta-A$ (see e.g. \cite{van2017pseudoinverse}).
Network clustering aims to organize nodes in the graph $G$ in such a way that
the majority of links in the graph connect nodes of the same cluster, while a
minority of links is shared by nodes from different clusters. This general
description is imprecise and part of the complication, which causes that
community detection in a network is a challenging problem, that has been
actively researched over the last few decades \cite{fortunato2010community}.
Before introducing our linear clustering process (LCP) in Section \ref{Sec_dynamic_clustering}, we briefly present basic graph partitioning concepts, while the overview of the more popular clustering methods is deferred to Appendix \ref{App_Clustering_Algorithms}.
\subsection{Network modularity}\label{Sec_2_1}
Newman and Girvan \cite{newman2004finding} proposed the modularity as a concept for a network partitioning,
\begin{equation}\label{Eq_modularity}
m = \frac{1}{2L}\cdot \sum\limits_{i=1}^{N} \sum\limits_{j=1}^{N} \left(a_{ij}-\frac{d_{i}\cdot d_{j}}{2L}\right) \cdot \mathbf{1}_{\{i\, \text{and}\, j \,\in
\text{ same cluster}\}},
\end{equation}
where $\mathbf{1}_{x}$ is the indicator function that equals 1 if statement $x$ is true, otherwise $\mathbf{1}_{x}=0$. The modularity $m$ compares the number of links between nodes from the same community with the expected number of intra-community links in a network with randomly connected nodes. When the modularity $m$ close to $0$, the estimated partition is as good as a random partition would be. On the contrary, a modularity $m$ close to $1$ indicates that the network can be clearly partitioned into clusters. Defining the $N\times N$ modularity matrix $C$,
\begin{equation}\label{Modularity_matrix_C}
C_{ij}= \begin{cases}
1 & \text{if nodes $i$ and $j$ belong to the same cluster}\\
0 & \text{otherwise},
\end{cases}
\end{equation}
allows us to rewrite the modularity (\ref{Eq_modularity}) as a quadratic form,
\begin{equation}\label{Eq_modularity_m_using_C}
m = \frac{1}{2L}\cdot u^{T}\cdot \left(A\circ C - \frac{1}{2L}\cdot \left(d\cdot d^{T}\right)\circ C\right)\cdot u,
\end{equation}
where $\circ$ denotes the Hadamard product \cite{horn2012matrix}. The number of clusters in a network is denoted by $c$, while the $c\times 1$ vector $n$ defines the size of each cluster:
\begin{equation}\label{Eq:Cluster_Size}
n = \begin{bmatrix}
n_{1} & n_{2} & \dots & n_{c}
\end{bmatrix},
\end{equation}
where the number of nodes in cluster $i$ is denoted as $n_{i}$.
\subsection{Stochastic block model}\label{SubSec:SBM}
The performance of the clustering methods in this paper are benchmarked on random graphs, generated by the Stochastic Block Model (SBM), proposed by Holland \cite{holland1983stochastic}. The SBM model generates a random graph with community structure, where a link between two nodes exists with different probability, depending on whether the nodes belong to the same cluster or not.
In this paper, we focus on the symmetric stochastic block model (SSBM), where only two different link probabilities are defined. Two nodes are connected via a link with probability $p_{in}$ if they belong to the same cluster, otherwise the direct link exists with probability $p_{out}$. Communities emerge when the link density within clusters is larger than the inter-community link probability $p_{in}>p_{out}$. Furthermore, we restrict clusters to be of the same size: \[n_{i} = \frac{N}{c}, i\in \{1,\,2,\,\dots \,,c\},\]
which causes the expected degree to be the same for each node:
\begin{equation}\label{Eq:Expected_Degree}
E[D] = \frac{b_{in}+\left(c-1\right)\cdot b_{out}}{c},
\end{equation}
irrespective of its cluster membership. We further consider a sparse and assortative SSBM. The SBMM is called sparse and assortative when the link probabilities $p_{in}=\frac{b_{in}}{N}$ and $p_{out}=\frac{b_{out}}{N}$ are defined upon positive constants $b_{in}>b_{out}$ that stay constant when $N \rightarrow \infty$. Decelle \textit{et al.} \cite{decelle2011inference,decelle2011asymptotic} found that when the difference $b_{in}-b_{out}$ is above the detectability threshold
\begin{equation}\label{Eq:SBM_treshold}
b_{in} - b_{out} > c\cdot \sqrt{E[D]},
\end{equation}
it is theoretically possible to recover cluster membership of the nodes, otherwise, the community structure of a network is not distinguishable from randomness. The threshold (\ref{Eq:SBM_treshold}) marks a phase transition between the undetectable and the theoretically detectable regime of the SSBM.
\section{Linear clustering process (LCP) on a graph}
\label{Sec_dynamic_clustering}
\subsection{Concept of the clustering process}
Each node $i$ in the graph $G$ is assigned a position $x_{i}[k]$ on a line
(i.e. in one-dimensional space) at discrete time $k$. We define the $N\times1$
position vector $x[k]$ at discrete time $k$, where the $i$-th vector component
consists of the position $x_{i}[k]$ of node $i$ at time $k$. We initialize the
$N\times1$ position vector $x[0]$ by placing nodes equidistantly on the line
and assign integer values from $1$ to $N$ to the nodes, thus, $x[0]=%
\begin{bmatrix}
1 & 2 & \dots & N
\end{bmatrix}
^{T}$. At last, we restrict the position $x_{i}[k]$ to positive real values.
We propose a dynamic process that determines the position of nodes over time.
The position difference between nodes of the same cluster is relatively small.
On the contrary, nodes from different clusters are relatively far away, i.e.
their position difference is relatively high. Based on the position vector
$x[k]$, we will distinguish clusters, also called communities, in the graph
$G$.
The proposed clustering process consists of two opposite and simultaneous
forces that change the position of nodes at discrete time $k$:
\begin{description}
\item[Attraction] Adjacent nodes sharing many neighbours are
mutually attracted with a force proportional to the number of common
neighbours. In particular, the attractive force between node $i$ and its
neighboring node $j$ is proportional to $\alpha\cdot\left( \left\vert
\mathcal{N}_{j}\cap\mathcal{N}_{i}\right\vert +1\right) $, where $\alpha$ is
the attraction strength and $\left( \left\vert \mathcal{N}_{j}\cap
\mathcal{N}_{i}\right\vert +1\right) $ equals the number of common neighbors
plus the direct link, i.e. $a_{ij}=1$.
\item[Repulsion] Adjacent nodes sharing a few neighbours are
repulsed with a force proportional to the number of different neighbours. The
repulsive force between node $i$ and its neighboring node $j$ is proportional
to $\delta\cdot\left( \left\vert \mathcal{N}_{j}\setminus\mathcal{N}%
_{i}\right\vert -1\right) $, where $\delta$ is the repulsive strength and
$\left( \left\vert \mathcal{N}_{j}\setminus\mathcal{N}_{i}\right\vert
-1\right) $ equals the set of neighbours of node $j$ that do not belong to
node $i$ minus the direct link (that is included in $\left\vert \mathcal{N}%
_{j}\setminus\mathcal{N}_{i}\right\vert )$. Since the force should be
symmetric and the same if $i$ and $j$ are interchanged, we end up with a
resultant repulsive force proportional to $\frac{1}{2}\cdot\delta\cdot\left(
\left\vert \mathcal{N}_{j}\setminus\mathcal{N}_{i}\right\vert +\left\vert
\mathcal{N}_{i}\setminus\mathcal{N}_{j}\right\vert -2\right)$.
\end{description}
\subsection{LCP in discrete time}
Since computers operate with integers and truncated real numbers, we concentrate
on discrete-time modeling. The continuous-time description is derived in Appendix \ref{Sec_Continuous_time}. We denote the continuous-time
variables by $y\left( t\right) $ and the continuous time by $t$, while the
discrete-time counterpart is denoted by $y\left[ k\right] $, where the
integer $k$ denotes the discrete time or $k$-th timeslot. The transition from
the continuous-time derivative to the discrete-time difference is
\begin{eqnarray*}
\frac{dx_{i}(t)}{dt} = \lim_{\Delta t\rightarrow 0}\frac{x_{i}(t+\Delta t)-x_{i}(t)}{\Delta t}
\rightarrow &
\left.\frac{x_{i}(t+\Delta t)-x_{i}(t)}{\Delta t}\right\vert_{\Delta t=1} \overset{\text{def}}{=} x_{i}\left[k+1\right] -x_{i}\left[k\right]
\end{eqnarray*}
Corresponding to the continuous-time law in Appendix \ref{law_continuous_time} and
choosing the time step $\Delta t=1$, the governing equation of position
$x_{i}[k]$ of node $i$ at discrete time $k$ is
\begin{equation}\label{clustering_law_discrete_time}%
x_{i}[k+1]= x_{i}[k]+\sum\limits_{j\in\mathcal{N}_{i}} \bigg(\frac{\alpha\cdot\left(\left\vert \mathcal{N}_{j}\cap\mathcal{N}_{i}\right\vert +1\right)}{d_{j}d_{i}}- \frac{\frac{1}{2}\cdot\delta\cdot\left( \left\vert \mathcal{N}_{j}\setminus \mathcal{N}_{i}\right\vert +\left\vert \mathcal{N}_{i}\setminus\mathcal{N}_{j}\right\vert -2\right)}{d_{j}d_{i}} \bigg)\cdot\Big(x_{j}[k]-x_{i}[k]\Big)
\end{equation}
where $\alpha$ and $\delta$ are, in the discrete-time setting, the strength
(in dimensionless units) for attraction and repulsion, respectively. The
maximum position difference at the initial state is $x_{N}[0]-x_{1}[0]=N-1$.
Node $j$ attracts an adjacent node $i$ with force proportional to their
position difference ($x_{j}[k]-x_{i}[k]$). The intensity of the attractive
force decreases as nodes $i$ and $j$ are closer on a line. The attraction is
also proportional to the number common neighbours $\left\vert \mathcal{N}%
_{j}\cap\mathcal{N}_{i}\right\vert $ of node $i$ and node $j$ plus the direct
link, as nodes tend to share most links with other nodes from the same
cluster. On the contrary, node $j$ repulses node $i$ with a rate proportional
to their position difference ($x_{j}[k]-x_{i}[k]$) and the average of the
number of node $j$ neighbours $\left\vert \mathcal{N}_{j}\setminus
\mathcal{N}_{i}\right\vert $ that are not connected to the node $i$ and,
similarly, the number of node $i$ neighbors, $\left\vert \mathcal{N}%
_{j}\setminus\mathcal{N}_{i}\right\vert $ that are not connected to the node
$j$. The repulsive and attractive force are, as mentioned above, symmetric in
strength, but opposite, if $i$ is interchanged by $j$.
\begin{figure*}[!]
\begin{center}
\includegraphics[angle=0, scale=0.7]{Forces_v1.pdf}
\caption{Dependence of the attractive and repulsive force on the number of common neighbours of adjacent nodes $i$ and $j$ (left-figure). Directions of the attraction and repulsion forces between the adjacent nodes (middle-figure). Dependence of the attractive and repulsive force on the absolute position distance between adjacent nodes $i$ and $j$ (right-figure).}
\label{Fig_Forces_1}
\end{center}
\end{figure*}
The directions of both attractive and repulsive forces between two adjacent
nodes $i$ and $j$ as well the dependence of both forces on the number of
common neighbours $\left| \mathcal{N}_{j}\cap\mathcal{N}_{i}\right| $ and
the absolute position distance $\left| x_{j}[k]-x_{i}[k] \right| $ are
illustrated in Fig \ref{Fig_Forces_1}.
In the continuous-time setting, as provided in Appendix \ref{law_continuous_time}, we eliminate one parameter by scaling the time $t^{\ast}=\delta t$. Because the time step
$\Delta t=1$ is fixed and cannot be scaled, the discrete-time model
consists of two parameters $\alpha\geq0$ and $\delta\geq0$.
So far, we have presented an additive law, derived in the common Newtonian
approach. The corresponding multiplicative law in discrete time is%
\begin{equation}\label{clustering_law_multiplicative}
\resizebox{0.9 \textwidth}{!}{$
x_{i}[k+1] = x_{i}[k]\cdot \Bigg( 1+ \sum\limits_{j\in\mathcal{N}_{i}} \bigg(\frac{\alpha\cdot\left( \left| \mathcal{N}_{j} \cap\mathcal{N}_{i} \right| + 1\right)}{d_{i}\cdot d_{j}} - \frac{\frac{1}{2} \cdot\delta\cdot\left( \left| \mathcal{N}_{j} \setminus\mathcal{N}_{i} \right| + \left| \mathcal{N}_{i} \setminus\mathcal{N}_{j} \right| -2 \right)}{d_{i}\cdot d_{j}} \bigg)\cdot \Big(x_{j}[k] - x_{i}[k]\Big) \Bigg)$}
\end{equation}
Although the physical intuition is similar, the multiplicative process in
(\ref{clustering_law_multiplicative}) behaves different in discrete time than
the additive law in (\ref{clustering_law_discrete_time}). Since also the
analysis is more complicated, we omit a further study of the multiplicative law.
We present the analogon of (\ref{clustering_law_discrete_time}) in a matrix form:
\begin{theorem}
\label{clustering_law_additive_matrix_form} The discrete time process
(\ref{clustering_law_discrete_time}) satisfies the linear matrix difference
equation
\begin{equation}
x[k+1]=\left( I+W-\text{diag}\left( W\cdot u\right) \right) \cdot
x[k],\label{clustering_law_matrix_equation}%
\end{equation}
where the $N\times1$ vector $u$ is composed of ones, the $N\times N$ identity matrix is denoted by $I$, while the $N\times N$ topology-based matrix $W$ is defined as
\begin{equation}\label{def_martrix_W}
W = \left( \alpha+\delta\right) \Delta^{-1}\cdot\left( A\circ A^{2}+A\right) \cdot \Delta^{-1}- \frac{1}{2}\cdot\delta\left( \Delta^{-1}\cdot A+A\cdot \Delta^{-1}\right)
\end{equation}
where $\circ$ denotes the Hadamard product. In particular,
\begin{equation} \label{def_matrix_W_element}
w_{ij}=a_{ij}\frac{\alpha \left( \left\vert \mathcal{N}_{j} \cap\mathcal{N}_{i}\right\vert +1\right) - \delta \left( \frac{\left\vert \mathcal{N}_{j}\setminus\mathcal{N}_{i}\right\vert +\left\vert \mathcal{N}_{i}\setminus\mathcal{N}_{j}\right\vert }{2}-1\right) }{d_{i}d_{j}}
\end{equation}
The explicit solution of the difference equation
(\ref{clustering_law_matrix_equation}) is
\begin{equation}
x[k]=\left( I+W-\text{diag}\left( W\cdot u\right) \right) ^{k}%
x[0]\label{solution_positionvector}%
\end{equation}
where the $k$-th component of the initial position vector is $(x[0])_{k}=k$.
\end{theorem}
\textit{Proof:} Appendix \ref{App_A}.
Theorem \ref{clustering_law_additive_matrix_form} determines the position of the nodal vector $x[k]$ at time $k$ and shows convergence towards a state, where the sum of attractive and repulsive forces (i.e. the resulting force) acting on a node are in balance. Nodes with similar neighbourhoods are grouped on the line, i.e. in the one-dimensional space, while nodes with a relatively small number of common neighbours are relatively far away. A possible variant
of the proposed linear clustering process may map the nodal position into a
higher dimensional space, like a circular disk or square in two dimensions,
and even with a non-Euclidean distance metric.
\subsection{Time-dependence of the linear clustering process}
\label{model_time_dependence_subsection}
The $N \times N$ matrix $I+W - \text{diag}\left( W \cdot u \right) $ in the governing equation (\ref{clustering_law_matrix_equation}) has interesting properties. As shown in this section, the related matrix $W - \text{diag}\left( W \cdot u \right) $ belongs to the class of $M$-matrices, whose eigenvalues have a non-negative real part. The (weighted) Laplacian is another element of the $M$-matrix class.
\begin{property}
\label{prop_M_non_negative}
The matrix $I+W - \text{diag}\left( W \cdot u \right) $ is a
non-negative matrix.
\end{property}
\textit{Proof:} The governing equation (\ref{clustering_law_matrix_equation})
\[
x[k+1] = \left( I + W - \text{diag}\left( W \cdot u \right) \right) \cdot
x[k]
\]
holds for any non-negative vector $x[k]$. Let $x[0] = e_{m}$, the basic vector
with components $\left( e_{m}\right) _{i} = \delta_{mi}$ and $\delta_{mi}$
is the Kronecker delta, then we find that the $m$-th column
\[
x[1] = \left( I+W - \text{diag}\left( W \cdot u \right) \right) _{col(m)}%
\]
must be a non-negative vector. Since we can choose $m$ arbitrary, we have
established that $I+W - \text{diag}\left( W \cdot u \right) $ is a
non-negative matrix. $\hfill\square$
\begin{property}
\label{prop_M_eigenvalues}
The principal eigenvector of the matrix $I+W - \text{diag}\left( W \cdot u \right) $ is the all-one vector $u$ belonging to eigenvalue 1. All other eigenvalues of matrix $I+W - \text{diag}\left( W \cdot u \right) $ are real and, in absolute value, smaller than 1.
\end{property}
\textit{Proof:} Appendix \ref{sec_prop_M_eigenvalues}.
The linear discrete-time system in
(\ref{clustering_law_matrix_equation}) converges to a steady-state, provided
that $\lim_{k\rightarrow\infty}||x[k+1]|| = \lim_{k\rightarrow\infty
}||x[k]||=||x_{s}||$, which is only possible if the matrix $\left( I + W -
\text{diag}\left( W \cdot u \right) \right) $ has all eigenvalues in
absolute value smaller than 1 and the largest eigenvalue is precisely equal to
1. Property \ref{prop_M_eigenvalues} confirms convergence and indicates that the steady-state vector $x_{s} = u$ in which the position of each node is the same. However, the steady state solution
$x_{s}=u$ is a trivial solution, as observed from the governing equation in
(\ref{clustering_law_discrete_time}), because the sum vanishes and the
definition of the steady state tells that $x[k+1]=x[k]$, which is obeyed by
any discrete-time independent vector. In other words, the matrix equation
(\ref{clustering_law_matrix_equation}) can be written as
\[
x[k+1]-x[k]=\left( W-\text{diag}\left( W\cdot u\right) \right)
\cdot\left( x[k]-u\right)
\]
which illustrates that, if $x[k]$ obeys the solution, then $r[k]=x[k]+s\cdot
u$ for any complex number $s$ is a solution, implying that a shift in the
coordinate system of the positions does not alter the physics.
Let us denote the eigenvector $y_{k}$ belonging to the $k$-th
eigenvalue $\beta_{k}$ of the matrix $W-\text{diag}\left( W\cdot u\right) $,
where $\beta_{1}\geq\beta_{2}\geq\cdots\geq\beta_{N}$, then the eigenvalue
decomposition of the real, symmetric matrix is
\[
W-\text{diag}\left( W\cdot u\right) =Y\text{diag}(\beta)Y^{T}%
\]
where the eigenvalue vector $\beta=(\beta_{1},\beta_{2},\cdots,\beta_{N})$ and
$Y$ is the $N\times N$ orthogonal matrix with the eigenvectors $y_{1}%
,y_{2},\cdots,y_{N}$ in the columns obeying $Y^{T}Y=YY^{T}=I$. Since
$\beta_{1}=0$ and $y_{1}=\frac{u}{\sqrt{N}}$, it holds for $k>1$ that
$u^{T}y_{k}=0$, which implies that the sum of the components of eigenvector
$y_{k}$ for $k>1$ is zero (just as for any weighted Laplacian \cite{van2017pseudoinverse}). The position vector in (\ref{solution_positionvector}) is
rewritten as
\[
x[k]=Y\text{diag}(1+\beta)^{k}Y^{T}x[0]=\sum_{j=1}^{N}(1+\beta_{j})^{k}%
y_{j}\left( y_{j}^{T}x[0]\right)
\]
Hence, we arrive at
\begin{equation}
x[k]-\frac{u^{T}x[0]}{\sqrt{N}}u=\sum_{j=2}^{N}(1+\beta_{j})^{k}(y_{j}%
^{T}x[0])\;y_{j}\label{solution_positionvector_in_terms_eigenvalues}%
\end{equation}
As explained above, the left-hand side is a translated position vector and
physically not decisive for the clustering process. Since $-1<\beta_{j}<0$ for
$j>1$, relation (\ref{solution_positionvector_in_terms_eigenvalues}) indicates
that, for $k\rightarrow\infty$, the right-hand side tends to zero and the
steady-state solution is clearly uninteresting for the clustering process. We
rewrite (\ref{solution_positionvector_in_terms_eigenvalues}) as
\[x[k]-\frac{u^{T}x[0]}{\sqrt{N}}u = (1+\beta_{2})^{k}\bigg( (y_{2}^{T}x[0])\;y_{2}+ \sum_{j=3}^{N}\left( \frac{1+\beta_{j}}{1+\beta_{2}}\right)^{k}(y_{j}^{T}x[0])\;y_{j}\bigg).
\]
Since $|1+\beta_{2}|>|1+\beta_{3}|$, we observe that
\begin{equation}\label{position_vector_versus_y_2}
\frac{x[k]-\frac{u^{T}x[0]}{\sqrt{N}}u}{(1+\beta_{2})^{k}\;(y_{2}^{T} x[0])}=y_{2}+O\left( \frac{1+\beta_{3}}{1+\beta_{2}}\right) ^{k},
\end{equation}
which tells us that the left-hand side, which is a normalized or scaled,
shifted position vector, tends to the second eigenvector $y_{2}$ with an error
that exponentially decreases in $k$. Hence, for large enough $k$, but not too
large $k$, the scaled shifted position vector provides us the information on
which we will cluster the graph.
The steady state in Property \ref{prop_M_eigenvalues} can be regarded as a reference position of the nodes and does not affect the LCP process nor the $N\times 1$ eigenvector $y_2$, belonging to the second largest eigenvalue $(1 + \beta_2)$ of the $N\times N$ “operator” matrix $I + W - \text{diag}(W\cdot u)$, which is analogous to Fiedler clustering based on the $N\times N$ Laplacian $Q$. While the Laplacian matrix $Q$ essentially describes diffusion and not clustering, our operator $I + W - \text{diag}(W\cdot u)$ changes the nodal positions, based on attraction and repulsion, from which clustering naturally arises.
\subsection{Influence of $\alpha$ and $\delta$ on the eigenvalues $\beta_{k}$ and the eigenvector $y_2$}
\begin{property}
\label{prop_alpha_delta}
The two parameters in the matrix $W$ in (\ref{def_martrix_W}) satisfy the bounds \begin{eqnarray}
0\leq\alpha &\leq&\frac{d_{\max}-1}{d_{\max}-\frac{1}{2}\left( 1+\frac{d_{\min}%
}{d_{\max}}\right) }\leq1 \label{bounds_alpha}\\
0\leq\delta& \leq& \frac{1}{d_{\max}-\frac{1}{2}\left( 1+\frac{d_{\min}}{d_{\max
}}\right) } \label{bounds_delta}
\end{eqnarray}
\end{property}
\textit{Proof:} Appendix \ref{sec_prop_alpha_delta}.
\begin{figure}[!h]
\begin{center}
\includegraphics[angle=0, scale=0.7]{Eigenvalue_lambda_2_vs_Alpha_and_Delta.pdf}
\caption{Gap $\beta_2 - \beta_3$ between the second and the third largest eigenvalue of the $N\times N$ matrix $W-\text{diag}\left(W\cdot u\right)$, for different values of the attractive $\alpha$ and repulsive $\delta$ strength (left figure). Relative difference $\frac{\beta_3 - \beta_2}{\beta_3}$, for different values of the attractive $\alpha$ and repulsive $\delta$ strength (right figure). An SSBM network with $N=1000$ nodes, $c=5$ clusters is used for both plots, with $b_{in} = 25$ and $b_{out}=2.5$.}
\label{Fig_influence_eigenvaluegap}
\end{center}
\end{figure}
Figure \ref{Fig_influence_eigenvaluegap} shows that influence of the attractive and repulsive strength $\alpha$ and $\delta$ on the eigenvalue gap $\beta_2-\beta_3$ is relatively small if $\alpha$ and $\delta$ are not too small and obeying the bounds (\ref{bounds_alpha}) and (\ref{bounds_delta}). While the difference increases when the attraction strength $\alpha$ is increasing, the repulsive strength $\delta$ has no visible influence on the eigenvalue gap.
The eigenvalue $\beta_{2}$ depends on the community structure of a graph. Figure \ref{Fig_beta_2_versus_modulairty} reveals positive correlation between the eigenvalue $\beta_{2}$ and the modularity index $m$ of a graph. As the modularity index increases, the eigenvalue $\beta_{2}$ approaches value $1$. In the limit case, when there are only intra-community links in the network, $\beta_{2}=1$, indicating the eigenvector $y_{2}$ represents a steady state.
\begin{figure}[!h]
\begin{center}
\includegraphics[angle=0, scale=0.75]{Beta_2_versus_modularity_v1.pdf}
\caption{The eigenvalue $\beta_{2}$ versus the modularity index $m$ of an SSBM graph of $N=999$ nodes and $c=3$ clusters, and an SSBM graph of $N=1000$ nodes, with $c=4,10$ clusters, respectively. The parameters $b_{in}$ and $b_{out}$ are varied, while keeping average degree $d_{av}=7$ fixed. For each combination of $b_{in}$ and $b_{out}$, the modularity index $m$ and the eigenvalue $\beta_{2}$ are computed. The correlation is presented in case only interactions between direct neighbours are allowed (left figure) and in case interactions between each pair of nodes are allowed (right figure).}
\label{Fig_beta_2_versus_modulairty}
\end{center}
\end{figure}
Figure \ref{Fig_influence_eigenvectory2} reveals that the repulsive strength $\delta$ does not affect the eigenvector $y_2$ components significantly. Eigenvector $y_{2}$ components of nodes from the same cluster are better distinguished from the remaining components of $y_{2}$ for smaller values of repulsive strength $\delta$.
\begin{figure}[!h]
\begin{center}
\includegraphics[angle=0, scale=0.6]{Eigenvector_y_2_vs_Delta_v2.pdf}
\caption{ Sorted eigenvector $\hat{y_{2}}$ components for different values of the repulsive strength $\delta$, in case of a SSBM network of $N=100$ nodes, $c=4$ clusters and with parameters $b_{in} = 25$ and $b_{out} = 1$. The attraction rate equals $\alpha = 0.05$ (left figure) and $\alpha = 0.5$ (right figure), while the repulsive strength $\delta$ obeys bounds in (\ref{bounds_delta}).}
\label{Fig_influence_eigenvectory2}
\end{center}
\end{figure}
\section{From the eigenvector $y_2$ to clusters in the network}\label{Sec:Position_Vector_Analysis}
The interplay of the attractive and repulsive force between nodes drives the nodal position in discrete time $k$ eventually towards a steady state $\lim\limits_{k\to \infty}x[k]=u$. However, the scaled and shifted position vector $x[k]$ in (\ref{position_vector_versus_y_2}) converges in time towards the second eigenvector $y_{2}$ with an exponentially decreasing error. In this section, we estimate the clusters in network, based on the eigenvector $y_{2}$.
By sorting the eigenvector $y_{2}$ to $\hat{y_2}$, the components of $y_{2}$ are reordered and the corresponding relabeling of the nodes of the network reveals a block diagonal structure of the adjacency matrix $A$. We define the $N\times N$ permutation matrix $R$ in a way the following equalities hold:
\begin{equation}\label{Eq_Permuation_Matrix_R}
\begin{aligned}
\hat{y_2} & = R\cdot y_2, \\
\hat{(y_2)_i} = \left(y_{2}\right)_{r_{i}} &\leq \hat{(y_2)_j} = \left(y_{2}\right)_{r_{j}}, && i<j,
\end{aligned}
\end{equation}
where the $N\times 1$ ranking vector $r = R\cdot w$ and $w = [1,2,\dots , N]$, with $r_{i}$ denoting the node $i$ ranking in the eigenvector $y_2$. The permutation matrix $R$ allow us to define the $N\times N$ relabeled adjacency matrix $\hat{A}$, the $N\times 1$ relabeled degree vector $\hat{d}$ of $G$, and the $N\times 1$ sorted eigenvector $\hat{y_{2}}$ as follows:
\begin{equation}\label{Eq_Relabeled_A_and_d}
\begin{cases}
\hat{A} &= R^{T}\cdot A \cdot R \\
\hat{d} &= R \cdot d \\
\hat{y_{2}} &= R \cdot y_2.
\end{cases}
\end{equation}
\begin{figure*}[!]
\begin{center}
\includegraphics[angle=0, scale=0.7]{Steady_State_Example_v3.pdf}
\caption{Adjacency matrix $A$ of an SSBM network of $N=1000$ nodes, $c=4$ clusters and parameters $b_{in}=26$, $b_{out}=0.67$ (top-left). Eigenvector $y_{2}$ components (top-right). Sorted eigenvector $\hat{y_{2}}$ components (bottom-right). Relabeled adjacency matrix $\hat{A}$ based on the sorted eigenvector $\hat{y_{2}}$ (bottom-left).}
\label{Fig_Model_Procedure_1}
\end{center}
\end{figure*}
Groups of nodes that have relatively small difference in the eigenvector $y_{2}$ components, while relatively large difference compared to other nodes in the network, compose a cluster. Therefore, the community detection problem transforms into recognizing intervals of similar values in the sorted eigenvector $\hat{y_{2}}$.
Figure \ref{Fig_Model_Procedure_1} exemplifies the idea, where the adjacency matrix $A$ of a randomly labeled SSBM network of $N=1000$ nodes and $c=4$ clusters is presented in the upper-left part, as a heat map. The eigenvector $y_{2}$ is drawn in the upper-right part, while the sorted eigenvector $\hat{y_{2}}$ is drawn on the bottom-right side. Finally, the relabeled adjacency matrix $\hat{A}$, based on nodal ranking of $y_{2}$ is depicted on the lower-left side. The sorted eigenvector $\hat{y_{2}}$ reveals a stair with four segments, equivalent to four block matrices on the main diagonal in relabeled adjacency matrix $\hat{A}$.
The eigenvector $y_{2}$ represents a continuous measure of how similar neighbours of two nodes are. There are two different approaches to identify network communities for a given eigenvector $y_{2}$:
\begin{itemize}
\item Cluster identification based on the sorted eigenvector $\hat{y_{2}}$. This approach is explained in subsection \ref{K_means_cluster_estimation}.
\item Cluster identification based on the ranking vector $r$. This approach does not rely on the eigenvector $y_{2}$ components, but solely on nodal ranking, as explained in subsection \ref{Recursive_algorithm_sec}.
\end{itemize}
\subsection{Community detection based on nodal components of the eigenvector $y_{2}$}\label{K_means_cluster_estimation}
To identify clusters, we observe the difference in eigenvector $y_{2}$ components between nodes with adjacent ranking. If $\left(\hat{y_{2}}\right)_{i+1}-\left(\hat{y_{2}}\right)_{i} < \theta$, where $\theta$ denotes a predefined threshold, then the nodes $r_{i}$ and $r_{i+1}$ belong to the same cluster, else the nodes $r_{i}$ and $r_{i+1}$ are boundaries of two adjacent clusters. The resulting cluster membership function is
\begin{equation}\label{Eq:Modularity_matrix_C:position}
C_{r_{i+1},r_{i}}= \begin{cases}
1 & \text{$\left(\hat{y_{2}}\right)_{i+1}-\left(\hat{y_{2}}\right)_{i} < \theta$}\\
0 & \text{otherwise},
\end{cases}
\end{equation}
where the threshold value $\theta$ is determined heuristically, as provided in section \ref{Sec_Results}. The cluster estimation in (\ref{Eq:Modularity_matrix_C:position}) can be improved by using other more advanced approaches, such as the K-means algorithm.
\subsection{Modularity-based community detection}\label{Recursive_algorithm_sec}
By implementing (\ref{Modularity_matrix_C}) and (\ref{Eq_Relabeled_A_and_d}) into (\ref{Eq_modularity}) we obtain:
\begin{equation}\label{Eq_modularity_2}
m = \frac{1}{2L}\cdot u^{T}\cdot \left(\hat{A}\circ \hat{C} - \frac{1}{2L}\cdot \left(\hat{d}\cdot \hat{d}^{T}\right)\circ \hat{C}\right)\cdot u,
\end{equation}
where $\hat{C} = R^{T}\cdot C \cdot R$. As shown in Figure \ref{Fig_Model_Procedure_1}, the network relabeling based on the ranking vector $r$ reveals block diagonal structure in $\hat{A}$. Thus, the relabeled modularity matrix $\hat{C}$ has the following block diagonal structure:
\begin{equation}\label{Eq_Modularity_Matrix_Hat_C}
\hat{C}=\begin{bmatrix}
J_{n_{1}\times n_{1}} & O_{n_{1}\times n_{2}} & \dots & O_{n_{1}\times n_{c}} \\
O_{n_{2}\times n_{1}} & J_{n_{2}\times n_{2}} & \dots & O_{n_{1}\times n_{c}} \\
\vdots & \vdots & \dots & \vdots \\
O_{n_{1}\times n_{1}} & O_{n_{c}\times n_{2}} & \dots & J_{n_{c}\times n_{c}} \\
\end{bmatrix},
\end{equation}
where $c$ denotes number of clusters in network, where the $i$-th cluster is composed of $n_{i}$ nodes. We define the $N\times 1$ vectors $\hat{e_{i}}$ for $i=\{1,2,\dots ,c\}$ as
\begin{equation}\label{Eq_vectors_e_i}
\hat{e_{i}}=\begin{bmatrix}
O_{\left(1 \times \sum_{j=1}^{i-1}n_{j}\right)} &
u_{\left(1 \times n_{i}\right)} & O_{\left(1\times \sum_{j=i+1}^{N}n_{j}\right)} \end{bmatrix}^{T},
\end{equation}
that allows us to redefine $\hat{C} = \sum_{i=1}^{c}\hat{e_{i}}\cdot \hat{e_{i}}^{T}$ and further simplify (\ref{Eq_modularity_2}):
\begin{equation}\label{Eq_Modulaity_m_with_E_vectors}
m = \frac{1}{2L}\cdot \sum\limits_{i=1}^{c}\hat{e_{i}}^{T}\cdot \left(\hat{A}-\frac{1}{2L}\cdot \left(\hat{d}\cdot \hat{d}^{T}\right)\right)\cdot \hat{e_{i}}.
\end{equation}
Since the vector $\hat{e_{i}}$ consists of zeros and ones, the equation (\ref{Eq_Modulaity_m_with_E_vectors}) represents the sum of elements of the matrix $\left(\hat{A}-\frac{1}{2L}\cdot \left(\hat{d}\cdot \hat{d}^{T}\right)\right)$ corresponding to each individual cluster.
\begin{algorithm}[H]
\caption{Recursive algorithm for cluster estimation
\label{alg:cluster-estimation}}
\begin{algorithmic}[1]
\Require{$\hat{A}$ and $\hat{d}$ are the relabeled adjacency matrix $A$ and the degree vector $d$ (\ref{Eq_Relabeled_A_and_d}), while $L$ denotes number of links. The modularity threshold is denoted by $\theta$. The function returns the $c\times 1$ vector $b$, whose elements are cluster borders in a relabeled graph.}
\Statex
\Function{EstimateClusters}{$\hat{A},\, \hat{d},\, N,\, L,\, \theta$}
\State $d_{f},d_{b},p,q \gets O_{N\times 1}$
\State $\left(d_{f}\right)_{1} \gets \hat{d}_{1}$
\State $\left(d_{b}\right)_{N} \gets \hat{d}_{N}$
\State $p_{1} \gets -\frac{\hat{d}_{1}^{2}}{(2L)^{2}}$
\State $q_{N} \gets -\frac{\hat{d}_{N}^{2}}{(2L)^{2}}$
\For{$i \gets 2 \textrm{ to } N$}
\State $l \gets N - i$
\State $\left(d_{f}\right)_{i} \gets \left(d_{f}\right)_{i-1} + \hat{d}_{i}$
\State $\left(d_{b}\right)_{N-i+1} \gets \left(d_{b}\right)_{N-i+2} + \hat{d}_{N-i+1}$
\State $ s \gets \frac{ \sum_{j=1}^{i}\hat{a}_{ij}}{L} - \frac{ 2\cdot \hat{d}_{i}\cdot (d_{f})_{i-1} + \hat{d}_{i}^{2}}{(2L)^{2}}$
\State $ t \gets \frac{\sum_{j=1}^{i}\hat{a}_{N-j+1,l+1}}{L} - \frac{2\cdot \hat{d}_{l+1}\cdot (d_{b})_{l+2} + \hat{d}_{l+1}^{2}}{(2L)^{2}} $
\State $ p_{i} \gets p_{i-1} + s$
\State $q_{l+1} \gets q_{l+2} + t$
\EndFor
\State $r \gets \operatorname*{arg\,max}_{\mathcal{N}} \left(p+q\right)$
\If{$(p+q)_{r} > \theta$}
\State $\hat{A_{1}},\, \hat{d_{1}},\,N_{1} \gets$ sub-matrix(vector) corresponding to the first cluster $\{1,2, \dots , r\}$
\State $\hat{A_{2}},\, \hat{d_{2}},\,N_{2} \gets$ sub-matrix(vector) corresponding to the second cluster $\{r+1,r+2, \dots , N\}$
\State \Return $\hat{b} \gets \begin{bmatrix}
\text{EstimateClusters(}\hat{A_{1}},\hat{d_{1}},N_{1},L,p_{r}\text{)} \\
r, \\
\text{EstimateClusters(}\hat{A_{2}},\hat{d_{2}},N_{2},L,q_{r}\text{)}
\end{bmatrix}$
\Else
\State \Return{$\hat{b} \gets\emptyset$}
\EndIf
\EndFunction
\end{algorithmic}
\end{algorithm}
We estimate clusters for a given ranking vector $r$ by optimising the modularity $m$ recursively.
In the first iteration, we examine all possible partitions of the network in two clusters and compute their modularity. The partition that generates the highest modularity is chosen.
In the second iteration, we repeat for each subgraph the same procedure and find the best partitions into two clusters. Once we determine the best partitions for both subgraphs, we adopt them if the obtained modularity of the generated partition exceeds the modularity of a parent cluster from the previous iteration. The recursive procedure stops when the modularity $m$ cannot be further improved, as described by pseudocode (\ref{alg:cluster-estimation}). This version of the proposed process is denoted as LCP in section \ref{Sec_Results}.
\subsection{Modularity-based community detection for a known number of communities}\label{Recursive_algorithm_given_N_sec}
The algorithm \ref{alg:cluster-estimation} also applies for graph partition with a known number of communities $c$. In that case, instead of stopping the recursive procedure described in algorithm \ref{alg:cluster-estimation} when the modularity $m$ cannot be further improved, we stop at iteration $(\log_{2}c+1)$. In each iteration, the partition with the maximum modularity is accepted, even if negative.
As a result, we obtain $2c$ estimated clusters with the $2c\times 2c$ aggregated modularity matrix $M_{c}$:
\begin{equation}\label{Eq_Modularity_Matrix_M_c}
\left(M_{c}\right)_{gh} = \sum\limits_{i\in g,j\in h}\left(\hat{A}-\frac{1}{2L}\cdot \hat{d}\cdot \hat{d}^{T}\right)_{ij},
\end{equation}
where $g,h \in \{1,2, \dots , 2c\}$ denote estimated communities. The aggregated modularity matrix $M_{c}$ allows us to merge adjacent clusters, until we reach $c$ communities in an iterative way. We observe the $(2c-1\times 1)$ vector $\mu$, where $\mu_{g} = \left(M_{c}\right)_{g,g+1}$. The maximum element of $\mu$ indicates which two adjacent clusters can be merged, so that modularity index $m$ is negatively affected the least. By repeating this procedure $c$ times, we end up with the graph partition in $c$ clusters. This version of the proposed process is denoted as LCP$_{n}$ in Section \ref{Sec_Results}.
\subsection{Non-back tracking method versus LCP}\label{Non_back_tracking_as_a_process}
Angel \textit{et al.} \cite[p.12]{angel2015non} noted that the $2N$ non-trivial eigenvalues of the $2L\times 2L$ non-back tracking matrix $B$ from (\ref{Eq:Non-back tracking B}) are contained in eigenvalues of the $2N\times 2N$ matrix $B^{*}$:
\begin{equation}\label{Eq:Non_back_tracking_2N}
B^{*} = \begin{bmatrix}
A & I - \Delta \\
I & O
\end{bmatrix},
\end{equation}
where the $N\times N$ matrix with all zeros is denoted as $O$. The $2N\times 2N$ matrix $B^{*}$, written as
\[B^{*} = \begin{bmatrix}
I + \left(A - \Delta\right) + \left(\Delta - I\right) & -\left(\Delta - I\right) \\
I & O
\end{bmatrix}\]
can be considered as a state-space matrix of a process on a network, similar to our LCP process we in (\ref{clustering_law_discrete_time}), with the last $N$ states storing delayed values of the first $N$ states. The $2N\times 2N$ matrix $B^{*}$ defines the set of $N$ second-order difference equations, where the governing equation for the node $i$ position is
\begin{eqnarray}
x_{i}[k+1] = x_{i}[k] + \sum\limits_{j\in \mathcal{N}_{i}}(x_{j}[k]-x_{i}[k])\nonumber+ \left(d_{i}-1\right)\cdot \left(x_{i}[k]-x_{i}[k-1]\right) \label{Eq_nbtm_process}
\end{eqnarray}
We recognize the second term in (\ref{Eq_nbtm_process}) as an attraction force between neighbouring nodes with uniform intensity, while in our LCP (\ref{clustering_law_discrete_time}) the attraction force intensity is proportional to the number of neighbours two adjacent nodes share. Further, while we propose a repulsive force between adjacent nodes in (\ref{clustering_law_discrete_time}), node $i$ in (\ref{Eq_nbtm_process}) is repulsed from its previous position $x_{i}[k]$ in direction of the last position change $(x_{i}[k]-x_{i}[k-1])$.
We implement the weighted intensity of the attractive force as in (\ref{clustering_law_discrete_time}), ignoring the repulsive force by letting $\delta = 0$, and define the $2N\times 2N$ matrix $W^{*}$, corresponding to $B^{*}$,
\begin{equation}\label{Eq_Non_back_tracking_Matix_W}
W^{*} = \begin{bmatrix}
I + \alpha\cdot \Big(A\circ A^{2} +A - \text{diag}\left(\left(A\circ A^{2} + A\right)\cdot u\right)\Big) + \left(\Delta - I\right) & -\left(\Delta - I\right) \\
I & O
\end{bmatrix}.
\end{equation}
We estimate the number of clusters $c$ in a network from $W^{*}$ similarly as in the non-back tracking method in Sec. \ref{Sec_non_back_tracking_matrix} by counting the number of eigenvalues in $W^{*}$ with real component larger than $\sqrt{\lambda_1(W^{*})}$. This approach is denoted as LCP$_{c}$ in Section \ref{Sec_Results}.
\section{Reducing intensity of forces between clusters}\label{Sec_Iterative_Computation}
The idea behind a group of methods in community detection, called divisive algorithms, consists of determining the links between nodes from different clusters. Once these links have been identified, they are removed and thus only the intra-community links remain \cite{girvan2002community}.
We invoke a similar idea to our linear clustering process.
An outstanding property of our approach is that the LCP defines the nodal position as a metric, allowing us to perform clustering in multiple ways. The position distance between any two, not necessarily adjacent nodes indicates how likely the two nodes belong to the same cluster. Then, the position metric also allows us to classify links as either intra- or inter-community.
Thus, we iterate the linear clustering process (\ref{clustering_law_discrete_time}) and, in each iteration, we identify and scale the weights of the inter-community links.
The attraction and repulsive forces are defined as linear functions of the position difference between two neighbouring nodes, as presented in Figure \ref{Fig_Forces_1}. While linear functions greatly simplify the complexity and enable a rigorous analysis, the linearity of forces introduces some difficulties in the process. Firstly, as two adjacent nodes are further away, both the attractive and the repulsive force between them increase in intensity. Similarly, as the neighbouring nodes are closer on a line, both forces decrease in intensity and converge to zero as the nodes converge to the same position. Secondly, the attractive force between any two neighbouring nodes is always of higher intensity than the repulsive force, causing the process to converge towards the trivial steady-state.
Non-linearity in the forces can be introduced in the proposed linear clustering process iteratively by scaling the weights of inter-community links between iterations, that artificially decreases the strength of forces between the two nodes from different clusters. In other words, we reduce the importance of links between nodes from different clusters, based on the partition from previous iteration.
\subsection{Scaling the weights of inter-community links}\label{removing_links_sub_section}
The difference $|\left(y_{2}\right)_{i} - \left(y_{2}\right)_{j}|$ in the eigenvector $y_{2}$ components of nodes $i$ and $j$ indicates how similar neighbourhoods of these nodes are. A normalized measure for the difference in neighbouring nodes $i$ and $j$ is the difference $\left(|r_{i} - r_{j}|\right)$ of their rankings in the sorted eigenvector $\hat{y_{2}}$. Thus, links that connect nodes with the highest ranking difference are most likely inter-community links.
We define the $N\times N$ scaling matrix $S$ as follows:
\begin{equation}\label{Ranking_Distance_eq}
s_{ij} =\begin{cases}
1, & \text{if } \lvert r_{j} - r_{i}\rvert < \theta_{r} \\
\gamma, & \text{otherwise },
\end{cases}
\end{equation}
where the $ij$-th element equals 1 if the absolute value of the ranking difference between nodes $i$ and $j$ is below a threshold $\theta_{r}$, otherwise some positive value $0\le \gamma \le 1$. Based on the $N\times N$ scaling matrix $S$ in (\ref{Ranking_Distance_eq}), we update the governing equation as follows:
\[
x[k+1] = \left( I + \tilde{W} - \text{diag}\left( \tilde{W} \cdot u \right) \right) \cdot
x[k],
\]
where $\tilde{W} = S\circ W$.
\begin{figure*}[!]
\begin{center}
\includegraphics[angle=0, scale=0.85]{Removing_Links_v4.pdf}
\caption{Adjacency matrix $A$ of an SSBM network of $N=1000$ nodes, $c=5$ clusters of equal size, with parameters $b_{in}=26$ and $b_{out}=2.25$ (top-left). Following 4 subfigures present the relabeled adjacency matrix based on the ranking vector $r$ in iterations $1,5,7$ and $12$, respectively. In each iteration, the weights of $2\%$ links are scaled (red colour). The weight of each link is allowed to be scaled once. The relabelled adjacency matrix $\hat{A}$ after $15$ iterations of scaling weights of links between clusters (bottom-right).}
\label{Removing_Links_Fig}
\end{center}
\end{figure*}
Scaling the link weights in (\ref{Ranking_Distance_eq}) only impacts the clustering process in (\ref{clustering_law_matrix_equation}), as defined in the equation above. However, modularity-based community detection, explained in Section \ref{Recursive_algorithm_sec}, operates on the $N\times N$ adjacency matrix $A$ in each iteration. Therefore, our implementation of scaling the weights of inter-community connections in network helps the process to better distinguish between clusters (i.e. eventually provides better relabeling in (\ref{Eq_Relabeled_A_and_d})), without modifying the $N\times N$ adjacency matrix $A$ and, hence, without negatively affecting the modularity $m$ optimisation in Algorithm \ref{alg:cluster-estimation}.
An example of removing links (i.e. $\gamma = 0$) is depicted on Figure \ref{Removing_Links_Fig}, where in each iteration weights of $\frac{15}{4}\%$ identified inter-cluster links are scaled. Scaling the weights of links between clusters significantly improves the quality of the identified graph partition.
\section{Benchmarking LCP with other clustering methods}\label{Sec_Results}
In this section, we benchmark the linear clustering process (\ref{clustering_law_discrete_time}) against popular clustering algorithms in Appendix \ref{App_Clustering_Algorithms} on random networks.
\subsection{Complexity of LCP}\label{SubSec:Complexity}
The computational complexity of LCP consists of three parts: the computation of (i) the $N\times N$ matrix $W$ in (\ref{def_martrix_W}), (ii) the $N\times 1$ eigenvector $y_{2}$ of the matrix $W-$diag$(Wu)$ and (iii) the identification of the clusters based on the sorted eigenvector $\hat{y_{2}}$.
\subsubsection{Computing the $N\times N$ matrix $W$}\label{SubSubSec:A^2}
The $N\times N$ matrix $A\circ A^{2}$ in (\ref{def_martrix_W}) requires the highest computational effort. Generally, computing the square of a matrix involves $O(N^{3})$ elementary operation, but the zero-one structure of the adjacency matrix significantly reduces the operations. We provide below an efficient algorithm for the computation of $A\circ A^{2}$, whose entries determine the number of $2$-hop walks between any two direct neighbours in the network.
\begin{algorithm}[H]
\caption{Computation of the $N\times N$ matrix $A\circ A^{2}$
\label{alg:adjacency-squared}.}
\begin{algorithmic}[1]
\Require{$A$ denotes the adjacency matrix, $N$ denotes number of links, while the set of node $i$ neighbours is denoted by $\mathcal{N}_{i}$.}
\State $A_{s} \gets O_{N\times N}$
\For{$i \gets 1 \textrm{ to } N$}
\For{$j \gets \mathcal{N}_{i}$}
\For{$m \gets \left(\mathcal{N}_{j}\setminus \{1,2,\dots ,i\}\right)\cap \mathcal{N}_{i}$}
\State $\left(A_{s}\right)_{i,m} \gets \left(A_{s}\right)_{i,m} + 1$ \Comment{Account for the 2-hop walk $(i\rightarrow j \rightarrow m)$}
\EndFor
\EndFor
\EndFor
\State $A_{s} \gets A_{s} + A_{s}^{T}$
\State \Return{$A_{s}$}
\end{algorithmic}
\end{algorithm}
We initialize the $N\times N$ matrix $A\circ A^{2}$ with zeros and only compute elements above the main diagonal, because $A\circ A^{2}$ is symmetric. The algorithm identifies all 2-hop walks between any two direct neighbours and accordingly updates the matrix. Let us consider a node $i$ with $d_{i}$ neighbours, denoted as $\mathcal{N}_{i}$. For a neighbouring node $ j\in \mathcal{N}_{i}$, we increment the elements $\left(A\circ A^{2}\right)_{im}$ by 1, where $m\in \left(\mathcal{N}_{j}\setminus \{1,2,\dots ,i\}\right)\cap \mathcal{N}_{i}$, accounting for 2-hop walks $i\rightarrow j \rightarrow m$. By repeating the procedure for each node, we compute all the elements above the main diagonal. Finally, we sum the generated matrix with its transpose to obtain $A\circ A^{2}$. Since the algorithm \ref{alg:adjacency-squared} is based on incrementing the matrix entries per each 2-hop walk between direct neighbours, the number of operations equals the sum $s = u^{T}\cdot \left(A\circ A^{2}\right)\cdot u$ of all elements of $A\circ A^{2}$
\begin{equation}\label{Eq:A_squared_complexity}
s = \sum\limits_{i=1}^{N}\sum\limits_{j=1}^{N} \lambda_{i}\cdot \lambda_{j}^{2}\cdot
u^{T}\left(x_{i}\circ x_{j}\right)\cdot \left(x_{i}\circ x_{j}\right)^{T} u
\end{equation}
The eigenvectors of the adjacency matrix $A$ are orthogonal. Therefore $\left(x_{i}\circ x_{j}\right)^{T}\cdot u = x_{i}^{T}\cdot x_{j} = 0$ if $i\ne j$, otherwise it equals 1 and (\ref{Eq:A_squared_complexity}) further simplifies to
\begin{equation}\label{Eq:A_squared_complexity_solution}
s = \sum\limits_{i=1}^{N}\lambda_{i}^{3},
\end{equation}
which equals 6 times number of triangles in the network \cite[p. 31]{van2010graph}, because a 2-hop walk between adjacent nodes $i$ and $j$ over a common neighbour $m$ is equivalent to a triangle $i \rightarrow m \rightarrow j \rightarrow i$. The computational complexity of $A\circ A^{2}$ thus reduces to $O\left(d_{av}\cdot L\right)$, as presented in Figure \ref{Squared_A_Fig}. For a given matrix $A\circ A^{2}$, the computational complexity of the $N\times N$ matrix $W$ is $O(L)$, because (\ref{def_martrix_W}) can be transformed into Hadamard product terms (i.e. element-based operations).
\begin{figure}[!h]
\begin{center}
\includegraphics[angle=0, scale=0.82]{Squared_A_v1.pdf}
\caption{Sum of the cubed eigenvalues $\lambda$ of the adjacency matrix $A$ (blue circles) and product of the average degree $d_{av}$ and number of links $L$ (red line), for an Erdős–Rényi random graph with $N=300$ nodes, versus the relative mean degree $\frac{d_{av}}{N}$ (left-figure) and $d_{av}\cdot L$ (right-figure).}
\label{Squared_A_Fig}
\end{center}
\end{figure}
\subsubsection{Computing the $N\times 1$ eigenvector $y_{2}$}\label{SubSubSec:y_2}
The eigenvector $y_{2}$ corresponds to the second largest eigenvalue $\beta_{2}$ of the $N\times N$ matrix $W-\text{diag}(W\cdot u)$. The largest eigenvalue $\beta_{1}=1$ corresponds to the eigenvector $y_{1}=\frac{1}{\sqrt{N}}u$. Computing the eigenvector $y_{2}$ is equivalent to computing the largest eigenvector of the matrix $W-\text{diag}(W\cdot u)-\frac{1}{N}\cdot u\cdot u^{T}$, which can be executed using the power method \cite{van2010graph}, for a given matrix $W$, with computational complexity $O\left(L\right)$.
\subsubsection{Computing the cluster membership function}\label{SubSubSec:C}
We apply the recursive algorithm \ref{alg:cluster-estimation} to identify communities based on the $N\times 1$ eigenvector $y_{2}$. The number of iterations of the algorithm ideally equals $T=\log_{2}c$, while in worst case scenario there are $c$ iterations.
Given a fixed number $c$ of communities, the computational complexity within an iteration is $O\left(L\right)$, as shown in pseudocode \ref{alg:cluster-estimation}. The number of clusters $c$ may depend upon $N$ and is in worst case equal to $N$. Thus, computational complexity increases in worst case to $O(N\cdot L)$.
\subsubsection{Scaling the inter-community links}\label{SubSubSec:Scaling_Links}
Between two iterations of the linear clustering process, we identify inter-community links and scale their weights, as defined in (\ref{Ranking_Distance_eq}). The computational complexity of this step is $O\left(L\right)$, as the ranking difference of neighbouring nodes is computer over each link.
Finally, computational complexity of the entire proposed clustering process equals $O(N\cdot L)$, because $d_{av} = O(N)$.
\subsection{Clustering performances on stochastic block generated graphs }\label{SubSec:SBMM_Results}
\begin{figure*}[!]
\begin{center}
\includegraphics[angle=0, scale=1.00]{Results_2_clusters_1e2_it.pdf}
\caption{ The estimated number of clusters (left-figure) in SSBM graphs with $N=1000$ nodes, average degree $d_{av}=7$ and $c=2$ clusters, for different values of parameters $b_{in}$ and $b_{out}$. The modularity of the estimated partitions is presented in right figure. The vertical dashed line indicates the clustering detectability threshold.}
\label{Fig_Results_2_clusters}
\end{center}
\end{figure*}
We compare the clustering performance of our LCP with that of clustering methods introduced in Appendix \ref{App_Clustering_Algorithms}, on a same graph generated by the symmetric stochastic block model (SSBM) with clusters of equal size. All graphs have $N=1000$ nodes. We vary the parameters $b_{in}$ and $b_{out}$ using (\ref{Eq:Expected_Degree}) in a way to keep the average degree $d_{av}=7$ fixed. For each SSBM network, we execute the clustering methods $10^{2}$ times and present the mean number of estimated clusters and mean modularity of produced partitions in Figures (\ref{Fig_Results_2_clusters}-\ref{Fig_Results_8_clusters}). While the non-back tracking algorithm (Sec \ref{Sec_non_back_tracking_matrix}) and our LCP$_{c}$ (Sec \ref{Non_back_tracking_as_a_process}) estimate only number of clusters, Newman's method (Sec \ref{Sec_Newman_modularity}), the Louvain method (Sec \ref{Sec_Louvain}) and our LCP (Sec \ref{Recursive_algorithm_sec}) and LCP$_{n}$ (Sec \ref{Recursive_algorithm_given_N_sec}) estimate both number of clusters and the cluster membership of each node. The attractive strength $\alpha=0.95$ and the repulsive strength $\delta = 10^{-3}$ are used in all simulations. Weights of $60\%$ links in total are scaled using (\ref{Ranking_Distance_eq}), evenly over $30$ iterations, where in $i$-th iteration scaled weight is $\frac{0.05\cdot i}{30}$.
The clustering performance on an SSBM graph with $c=2$ clusters is presented in Figure \ref{Fig_Results_2_clusters}. The non-back tracking algorithm and our LCP$_{h}$ achieve the best performance in estimating the number of communities $c$, as shown in left-part of Figure \ref{Fig_Results_2_clusters}. Further, our LCP outperforms both modularity-based methods, in identifying the number of communities as well as in modularity $m$.
Figure \ref{Fig_Results_2_clusters} illustrates a significant difference in performance between our LCP and the non-back tracking matrix method. The modularity $m$ of our LCP is plotted in blue, while LCP$_c$, when the number $c$ of communities is known, is presented in magenta. Our LCP and the other two modularity-based methods perform poorly in recognising number $c$ of clusters for a wide range of $b_{in}-b_{out}$, because modularity-based methods generate partitions of higher modularity than that of the original network (in black colour), but with different number of communities!
Our LCP$_{n}$ shows that even for a given number of clusters $c$, there exists a graph partition with a higher modularity than the modularity of the original network, within theoretically detectable regime.
\begin{figure*}[!]
\begin{center}
\includegraphics[angle=0, scale=1.00]{Results_3_clusters_1e2_it.pdf}
\caption{ The estimated number of clusters (left-figure) in SSBM graphs with $N=999$ nodes, average degree $d_{av}=7$ and $c=3$ clusters, for different values of parameters $b_{in}$ and $b_{out}$. The modularity of the estimated partitions is presented in right figure. The vertical dashed line indicates the clustering detectability threshold.}
\label{Fig_Results_3_clusters}
\end{center}
\end{figure*}
In a graph with $c=3$ communities, Figure \ref{Fig_Results_3_clusters} shows that the non-back tracking matrix method again outperforms other three methods in identifying number of communities. Our LCP produces better partitions over the entire range $b_{in}-b_{out}$ than Newman's and the Louvain method. As in the case with $c=2$ clusters, by optimising modularity, the exact number of communities in SSBM graph cannot always be recovered, because there exist other partitions with higher number $c$ of clusters with a higher modularity.
\begin{figure*}[!]
\begin{center}
\includegraphics[angle=0, scale= 1.00]{Results_8_clusters_1e2_it.pdf}
\caption{ The estimated number of clusters (left-figure) in SSBM graphs with $N=1000$ nodes, average degree $d_{av}=7$ and $c=8$ clusters, for different values of parameters $b_{in}$ and $b_{out}$. The modularity of the estimated partitions is presented in right figure. The vertical dashed line indicates the clustering detectability threshold.}
\label{Fig_Results_8_clusters}
\end{center}
\end{figure*}
Finally, Figure \ref{Fig_Results_8_clusters} shows a similar view for an SSBM graph with $c=8$ clusters. Our LCP again outperforms the other three methods in estimating number of communities over the entire range of $b_{in}-b_{out}$ values. In addition, the modularity of our LCP is superior to the other two methods.
\section{Conclusion}\label{Sec:Conclusion}
In this paper, we propose a linear clustering process (LCP) on a network consisting of an attraction and repulsion process between neighbouring nodes, proportional to how similar or different their neighbours are. Based on nodal positions, we are able to estimate both the number $c$ and the nodal membership of communities. Our LCP outperforms modularity-based clustering algorithms, such as Newman's and the Louvain method, while being of the same computational complexity. The proposed LCP allows estimating the number $c$ of clusters as accurately as the non-back tracking matrix.
The linear clustering process LCP is described by a matrix $I + W - \text{diag}(W\cdot u)$, which can be regarded as an operator acting on the position of nodes, comparable to quantum mechanics (QM). In QM, an operator describes a dynamical action on a set of particles. Since quantum mechanical operators are linear, the dynamics are exactly computed via spectral decomposition. In a same vein, our operator $I + W - \text{diag}(W\cdot u)$ is linear and describes via attraction and repulsion a most likely ordering of the position of nodes that naturally leads to clusters, via spectral decomposition, in particular, via the eigenvector $y_2$ in Section \ref{model_time_dependence_subsection}.
\section*{Acknowledgements}
The authors are grateful to S. Fortunato for useful comments.
This research is part of NExTWORKx, a collaboration between TU Delft and KPN on future communication networks.
|
\section*{Supplementary Material}
Video: \url{https://youtu.be/LlW-ylcU0mk}
\vspace{8pt}
\section{Introduction}
\begin{figure*}[ht!]
\centering
\subfigure[]
{\includegraphics[height=1.1in, width=1.66in]{Figs/title_a.png}
\label{fig:title_a}}
\subfigure[]
{\includegraphics[height=1.1in, width=1.66in]{Figs/title_b.png}
\label{fig:title_b}}
\subfigure[]
{\includegraphics[height=1.1in, width=1.66in]{Figs/title_c.png}
\label{fig:title_c}}
\subfigure[]
{\includegraphics[height=1.1in, width=1.66in]{Figs/title_d.png}
\label{fig:title_d}}
\caption{Proposed UAV-miniUGV hybrid system. (a) shows the developed miniUGV with an EPM at the back and a monocular camera at the front. (b) shows the CAD model of the complete system. (c) shows the real system. (d) shows the replication of a hidden space in the lab. We use this hidden space to understand the behavior of the developed methods in the real world.
}
\label{fig:hybrid-system}
\end{figure*}
To determine the fitness of anything made and to guarantee a safe workplace, a careful examination is required to identify and record hazards for corrective actions. Therefore, a variety of autonomous robots have been developed by researchers and the industry
{for applications such as area surveillance \cite{Stolfi2021, MISHRA20201, Ganesh2015}, generating detailed maps \cite{Qin2019}, performing complex manipulation tasks \cite{aerial_grasping}, \cite{Srivastava_2021} and retrieval tasks \cite{Shankar2018, Corke2004}, which require robots to carry objects, deploy sensors, avoid obstacles, move in steep and rough terrains. }
For a long time, UAVs have been a center of attraction for inspection and exploration tasks because of their ability of rapid areal scanning. For the scenarios like Search and Rescue (SAR) or natural disasters, the information collected by the flying sensors is not sufficient. In some cases, we specifically want to look inside the areas which remain hidden from the UAV's sensors. We call this space a {\em hidden space} and in this work, we propose a hardware architecture equipped with an autonomy software system to collect information from hidden space as well as to perform some manipulation tasks inside, like eliminating a potential threat from that space.
Our aim {is to} design a hybrid system that leverages the strengths of heterogeneous autonomous platforms in order to accomplish a complex mission. Fig. \ref{fig:hybrid-system} shows the proposed hybrid system that can efficiently perform hidden region exploration and manipulation.
In greater detail, let us assume that a UAV has scanned a given area and selected the locations where deep scanning is required. By deep scanning, we mean there exists a hidden space that cannot be explored using a UAV but may be an ideal navigable space for other types of robots, for instance, a ground robot or a quadruped.
The hidden area inspection clearly demands a new hybrid system that must use a UAV as the {\em Primary Robot (PR)} despite having limitations on maximum payload and total flight time. Our proposed system uses a UAV as the PR and a UGV as the {\em Secondary Robot (SR)}.
In this paper, we provide our system design details for both hardware and software to achieve the required autonomy capability.
The primary technical challenges identified to make this system function in the real world are as follows:
\begin{itemize}
\item Design a UAV-UGV system physically coupled using a passive tether;
\item Design a miniature robot with autonomous navigation, exploration and manipulation capabilities, even though the miniature robot is equipped only with limited low-end sensors and computation board;
\item Design a gripper that can be installed on {the UGV} to pick and place objects in a hidden space;
\item Design a robust software-system architecture to support repetitive operations.
\end{itemize}
\subsection{Our Contribution}
\label{contribution}
Since the entire architecture involves numerous components, there are a lot of challenges to overcome before being fully autonomous in any real-life scenarios. All of these challenges cannot be addressed in one attempt. Therefore, in this work, we mainly focus on the overall functionality of the proposed hardware and software system to accomplish the hidden object searching and retrieving mission.
We have designed a miniUGV under the constraints imposed by the UAV, capable of autonomous exploration and manipulation.
{We are using a tether-controlled mechanism on the UAV to retract and release the miniUGV.}
An Electropermanent Magnet (EPM) has been installed on the miniUGV to grab magnetic materials if it finds necessary to take that object out of the hidden space during an inspection. A different type of gripper can be designed and installed on the miniUGV to grab non-magnetic materials. We have used EPM to prove the concept that the miniUGV can also be used as a manipulator.
To be explicit, the main contributions are
\begin{itemize}
\item We propose a hybrid system with a UAV as the PR and a miniUGV as the SR with autonomy software architecture for area inspection and manipulation.
The architecture supports periodical area inspection where the UAV performs aerial inspection and it uses miniUGV to inspect the hidden space.
\item Our in-house made miniUGV is equipped with autonomous navigation and visual servoing capabilities that can easily escape through a $15 cm$ wide and $7 cm$ high narrow entrance. With an EPM at the back, the miniUGV can also act as a manipulator to grab light-weight magnetic objects.
\item To accomplish the mission, we also designed a curiosity-driven exploration heuristics for miniUGV to find a target object in a hidden space using minimal sensing resources i.e., IR sensors and camera.
\item
{Finally, the hardware schematics, architecture design and the software code for the proposed system is made available on the public domain.}
\end{itemize}
\section{Related Work}
Recent advancement in embedded platforms has equipped unmanned robots with low form-factor and high computational capability, leading to an increase in the number of autonomous applications. For fast area search and mapping, the UAV is an ideal robot but faces difficulty to maneuver in a constrained environment, whereas the UGV has a longer duration of service, can carry a large battery, better computation devices, and sensors. Over the decades, their cooperative efforts are seen in tasks such as landing of UAV over UGV \cite{landing1}, precision agriculture \cite{agri1}, area mapping where a UAV searches the space and safely lands on UGV for long-duration travel \cite{area_map2}, path discovery and planning \cite{pathplan1}, civil infrastructure inspection \cite{inspect_app3}, etc.
In \cite{Papachristos2014}, the authors employed a UGV-UAV system to use a powered tether mechanism, which allows heavy batteries to be placed on the UGV and carry power via cables to the UAV platform, in order to achieve longer flight time. Further, a combined path planning strategy is proposed that considers the UAV’s tether constraint kinematics along with the UGV kinematics to find an optimal path in a map based on RRT* algorithm. A passive tether like a rope between the UGV and UAV is used \cite{Miki2019}, with an objective to enable a UGV platform to climb a vertical cliff.
A similar study by \cite{Shankar2018}, uses tethered hooks attached to UAV platforms to re-capture parachute deployed sensor nodes for surveillance.
Scenarios of high-rise building’s spray painting where drone sprays the paint and UGV powers the drone \cite{tether_app3}, as well as cleaning glass and facade where UGV rolls on rails to support the UGV-UAV cleaning system~\cite{tether_app2}, have helped in reducing the human life risk. Although in these studies, researchers developed mechanisms or techniques to physically connect UAV and UGV together, they generally only focus on considering UAV as a secondary platform.
In contrast, we used UAV as a primary platform that makes it move independently to reach a target location inaccessible to UGVs.
Also related is the infrastructure inspection which requires sensors transported to the test site with teams of trained inspectors. With the help of UAVs, this task can be deployed more frequently and consistently than human operators, thus provides a huge decrease in inspection cost (time, effort, manpower, repair expense, life risk involved), covers a large area, and can be used in hazardous situations. The UAV provides a quick visual scan of infrastructure while the UGV is required to access the narrow regions which are not accessible to the UAV. This collaborative UAV-UGV team has the potential to deliver higher fidelity data and a better scanning coverage than either platform could provide independently.
In \cite{inspect_app3}, an approach for civil inspection has a permanent UGV housing mounted in UAV's landing gear.
{To involve UGV in the mission, the UAV has to land, thus a pick and drop mechanism is required to increase the effective operation time of the secondary robot. In contrast, we physically couple UGV with UAV using a passive tether that not only makes it reusable but also enables the UGV to act as an extended manipulator.}
To organize the paper, Section \ref{system_overview} explains the system overview, and Section \ref{hardware} provides the hardware design including realization details of a miniUGV system. Section \ref{CDOS} discusses the autonomy software architecture and methods
using IR sensors and cameras as main onboard sensors. Section \ref{experimental_results} describes the experimental setup and results followed by a conclusion in Section \ref{conclusion}.
\section{System Architecture Overview}
\label{system_overview}
\begin{figure}[t!]
\centering
\includegraphics[width=0.98\linewidth]{Figs/softwareArchitecture_new.pdf}
\caption{An overview of the system architecture. Red dotted lines represent the wireless communication and red solid lines represent wired communication. COMM represents the radio signal transmitter module which is used to take manual control of the mission in case of emergency.
\vspace{-0.5cm}
}
\label{architecture}
\end{figure}
As shown in Fig. \ref{architecture}, the proposed hybrid system consists of a UAV, a miniUGV, and a tether-controller to retract and release the miniUGV at the entry point of a hidden space. The UAV Mission Controller module (UAVMC) triggers the Aerial Exploration, and the UAV being the PR starts a search operation. Simultaneously, a hidden space detection algorithm runs inside the UAVMC. Identifying the hidden space is an interesting and challenging problem and it needs to be addressed separately. In this work, we assume that a visual marker \cite{marker} is present at the opening of each hidden space. Once a hidden space is detected, UAVMC stops the exploration and triggers the Landing module which detects the landing zone and starts releasing miniUGV after aligning the UAV over the landing point.
The Inertial Measurement Unit (IMU) installed on the miniUGV continuously gives feedback during the landing process. Fig. \ref{fig:Tether_agv} shows that there are two tether attachment points on the miniUGV and it is selected in a way that it always prevents a vertical touch down angle to make sure a successful landing.
\subsection{Mode Switching
After performing a successful landing, the miniUGV performs two operations - hidden aerial exploration and object grasping. The UAVMC takes the decision for switching the mode from aerial to ground operations based on the released string length while the sequential decisions to change the mode from exploration to object tracking and grabbing is performed by the miniUGV Mission Controller (miniUGVMC).
\subsubsection{Aerial Operations}
To explain the mode switching, let us assume that $x=[x_\mathcal{W}, y_\mathcal{W}]$ be the pose of miniUGV, where $x_\mathcal{W}$ and $y_\mathcal{W}$ are the position coordinates in world frame $\mathcal{W}$.
Let $\mathcal{X}=[\mathcal{X}_\mathcal{W}, \mathcal{Y}_\mathcal{W}, \mathcal{Z}_\mathcal{W}]$ be the pose of UAV in $\mathcal{W}$, where $\mathcal{Z}_\mathcal{W}$ represents the height of center of mass (COM) of the UAV. Let $L$ be the total string length, $l_0$ is the distance of entry point of hidden space from the COM of the UAV, and $l_t$ is the current length of the released string. The ground operations are constrained by the string length and therefore, we use $l_t$ to switch modes as $l_t= 0$ for aerial exploration, $l_t \in (0, \mathcal{Z}_\mathcal{W}]$ for miniUGV landing and, $l_t \in [l_o, L]$ for hidden area exploration. The miniUGV lands near to the entry point of hidden space that allows us to assumed $(l_o - \mathcal{Z}_\mathcal{W}) \approx 0$.
\subsubsection{Ground Operations}
The SR performs three tasks in deep exploration mode - exploration to find a target object, object tracking, and object grasping. The miniUGVMC triggers a sequential decision making process by initiating hidden area exploration. If the target object is discovered inside the hidden space, Object Tracking (OT) gets triggered. The OT uses visual servoing to reach to the target object and a pre-defined grabbing maneuver is performed to grab the object using EPM.
{Considering the low computational capability of miniUGV, we designed and tested an exploration algorithm to make the system functional in the real world.}
The proposed system design along with the code, CAD files and PCB design are open-sourced\footnote{
{https://github.com/scifiswapnil/UAV-uGV-system}}.
\begin{figure}[t!]
\subfigure[miniUGV touch down angle]{\label{fig:Tether_agv}\includegraphics[width=0.46\linewidth]{Figs/tetheragv.png}}\quad
\subfigure[Sensor Coverage]{\label{fig:sensor_cov}\includegraphics[width=0.48 \linewidth, height=0.405\linewidth]{Figs/sensor_cov.pdf}}
\caption{(a) Shows the tethering mechanism and (b) shows Dual-Sensor coverage. Green-colored grids are frontiers in IR's field of view. Red grids are the two frontiers selected on different parameters.
\vspace{-15pt}}
\label{fig:Tether mechanism}
\end{figure}
\section{Hardware System Realization}
\label{hardware}
The choice of SR highly depends on the type of hidden space. For instance, a miniUGV is the best suitable SR for exploring the narrow and closed space shown in Fig.~\ref{fig:hybrid-system}.
Tether-controlled release and retract system can also be replaced with a pick and drop mechanism to make the SR move freely inside the hidden space. Removing a physical connection between the two robots can have a huge advantage in the exploration of a more complex hidden space with obstacles but at the same time, picking the SR becomes more challenging. Hence, we opt for the tether-controlled pick and place of the SR in this work.
\subsubsection{miniUGV Design}
We have developed a custom-made miniUGV platform with onboard computing for path planning and vision processing. The platform is a 4-wheel drive mobile robot with a footprint of $130~mm \times 120~mm \times 55~mm$. The chassis of the mobile robot is designed to house all the hardware components inside and manufactured using an additive manufacturing technique. The chassis is manufactured using PLA (Poly-lactic Acid) filament that is durable and lightweight material for the mobile platform. Fig.\ref{fig:ugv_arch} shows the hardware architecture diagram of the miniUGV with various components and their interfaces.
The onboard compute unit is raspberry pi 3 Model B+, with an 8-megapixel camera and a peripheral interface for 9 degrees of freedom IMU, closed-loop motor drive control, and IR obstacle sensor. The DC motors with quadrature hall-effect encoders are used and controlled using the two H-bridge driver boards and the quadrature encoder feedback to the compute unit. The motor drives have a closed-loop PID controller for velocity and position control of the motors. The IMU is interfaced with the compute unit via I2C, which gives 3 axes acceleration, 3 axes gyro rate, and 3 axes magnetometer data, for estimating the orientation of the miniUGV in 3D space. The IMU unit houses an onboard digital motion processor that performs independent step counting, quaternion calculation, and orientation estimation, effectively removing the communication delays and offloading the computation task off the compute unit. The platform is equipped with an electro-permanent magnet (EPM) by OpenGrab, where the external magnetic field can be controlled using an electric pulse around a permanent magnet acting as a magnetic gripper. The miniUGV platform is powered by a 2S 1000mAh lithium-polymer battery with a battery management system (BMS) that monitors and controls the charge/discharge rate. Effectively, the miniUGV can operate for a {\em continuous 18 minutes}, where the processor is running at 95\% capacity and all the sensors and drive motors are actively operational. Fig.\ref{fig:mini-UGV} shows the CAD model of the miniUGV and the manufactured unit. The miniUGV is connected via a tether to the UAV platform. For communication, the miniUGV uses the raspberry pi's onboard WiFi network.
\begin{figure}[t!] \vspace{6pt}
\centering
\includegraphics[width=0.8\linewidth, height=1.7in]{Figs/miniUGV_hardware.pdf}\quad
\caption{Hardware overview of our custom-made miniUGV. Power Distribution Board (PDB) distributes the power from battery to all the components.}
\label{fig:ugv_arch}
\end{figure}
\subsubsection{Object Grabbing}
The miniUGV is equipped with an EPM for object grabbing. When the miniUGV reaches close to the target object after a particular minimum distance d\textsubscript{min}, it rotates a full 180 degrees in place that points the EPM to face the target object. Then the EPM is triggered and, the miniUGV
moves backward to push the object. This push confirms the contact between the EPM and the object.
\begin{figure}[t!]
{
\centering
\includegraphics[scale=0.15]{Figs/super_final_final.png}
\caption{The miniUGV platform (Left: real robot, Right: CAD model) \vspace{-10pt}
}
\label{fig:mini-UGV}
} \vspace{-10pt}
\end{figure}
\subsubsection{Tether mechanism}
The tethering module is a setup responsible for maintaining the tether distance between the UAV and the miniUGV. The tether mechanism is made of a Robotis-Dynamixel MX-38 motor. This motor is interfaced via a driver module which is connected to a micro-controller that drives it at a particular velocity to a position. The Dynamixel motor has a rotatory spindle that unwinds the tether with a certain velocity based on the tether required or winds the tether to retract the miniUGV.
The miniUGV communicates its current pose, desired velocity and desired acceleration information to the tethering module over WiFi.
The tether controller is a feedback control loop implemented on the micro-controller unit that continuously tracks the miniUGV's and the UAV's positions, computing the tether distance requirement to drive the rotor.
\textit{Controller:} Based on the miniUGV's position feedback, a taut control loop \cite{Nicotra2014} either releases or retracts the tether to maintain a positive tension on the cable. Inspired by Tal Glick {\em et al's} work \cite{Glick2018} on an unified model approach for tether control, a model predictive control (MPC) loop is developed for the taut control of the tether. In \cite{Glick2018} and \cite{Lee2015}, the UAV platform is connected via a tether cable to a ground station. In our proposed system, the drone is considered as a stable (position-controlled) platform having a winching mechanism that controls the length of the tether. The miniUGV on the other hand has fixed knot. The MPC controller takes input as the desired velocity and acceleration, based on which the winching mechanism either releases or retracts the tether.
\section{Autonomy Software System}
\label{CDOS}
To search an object in an unknown environment, a robot needs to first know where it is located in that environment. This can be achieved through the matured simultaneous localization and mapping (SLAM) technique~\cite{slam}.
In addition to that, since our task is the object searching with onboard sensors, we will also need to perform state estimation for the target object, and further drive the miniUGV to quickly discover and fetch the object.
Fig.\ref{fig:sensor_cov} shows that the miniUGV is equipped with two types of sensors - IR sensors of maximum sensing range $d_{ir}$ and a field of view (fov) $\beta$, and a monocular camera of fov $\alpha$.
{Typically, $\beta$ is much smaller than $\alpha$. If the miniUGV relies on IR sensors for space exploration and uses a camera only for object detection, it takes more time to find the object. In this section, we will discuss a way to include camera observations in the exploration process.}
\subsection{{Object Occupancy Mapping with Camera}}
\label{sect:cam_model}
We estimate the state of the object through a probabilistic way.
Similar to the design of {\em occupancy grids} \cite{PRbook}, we assume another occupancy map $\mathcal{O} = \{{{\omicron}}\}$ that stores the probabilities of grid cells being occupied with any given object of interest (instead of the whole environment). Let us call it an {\em object map}. In order to find the probability of occupancy of a given object in $\mathcal{O}$, we assume that the object detection algorithm provides the detection status as $true$ or $false$ along with a confidence value. We consider this confidence as the probability of occupancy of a grid with the given object.
{However, we must find the distance of the detected object in the camera frame.} The confidence in object detection is the highest when all the features describing that object are clearly visible. The visibility decreases with the increasing distance of the object from the camera. Hence, for any image based object detection algorithm, we assume that
the confidence value $conf = \eta/\left| d \right|$, where
$d$ is the distance of the object in the camera frame and $\eta$ is the constant of proportionality.
Let $\mathcal{D}_{cam}$ be the maximum distance an object can be detected by an object detection algorithm (e.g., \cite{Anand2019}). With that, we assume a monocular camera acts like a 2D range finder with a fov $\alpha$ (same as the fov of the camera) and with a maximum sensing range $\mathcal{D}_{cam}$, that measures the probability of occupancy of the object in $\mathcal{O}$.
The distance of an object of known size can be obtained as
$d = \frac{k_c}{\gamma}$
where $k_c$ is an intrinsic constant of the camera and $\gamma$ is the pixel per meter (ppm) value. The ppm is defined as the length of an object in an image divided by the length of the object in the real world \cite{Anand2019iros}.
Let $\zeta_{1:t}$ be the set of all measurements up to time $t$, and $x_{1:t}$ is the path of the miniUGV defined as the sequence of all poses.
The probability that a grid $i$ in $O$ is occupied with the given object is denoted as $p(\omicron_i|\zeta_{1:t}, x_{1:t})$ which is updated incrementally in the fashion of a Bayesian filter \cite{PRbook},
\begin{equation}
\label{p_o_update}
p(\omicron_i|\zeta_{1:t}, x_{1:t}) = \frac{p(\omicron_i|\zeta_t, x_t) p(\zeta_t|x_t) p(\omicron_i|\zeta_{1:t-1}, x_{1:t-1})}{p(\omicron_i) p(\zeta_t| \zeta_{1:t-1}, x_{1:t}) },
\end{equation}
where $p(\omicron_i|\zeta_{1:t-1}, x_{1:t-1})$ represents the occupancy probability in the previous time step, and prior $p(\omicron_i)$ is set with probability 0.5 in the initial step.
Furthermore, the grids in $\mathcal{O}$
can be represented as
$\mathcal{O} = \mathcal{O}_{free} \cup \mathcal{O}_{occ} \cup \mathcal{O}_{unk}$,
where $\mathcal{O}_{free}$, $\mathcal{O}_{occ}$ and $\mathcal{O}_{unk}$ are the set of $free$, $occupied$, and $unknown$ grid cells respectively.
In practice we determine the occupancy probabilities as follows,
\begin{equation}
\label{cap_O}
p(\omicron_i |\zeta_{1:t}, x_{1:t})=
\begin{cases}
0,[Free]& \text{if } p(\omicron_i|\zeta_{1:t}, x_{1:t}) \in (0, \lambda_1)\\
0.5,[Unk]& \text{if } p(\omicron_i|\zeta_{1:t}, x_{1:t}) \in [\lambda_1, \lambda_2]\\
p(\omicron_i |\zeta_{1:t}, x_{1:t}),& \text{if } p(\omicron_i|\zeta_{1:t}, x_{1:t}) \in (\lambda_2, 1)\\
\end{cases}
\end{equation}
Here, $\lambda_1, \lambda_2 \in [0, 1]$ are the confidence in object detection. (In our experiments we empirically set $\lambda_1=0.10$ and $\lambda_2=0.95$.)
\begin{figure*}[t]
{
\centering
\includegraphics[scale=0.56]{Figs/exp_.png}
\vspace{-0.6cm}
\caption{Experimental setup. First and second images from the left shows snapshots of the real experiment conducted for two different types of hidden spaces with the proposed system. The last image is the detected landing zone. \vspace{-8pt}}
\label{fig:exp-images}
}
\end{figure*}
\subsection{Occupancy Grid Mapping with IR Sensors}
Since compared to the camera the IR has a shorter range and a limited fov, we create a different occupancy grid map with only the IR perception. This shall help define the curiosity-based exploration which will be discussed later.
Specifically, we approximate the hidden space $h$ with $N$ number of 2D grids $G = \{{{g}}\}$, where each grid $g_i$ is represented by its center. The 2D grids are also probabilistic representation of the space
and we assume the miniUGV’s position is known and unobserved grids have a uniform prior of being occupied. With the Bayesian filtering mechanism described in Sect.~\ref{sect:cam_model} and the {\em log odds} conversion~\cite{PRbook}, the probability that an individual grid $g_i$ is occupied given sensor measurements from time 1 to $t$ (i.e., $z_{1:t}$) can be obtained:
\begin{equation}
p(g_i|z_{1:t}) = \Bigg(1+\frac{1-p(g_i|z_t)}{p(g_i|z_t)} \frac{1-p(g_i|z_{1:t-1})}{p(g_i|z_{1:t-1})}\Bigg)^{-1}.
\label{occ}
\end{equation}
Similarly, the grids are labeled free, occupied, or unknown based on the depth sensor input as $G = G_{free} \cup G_{occ} \cup G_{unk}$.
We also define $F_g$ to be the global set of frontiers and $F_v$ to be the local set of frontiers limited to the fov of the IR sensors, such that $F_v \subseteq F_g$ as described in \cite{Cieslewski2017}.
\begin{comment}
Now we update \ref{ob_occ} update \ref{occ} as
\begin{equation}
\begin{split}
p(\omicron_i|\zeta_{1:t}, z_{1:t}, x_{1:t}) =
p(\zeta_{1:t}|\omicron_i, z_{1:t}, x_{1:t})\cdot p(z_{1:t}|\omicron_i, x_{1:t}) \\
\cdot p(x_{1:t}|\omicron_i)
\cdot p(\omicron_i) / \big(p(z_{1:t}|\zeta_{1:t}, x_{1:t})\cdot p(x_{1:t}|\zeta_{1:t}) \cdot p(\zeta_{1:t})\big)
\end{split}
\end{equation}
\textcolor{red}{Simplify this equation. Add update from ir sensor, talk about dependency}
\end{comment}
\subsection{Curiosity Heuristic for Hidden Object Exploration
}
Philosopher and psychologist William James called curiosity ``the impulse towards better cognition", meaning that it is the desire to understand unknowns \cite{James_def}. We give the miniUGV a curiosity drive to know about the object hidden in the given space. This curiosity aims to answer two primary questions about the hidden space: does it contain the given object? And where is the object located?
To address this issue, we attach a curiosity value to each grid in $F_v$.
According to the data from \cite{Kang2009}, decision-makers were least curious when they had no clue about the answer and also when they were extremely confident. They were most curious when they had some idea about the answer but lacked confidence.
A study shows that the curiosity about the answer to a question is a U-shaped function of confidence about knowing that answer~\cite{Kidd2015}.
And therefore, we define $c_i$ as the curiosity attached to $g_i, \forall \,\, i \in N$, i.e.,
\begin{equation}
\label{eq:curiosity}
c_i= - \frac{(p(\omicron_i |\zeta_{1:t}, x_{1:t}) +a)^2}{4b} + \mathcal{K},
\end{equation}
where $a$, $b$ and $\mathcal{K}$ are constants determining the stiffness of the curiosity curve. These value must be set in a way that $c_i$ gets maximum value when $(p(\omicron_i |\zeta_{1:t}, x_{1:t}) \gets 0.5$, e.g., $a = -0.5, b = 0.1, \mathcal{K} = 0.62$ bounds $\mathcal{C}_i$ within $[0, 1]$.
Remember that we have initialised $\mathcal{O}$ with a prior $p(\omicron_i) = 0.5$. This is an optimistic initialization as the miniUGV considers the probability of finding the object at each grid is $50\%$ prior to the start of a search operation. This initialization makes the miniUGV highly curious to look at each grid to find the given object. With each observation, miniUGV reduces its curiosity by finding that the object is either not present at a grid or the probability of the object's occupancy at $\omicron_i$ is higher than $0.5$ as considered in Eq. \eqref{cap_O}.
\begin{algorithm} [h!]
\label{alg:explore}
{\small
\KwResult{Explore the search space to find an object}
$t \gets time_{now}$\;
$a \gets -0.5$,
$b \gets 0.1$
$\mathcal{K} \gets 0.62$\;
$G \gets 0.5$,
$O \gets 0.5$\;
Find $\mathcal{C}_t, F_g, F_v$\;
\While{$F_v \neq \emptyset \And F_g \neq \emptyset$}
{
Update $G, \mathcal{O}$\;
Find $F_g, F_v$\;
\eIf{$F_v \neq \emptyset $}
{
Find $\mathcal{C}_t$\;
Find $p(\omicron_i |\zeta_{t+1}, x_{t+1})$~ $\forall~x_{t+1} \in F_v$\;
Find $\mathcal{L(C)}$ $\forall F_v$\;
Find $g*$, // using Eq. \eqref{eq:action_select}\;
Execute $goto $ $g*$\;
}
{
\eIf{$F_g \neq \emptyset $}
{
$g^n = nearest \, (g_i \in F_g$)\;
Execute $goto$ $g^n$\;
}
{
$\delta t = t - time_{now}$\;
\textbf{return} $O, G, \delta t$\;
}
}
}
}
\caption{Curiosity-Driven Hidden Space Exploration}
\end{algorithm}
As the grids are independent, we assume the curiosity value at each grid is also independent of another and we measure the total curiosity attached to $\mathcal{O}$ at time $t$ as
\begin{equation}
\mathcal{C}_t = \sum_{i=0}^{N} c_i = \sum_{i=0}^{N} - \frac{(p(\omicron_i |\zeta_{1:t}, x_{1:t}) +a)^2}{4b} + \mathcal{X}.
\end{equation}
As mentioned in Sect. \ref{sect:cam_model}, the confidence in object detection varies with the distance of that object from camera, we estimate the observation value at time $t+1$ for miniUGV pose $x_{t+1}$ (pointing towards $\omicron_i$) as
\begin{equation}
p(\omicron_i|\zeta_{t+1}, x_{t+1}) = \frac{p(\omicron_i|\zeta_t, x_t)d_t}{d_{t+1}},
\end{equation}
where $d_t$ and $d_{t+1}$ are the distance of grid $\omicron_i$ from $x_t$ and $x_{t+1}$ respectively.
Now, with this estimated observation, we get the $\mathcal{C}_{t+1}$ by applying equations Eq. \eqref{p_o_update}, Eq. \eqref{cap_O} and Eq. \eqref{eq:curiosity} sequentially. If the miniUGV performs an action to go to $x_i$, the expected curiosity loss $\mathcal{L(C)}$ can be calculated as
\begin{equation}
\label{eq:C_loss}
\mathcal{L(C)} = \mathcal{C}_t - \mathcal{C}_{t+1}.
\end{equation}
Here, we only consider those grids which have updated values from camera measurements,
as the other grids remain the same.
The miniUGV always executes the action that can take it to the grid with the
{highest curiosity loss} given by
\begin{equation}
\label{eq:action_select}
\begin{aligned}
g_i^* = \operatorname*{argmax}_{g_i \in F_v}\, \{\mathcal{L(C)}\}\ \
\textrm{s.t.}\, \, {p(o_i) > 0.5, \ \forall \omicron_i \in \mathcal{O}}. \\
\end{aligned}
\end{equation}
In essence, the mystery is uncovered (i.e., object discovered) when the curiosity becomes very low.
{Alg. \ref{alg:explore} shows the
exploration algorithm that attaches curiosity to the occupancy grid based on the confidence in object detection. Note that this curiosity can be added to any existing space exploration algorithms.}
\section{Experimental Results}
\label{experimental_results}
The experimental setup consists of a DJI Matrice 100 UAV with DJI manifold onboard computer and DJI guidance kit for obstacle avoidance and vision-based localization. The UAV is equipped with a downward-facing Realsense D415 stereo vision camera and the miniUGV attached via the tethering mechanism. A major challenge faced in the setup was the communication between UAV and miniUGV system over WiFi, primarily because of the design of the miniUGV enclosure that made the signal reception a challenge. Also, since the magnetic encoders for the motors are mounted close to the compute board in the miniUGV, it affected the communication signal. This problem was tackled by limiting the communication bandwidth for only sharing the miniUGV position back to the UAV.
\begin{figure}[ht!] \vspace{-5pt}
{
\centering
{
\includegraphics[width=0.6\linewidth]{Figs/zone_map.png}} \caption{Shows five different zones in the map to put the object for performance analyses. Blue points are the randomly selected position of the object in zone $5$ for evaluating the proposed algorithms.
}
\label{fig:zone_map}
}
\end{figure}
\begin{figure*}[ht!]
{
\subfigure[Occupancy Grid Map1 ($G_{ir}$)]
{\label{fig:G_map1}\includegraphics[height=1.in]{Figs/G_1_map.png}} \
\subfigure[Object Map1 ($O_{cam}$)]
{\label{fig:O_map1}\includegraphics[width=2.45in,height=1.in]{Figs/O_map.png}}
\subfigure[$G$ and $O$ visualised togather]
{\label{fig:G_O_map1}\includegraphics[height=1.in]{Figs/G_O_1_map.png}
}\vspace{-5pt}
\subfigure[Occupancy Grid Map2 ($G_{ir}$)]
{\label{fig:G_map2}\includegraphics[height=1.in]{Figs/G_2_map.png}} \
\subfigure[Object Map2 ($O_{cam}$)]
{\label{fig:O_map2}\includegraphics[width=2.4 in, height=1.in]{Figs/O_2_map.png}}
\subfigure[$G$ and $O$ visualised togather]
{\label{fig:G_O_map2}\includegraphics[height=1.in]{Figs/G_O_4_map.png}} \vspace{-10pt}
\caption{Different maps generated after exploration. Yellow color shows the grids occupied with the object. Green color shows the frontiers. \vspace{-15pt}}
\label{fig:two-maps} \vspace{-5pt}
}
\end{figure*}
\begin{figure}[t!]
{
\centering
\subfigure[]{\label{fig:exp_res1}\includegraphics[width=4.2cm, height=3.1cm]{Figs/com1.png}}
\subfigure[]{\label{fig:exp_res2}\includegraphics[width=4.2cm, height=3.1cm]{Figs/com2.png}}
\caption{Performance evaluation on two different maps between the rapid frontier method and the proposed {curiosity driven object search (CDOS)} method. (a) Shows the performance evaluation on sparse map shown in Fig.~\ref{fig:G_map1}. (b) Shows the performance evaluation on dense map shown in Fig.~\ref{fig:G_map2}.
}
\label{fig:maps}
}
\end{figure}
\begin{figure}[t!]
{
\centering
\subfigure[]{\label{fig:exp_fov_a}\includegraphics[width=4.1cm, height=2.6cm]{Figs/Figure_1.png}}\quad
\subfigure[]{\label{fig:exp_fov_b}\includegraphics[width=4.1cm, height=2.6cm]{Figs/Figure_2.png}}
\caption{(a) Shows the performance evaluation by changing camera fov ($\alpha$). (b) Shows the performance evaluation by changing IR fov ($\beta$).
}
\label{fig:maps1}
} \vspace{-15pt}
\end{figure}
\subsection{System Evaluatio
}
All the system's functionalities are verified by conducting repetitive experiments in which the proposed system is given a task to eliminate an object (that might be a potential threat) from a hidden space as shown in Fig. \ref{fig:exp-images}. The UAV takes off and goes to the hidden space whose position is assumed to be known and detects a safe landing zone using the downward-facing depth camera.
Fig. \ref{fig:exp-images} shows the detected landing zone during the autonomous mission. After landing, the system switches the mode of explorations from aerial to the ground and then the miniUGV performs curiosity-driven exploration for object search. The miniUGV creates two different types of map during exploration. Fig. \ref{fig:G_map1} and \ref{fig:G_map2} show the occupancy grid maps generated with the IR depth sensors and, Fig. \ref{fig:O_map1} and \ref{fig:O_map2} show the object map created with the monocular camera modelled as a range finder. The final map is the combination of both as shown in Fig. \ref{fig:G_O_map1} and \ref{fig:G_O_map2}. To reduce the complexity of the mission, we used a simple scenario to evaluate the system performance in the real world. And therefore, the performance of the exploration algorithm has been rigorously tested in a simulated environment on two different maps shown in Fig. \ref{fig:G_map1} and \ref{fig:G_map2}.
\subsection{Exploration results}
We divide the map into five different zones as shown in Fig. \ref{fig:zone_map}.
The miniUGV always starts exploration from zone $1$. Zone $2$ and $4$ are out of the IR sensors' reach and are only covered by the camera. Zone $3$ and $5$ are remaining spaces in the map. {Initially,} zone $2$ and $4$ are out of the IR sensors' reach and are only covered by the camera.
We sample $N$ points from each zone uniformly and place the object at each point one by one.
We choose the state-of-the-art {\em rapid frontier exploration} \cite{Cieslewski2017} method as a comparison baseline. We repeat this process for both maps.
{The baseline method uses camera to detect the object and it relies on IR sensors for space exploration. However, the proposed
method is advantageous over the baseline since it uses the confidence in object detection while selecting the next frontier.}
Results shown in Fig. \ref{fig:exp_res1} and \ref{fig:exp_res2} are generated with $max_{velocity}=2~m/s$, $D_{IR}=D_{cam}$ and $\alpha>\beta$. The target object is considered to be detected only when $conf>0.95$. Fig. \ref{fig:exp_res1} and \ref{fig:exp_res2} show the plots for average time taken to find the object in the maps shown in Fig. \ref{fig:G_map1} and \ref{fig:G_map2} respectively. Both the baseline method and our approach take the almost same time to find the object in Zone $1$ as the observations are available for both the sensors in Zone $1$. For all other zones, observations available for the camera are more than the IR sensors, and it is true for all $\alpha > \beta$. The curiosity values attached to each frontier drive the robots to deviate towards the object of interest even if the confidence in object detection is very low. As a result, the proposed algorithm always takes lesser time except for Zone $1$.
Note that, the environment in Fig.~\ref{fig:G_map2} is more cluttered than in Fig.~\ref{fig:G_map1}.
The chances of the object getting captured by the camera become lower in a more cluttered environment. If there is no detection (even with minimum probability), there would not be any curiosity to know about the object. Therefore, the performance gap between the two methods is low in a cluttered map.
We then evaluated the proposed method on different values of $\alpha$ by keeping IR fov ($\beta$) constant at $30^o$. Fig. \ref{fig:exp_fov_a} clearly shows the proposed method is faster for all $\alpha > \beta$. Usually, the camera fov is larger than the IR fov (at least this is the case for miniUGV), these results show that considering the probability values obtained from object detection helps to find any object in lesser time than using existing algorithms for unknown area exploration by considering deterministic object detection. As a final comparison, we also tested our method for different values of $\beta$ and a constant $\alpha$. Fig. \ref{fig:exp_fov_b} shows that the time taken to find the abject using CDOS is at least equal to or less than the time taken by the rapid frontier method with deterministic object detection.
\section{Conclusion}
\label{conclusion}
In this work, we presented a novel UAV-miniUGV hybrid system for hidden area exploration. We have designed a miniUGV capable of autonomous exploration and tested its functionalities like object grabbing, visual servoing navigation in several scenarios. The overall performance of the system has been tested by eliminating an object from a hidden space. We have designed and implemented a curiosity-driven object search heuristics, and validated its performance on various parameters. The proposed system takes advantage of a wider fov of the camera as well as the stochastic behavior of object detection.
For future work, we would like to analyze the pros and cons of using a non-tethered pick and release of SR over a tethered counterpart discussed in this paper. Also, we plan to focus on the development of minimalistic hardware and software for lightweight exploration robots that can be dropped and retrieved by a UAV system. This is the first version and hence it has a huge potential to further improve.
|
\section{Introduction}
The problem of the geometric representation of the universe and of how to incorporate it in the scheme of the general relativity theory is far from being solved. Modern ideas (rather hypotheses) about the cause of the formation of the state of the universe suggest the instability of some scalar field associated with the quantum nature of the matter \cite{LIN}. The causes and physical mechanism of the appearance of this field, and hence the origin of the universe, remain open for discussion for many years. Earlier, an approach was proposed \cite{Lev6} to describe the origin and evolution of the universe in terms of the first principles of statistical mechanics and quantum field theory. In this approach, we can answer the question of probable emergence of such a field, but nothing can be said about its physical and geometric nature. The purpose of this paper is to describe it in terms of some physical entity that has a simple geometric interpretation and to use a known mathematical apparatus that may describe its probable changes.
We have to define a mathematical representation of the physical entity with which we could describe the initial state of the universe. The spinor representation of the wave function is rather unsuitable for the description of the origin of the universe as a quantum object \cite{lev}. It was shown \cite{kar} that there is no finite-dimensional representation of the complete linear coordinate group of transformations for spinors. Furthermore, Dirac spinors preserve the structure of a linear vector space, but not the structure of a ring. It was also shown \cite{fro} that there can exist only separate associative algebras with a partition on the field of real numbers the real number, the complex number, and the Clifford number. Just these algebras possess the ring structure \cite{kas}.
The Clifford algebra is a vector space over the field of real numbers and is presented as the additive group where multiplication of elements is distributive rather than commutative with respect to addition. This ring has ideals that may be obtained by multiplying the separated element on the right or on the left by ring elements \cite{kas}. The ideals after this procedure are just the Dirac spinors in the standard approach. Thus the representation of the Clifford algebra contains more information on the physical properties than spinor representations. The Clifford algebra may be extended \cite{hil}, \cite{dor} to include in the description the initiation and evolution of the Universe.
As it was shown earlier \cite{con} - \cite{cin}, the application of the Clifford algebra covers all standard functions of quantum mechanics and provides \cite{lev} a unifying basis for the physical knowledge including the general relativity and electromagnetism. When we introduce the Clifford algebra in the scheme of quantum mechanics \cite {con2}, we should not ignore the specifics of this formulation. Actually, in this case we obtain a quantum-mechanical theory that provides only an algebraic structure and does not contain any further specific requirements. It is possible to show \cite{con} - \cite{cin} that Clifford algebra formalism is completely equivalent to the traditional approach to quantum mechanics.
First of all in cosmology, the question arises about the geometric nature of the fundamental field. It may be scalar as well as have other geometric images. It is natural that its geometric characteristics should follow from the space that would be created due to the distribution of the matter. In terms of relevant physical characteristics, the Clifford number is the most suitable at the moment \cite{lev,Lev6}. The main idea of this paper is to describe the origin of the universe in terms of the Clifford number and to find an explanation for the dark matter and energy, as well as to explain the observed behaviour for the visible matter. To do this, we first focus on the basic properties of the Clifford algebra and show its advantages for the physical subject under consideration.
In this paper we also show that the Hamiltonian of the universe in terms of the geometric interpretation of the wave function fully corresponds to the theory of supersymmetry. Thus it is possible to explain the asymmetry between the Bose and Fermi degrees of freedom of the universe and to obtain non-standard conditions of spontaneous symmetry breaking with the condensation of fields of different tensor dimensions. This opens up the possibility of another interpretation of the quantum phenomena in the early universe with the emergence of the dark matter.
\section{Clifford algebra}
We use the basic idea of correspondence between the matrices and elements of the external algebra and thus determine the space for the Clifford algebras. In the case of space-time in the special theory of relativity, the Clifford algebra is constructed by Dirac matrices $\gamma_{\mu}$ that are associated with unit vectors. The linear combination product of this matrices possesses all the properties of the structure of the Clifford algebra with three imaginary unit vectors because one time matrix $\gamma^{2}_{0}=1$ and three space matrices $\gamma^{2}_{\mu}=-1$. Therefore, we can reproduce any element belonging to the induced vector space in the form of a direct sum through all probable tensor representations. The existence of a unique set of linear independent forms defined at an arbitrary point of the space suggests that the nature of the forms translated over the manifold is similar to the nature of forms that characterize such manifold \cite{lev}. This may be also determined by the similar form of the geometric entities as functions of elements of the induced space. Making use of this basis, we consider the realization of any geometric object in the ordinary Euclidean space. In this case the full geometrical entity may be written in terms of the direct sum of a scalar, a vector, a bi-vector, a three-vector, and a pseudo-scalar, i.e., $\Psi=\Psi_{s}\oplus\Psi_{v}\oplus\Psi_{b}\oplus\Psi_{t}\oplus\Psi_{p}$,
that is given by
\begin{equation}
\Psi=\psi_{0}\oplus\Psi_{\mu}\gamma_{\mu}\oplus\Psi_{\mu \nu
}\gamma_{\mu}\gamma_{\nu}\oplus\Psi_{\mu\nu\lambda}\gamma_{\mu}\gamma_{\nu}\gamma_{\lambda}\oplus\Psi_{\mu
\nu \lambda \rho}\gamma^{\mu}\gamma_{\mu}\gamma_{\nu}\gamma_{\lambda}\gamma_{\rho}
\end{equation}
With the reverse order of composition we may change the direction of each basis vector, and thus obtain $\overline{\Psi}=\Psi_{s}\ominus\Psi_{v}\oplus\Psi_{b}\ominus\Psi_{t}\oplus\Psi_{p}$.
Another element of the symmetry is the change of multiplication of the basis vectors to the inverse in the representation of the Clifford numbers, that turns its into $\tilde{\Psi}=\Psi_{s}\oplus\Psi_{v}\ominus\Psi_{b}\ominus\Psi_{t}\oplus\Psi_{p}$. As long as the symmetry element is introduced, there should be present a mathematical operation on the field of Clifford numbers. The direct sum of the tensor subspaces may be attributed with a ring structure with the use of a direct tensor product in the symbolic notation given by $ \Psi\Phi=\Psi\cdot\Phi+\Psi\wedge\Phi$, where $\Psi\cdot\Phi$ is an inner product or convolution that decreases the number of basis vectors and $\Psi\wedge\Phi$ is an external product that increases the number of basis vectors. If each Clifford number is multiplied by a spatial fixed matrix that has one column with elements and other zeros, then we obtain the Dirac spinor with four elements. Making use of this column we can reproduce the spinor representation of each Clifford number. There occurs full correspondence between the spinor column and the elements of the exterior algebra.
Now let us determine the rule of comparing two Clifford numbers in different points of the manifold. For this purpose we have to consider the deformation of the coordinate system and the rule of parallel tanslations on different manifolds. An arbitrary deformation of the coordinate system may be set in terms of the basis deformations $e_{\mu}=\gamma_{\mu}X$, where $X$ is the Clifford number that describes arbitrary changes of the basis (including arbitrary displacements and rotations), that do not violate its normalization, i.e., provided $\tilde{X}X=1$. It is not difficult to verify that $e^{2}_{\mu}=\gamma_{\mu}\tilde{X}\gamma_{\mu}X=\gamma^{2}_{\mu}\tilde{X}X=I$, and this relation does not violate the normalization condition \cite{kas}. Now, for an arbitrary basis, we may set, at each point of the space, a unique complete linearly independent form as a geometric entity that characterizes this point of the manifold. For a four-dimensional space, such geometric entity may be given by
\begin{equation}
\Psi=\Psi_{0}\oplus\Psi_{\mu}e_{\mu}\oplus\Psi_{\mu \nu}e_{\mu}e_{\nu}\oplus\Psi_{\mu\nu\lambda}e_{\mu}e_{\nu}e_{\lambda}\oplus\Psi_{\mu
\nu \lambda \rho }e_{\mu}e_{\nu}e_{\lambda}e_{\rho} \label{41}
\end{equation}
If this point of the manifold is occupied by matter, then its geometric characteristics may be described by the coefficients of this representation. A product of arbitrary forms of this type is given by a similar form with new coefficients, thus providing the ring structure. This approach makes it possible to consider the mutual relationship of fields of different physical nature \cite{lev},\cite{tak}. To determine the characteristics of the manifold as a point function implies to associate each point of the set with a Clifford number and to find its value. If this function is differentiable with respect to its argument, we can introduce the differentiation operation. To define a transfer operation on an arbitrary manifold, you have to define a derivative operator, e.g., as given by $ D = \gamma_{\ mu} \frac {\partial}{\partial x_{\mu}} $ where $ \frac {\partial} {\partial x_{\mu}} $, represents the change along the curves passing through a given point in the space. The result of action this operator for any Clifford number may be represented as
\begin{equation}
D\Psi=D\cdot\Psi+D\wedge\Psi
\end{equation}
where $ D \cdot \Psi $ and $ D \wedge \Psi $ may be called the "divergence" and the "rotor" of the relevant Clifford number. Within the context of the definition of a differentiated variety, it is not enough to have one non-special coordinate system covering a variety whose topology differs from the topology of an open set in the Euclidean space. The structure of such a geometric construction should be supplemented by a correlation between the values of the transferred forms at different points of the manifold. When assigning internal values to the characteristics of the manifold, we have to introduce the transformation of Clifford numbers by changing the coordinate system. It may be identified by displaying the corresponding Clifford numbers under the action of a certain group that causes the relevant transformation. Conversion is possible if caused by any geometric characteristic changing the coordinate system accordingly, as well as by transforming the geometric objects. This requires the use of Clifford algebras as elements of a group of the internal vector space (simple group $ Sp(n) $) $X Y=fZ$, where $X$,$Y$,$Z$ have representation constructed from similar elements of the group, $f$ are the structure constants. A certain transformation group transforms each form according to the law $ \Psi'=\Psi X$, where $X$ determines the mapping elements and satisfies the condition $\tilde{X}X=1$. For such an algebra, we may write the first structure equation that define the co-variant derivative as given by \cite{lev}
\begin{equation}
\Omega=d\Psi-\omega \Psi\label{WF}
\end{equation}
with the gauge transformation law for the connectivity $\omega$
\begin{equation}
\omega'=X\omega\tilde{X}+Xd\tilde{X}
\end{equation}
for the conservation co-variant transformation according to the same law $\omega'=\omega X$. Here the tensor representation is similar to that of an arbitrary form of the Clifford algebra. The latter equation is referred to as the first structure equation but now the form assumes the value in the Clifford algebra. In this case the arbitrary Clifford number may always be reduced to the canonical form, but local deformations of the proper basis become, however, not observable since the Tetroude form $Xd\tilde{X}$ corresponds to the second term of the gauge transformation. Then the second structure equation that defines the "curvature" form may be written as
\begin{equation}
F=d\omega-\omega\omega\label{EF}
\end{equation}
with the law of transformation under the algebra being given by $F'=X F\tilde{X}$. The transfer equation for the curvature tensor with the present transformation law may be written as
\begin{equation}
dF-F\omega +\omega F=J
\end{equation}
where $J$ is the source form with analogous general representation that complies the transformation $J'=X J\widetilde{X}$. The equation thus obtained may be regarded as the field equation, its form is apparently similar to the analogous equation for the connectivity form obtained in the Lie algebra. Those equations have a more general character as their structure contains interrelation of geometric characteristics of different tensor nature. In this approach we may write the fourth structure equation that demonstrates the relationship between the covariant derivation and the form curvature, i.e., $ d\Omega-\Omega\omega +\Psi F=0$. It is the Bianca relation but in the form of Clifford numbers.
\section{Wave function of the Universe}
We may assume that each elementary formation at an arbitrary point of the manifold may be described in terms of a Clifford number. Then the particle wave function is represented by a complete geometric object - the sum of probable direct forms of the induced space of the Clifford algebra \ref{41}. Moreover, by attributing to the wave function a geometric interpretation we can obtain the correct transfer rules for an arbitrary variety \cite{lev} and come up with new results associated with the geometric nature of the wave function. For the wave function as a geometric entity, we may write the first structure equation \ref{WF} in the standard form given by
\begin{equation}
d\Psi-\omega \Psi=m \Psi
\end{equation}
The latter reproduces the form of the Dirac equation but with fuller meaning than in the spinor representation. Such Dirac equation for the wave function \ref{WF} may be obtained by minimizing the action constructed from the geometric invariant
\begin{equation}
S=\int d\tau \left\{\Omega\widetilde{\Omega}+m\Psi\widetilde{\Psi}\right\}
\end{equation}
The Lagrange multiplier $m$ provides the normalization condition for the wave function $\int d\tau \left\{\Psi\widetilde{\Psi}\right\}=1$. The above action is nondegenerate on the solution of the Dirac equation in contrast to the standard approach. In our approach the dynamic equation for the wave function is presented as a rule for the parallel translations for the characteristic on an arbitrary manifold.
To apply this equation to the description of the birth of the universe, we have to make another assumption. The new vacuum contains only the born formation. For this reason, all changes associated with the wave function are due only to its changes in vacuum presence. Therefore, its behavior may be influenced only by one characteristic of the new vacuum, namely, the wave function. In this case, the wave function itself acts as a field that changes its characteristics, or as a connectivity of the space with the new vacuum. The equation for the wave function looks natural as given by
\begin{equation}
d\Psi-\lambda \Psi \Psi=F
\end{equation}
when the first structure equation is at the same time the second structure equation \ref{EF} with the equation for the "curvature" given by
\begin{equation}
dF-F \lambda \Psi +\lambda\Psi F=J
\end{equation}
It is assumed that $\omega \sim \lambda \Psi $. Among the new results, we indicate that the Dirac equation in the geometric representation in the general theory of relativity is just the equation of transfer on an arbitrary variety, therefore, its solution may be interpreted purely geometrically. Moreover, the geometric representation of the wave function yields other results that may simply reveal the geometric nature of the wave function \cite{kle}. Later, these equations will be derived from the least action principle in the geometric interpretation. As follows from the previous analysis, a complete coordinates transformation group associated with the structure equation exists only in the Clifford-number representation of the wave function. The first structure equation for the wave function reproduces the form of the Dirac equation and, as it was shown in \cite{kas}, its solutions are similar to those for the spinor representation. The latter observation solves the problem of the finite-dimensional representation of the wave function under the complete linear group of coordinate transformations.
It was proved in \cite{kas} that each even number $\Psi=\overline{\Psi}$ for $\Psi\tilde{\Psi}\neq 0$, and the Clifford number in the Euclidean space may be reduced to the canonical form, i.e.,
\begin{equation}
\Psi=\left\{\mu(x)\exp(i\beta)\right\}^{\frac{1}{2}}X
\end{equation}
where $X \tilde{X}=1$ describes all the coordinate transformations associated with the translation and rotation of coordinates and with the Lorentz transformation in the Euclidean space. It is not difficult to show that $\int \widetilde{\Psi}\Psi d\tau$ is scalar and in the physical interpretation of this geometric entity it is rather evident since $\mu(x)$ may be associated with the probability density of finding a particle in an arbitrary spatial point, and $\beta$ is the angle that determines the eigenvalue of a particle with positive or negative energy. We can take $\beta=0$ for the matter and $\beta=\pi$ for the antimatter and thus provide a possibility to describe the intermediate states of the particle since the form of the wave function of an arbitrary ensemble of particles is analogous \cite{dor}. According to the proof in the book \cite{kas}, the odd part of the general Clifford number may be presented as the even part multiplied by an separate element of this algebra and thus there is no problem with the manipulation of the full Clifford number. For an arbitrary wave function we take into account only the scalar part of this product that may be explicitly written as
\begin{equation}
\rho=\Psi^{2}_{0}+\Psi_{\mu}\Psi^{\mu}+\Psi_{\mu \nu
}\Psi^{\mu \nu}+\Psi_{\mu\nu\lambda}\Psi^{\mu\nu\lambda}+\Psi_{\mu \nu \lambda \rho}\Psi^{\mu \nu \lambda \rho}
\end{equation}
Thus it becomes possible to describe the intermediate states of the particle inasmuch as the form of the wave function of an arbitrary particle ensemble is analogous \cite{dor},\cite{tre}-\cite{Sch}. The structure equation thus obtained is written in terms of the introduced variable and is completely equivalent to the Dirac equation that has the well known solutions both for the calculation of the hydrogen atom spectrum and for the interpretation of the electron states \cite{kas}.
\section{Modernization of the standard cosmological model}
In the case of spontaneous generation of an additional field in vacuum, the ground state energy of the "new" vacuum for fields of different nature should be lower than the ground state energy of the "initial" vacuum \cite{Lev6}. Moreover, the interaction of this field with the fluctuations of any other field provides energy conservation for the new state of the system. We assume that occurrence in vacuum of the fundamental field that is generated spontaneously and interacts with the fluctuations of all other fields may be described in terms of the Clifford number \cite{lev}. The probable stationary states of the fundamental field are generated by the multiplicative noise produced by the nonlinear self-interaction with fluctuations of this field. The generator of these fluctuations is the vacuum itself for each point of the Planck size on the manifold.
The present model differs from the widely studied scenario of the stochastic inflation of the Universe \cite{LIN} that takes into account fluctuations of the fundamental field but disregards the fluctuations of the unstable vacuum. The internal fluctuations generate the stochastic behavior of the system that may cause changes of its stationary state. The most significant point is that now the fundamental field is not a scalar but a Clifford number and contains all the geometric characteristics of the space that may be born as a result of the emergence of the matter. Only the distribution of the matter can describe the space that arises.
We start with the assumption that the phase transition from the "initial" vacuum to the new state of the vacuum generates a new non-zero field. This means that the new field generates the "new" vacuum different from the "primary" vacuum for any field of an arbitrary geometric characteristic that may exist. The resulting field must reduce the energy of the "new" vacuum with respect to the energy of the "primary" vacuum. Therefore, the energy density of the ground state of the "new" vacuum may be supplied through
\begin{equation}
\epsilon =\epsilon_{v}-\frac{\mu_{0} ^{2}}{2} \widetilde{\Psi}\Psi, \label{10}
\end{equation}
where the second part is the field energy in terms of the wave function with the geometric representation as the Clifford number; the coefficient $\mu_{0} ^{2}$ describes the coupling of the new field and the "primary" vacuum, i.e., the self-consistent interaction of the new field with the probable fluctuations that may exist in the "primary" vacuum. Here we should make two remarks. The first concerns the decrease in the initial energy of the ground state with the appearance of the new field, and the second is related to the coupling coefficient that is now positive, so that explanations of the appearance of such a sign used in the standard approach are not needed. The energy of the new system may be presented in the form
\begin{equation}
E=E_{v}-\int \frac{\mu_{0} ^{2}}{2} \widetilde{\Psi}\Psi d\tau, \label{11}
\end{equation}
If we want to describe the evolution of the system $\langle out| \exp iHt |in \rangle$, we still need to average all probable fluctuations with which the new field can interact. For this purpose it is sufficient to present the nonlinear coupling in the form $\mu_{0} ^{2}=\mu ^2+\xi $, where $ \langle \xi(t)\xi(0) \rangle=\sigma ^2$ and $\sigma ^2$ is the dispersion of the coupling coefficient fluctuations, and to carry out averaging
\begin{equation}
\langle out| \exp iHt |in
\rangle \sim \langle out|\int D \Psi \int D\xi \exp i\left\{ E_{v}-\frac 12\mu ^2\widetilde{\Psi}\Psi+\frac 12 \xi\widetilde{\Psi}\Psi +\frac{\xi ^2}{\sigma ^2}\right\}|in
\rangle \label{12}
\end{equation}
Integration over fluctuation fields yields
\begin{equation}
\langle out| \exp iHt |in
\rangle \sim \langle out|\int D \Psi \exp i\left\{ E_{v}-\frac 12\mu ^2\widetilde{\Psi}\Psi+\frac{\sigma ^2}{4} (\widetilde{\Psi}\Psi)^2 \right\}|in
\rangle \label{13}
\end{equation}
This implies that we have a system with the effective energy (averaged over the fluctuations of the other field coupled with the scalar field) given by
\begin{equation}
E=E_{v}-\int \left[\frac 12\mu ^2\widetilde{\Psi}\Psi-\frac{\sigma ^2 }{4}(\widetilde{\Psi}\Psi )^2 \right] d\tau \label{14}
\end{equation}
where $V(\Psi)=-\frac 12\mu ^2\widetilde{\Psi}\Psi+\frac{\sigma ^2 }{4}(\widetilde{\Psi}\Psi)^2$ is the well-known expression for the energy of the fundamental field with the nonlinear coupling coefficient determined by the dispersion of fluctuations. This implies that with no field $\Psi=0$, $E=E_{v}$ while for $ \widetilde{\Psi}\Psi=\frac{\mu ^2}{\sigma^{2}}$ the expression for the effective ground state energy of the "new" vacuum reduces to $ E=E_{v}-\frac{\mu ^4}{4 \sigma^{2}}\tau $. As follows from the latter relation, the energy of the "new" vacuum is lower than the energy of the primary vacuum, i.e. the phase transition results in the formation of a new vacuum ground state. If $\sigma^{2}$ tends to infinity, then the energy of the new state tends to the initial energy of the ground state. The energy of the new state may vanish for $E_{v}=\frac{\mu ^4}{4 \sigma^{2}}$. This relation may be applied to estimate the maximum dispersion of the field fluctuations if the vacuum temperature is given by $\Theta_{v}=\frac{\mu ^4}{2 \sigma^{2}}$.
The effective potential may now be given in terms of the probability density $\widetilde{\Psi}\Psi =\rho$ of the material field $V(\rho)=-\frac 12\mu ^2\rho+\frac{\sigma ^2 }{4}\rho^2$ that may be useful for the interpretation of different compositions of the energy and matter produced by the spontaneous symmetry breaking. It should be noted that this is the total probability density of the material field, and whether it is "dark" depends on the tensor characteristics of the field in which we feel it. It may be invisible in the vector electromagnetic field but should definitely be felt in the gravitation field and probably in the fields of other tensor representations. As will be shown below, this value $\rho$ is at the same time the supercharge and characterizes the transformation between the bosonic and femionic degrees of freedom in the general representation of the wave function.
\section{Supersymmetry in the geometric representation and probable predictions}
We can now propose a slightly different scenario for the birth of the Universe based on the representation of its wave function as a geometric entity. What appears as the result of the birth of the matter should contain a geometric image. Only the birth of the matter and its distribution may be interpreted in terms of geometry. Such an entity might be a Clifford number with an appropriate physical interpretation. An additional field is required for the emergence of the matter whose spontaneous excitation leads to the emergence of elementary particles. In our case, such a field is the wave function $\Psi$ in terms of different tensor representations, i.e., it has all probable tensor representations with the dimension of the space to be created. That is, the geometry is laid down from the very beginning in the characteristics of the point of the variety on which we describe it.
Minimizing the expression for the energy of the system \ref{11} by independent functions $\Psi$ and $\widetilde{\Psi}$, we obtain for the wave function in the homogeneous case the Gross-Pitaevskii equation with the physical consequences on its solution, i.e.,
\begin{equation}
\frac{\delta E}{\delta \widetilde{\Psi} }=\left[-\mu ^2+\sigma ^2(\widetilde{\Psi}\Psi )\right] \Psi=0 \label{17}
\end{equation}
Similar in content, but richer in nature one can obtain equation by the dynamics of changes of the wave function in the geometric interpretation. To do this, we consider the action written for the wave function of the universe in the presence of the matter. As it was mentioned earlier, the action in terms of the geometric invariant may be given by
\begin{equation}
S=\int d\tau \left\{ \frac{1}{2}\left(F\widetilde{F}+\widetilde{F}F\right) +m\Psi\widetilde{\Psi}\right\} \label{18}
\end{equation}
The Lagrange multiplier $m$ provides the normalization condition for the wave function $\int d\tau \left\{\Psi\widetilde{\Psi}\right\}=1$ and the "general" curvature in the representation of Clifford numbers takes the form; $F=d \Psi-\Psi \Psi$. Minimization of this functional yields an equation that is at the same time the second structure equation, i.e.,
\begin{equation}
dF-F \lambda \Psi +\widetilde{\lambda}\Psi F=J\label{19}
\end{equation}
for the change of the "curvature" at the parallel transfer under the influence of the complete group of transformations of the coordinate system. In the homogeneous case $d\Psi=0$ and since $dF=0$ such equation is transformed to the previous Gross- Pitaevskii equation, i.e.,
\begin{equation}
\left[-\mu ^2+\sigma ^2(\widetilde{\Psi}\Psi )\right] \Psi=0 \label{20}
\end{equation}
with $-\mu ^2=m$ and $\sigma ^2 =\widetilde{\lambda}-\lambda $
We have suggested that in the presence of the spontaneously generated fundamental field the energy of the vacuum state for any other field is lower than the ground state energy of the primary vacuum and that the energy of the fundamental field is influenced by its nonlinear interactions with the fluctuations of the physical fields of different nature. To avoid the question about the influence of the gravitation field on the Universe evolution at the stage of the spontaneous generation of the fundamental field we notice that the energy of the primary vacuum is not contained in the Einstein equation and the dynamics of the Universe is determined only by the potential energy of the fundamental field that produces the matter. The distribution of the matter in its turn determines the geometry.
According to the Dirac theory we may move from the classical Poisson brackets to the quantum ones and rewrite the Hamiltonian in terms of the secondary quantization, where instead of classical geometric representations of the wave function we introduce the operators of creation and destruction of quanta of this field. In the operator form the Hamiltonian of the Universe may be written as:
\begin{equation}
H=E_{v}-\frac 12\mu ^2\hat{\Psi}^{+}\hat{\Psi}+\frac{\sigma ^2 }{4}\hat{\Psi}^{+}\hat{\Psi} \hat{\Psi}^{+}\hat{\Psi} \label{34}
\end{equation}
For the field operators of the general form thus introduced, the commutation relations may be unusual (q-deformed). We return to this question somewhat later, and here we just note that this field by definition has a representation of both bosons and fermions fields that should include probable transformations of bosons into fermions and vice versa. For this reason, we may assume that the product of the birth and destruction operators of the field in the general representation may be written in the form
\begin{equation}
\hat{\Psi}^{+}\hat{\Psi}=Q=\hat{B}^{+}\hat{F}+\hat{F}^{+}\hat{B}\label{35}
\end{equation}
where $\hat{B}^{+}$ and $\hat{B}$ denote Bose creation and annihilation operators and $\hat{F}^{+}$ and $\hat{F}$ denote Fermi creation and annihilation operators with the (anti)commutation relation $[\hat{B},\hat{B}^{+}]={\hat{F},\hat{F}^{+}}=1$, $[\hat{B},\hat{F}^{+}]=[\hat{F},\hat{B}^{+}]=\hat{F}^{2}=(\hat{F}^{+})^{2}=0$. Now it is not difficult to verify that the required Hamiltonian may be written in the supersymmetry form given by
\begin{equation}
H=E_{v}-\frac 12\mu ^2 Q+\frac{\sigma ^2 }{4}Q^{2} \label{34}
\end{equation}
where the square of the supercharge is $Q^{2}=\hat{B}^{+}\hat{B}+\hat{F}^{+} \hat{F}^{+}=n_{B}+n_{F}$ and presents the total number of bosons and fermions. It is obvious that all the elements of the supersymmetry with the commutation relations $[\hat{Q},\hat{Q}^{2}]=0, \hat{Q}=\hat{Q}^{+} $ are contained in the presented form, where individual parts of the Hamiltonian are associated with the integrals of motion and are preserved both separately and together. The interpretation of the supercharge in our case is that the geometric representation of the wave function makes it possible to consider a physical mixture of bosons and fermions, and the charge itself describes the probable transformation of particles into each other. That is, the initial wave function describes a mixed state of bosons and fermions with probable mutual transformations of individual components. If we calculate the partition function \cite{Kap}
\begin{equation}
Z=Tr \exp\beta [\frac 12\mu ^2 Q-\frac{\sigma ^2 }{4}Q^{2}] \label{36}
\end{equation}
then we obtain the thermodynamic quantities of interest as given by
\begin{equation}
\langle Q \rangle=\frac{2}{\beta Z}\frac{\partial Z}{\partial \mu^{2}}
\end{equation}
\begin{equation}
\langle Q^{2} \rangle=\frac{1}{ Z}\frac{\partial Z}{\partial \beta}+\mu^{2}\langle Q \rangle
\end{equation}
with the average value of the number of fermions in the universe being given by
\begin{equation}
\langle n_{F}\rangle=\frac{1}{ 2}(1-Z^{-1})
\end{equation}
and for bosons
\begin{equation}
\langle n_{B}\rangle=\frac{\langle Q^{2} \rangle}{E}-\langle n_{F}\rangle
\end{equation}
As it was shown in the paper \cite{Kap} for small $\mu^{2}$, the relation between bosons and fermions in the universe may be presented as
\begin{equation}
\frac{\langle n_{B}\rangle}{\langle n_{F}\rangle}=\frac{\exp\beta \sigma^{2} +1}{\exp\beta \sigma^{2} -1}
\end{equation}
and may take arbitrary predetermined values of this relation, depending on when the phase transition of the condensation of the bosonic part of the general representation of the wave function occurs. This observation indicates the reason for the bosonic asymmetry of the universe that is closely related to the baryonic asymmetry
\section{Conclusion}
We consider the Clifford algebra formalism as a suitable method for describing the initial state of a vacuum with probable birth of a fundamental field. This field should contain probable geometric characteristics and be completely equivalent to the traditional approach in the quantum field theory. The approach makes it possible to explain the existence of supersymmetric properties of the initial fundamental field, as well as the spontaneous violation of symmetry between bosons and fermions in the universe. Moreover, it provides a possibility to explain the appearance of the "dark matter" due to the influence of fields of tensor dimensions different from the electromagnetic field. Unfortunately, a rigorous mathematical proof of such an approach does not exist at the moment, but for purely physical reasons such representations may favor better understanding of the scenario of the birth and evolution of the universe.
|
\section{Introduction}
Brain and its activity is a complex network consisting of neuronal populations, coactivity of which have to be consistent across large populations with individual differences and dependence on variables of interest such as age, socioeconomic status, and presence or absence of various neurodevelopmental and neurodegenerative deceases and conditions \cite{Geerligs2016}. This activity known as the functional connectivity (FC), opposed to the structural connectivity which is the actual physical network of the neuronal populations and their interconnections, is estimated from the resting state functional magnetic resonance imaging (rs-fMRI).
Deceases and conditions are shown to be related to the increased, decreased, abnormal or generally altered functional connectivity, with the autism spectrum disorder (ASD) being prominent example. ASD is a condition related to the brain development, which affects the individual's social behavior. ASD typically develops in infancy and early childhood and screened by the behavioral tests which may be subjective. Therefore, it is important to bring the quantitative, interpretable methods to screen autism, with fMRI being the promising modality to address that purpose.
The common approach to estimate FC consist of dividing the raw fMRI image time series into ROI groups defined by parcellation atlas. Next the representative time series of each ROI are obtained by averaging or principal component analysis (PCA), and some measure of dependence between each pair of representatives, most commonly the Pearson correlation coefficient is computed. We refer to this pipeline as the univariate approach, based on univariate time series representatives, assuming the homogeneity of the time series in each ROI.
Although as commonly raw time series are grouped into ROI based not on their similarity, but according to atlases based on neuroanatomy or brain's cytoarchitecture, one could not expect time series homogeneity within a ROI or single mode of their distribution. Recently introduced multivariate approach \cite{Geerligs2016,Walther2016,Yoo2019,Basti2020} addresses this issue, with the most popular test for multivariate independence characterizing the relation strength is the distance correlation \cite{Szekely2007}.
In our work we propose and assess a new multivariate measure of the functional connectivity based on the Wasserstein distance \cite{Kantorovich1958,Vaserstein1969} with the application to ASD screening.
\section{Materials and methods}
\subsection{Functional connectivity}
The functional connectivity is estimated from fMRI data using some measure of dependence between time series of each ROI, most often Pearson correlation coefficient between the represetative time series of each ROI obtained by averaging or taking the first principal component.
More formally, given $\{X_i\}_{i=1}^N$ random variables with equal number of observations $t$, but of different dimensions $\{n_i\}_{i=1}^N$ where $N$ is the number of ROIs, one seeks to estimate some measure of statistial dependence $\rho_{ij}$ for each pair $(X_i, X_j)$. If such estimating is done using $1$-dimensional summaries of each $X_i$ we call that \emph{univariate} approach, otherwise if $X_i$ and $X_j$ are compared directly we call that \emph{multivariate} approach.
\input{figures/figure1}
\subsubsection{Pearson correlation}
Most often used similarity metric for comparing fMRI time series is \emph{Pearson correlation}, which is defined for population as
\vspace{-1em}
\begin{equation}
\rho(X, Y) = \frac{\mathrm{Cov}(X, Y)}{\sqrt{\mathrm{Var}(X) \mathrm{Var}(Y)}}.
\end{equation}
Empirical Pearson correlation is given by
\vspace{-0.5em}
\begin{align}
r(\mathbf{x}, \mathbf{y})& = \frac{\sum_{i=1}^n (x_i - \bar{\mathbf{x}}) (y_i - \bar{\mathbf{y}})}{\sqrt{\sum_{i=1}^n x_i^2}\sqrt{\sum_{i=1}^n y_i^2}}\\
&= \frac{\langle \mathbf{x}, \mathbf{y} \rangle}{\|\mathbf{x}\|_2\|\mathbf{y}\|_2},
\end{align}
with (3) valid for z-transformed vectors, standardized to the zero mean and unit variance.
\subsubsection{Distance correlation}
Pearson's correlation coefficient captures only linear association, with rank correlation coefficients traditionally used to address this issue. The relatively recently introduced distance correlation \cite{Szekely2007} in addition to being able to capture nonlinear dependencies is a multivariate measure of association between the groups of random variables of arbitrary, not of necessarily equal size. The idea of distance correlation is based on measuring the correlation of $t^2$ pairwise distances between realizations of random variables $X \in \mathbb{R}^n$ and $Y \in \mathbb{R}^m$, with population \emph{distance correlation} is defined analogously to the Pearson correlation
\begin{equation}
dCor(X, Y)^2 = \frac{\mathrm{dCov}^2(X, Y)}{\sqrt{\mathrm{dVar}^2(X) \mathrm{dVar}^2(Y)}}.
\end{equation}
For an observed random sample, given by the ROI matrices $\mathbf{X}, \mathbf{Y} = \{(X_k, Y_k) \mid 1, \dots, n \}$ from the joint distribution of random vectors $X \in \mathbb{R}^p$ and $Y \in \mathbb{R}^q$ define
\begin{align}
&a_{kl} = |X_k - X_l|_p,\nonumber\\
\bar{a}_{k\cdot} = \frac{1}{n} \sum_{l=1}^n a_{kl},~~~~~~&\bar{a}_{\cdot k} = \frac{1}{n} \sum_{k=1}^n a_{kl},&\bar{a}_{\cdot\cdot} = \frac{1}{n^2} \sum_{k,l=1}^n a_{kl}\nonumber
\end{align}
$k,l = 1, \dots, n.$ to construct double centered pairwise distance matrix $\mathbf{A}$ (and simirarly for $\mathbf{B}$) with entries given by
\begin{equation}
(\mathbf{A})_{kl} = a_{kl} - \bar{a}_{k\cdot} - \bar{a}_{\cdot k} + \bar{a}_{\cdot\cdot}.
\end{equation}
The empirical distance covariance $dCov_n(\mathbf{X}, \mathbf{Y})$ is defined by
\begin{equation}
dCov_n^2(\mathbf{X}, \mathbf{Y}) = \frac{1}{n^2} \sum_{k,l=1}^n A_{kl}B_{kl}.
\end{equation}
The empirical distance variance $dVar_n(\mathbf{X})$ is defined by
\begin{equation}
dVar_n^2(\mathbf{X}) = \frac{1}{n} \sum_{k,l=1}^n A_{kl}^2.
\end{equation}
Finally, the empirical \emph{distance correlation} $dCor_n(\mathbf{X}, \mathbf{Y})$ is defined
\vspace{-1em}
\begin{equation}
dCor_n(\mathbf{X}, \mathbf{Y}) = \frac{dCov_n^2(\mathbf{X})}{\sqrt{dVar_n^2(\mathbf{X})dVar_n^2(\mathbf{Y})}},
\end{equation}
for $dVar_n^2(\mathbf{X})dVar_n^2(\mathbf{Y}) > 0$ and zero otherwise.
\subsubsection{Wasserstein distance}
Optimal transport considers comparing the measures over the domain $X$. Given a ground metric $d: X \times X \rightarrow \mathbb{R}$ optimal transport equips the space of measures $\mathcal{P}(X)$ with a metric referred to as the \emph{Wasserstein distance} \cite{Kantorovich1958,Vaserstein1969}, which for any $\mu, \nu \in \mathcal{P}(X)$ and $p \geq 1$ is defined as \cite{Peyre2019}
\begin{equation}
W_p^p(\mu, \nu) = \inf_{\pi \in \Pi(\mu, \nu)} \int_{X \times X} d^p(x, y) d \pi(x, y)
\end{equation}
where $W_p^p$ denotes the $p$-th power of $W_p$ and $\Pi(\mu, \nu)$ is the set of probability measures on the product space $X \times X$ whose marginals coincide with $\mu$ and $\nu$; namely
\vspace{-1.25em}
\begin{equation*}
\small
\Pi(\mu, \nu) = \{ \pi \in \mathcal{P}(X \times X) \mid P_1\#\pi = \mu, P_2\#\pi = \nu \}.
\end{equation*}
Given ROI matrices $\mathbf{X} \in \mathbb{R}^{n \times t}$ and $\mathbf{Y} \in \mathbb{R}^{m \times t}$, consider the discrete measures $\mu_{\mathbf{X}}, \nu_{\mathbf{Y}} \in \mathcal{P}(X)$ that can be written as linear combinations $\mu_{\mathbf{X}} = \sum_{i=1}^n a_i \delta_{\mathbf{x}_i}$ and $\nu_{\mathbf{Y}} = \sum_{j=1}^m b_j \delta_{\mathbf{y}_j}$ of Dirac's deltas centred at a finite number $n$ and $m$ of points $(\mathbf{x}_i)_{i=1}^n \in \mathbb{R}^t$ and $(\mathbf{y}_j)_{j=1}^m \in \mathbb{R}^t$. In order for $\mu$ and $\nu$ to be probabilities, set the vector weights $\mathbf{a} = (1/n,\dots,1/n)^T \in \mathbb{R}^n$ and $\mathbf{b} = (1/m,\dots,1/m)^T \in \mathbb{R}^m$.
We define the \emph{Wasserstein functional connectivity (WFC)} as the Wasserstein distance between the two discrete measures $\mu_{\mathbf{X}}$ and $\nu_{\mathbf{Y}}$ with corresponding weight vectors $\mathbf{a}$ and $\mathbf{b}$ corresponds to
\begin{equation}
W_p^p(\mu_{\mathbf{X}}, \nu_{\mathbf{Y}}) = \min_{\mathbf{T} \in \Pi(\mathbf{a}, \mathbf{b})} \langle \mathbf{T}, \mathbf{M} \rangle_F,
\end{equation}
where $\mathbf{M} \in \mathbb{R}^{n \times m}$ is the cost matrix with entries $(\mathbf{M})_{ij} = d^p(\mathbf{x}_i, \mathbf{y}_j)$ are set to the $\ell^p$-distance $d^p(\cdot, \cdot)$, $\langle \mathbf{T}, \mathbf{M} \rangle_F$ is the elementwise Frobenius inner product, and $\Pi(\mathbf{a}, \mathbf{b})$ denotes the transportation polytope
\vspace{-1.25em}
\begin{equation*}
\Pi(\mathbf{a}, \mathbf{b}) = \{ \mathbf{T} \in \mathbb{R}^{n \times m}_+: \mathbf{T}\mathbf{1}_m = \mathbf{a}, \mathbf{T}^T\mathbf{1}_n = \mathbf{b} \}.
\end{equation*}
\subsection{Dataset}
Autism Brain Imaging Data Exchange (ABIDE) \cite{DiMartino2014} is a collaboration of $16$ international imaging sites that have aggregated and are openly sharing neuroimaging data. The dataset is composed of structural and resting state functional MRI (rs-fMRI) data along with an extensive array of phenotypic information. We used the preprocessed version\footnote{http://preprocessed-connectomes-project.org/abide/} with $879$ subjects total, with $474$ individuals suffering from ASD and $405$ typical controls. Functional preprocessing was performed using the Configurable Pipeline for the Analysis of Connectomes (CPAC) including the band-pass filtering and global signal regression.
\subsection{Experiment}
Functional connectivity matrices were computed using the Pearson correlation between the ROI representative time series obtained by averaging, and the distance correlation in the voxel space and the Wasserstein distance in the time space for full ROIs. Being symmetric with the same values on the diagonal, matrices were further vectorized by taking the upper half triangle for each matrix.
Following the best practices in fMRI data analysis described in Dadi et al. \cite{Dadi2019}, we used logistic regression with L$1$ and L$2$ regularization with the grid search for the regularization parameter on the $[0.2, 5]$ range. The data was split into the train and test subsets with the $80/20$ ratio, with the random downsampling of the largest class performed to mitigate the class imbalance. The classification accuracy mean and standard deviation in percents for the best pair of the regularization norm and regularization parameter value averaged over $20$ runs are reported in the Table \ref{table:results}.
Also we conducted the analysis of how considered connectivity measures are correlated between themselves. For that, first, mean connectivity matrices were computed for each measure, averaged for all subjects in the dataset. Second, the absolute differences between mean connectivity matrices of the ASD-affected and control patients were computed. Then, for each set of matrices the correlation was computed for each pair in the set, summarizing the values for each pair in the Table \ref{table:correlations}.
\section{Results}
\vspace{-0.75em}
\input{tables/table1_all}
By the experimental evaluation on the ABIDE dataset -- the benchmark for autism diagnosis, we obtained the following results:
\begin{itemize}
\item WFC outperforms the univarate and multivariate methods (Table \ref{table:results}) for the whole dataset composed of the data from all $16$ sites.
\item WFC is competitive sitewise, quantified by the best average rank, including being ranked first on the $3$ largest sites -- NYU (171 subjects), UM (82), USM (61). For the Stanford site using the WFC metric leads to mean accuracy increase of $38.19\%$ compared to the Pearson correlation.
\item WFC is more similar to dCor than to Pearson r, as shown with the visual inspection (Fig. \ref{figure:figure1}, right) and high correlation between them obtained for the averaged connectivity matrices as reflected in the Table \ref{table:correlations}, despite WFC is being defined in the time space, not the voxel space contrary to dCor.
\end{itemize}
\vspace{-0.75em}
\input{tables/table_correlations}
\section{Discussion}
Albeit shown to be beneficial for the downstream tasks, the main limitation of the Wasserstein distance connectivity measure is the increased computational complexity compared to both Pearson and distance correlations. This limitation can be addressed with several complementary approaches.
First is the use of dimensionality reduction -- we observed that applying PCA on the time domain, taking $25$\% principal components with the subsequent computation of the Wasserstein distance does affect the value of WFC measure less than $1$\% for the worst case over the population. Second, one can replace exact computation of the Wasserstein distance with the computation-efficient approximations, making use of the entropic regularization \cite{Peyre2019} or tree metrics \cite{Le2019}. Third, while it is not affecting computational complexity, the computation of the connectivity matrix entries is independent of each other, so is amenable to the distributed and parallel computation.
The key suggestion that Basti et al. \cite{Basti2020} make is that multivariate connectivity analysis benefits from increased functional inhomogeneity of time series constituting a ROI. The corresponding analysis of the dependence of the inhomogeneity and the performance of the classifiers using Wasserstein distance as connectivity measure relative to the univariate methods based on average and principal component analysis-based representatives is left for the future work.
\vspace{-0.25em}
\subsubsection*{Conclusions}
\vspace{-0.25em}
We have proposed the new Wasserstein distance-based measure of the multivariate brain functional connectivity and demonstrated that it is preferable over the commonly used univariate and multivariate connectivity measures given by the Pearson and distance correlations respectively for the task of autism diagnosis.
\section{Compliance with Ethical Standards}
This is a numerical simulation study conducted retrospectively using human subject data made available by Di Martino et al. \cite{DiMartino2014} for which no ethical approval was required.
\section{Acknowledgements}
No funding was received for conducting this study. The authors have no relevant financial or non-financial interests to disclose.
\section{Tables}
\input{tables/table1}
\input{tables/table2}
|
\section{Supplementary material to ``Nonergodic dynamics of
interacting bosonic dipoles in optical lattice''}
\subsection{Parameters of the extended Bose-Hubbard model}
The calculation of the parameters of the extended Bose-Hubbard
model closely parallels the technique described in detail in Ref.~\cite{Kraus22}. We assume
a quasi one-dimensional model with an optical lattice along $x$, and a tight harmonic confinement in the transversal directions, leading to the single-particle trapping potential
\begin{equation}
V_t(\mathbf{r}) = \frac{m\omega^2}{2} \left(y^2 + z^2 \right)+ V_0\cos^2(k x)\ ,
\label{pot1}
\end{equation}
where $m$ is the particle mass, $\omega$ is the harmonic trapping frequency along $y$ and $z$, and $k$ is the wavevector of the laser that forms the optical lattice.
The Hamiltonian of the system may be expressed as~(see e.g. \cite{Dutta15}):
\begin{eqnarray}
\hat{H}&=&\int d^3\mathbf{r} \hat{\Psi}^\dagger (\mathbf{r})\left[-\frac{\hbar^2}{2m}\nabla^2+V_{\text{t}}(\mathbf{r})\right]\hat{\Psi}(\mathbf{r})\\
&+&\frac{1}{2}\int d^3\mathbf{r} \int d^3\mathbf{r}'\hat{\Psi}^\dagger(\mathbf{r})\hat{\Psi}^\dagger (\mathbf{r}')V_{\rm int}(\mathbf{r}-\mathbf{r}')\hat{\Psi}(\mathbf{r}')\hat{\Psi}(\mathbf{r}) \nonumber
\ , \label{secquant}
\end{eqnarray}
with the bosonic field operators $\hat{\Psi}(\mathbf{r})$ and $\hat{\Psi}^\dagger(\mathbf{r})$ that obey the commutation relation $\left[\hat{\Psi}(\mathbf{r}), \hat{\Psi}(\mathbf{r}')^\dagger \right]=\delta ^3\left(\mathbf{r}-\mathbf{r}' \right)$. $V_{\rm int}(\mathbf{r}-\mathbf{r}')$ describes interactions between bosons that is conveniently split into the contact and dipole-dipole terms:
\setcounter{equation}{3}
\begin{equation}V_{\rm int}(\mathbf{r})=V_c(\mathbf{r})+V_d(\mathbf{r})\,.
\label{pot2}
\end{equation}
The contact term is characterized by the $s$-wave scattering length, $a_s$. Using the customary notation, $V_c(\mathbf{r})=g\delta^{(3)}(\mathbf{r})$, with $g=4\pi \hbar ^2 a_s/m$. Dipole-dipole interactions give a second interaction term, $V_d(\mathbf{r})$. We consider dipoles polarized by an external field along the $z$ axis (perpendicular to the axis of the optical lattice) with
\begin{equation}
V_d(\mathbf{r})=C \frac{1-3\cos^2(\theta)}{r^3} \,,
\end{equation}
where $\theta$ is the angle between the dipole and $\mathbf{r}$.
The dipole-dipole interaction is anisotropic in space since the force depends on the dipole orientation.
The strength of the dipole-dipole interactions $C=\mu_0\mu^2/4\pi$ ($C=d^2/(4\pi\epsilon_0)$) for magnetic (electric) dipoles with moment $\mu$~($d$) where $\mu_0$ ($\epsilon_0$) are the magnetic (electric) permeability, respectively.
For sufficiently deep optical lattice ($s=V_0/E_R > 3$, where $E_R=\hbar^2k^2/(2m)$) we may expand the field operator as
\begin{align}\label{exp}
\hat \Psi(\mathbf{r})=\sum_{j=1}^L {\cal W}_j(\mathbf{r}) \hat{a}_{j} = \sum_{j=1}^L \phi_0(y)\phi_0(z)W_j(x)\, \hat{a}_{j}\,,
\end{align}
where $j=1,\ldots,L$ denotes the site index, and the operator $\hat{a}_{j}$ annihilates boson at site $j$. The corresponding basis function ${\cal W}_j(\mathbf{r})$ is the product of the ground states of the harmonic oscillators along $y,z$,
and the Wannier function~(of the lowest band) along the lattice (shallower lattices may implicate the necessity of taking higher bands into account). Plugging Eq.~\eqref{exp} into Eq.~\eqref{secquant}, one expresses the Hamiltonian in as a polynomial of the annihilation and creation operators. Employing the orthogonality of the Wannier functions one arrives at the form:
\begin{align}
\hat{{H}} = -t\sum_{j=1}^{L-1}(\hat{a}^{\dagger}_j\hat{a}_{j+1}+H.c.)
+\frac{1}{2}\sum_{i,j,k,l}^LV_{ijkl}\hat{a}^{\dagger}_i\hat{a}^{\dagger}_j\hat{a}_{k}\hat{a}_{l}.
\label{eqgen}
\end{align}
where the integrals $V_{ijkl}$ are explicitly given as
\begin{align}
V_{ijkl}=\int d^3\mathbf{r} d^3\mathbf{r}'{\cal W}_i(\mathbf{r}) {\cal W}_j(\mathbf{r}')V_{\rm int}(\mathbf{r}-\mathbf{r}'){\cal W}_k(\mathbf{r}'){\cal W}_l(\mathbf{r}).
\label{coe1}
\end{align}
The single particle tunneling amplitude is obtained from the single particle part of the Hamiltonian
\begin{align}
t=-\int d^3\mathbf{r}{\cal W}_i(\mathbf{r} \left[-\frac{\hbar^2}{2m}\nabla^2+V_t(\mathbf{r})\right]{\cal W}_{i+1}(\mathbf{r})\nonumber \\
= \int dx W_i(x)\left[ \frac{\hbar^2}{2m}\frac{\partial^2}{\partial x^2}- V_0\cos^2(kx)\right]W_{i+1}(x) \ ,
\end{align}
where we limit ourselves to NN tunneling only assuming a sufficiently deep optical lattice
(for shallow lattices, with $s=V_0/E_R<4$, one might need to include next-to-NN tunnelings into the picture, see e.g.
Ref.~\cite{Trotzky12}.)
As it turns out the integrals over perpendicular directions can be explictly carried out \cite{Sinha07,Deuretzbacher10,Deuretzbacher13,Bartolo13} yielding
\begin{align}
V_{ijkl}=\int dx dx'W_i(x) W_j(x')V_{\rm 1D}(x-x')W_k(x)W_l(x'),
\label{coev}
\end{align}
where the effective quasi-one-dimensional potential is given by
\begin{widetext}
\begin{align}
V_{\rm 1D} = \left(g_{1D}-\frac{2C}{3l^2}\right)\delta(|x-x'|)+
\frac{C}{l^3}\left[\sqrt{\frac{\pi}{8}}e^{(x-x')^2/(2l^2)}\left(1+\frac{(x-x')^2}{l^2}\right){\textrm{Erfc}}\left(\frac{|x-x'|}{l\sqrt{2}}\right)-\frac{|x-x'|}{2l}\right],
\label{ugly}
\end{align}
\end{widetext}
in terms of the harmonic oscilator length in the perpendicular direction, $l=(\hbar/m\omega)^{1/2}$
and the effective 1D contact interaction strength $g_{1D}=g/(2\pi l^2)$. In Eq.~\eqref{ugly} $\delta$ stands for Dirac delta function and Erfc for the complementary error function. Note that the term proportional to the delta function contains contributions from both the contact and dipole-dipole interactions.
The largest contribution is given by the on-site
interaction term~(diagonal in $ijkl$ indices), traditionally denoted as $U\equiv V_{iiii}$. For contact interactions this is the dominant term. For long-range dipolar interactions the next important term has the form of a density-density
interaction $(V_{ijij}+V_{ijji}) {\hat n}_i{\hat n_j}$ (for $i\ne j$) where often only the NN term for $ j=i\pm 1$ is taken into account.
We mention parenthetically that while for contact interactions $V_{ijij}$ and $V_{ijji}$ are identical, for a dipolar potential one finds that $|V_{ijij}|\ll |V_{ijji}|$. Since Wannier functions are well localized on sites, for tight perpendicular binding and $1/r^3$ potential one may approximate $V_{ijij}=V/|i-j|^3$ recovering the
typical dipolar tail; here {$V= V_{0110}$} is the value of the integral for the NNs. A standard extended Bose-Hubbard model (see e.g. \cite{Rossini12}) considers just terms involving $U$ and
$V$ coefficients and neglects the dipolar tail. The latter may play an important role in the dynamics of the system \cite{Li21,Korbmacher22}, and may differ from the
standard $1/r^3$ decay if the transversal confinement
is not sufficiently strong~\cite{Korbmacher22}.
\begin{figure}[t!]
\includegraphics[width=\linewidth]{T_fit.pdf}
\caption{Relation between the DDT rate $T$ and $V$ assuming a constant $U/t=3$. The dash-dotted line shows the almost perfect correspondence with a linear function. The dashed lines emphasize that at $V/t=8.8$, $T/t = -1$.}
\label{fig:T_fit}
\end{figure}
Other important terms, introduced by Hirsch \cite{Hirsch94} for strongly-correlated spinful fermions, are density-dependent tunnelings~(DDT), also called correlated hoppings, coming from $V_{ijkl}$ terms with three equal indices. The most important corresponds to NN correlated tunneling, e.g. $V_{iii(i+1)} {\hat a}_i^\dagger {\hat n}_i {\hat a}_{i+1}$. For shortness of notation the corresponding amplitude is denoted by $T$ (or rather, due to some historical reasons $-T$ \cite{Dutta15}). Taking together contributions containing $U$, $V$, and $T$ terms, one arrives at the extended Bose-Hubbard Hamiltonian of Eq.~\eqref{eq:ext}.
Figure~\ref{fig:T_fit} shows that $T$ varies linearly with $V$
for a broad range of value of $V/t$~(this is so for the chosen geometry, but does not need to be the case in general). For a lattice depth $s=8$, which
we employ at many points in the main text, one obtains $T/t=-1$ at $V/t=8.8$. Changing $s$, displaces the critical value of $V/t$ to achieve this condition,
since the bare hopping changes exponentially with the lattice depth.
\begin{figure}%
\includegraphics[width=0.8\linewidth]{ent_sect.pdf}
\caption{ (a) Entanglement entropy $\mathcal{S}$ for the hard-core sector at $T/t=-1$ ($V/t=8.8$) for $N/L=9/18$. The dashed horizontal line corresponds to the random-matrix theory value, vertical lines indicate a single sector taken to calculate the average entropy~(shown in the inset for different system sizes). Panel (b) shows the entanglement entropy for the remaining decoupled sector for $N/L=7/14$. The average entropy does not seem to depend on the system size for this nonergodic sector.}
\label{fig:ent_sect}
\end{figure}
\subsection{Further spectral analysis}
Additional information about the eigenstates can be obtained from the evaluation of their half-chain entanglement entropy, defined as
the von Neumann entropy of the reduced density matrix
$\rho_{L/2}$,
$
\mathcal{S}=-Tr[\rho_{L/2}\ln{\rho_{L/2}}].
$
where $\rho_{L/2} = Tr_{1,\cdots,L/2} \vert \psi \rangle \langle \psi \vert $ is obtained after tracing out half of the system for a given eigenstate $ \vert \psi \rangle $. Under the condition $T/t=-1$, we may separately analyze the ergodic (hard-core) and nonergodic sectors. As shown in Fig.~\ref{fig:ent_sect}, the hard-core sector is characterized by large entropies, with the average growing linearly with the system size (volume law - compare inset). The characteristic finger-like structure is due to a partial Hilbert-space fragmentation. By comparison, the mean entropy of the states belonging to the nonergodic subspace seems to grow much more weakly with the system size, and the distribution of $\mathcal{S}$ is broad with many low entanglement states.
\end{document}
|
\section{ОСНОВНЫЕ ОПРЕДЕЛЕНИЯ И ПОСТАНОВКА ЗАДАЧИ}
\section{Introduction}
Locally finite $\mathbb{N}$-graded graphs,
often called Bratteli diagrams,
proved to be a useful instrument
both for algebraic and stochastic problems. It is important to stress
that the geometry of such graphs
is related to the theory of discrete
topological Markov chains, although
the usual questions about Markov chains differ from the questions of the theory of Bratteli diagrams. Combining these two theories
may yield interesting results.
It is well known in the theory of Markov chains that it is fruitful to consider
not finitely or countably many
states, but continuously many states.
This approach is well developed in probability
theory, although some generalizations
of notions and results are nontrivial.
But in the theory of graded
graphs, it is not yet developed.
In this paper we want to
consider first examples of results
concerning so-called continuous graded
graphs. We recall that the theory
of continuous graphs is a new and active branch \cite{L},
but the theory of graded graphs has its own problems which are somehow
different from the usual graph theory
problems.
A basic notion for a graph
is the path space, possibly structured.
For a graded graph, we consider
``graded'' paths which begin at the
first level and go down.
In this paper we consider graphs in which both the levels and the sets of edges between consecutive
levels correspond to
convex cones in Euclidean spaces. Such graphs are said to be of \emph{Gelfand--Tsetlin type}.
The simplest
such graph has the half-line $\mathbb{R}_+$ as the vertex set at each level,
and there is an edge from $x$ to $y$ iff $x\leqslant y$. This makes it possible
to identify a path in this graph with a series with nonnegative terms.
We call it the \emph{Cesàro} graph. We are especially interested in the following
two generalizations of the Cesàro graph. The first one is the
continuous Gelfand--Tsetlin graph: the vertices at level $n$ are
the increasing sequences $x_1\leqslant x_2\leqslant\ldots\leqslant x_n$
of real numbers, and there is an edge between two sequences
$x_1\leqslant x_2\leqslant\ldots\leqslant x_n$ and $y_1\leqslant y_2\leqslant\ldots\leqslant y_{n+1}$ iff they interlace:
$$y_1\leqslant x_1\leqslant y_2\leqslant x_2\leqslant\ldots\leqslant x_n\leqslant y_{n+1}.$$
This graph has the subgraph consisting of nonnegative sequences (``positive Gelfand--Tsetlin graph''). An important intimately related graph is the so-called \emph{rank $d$
Gelfand--Tsetlin graph}, in which the vertices at level $n$ are the
increasing sequences
$x_1\leqslant x_2\leqslant\ldots\leqslant x_d$ of $d$ nonnegative numbers,
and an edge corresponds to the interlacing condition
$$x_1\leqslant y_1\leqslant x_2\leqslant y_2\leqslant\ldots\leqslant x_d\leqslant y_{d}.$$
This graph may be regarded as a subgraph of the positive Gelfand--Tsetlin graph by considering only the levels $\geqslant d$ and sequences of the form $(0,0,\ldots,0,x_1,\ldots,x_d)$
(with $n-d$ initial zeros). These graphs correspond to the spectra of corners
of Hermitian matrices (arbitrary, or nonnegative definite, or nonnegative definite of rank at most $d$).
For an infinite nonnegative definite
Hermitian matrix $A$ of rank at most $d$,
we may define $A_n$ for $n\geqslant d$ as the principal
$n\times n$ minor of $A$, and let
$x_1\leqslant \ldots \leqslant x_d$
be the $d$ largest eigenvalues of $A_n$
(all the other eigenvalues are equal to 0).
Thus we get a path in
the Gelfand--Tsetlin graph of rank $d$.
The pushforward of measures
on nonnegative definite
infinite Hermitian matrices of
rank $d$ are, therefore, measures
on the path space of the Gelfand--Tsetlin graph of rank $d$, and $U(\infty)$-invariant
measures on matrices correspond to
central measures on the path space.
For matrices, we get the so-called
\emph{Wishart measures}. The Wishart measure of rank $d$
with frequencies $\lambda_1,\ldots,\lambda_d$ is defined as
follows: take $d$ i.i.d.\ standard complex Gaussian vectors
$\xi_1,\xi_2,\ldots,\xi_d\in \mathbb{C}^n$ and consider the matrix
$\sum \lambda_i \xi_i\cdot \xi_i^*$. Without rank restrictions, the $U(\infty)$-invariant
measures include, besides Wishart measures, which now may have infinitely many frequencies, the so-called Gaussian unitary ensemble (GUE), corresponding to Gaussian matrix entries (real on the main diagonal and complex otherwise; the entries on the diagonal and above the diagonal are independent).
Possibly the
first paper in which the Gaussian unitary ensemble
is considered
in the framework of
central measures on graded graphs is \cite{Bar}.
The complete list of invariant
measures for infinite Hermitian matrices
(and other similar series)
is known for a while and was obtained by different methods by Pickrell
\cite{Pic} and Olshanski and Vershik
\cite{VO96}
(the latter paper also uses the ergodic method but in a
very different way from what we mean here).
This problem was considered in an old
paper \cite{V74} of the first author
about the ergodic method, but the list of Wishart measures
was missed there.
The invariant measures on positive definite Hermitian
matrices are easier to find, this problem was
addressed in \cite{DS,Aus}.
Gaussian measures ($GUE, GOE$) on matrices
are especially popular. But we study these measures in terms of the graph of spectra of matrices rather than the graph of matrices themselves (for the $GUE$, this is done in \cite{Bar}). The spectral approach is important not only in itself, but also because it allows us to discover more subtle properties of invariant measures. Moreover, this point of view is also possible for other fields (see, for example, \cite{Meta}): for self-adjoint and
quaternionic matrices. For more
on the relation between
graphs of matrices and graphs of spectra, see
\cite{VP}.
Another important generalization is the graph of increasing sequences $x_1\leqslant \ldots
\leqslant x_d$ of nonnegative real numbers in which there is an edge between two sequences iff
$x_i\leqslant y_i$ for all $i=1,\ldots,d$. In other words, vertices
are continuous Young diagrams with at most $d$ rows, and an edge corresponds
to one diagram being contained in the other.
A probabilistic Markov chain
is determined by transition
probabilities, but in the theory of Bratteli diagrams, like in many other problems in
mathematics and physics \cite{V21}, the basic notion is that of \emph{cotransition probability}. Cotransition, or cocycle,
is the main additional structure on a graded graph, and here we chose the cotransition probability measures to be the \emph{normalized
Lebesgue measures on the phase spaces of
cotransitions}. These cotransitions
arise in the theory of
invariant measures on Hermitian matrices \cite{Bar}. The Lebesgue cocycle
is a continuous analog of the uniform cocycle for discrete graded graphs.
A measure on the path space of a graded graph with Lebesgue cocycle
is called a \emph{central measure}.
Next, a remarkable problem arises of
describing all central measures for a given graph.
We discuss it for specific cases. For many countable graphs
(the Pascal, Young, Schur graphs,
dynamical graphs on groups and random
walks), this problem has been considered
in the recent decades.
A natural method for solving this problem
is the ergodic method \cite{V74,V21},
see the details below. Applying it
to find central
measures on continuous
graphs is especially spectacular
because of the geometric nature of the cocycle.
In many cases, the list of ergodic central measures is conjecturally known (or even already found by other methods). For example, in our case, each such measure is given by a set of finite or countable frequencies of natural objects. The problem is
to explain why the frequencies uniquely determine an ergodic central measure, why they are determined by it, and, most importantly, how the frequency description of the measure is related to the corresponding product measure.
Up to now, there was no such description not only for continuous graphs, but even for the Young graph. This is what is done below for the chosen class of graded graphs and for some discrete graphs.
Now we outline the contents of the paper. We begin with the fundamental example of the continuous Cesàro graph. The cocycle here is the set of normalized
Lebesgue measures on simplices $\{x_i\geqslant 0,\sum x_i=S\}$ of
growing dimension with the sum $S$ growing linearly with the dimension.
It is easy to understand that a central measure on this graph is the weak limit of Lebesgue measures on simplices, but it is more difficult to prove that
there exists a unique central measure for each
growth rate, i.e., for each frequency.
The structure of the measure is clear: it is a measure on the set of Cesàro
sequences, which explains the name.
Cesàro measures play here the same role as Bernoulli measures for discrete graphs.
For graphs more general than the Cesàro graph, that is, for the finite rank
Gelfand--Tsetlin graphs and for the graph of continuous Young diagrams,
\emph{the central measures with distinct
frequencies are the restrictions of direct products of
Cesàro measures to the corresponding sets of
paths}. In the case where some frequencies are equal, some regularization of the weak limits is needed. But, quite unexpectedly, it turns out that exactly the same, and even simpler, answer can be obtained for the problem of describing the central measures of discrete type for many countable graphs, e.g., for the Young graph.
Note that in \cite{VK81} all central measures for the Young
graph were realized by means of a generalized RSK algorithm
as homomorphic images of Bernoulli measures,
and in the subsequent paper \cite{RomSni}
it was proved that this homomorphism is an isomorphism.
However, this approach, while important, does not prove the completeness of the list of central measures. It turns out that the completeness of the list in this and other cases follows from an entirely different relation between Bernoulli measures and central measures: {\it the latter are not only factors of Bernoulli measures, but also restrictions to their proper subsets}. To the authors' knowledge, for some reason this fact has been so far overlooked by
mathematicians.
The class of continuous graphs under consideration
is wide enough. It is convenient to regard infinite
paths in these graphs as \emph{positive vector series}.
The dimension of the vector being equal to $1$
corresponds to the case of the Cesàro graph.
We emphasize that the main conclusion from our considerations is that the ergodic central measures for a rather wide class of continuous graphs are determined by their frequencies,
i.e., growth rates of specific coordinates.
More precisely, a central measure is a Markov
measure on positive vector series, and the behavior of these random series reflects the asymptotic behavior of frequencies.
It is important to stress that for the Gaussian unitary ensemble (for the Plancherel measure in the Young graph case), the picture is quite different, since all the limiting frequencies are zero; $GUE$ (as well as the Plancherel measure) owes its existence to a very nontrivial interaction between the positive and negative parts of the
spectra, each of which has a sublinear (rather than linear) asymptotics.
A remarkable parallel between these examples will be discussed in another paper. We are grateful
to a referee who drew our attention
to the interesting paper \cite{Buf}
where the connection between these two measures
is established via the values of moments.
But we mean a different type of parallelism.
We emphasize that the calculations involved in the ergodic method are sometimes quite nontrivial. In this paper we observe the very fact that central measures are determined by frequencies and find them explicitly in the simplest case of the so-called Cesàro graph. Further calculations are the subject of a forthcoming publication.
\section{Setup}
The vertex set of an $\mathbb{N}$-graded graph is
$$\Gamma=\bigsqcup_{n=0}^{\infty}\Gamma_n,$$
where $\Gamma_n$ is the set of vertices of level $n$ and every edge joins two vertices of consecutive levels.
Here we assume that the following two conditions are satisfied.
\smallskip
1. Euclideness. All levels $\Gamma_n$, $n=0,1,2 \dots$, are identified
with closed finite-dimensional convex polyhedral cones $\Gamma_n\subset {\mathbb R}^{d_n}$
with nonempty interior (but the levels in a graph
are disjoint, so, rigorously speaking, we must write something like $\Gamma_n\subset
{\Bbb R}^{d_n}\times \{n\}$). The dimensions $d_n$ may or may not depend on $n$.
\smallskip
2. Convex incidence. Two vertices $x \in \Gamma_n$, $y \in \Gamma_{n+1}$
of consecutive levels are joined by an edge
iff the pair ${x,y}$
belongs to a certain closed convex polyhedral cone
$$D_n \subset \Gamma_n \times \Gamma_{n+1}, \qquad n=0,1,\dots.$$
We require the natural
projections $D_n\to \Gamma_n$ and $D_n\to \Gamma_{n+1}$ to be surjective; in
other words, every vertex in $\Gamma_0$ has a neighbor in $\Gamma_1$,
and every vertex in $\Gamma_n$, $n>0$, has neighbors in both $\Gamma_{n-1}$
and $\Gamma_{n+1}$.
An infinite path is defined as a sequence of vertices
$(v_0,v_1,\ldots)$, $v_i\in V_i$, such that $(v_i,v_{i+1})\in D_i$
for all $i=0,1,\ldots$. The path space is the set
of all infinite paths endowed with a natural (weak) topology.
\smallskip
3. Nondegeneracy. For all interior points
$x \in \Gamma_n$, $z \in \Gamma_{n+1}$, the section
$D^{+}_n(x)=\{y\in \Gamma_{n+1}: (x,y)\in D_n\}$ (the set of second
endpoints of the edges from $x$ to $\Gamma_{n+1}$)
is a nondegenerate convex polyhedral set that does not contain
lines;
and the section $D^{-}_n(z)=\{y\in \Gamma_n: (y,z) \in D_n\}$ (the set of starting
points of the edges from $\Gamma_n$ to $z$) is a nondegenerate
convex polytope. (The nondegeneracy here means the maximum dimension).
This property can be weakened, but it determines an
important special class of continuous graphs.
\smallskip
A certain class of continuous graphs is already defined.
We proceed to specify the class of graphs under consideration by imposing new conditions on the graph structure.
\smallskip
4. Homogeneity. This condition concerns
the case where all levels
are isomorphic: a~graph is said to be \emph{homogeneous}
if the incidence cones $D_n$ coincide for all $n$.
\smallskip
Graphs satisfying the above properties
1--3 are called {\it continuous graphs of Gelfand--Tsetlin type}, and if all levels of such a graph
coincide and property~4 holds, we call
it a {\it homogeneous continuous graph of Gelfand--Tsetlin type}.
\section{Cocycles}
We proceed to the definition of cotransition probabilities
and cocycle for graded graphs. They are introduced in the same
(and even simpler) manner as for discrete graded graphs.
For every interior vertex $x\in \Gamma_n$, $n>0$, the paths from
$\Gamma_0$ to $x$ form a compact set $\mathcal{P}(x)$
of dimension $d_0+d_1+\ldots+d_{n-1}$, which is
stratified in a natural way into sets $\mathcal{P}(y)$ for $y\in D_n^{-}(x)\subset \Gamma_{n-1}$.
Assume that we have already fixed a probability measure $\mu_y$ on each
$\mathcal{P}(y)$ for $y\in D_n^{-}(x)\subset \Gamma_{n-1}$. Then, if we fix a
probability measure on $D_n^{-}(x)$ (so-called cotransition probability measure),
it defines a probability measure $\mu_x$ on $\mathcal{P}(x)$. So, by an obvious induction,
a system of
cotransition measures defines a measure on the paths to $x$ for every vertex $x$
of $\Gamma$. More generally, it defines a~probability measure
on the paths from $y$ to $x$
for any two vertices $x$ and $y$ such that there exists a~path from $y$ to $x$.
Hereafter, we assume that the measures are absolutely continuous with
respect to the Lebesgue measures in the corresponding Euclidean spaces. Then,
for tail-equivalent paths $P_1,P_2$
(that is, infinite paths that coincide from some point on)
we can define a \emph{cocycle} $c(P_1,P_2)$ as follows. Take
a large number $n$ such that $P_1,P_2$ coincide after level $n$
and cut the paths $P_1,P_2$ at this level. We get two paths $P_1(n), P_2(n)$
belonging to the same set $\mathcal{P}(x)$ for certain $x\in \Gamma_n$.
The ratio of the densities of $\mu_x$ at $P_2(n)$ and $P_1(n)$ is denoted by
$c(P_1,P_2)$. It is straightforward that this number does not depend on
the choice of $n$.
In the discrete case, it is natural to consider the uniform distributions on the sets
$\mathcal{P}(x)$, which correspond to the \emph{central cocycle}
$c(P_1,P_2)\equiv 1$ (instead of considering densities, here we may simply divide probabilities).
Borel
measures on infinite paths with such a cocycle are called \emph{central measures}. As follows from the definition, they are Markov measures (with time corresponding
to the grading), but not vice versa.
The set of ergodic (indecomposable) central measures is called the
\emph{absolute} of the graph. Finding the absolute is the most important problem
in the theory of graded graphs.
For the class of continuous Gelfand--Tsetlin type graphs introduced above,
we introduce a natural analog of the central cocycle, which we call the
\emph{Lebesgue cocycle}. It corresponds to considering the normalized Lebesgue measure
on each $\mathcal{P}(x)$, in other words, again $c\equiv 1$. Measures with such a~cocycle are again called central measures, and the main problem is again to describe all
ergodic central measures, i.e., the \emph{absolute} of Gelfand--Tsetlin type graphs.
More general cocycles are to be considered separately.
\section{Ergodic method}
The so-called ergodic method
for describing central and invariant measures was suggested in \cite{V74}. It essentially follows from the individual ergodic theorem or the
martingale convergence theorem.
To state the method, we introduce the necessary notion of weak convergence of measures on
the path space.
A sequence of measures on the path space of a graph is weakly convergent if for any cylinder set the values of the measures of this set converge.
In our situation, the ergodic method can be described as follows
(cf.~\cite{V74,V21}).
\begin{theorem}\label{tt}
For every ergodic central measure $\mu$ on the
path space of a Gelfand--Tsetlin type graph~$\Gamma$
there exists a path, i.e., a sequence of vertices
$\{x_n\}_n$, $x_n \in D_n^-(x_{n+1})$, $n=1,2,\dots$,
for which the sequence of the Lebesgue measures on the sets $\mathcal{P}(x_n)$
weakly converges to $\mu$.
\end{theorem}
Although the
continuous setting is formally different from
what has been considered earlier,
the usual proof holds verbatim.
This theorem reduces the problem of describing the absolute to concrete calculations,
which may be laborious in practice. On the other hand,
the conditions of weak convergence in concrete cases may already be restrictive enough and
allow one to describe the absolute explicitly.
The ergodic method was used in \cite{VK81} to describe
the characters of the infinite symmetric group $S_\infty$, in \cite{VO96} to describe
the invariant measures on infinite Hermitian matrices, and in many other
problems about invariant or central measures.
\section{Frequencies}
Here we restate the main problem of describing the absolute of
a Gelfand--Tsetlin type graph in concrete coordinate terms.
Assume that each level has dimension $d$ and
coincides with the cone $$\mathcal{C}:=\{(x_1,\ldots,x_d)\in \mathbb{R}^d\colon 0\leqslant x_1\leqslant x_2\leqslant \ldots \leqslant x_d\}.$$
A convex cone $D_n$ (which is also the same for all
$n$) is defined by several linear inequalities
satisfied by two vertices $x=(x_1,\ldots,x_n)\in \Gamma_n$ and $y=(y_1,\ldots,y_n)\in \Gamma_{n+1}$.
Our two main examples are as follows.
\smallskip
(i) The {\it rank $d$ Gelfand--Tsetlin graph} is defined by the
interlacing inequalities $$
D_n=\mathcal{D}_{GT}:=\{x,y\colon 0\leqslant x_1\leqslant y_1\leqslant x_2\leqslant y_2\leqslant \ldots \leqslant x_d\leqslant y_d\}.
$$
This graph arises in random matrix theory.
Namely, consider an $n\times n$
Hermitian matrix
with eigenvalues $0$ (of multiplicity $n-d$),
$y_1,\ldots,y_d$, where $0\leqslant y_1\leqslant \ldots\leqslant y_d$.
Its principal minor of order $n-1$ has
eigenvalues $0$ (of multiplicity $n-d-1$) and
$x_1,\ldots,x_d$, where
$
0\leqslant x_1\leqslant y_1\leqslant x_2\leqslant y_2\leqslant \ldots \leqslant x_d\leqslant y_d$.
Moreover, a unitarily invariant
(in the natural sense) measure on the orbit of
matrices with fixed eigenvalues and fixed
eigenvalues of the minor of order $k$, $d\leqslant k\leqslant n$, is
the Lebesgue measure on the corresponding polytope.
This is proved by Baryshnikov \cite{Bar} in greater generality.
So, along with the Gelfand--Tsetlin graph, we may consider a ``covering''
graph of matrices: at level $k$, the vertices are the $k\times k$ Hermitian matrices of
rank at most $d$, and an edge joins a $k\times k$
matrix with its principal $(k-1)\times (k-1)$ minor.
The ordinary Gelfand--Tsetlin graph is defined similarly, but without
the rank restriction.
Note that the definition of our graphs uses only inequalities between elements, and thus it can be automatically considered over an arbitrary linearly ordered set (for example, over a segment instead of the half-line) and even a poset.
In the discrete case of the linearly ordered
set $\mathbb{Z}$, it is related to the
representation theory of unitary groups, see, for example,
\cite{BO}.
\smallskip
(ii) {\it The Young jumps graph}. Here $D_n$ is the following cone:
$$
\mathcal{D}_{Y}:=\{x,y\in \mathcal{C}\colon x_1\leqslant y_1, x_2\leqslant y_2,
\ldots ,x_d\leqslant y_d\}.
$$
If we identify a point $(x_1,\ldots,x_d)$
with the Young diagram with $d$
rows of (not necessarily integer) lengths
$x_1,\ldots,x_d$, then the condition $(x,y)\in \mathcal{D}_{Y}$
means that one diagram is
contained in the other. In the ordinary discrete
Young graph, an edge corresponds to the condition
``one diagram is contained in the other and differs
from it
by exactly one box.'' Here we may
remove (or add, if we consider
moving along a forward path)
arbitrarily many ``boxes.''
\medskip
Let $x(n)=(x_1(n),x_2(n),\ldots,x_d(n))\in \Gamma_n$, $n=1,2,\ldots$, be a path defining a central probability measure $\mu$
on a homogeneous Gelfand--Tsetlin type graph $\Gamma$
according to Theorem \ref{tt}.
If $\lim \frac{x_i(n)}n=\lambda_i\in [0,\infty]$, we say that the $\lambda_i$'s
are the \emph{frequencies} of the corresponding measure. We always may
pass to a subsequence of $\{x(n)\}$ for which the frequencies
exist. The importance of this notion is due to the fact that
in many cases (and we conjecture that for homogeneous Gelfand--Tsetlin type graphs, this is always the case) the frequencies uniquely
determine an ergodic central measure.
\section{Restricting and recovering}
Let $\Gamma$ be a graded graph with levels
$\Gamma_n$ and
$\tilde{D}_n\subset D_n$
be (measurable) subsets of the edge sets of $\Gamma$.
Denote by
$\tilde{\Gamma}$ the subgraph induced on
these subsets.
Let $\mu$
be a central
probability measure on the path space of
$\Gamma$. Assume that with positive
$\mu${\nobreakdash-}probability a random path is a path in
$\tilde{\Gamma}$, that is, it goes only along
the edges in $\sqcup \tilde{D}_n$.
Our crucial observation is the following
almost obvious theorem.
\begin{theorem}\label{l1}
The restriction of $\mu$ to the path space
of $\tilde{\Gamma}$ is also a central measure.
\end{theorem}
\begin{proof}
Fix a level $n_0$. Like every central measure, $\mu$
enjoys the Markov property, that is,
it is characterized by the distribution
at the level $\Gamma_{n_0}$
and the
conditional measures on the tails after
this level. The initial segments before
$\Gamma_{n_0}$ are distributed (Lebesgue) uniformly
on the corresponding polytopes
$\mathcal{P}(x)$, $x\in \Gamma_{n_0}$,
and do not depend on the tails.
All this is preserved under restriction,
since the restriction of a Lebesgue measure is
again a Lebesgue measure.
\end{proof}
Note that the theorem does not
give a recipe for obtaining
specific formulas for
the restricted measure, and this task can be quite
nontrivial. For example,
consider the ordinary (discrete) Young graph consisting
of Young diagrams with at most $k$
rows. It may be viewed as the subgraph consisting of decreasing
integer sequences $n_1\geqslant
n_2\geqslant \ldots \geqslant n_k\geqslant 0$
in the Pascal graph $\mathbb{Z}_{\geqslant 0}^k$.
Consider the Bernoulli measure with
frequencies $p_1>p_2>\ldots>p_k$
on $\mathbb{Z}_{\geqslant 0}^k$ (which means
that at each step we increase the
$i$th coordinate by 1 with probability $p_i$
and do not change the other coordinates). Its
restriction to the Young graph is the Thoma
measure with parameters $p_1,\ldots,p_k$:
the probability of a~diagram with
row lengths
$n_1\geqslant
n_2\geqslant \ldots \geqslant n_k\geqslant 0$
is the value of the corresponding
Schur function
at the point $(p_1,\ldots,p_k)$.
A counterpart of Theorem \ref{l1} states that
a finite measure $\nu$ on the path space of the
subgraph
$\tilde{\Gamma}$
can be extended to a measure
$\mu$ on the path space of the whole
graph $\Gamma$ such that the restriction of~$\mu$ to the path space of $\tilde{\Gamma}$
coincides with $\nu$ (but this
$\mu$ is not always finite, this is a
delicate issue).
Indeed, the centrality
property allows us to define $\mu$
on the images of the path space of
$\tilde{\Gamma}$ under transformations
that preserve central measures.
For example, we may consider transformations that fix the tail after level $n_0$
and change the initial segments accordingly
(preserving the Lebesgue measure).
This defines $\mu$ on the set of paths that are eventually in $\tilde{\Gamma}$. It is not
hard to see that this definition is consistent, and we
get a measure on the path space of $\Gamma$
supported on the paths that are eventually
in $\tilde{\Gamma}$.
If the resulting measure is finite, then the original measure
on the path space of $\tilde{\Gamma}$ can be obtained by restriction
(and normalization). For homogeneous Gelfand--Tsetlin type
graphs, as well as for the discrete Young graph, this corresponds
to the case of \emph{distinct frequencies}.
\section{One-dimensional case: Cesàro measures}
Here we deal with the $d=1$ case of the Gelfand--Tsetlin graph
(the Young jumps graph coincides with the
Gelfand--Tsetlin graph for $d=1$) and compute the
central measures of this graph.
A path in our graph is simply a sequence
$0\leqslant x_0\leqslant x_1\leqslant \ldots$
of nonnegative numbers, and the path space
$\mathcal{P}(a)$ for a vertex
$a\in \Gamma_n=[0,\infty)$ is the $n$-dimensional
simplex
$$\Delta_n(a):=\{0\leqslant x_0\leqslant x_1\leqslant \ldots\leqslant x_n=a\}.$$
We must study
the distribution of finitely many
coordinates $x_0,\ldots,x_{k-1}$ (where $k$
is fixed and $n$ is large) with respect
to the normalized Lebesgue measure $\mu_a$
on this simplex.
Consider an ergodic central measure
on the one-dimensional
Gelfand--Tsetlin graph. By Theorem~\ref{tt},
it corresponds to a certain sequence of vertices
(even to a path, but it is more convenient to consider
a~sequence: the difference is that in a
sequence some levels may be skipped).
Passing to a subsequence, we may assume that
the vertices $a_i\in \mathbb{R}_+$ at levels $n_i$
satisfy the relation $\lim a_i/n_i=\lambda$, where
$\lambda\in [0,\infty]$.
Thus, the following theorem implies that
every ergodic central measure
on the one-dimensional
Gelfand--Tsetlin graph is an exponential random
walk over the levels of the graph.
\begin{theorem}
Let $\lambda>0$ be constant
and $a=\lambda n+o(n)\in \Gamma_n$
be a vertex of level $n$ of the one-dimensional
Gelfand--Tsetlin graph. Then for every fixed
positive integer $k$, the measure induced
by~$\mu_a$
on the sequences $x_0<x_1<\ldots<x_{k-1}$
converges to the exponential random walk
with mean $\lambda$:\break $x_0,x_1-x_0,\ldots,x_{k-1}-x_{k-2}$ are i.i.d.\ with distribution
$\operatorname{Exp}(\lambda)$.
If $a=o(n)$, then the corresponding distributions weakly converge
to the $\delta$-distribution at the zero sequence $(0,0,\ldots)$;
if $a/n\to \infty$, then
the corresponding distributions weakly converge to the zero measure.
\end{theorem}
Because of the importance of this theorem
and its relation to generalizations,
we provide
three different proofs. We focus on the case $0<\lambda<\infty$,
the proofs for $\lambda=0$ and $\lambda=\infty$ are similar (and even
simpler).
\begin{proof}[Proof 1]
Denote $y_0=x_0$, $y_i=x_i-x_{i-1}$ for $i=1,\ldots,n-1$.
Then our simplex is defined as $\{y_i\geqslant 0,\sum y_i\leqslant a\}$.
Its volume equals $a^n/n!$.
If we restrict the distribution to the first $k$ coordinates
$(y_0,\ldots,y_{k-1})$, then
the density at the above point is proportional to the
volume of the corresponding section $\{y_k+\ldots+y_{n-1}\leqslant a-(y_0+\ldots+y_{k-1})\}$. Denoting $s=y_0+\ldots+y_{k-1}$,
we see that this volume equals $(a-s)^{n-k}/(n-k)!$. Thus, the density
equals
$$
\frac{(a-s)^{n-k}}{x^{n-k}\cdot a^k}\cdot{n!}{(n-k)!}=
(1-s/a)^{n-k}\cdot \frac{n(n-1)\ldots (n-k+1)}{a^k}
\sim e^{-s/\lambda}\lambda^{-k}=\prod_{i=1}^k \lambda^{-1}e^{-y_i/\lambda},
$$
which proves the claim.
\end{proof}
\begin{proof}[Proof 2]
It is more convenient to study the distribution of
$a-x_{n-1}$, $x_{n-1}-x_{n-2}$, $\ldots$,
$x_{n-k+1}-x_{n-k}$. Of course, this vector has
the same distribution as $x_0,x_1-x_0,\ldots,x_{k-1}-x_{k-2}$, as can be seen from the measure-preserving
automorphism $(x_0,\ldots,x_{n-1})\to
(a-x_{n-1},\ldots,a-x_0)$ of $\Delta_n(a)$.
Consider the following map from $\Delta_n(a)$
to the unit cube $[0,1]^n$, which maps
the normalized Lebesgue measure on the simplex to the Lebesgue measure on the cube:
$$
\Phi_n\colon (x_0,x_1,\ldots,x_{n-1})\to
\left(\frac{x_0}{x_1},\frac{x_1^2}{x_2^2},
\ldots,\frac{x_{n-2}^{n-1}}{x_{n-1}^{n-1}},
\frac{x_{n-1}^n}{a^n}\right)=:(t_0,\ldots,t_{n-1}).
$$
The Jacobi matrix of $\Phi_n$ is triangular,
which immediately implies that its Jacobian equals
$n!$. Thus, $\Phi_n$ indeed maps
the uniform distribution on $\Delta_n(x)$
to the uniform distribution on $[0,1]^n$.
Hence,
\begin{align*}
a-x_{n-1}&=a\left(1-t_{n-1}^{1/n}
\right)\sim -\frac{a}n
\log t_{n-1},\\
x_{n-1}-x_{n-2}&=x_{n-1}
\left(1-t_{n-2}^{1/(n-1)}\right)=
x\cdot t_{n-1}^{1/n}\cdot
\left(1-t_{n-2}^{1/(n-1)}\right)\sim
-\frac{x}n \log t_{n-2},\\
x_{n-2}-x_{n-3}&=x_{n-2}
\left(1-t_{n-3}^{1/(n-3)}\right)\sim
-\frac{a}n \log t_{n-3},\\
\ldots&\ldots \ldots,
\end{align*}
which yields the result, since $a/n$ goes to
$\lambda$ and $-\lambda \log T\in \operatorname{Exp}(\lambda)$
when $T\in \operatorname{Unif}(0,1)$. (In the above argument, $a\sim b$ means that $a/b$ converges to 1 in probability. We have used the fact that $t^{1/n}\sim 1$ for $t$ distributed uniformly on $(0,1)$.)
\end{proof}
\begin{proof}[Proof 3] This proof, based on the law of large
numbers, is ideologically close to the proof of the
Maxwell--Poincaré lemma, which states that the distribution
of coordinates of a Lebesgue-random point on
the $n$-dimensional sphere
of radius $\sqrt{n}$ is asymptotically standard normal.
We start with the distribution $\eta_n$ of a vector
$y\in (0,\infty)^n$
defined as the product of the $\operatorname{Exp}(\lambda)$
distributions
over all $n$ coordinates,
and prove that the distribution
of the first $k$ coordinates of this vector
is close to that of a random point
in our simplex $\Delta_n(a)$.
The density of $\eta_n$ at a point $y=(y_1,\ldots,y_n)$
equals $$p(y)=\prod_{i=1}^d \lambda^{-1}e^{-y_i/\lambda}=\lambda^{-n}
e^{-(y_1+\ldots+y_n)/\lambda}.$$
Thus, we see that $p(y)$ depends only on the sum $y_1+\ldots+y_n$ of the coordinates.
In other words, the random point
$y$ can be obtained by choosing the value of
$x=y_1+\ldots+y_n$ at random
(according to a certain distribution,
which is in fact a gamma distribution, but we do
not use this) and then choosing a uniformly distributed random point in the simplex
$S_x:=\{y_i\geqslant 0,\sum y_i=x\}$.
Now fix a set $\Omega\subset{\mathbb{R}^k}$ that is
a product of intervals. The probability that $(y_1,\ldots,y_k)\in \Omega$ is equal to the expectation
against~$x$ of the probability
that $(y_1,\ldots,y_k)\in \Omega$
for a random point $(y_1,\ldots,y_n)\in S_x$.
Note that $(y_1,\ldots,y_k)\in \Omega$
if and only if
$\frac{\lambda n}x(y_1,\ldots,y_k)\in
\frac{\lambda n}x \Omega$, and the point
$\frac{\lambda n}x(y_1,\ldots,y_n)$
is uniformly distributed
in $S_{\lambda n}$. Now observe that,
by the law of large numbers, for any
$\varepsilon>0$ the probability that
$\frac{\lambda n}x\notin [1-\varepsilon,
1+\varepsilon]$ tends to 0.
When $\tau\in [1-\varepsilon,
1+\varepsilon]$, the sets $\tau \Omega$
vary a little. In particular, their
intersection contains a set $\Omega_{-}(\varepsilon)$
and their union is contained in a set
$\Omega_+(\varepsilon)$ such that the
$\eta_k$-measures of both sets are close to
each other. This yields the required
weak convergence.
\end{proof}
Thus, any ergodic measure on the one-dimensional Gelfand--Tsetlin graph
is concentrated on the series $x_0+(x_1-x_0)+(x_2-x_1)+\ldots$
that are Cesàro summable to a certain number $\lambda>0$.
That is why we suggest to call it a \emph{Cesàro measure}.
\section{Multi-dimensional case}
Here we explain how combining the one-dimensional
Cesàro case and Theorem \ref{l1} allows us to
handle the case of a finite rank Gelfand--Tsetlin
graph and the Young jumps graph.
Note that the rank $d$
Gelfand--Tsetlin graph is a subgraph
of $\mathbf{Ces}^d$,
where $\mathbf{Ces}$ is the Cesàro
graph. On $\mathbf{Ces}^d$
we can consider a Bernoulli measure that
is the product of Cesàro measures
with frequencies $p_1<p_2<\ldots<p_d$.
With positive probability, any
vertex $(x_1,\ldots,x_d)$
is an increasing sequence and
any two consecutive
vertices $(x_1,\ldots,x_d)$ and
$(y_1,\ldots,y_d)$ interlace
(by the law of large numbers, with probability 1 this happens eventually, and it can easily
be seen that with positive probability this
``eventually'' is from the very beginning).
Thus, by our recovering procedure we get a unique
ergodic central measure on the path
space of the Gelfand--Tsetlin graph with given distinct
frequencies.
A measure for which some
frequencies are equal can be obtained as a weak limit
of measures with distinct frequencies, but we
skip the details here. All the same
holds for the Young jumps graph: the ergodic central measures
are in one-to-one correspondence with arrays of frequencies,
and for distinct frequencies the corresponding measures
are obtained by restricting Cesàro--Bernoulli measures.
To summarize, we provided a new way of studying central
measures, which is based not on computations and estimates,
but on the intrinsic properties of these measures. Of course,
it can also be applied to discrete graphs, but, quite surprisingly, it
was realized for continuous graphs. Applications
to discrete graphs are to be considered in a separate paper.
The calculations of transition probabilities also
get a new interpretation.
For the Gelfand--Tsetlin graph
of growing rank (the graph of interlacing sequences),
we have countably many frequencies with a finite
sum of absolute values, and if all these frequencies
are distinct, the
same restriction argument works.
This method reduces an explicit evaluation
of central measures with given
frequencies to a probabilistic problem of finding the ``Cesàro--Bernoulli'' measures
of the corresponding cones (in order to
compute the restriction).
An alternative, more standard,
way is to apply a continuous version of the Lindstrom--Gessel--Viennot lemma.
As to general Gelfand--Tsetlin
type graphs, we conjecture that for them central measures
are also uniquely determined by
limiting frequencies.
\medskip
We are grateful to the referees for numerous useful suggestions.
|
\section{Problem Definition}
\label{section:Problem_definition}
In this real-world application, we aim to solve the dynamic optimization problem of aligning a camera image that has been shifting over time (see Figure \ref{fig:Image_sequence_1}).
\begin{figure}[h]
\centering
\includegraphics[width=0.225\textwidth]{images/2_TUN-99122-lq_avg_frame.png}
\includegraphics[width=0.225\textwidth]{images/1_TUN-99122-lq_avg_frame.png}
\hspace{1cm}
\includegraphics[width=0.225\textwidth]{images/2_TUN-99141-lq_avg_frame.png}
\includegraphics[width=0.225\textwidth]{images/1_TUN-99141-lq_avg_frame.png}
\caption{\textbf{Left:} Images obtained from the camera at time $t-1$. \textbf{Right:} Images obtained from the camera at time $t$. A shift in the camera can be observed including translation and rotation distortions.}
\label{fig:Image_sequence_1}
\end{figure}
To do so, we optimize the matrix $H$ (see Equation (\ref{eq:homography})) containing 8 Degrees of Freedom (DoF) that correspond to image geometrical transformation (such as translation, rotation, skewing).
\begin{equation}
H=
\begin{pmatrix}
DoF_{11} & DoF_{12} & DoF_{13} \\
DoF_{21} & DoF_{22} & DoF_{23} \\
DoF_{31} & DoF_{32} & 1 \\
\end{pmatrix}
\label{eq:homography}
\end{equation}
The direct and inverse transformation are given in Equation (\ref{eq:transformation})
\begin{equation}
\begin{split}
& I_{1} \cdot H = I_{2} \\
& I_{1} = I_{2} \cdot H^{-1}
\end{split}
\label{eq:transformation}
\end{equation}
Therefore, given two images $I_{1}$ and $I_{2}$ that have been captured using the same camera (from the same position, just changing its viewpoint), and that have a view overlap, it should exist a matrix $H$ that let us transform the content of one image into the other one. Specifically, the values in $H$: $DoF_{11}$, $DoF_{12}$, $DoF_{21}$ and $DoF_{22}$ estimate the rotation. Parameters $DoF_{13}$, $DoF_{23}$ estimate the translation. Finally, $DoF_{31}$, $DoF_{32}$ estimate the skewing effect. Parameter $H_{33}$ is used for the scale magnitude. Given the problem setup it can be fixed to a constant value of 1.
To perform estimation
of the $H$ matrix parameters, a set of point have been initially detected and described in both images. These points were then pairwise matched based on the descriptors similarity (see Figure \ref{fig:matched_keypoints} for an illustration)
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{images/sequence_1/matched_visual_keypoints_TUN-99122-lq_avg_frame.png}
\hspace{1cm}
\includegraphics[width=0.45\textwidth]{images/sequence_2/matched_visual_keypoints_TUN-99141-lq_avg_frame.png}
\caption{Sets of images containing the detected and matched keypoints.}
\vspace{-1.5em}
\label{fig:matched_keypoints}
\end{figure}
We formulate the IA as a dynamic optimization problem by considering the loss function $\mathcal{L}_{loss}$ that provides the (partial) sum of distances between pairs of matched keypoints after apply the transformation $H$ on one point from the pair (see Equation (\ref{eq:loss_function})).
\begin{equation}
\begin{split}
& error_k = d_{L_1} \left ( keypoint_{I_2}^k, keypoint_{I_1}^k \cdot H\right ) \\
& \mathcal{L}_{loss} = \sum_{error_k < P_{i}({errors})} error_k
\end{split}
\label{eq:loss_function}
\end{equation}
More specifically, we use $L_1$ distance between pairs of points. When calculating the loss function we take into account lower the distances up to a certain $i^{th}$ percentile $P_{i}$. For the experiments we found empirically the value of $i=80$ to perform well. We use the percentile to smooth the optimization function (\emph{e.g.} filter non-properly matched keypoints).
\section{Introduction}
Image Alignment (IA) has become an important task to be taken into account in real-world applications. IA is the process of overlaying images to be able to further analyze them. It is a crucial step because many systems rely on obtaining the correct formatted data to take a posteriory decision. This concept comes from a more generic one called image registration. IA was initially introduced by \emph{Lukas, Kanade et al.} in \cite{lucas1981iterative}.
In this work, we tackle the Dynamic Optimization Problem (DOP) of IA in a real-world application using a Dynamic Optimization Algorithm (DOA) called Fractal Decomposition Algorithm (FDA), introduced in \cite{nakib2017deterministic} by \emph{Nakib et al.}. We used FDA to perform IA on CCTV camera feed from a tunnel. As the camera viewpoint can change by multiple reasons such as wind, maintenance, etc. the alignment is required to guarantee the correct functioning of video-based traffic security system.
The rest of this paper is organized as follows. First, Section \ref{section:Problem_definition} introduces the problem formulation. Then, Section \ref{section:Methodology} recalls the FDA foundations. Afterwards, \ref{section:Results} introduces the conducted experiments and explains the obtained results. Finally, in Section \ref{section:Conclusions} the conclusions are presented.
\section{Methodology}
\label{section:Methodology}
To solve the stated problem, we propose to use FDA to deal with this dynamic environment.
More specifically, FDA uses a fractal decomposition structure based on hyperspheres (see Figure \ref{fig:FDA_decomposition}) to explore the search space, and an Intensive Local Search (ILS) method to exploit the promising regions.
\begin{figure}[h]
\centering
\includegraphics[width=0.3\textwidth]{images/FDA_decomposition.png}
\hspace{1cm}
\includegraphics[width=0.3\textwidth]{images/FDA_inflation.png}
\caption{\textbf{Left:} 4-level decomposition using the hypersphere fractal structure. \textbf{Right:} Representation of the hypersphere fractal decomposition with inflation coefficient to ensure that the search space is fully covered given a particular dimension.}
\vspace{-1.5em}
\label{fig:FDA_decomposition}
\end{figure}
It was shown in \cite{souquet2020hyperparameters} that this approach is exceptionally beneficial in high dimensional space problems. Herein, our goal is to demonstrate that FDA can also provide accurate results in a low dimension dynamic optimization problems and can be successfully applied to solve a real-world task.
\section{Conclusions}
\label{section:Conclusions}
In this paper, we presented a real-life use case of applying FDA. The obtained results demonstrate the efficiency of the method for solving the problem of image alignment. For the future work, we plan to extend FDA to deal with higher dimensional real-world problems (\emph{e.g.} computer vision) to further validate the FDA's capabilities.
\section{Results}
\label{section:Results}
The Figure \ref{fig:loss_function_optimization} illustrates the optimization process. The red dashed line shows the best fitness error in the current period, and blue line gives the current fitness error obtained at a given evaluation.
\begin{figure}[h]
\centering
\includegraphics[width=0.5\textwidth]{images/camera_regristration_dfda_v1.png}
\caption{Graph providing the dynamic optimization experiments. In blue the fitness error (obtained from the loss function) can be seen. In red the best fitness error obtained so far during this period of time is shown. Particularly, it can be observed that a camera has moved over time 5 times (red peaks).}
\vspace{-1.5em}
\label{fig:loss_function_optimization}
\end{figure}
In Figure \ref{fig:blended_output} a pair of blended images is given in the same coordinate space showing that FDA has provided accurate results.
\begin{figure}[h]
\centering
\includegraphics[width=0.4\textwidth]{images/blend_image_TUN-99122-lq_avg_frame.png}
\hspace{2cm}
\includegraphics[width=0.4\textwidth]{images/blend_image_TUN-99141-lq_avg_frame.png}
\caption{Sets of blended images represented in the same space showing an accurate match.}
\vspace{-1.5em}
\label{fig:blended_output}
\end{figure}
\section{Problem Definition}
\label{section:Problem_definition}
In this real-world application, we aim to solve the dynamic optimization problem of aligning a camera image that has been shifting over time (see Figure \ref{fig:Image_sequence_1}).
\begin{figure}[h]
\centering
\includegraphics[width=0.225\textwidth]{images/2_TUN-99122-lq_avg_frame.png}
\includegraphics[width=0.225\textwidth]{images/1_TUN-99122-lq_avg_frame.png}
\hspace{1cm}
\includegraphics[width=0.225\textwidth]{images/2_TUN-99141-lq_avg_frame.png}
\includegraphics[width=0.225\textwidth]{images/1_TUN-99141-lq_avg_frame.png}
\caption{\textbf{Left:} Images obtained from the camera at time $t-1$. \textbf{Right:} Images obtained from the camera at time $t$. A shift in the camera can be observed including translation and rotation distortions.}
\label{fig:Image_sequence_1}
\end{figure}
To do so, we optimize the matrix $H$ (see Equation (\ref{eq:homography})) containing 8 Degrees of Freedom (DoF) that correspond to image geometrical transformation (such as translation, rotation, skewing).
\begin{equation}
H=
\begin{pmatrix}
DoF_{11} & DoF_{12} & DoF_{13} \\
DoF_{21} & DoF_{22} & DoF_{23} \\
DoF_{31} & DoF_{32} & 1 \\
\end{pmatrix}
\label{eq:homography}
\end{equation}
The direct and inverse transformation are given in Equation (\ref{eq:transformation})
\begin{equation}
\begin{split}
& I_{1} \cdot H = I_{2} \\
& I_{1} = I_{2} \cdot H^{-1}
\end{split}
\label{eq:transformation}
\end{equation}
Therefore, given two images $I_{1}$ and $I_{2}$ that have been captured using the same camera (from the same position, just changing its viewpoint), and that have a view overlap, it should exist a matrix $H$ that let us transform the content of one image into the other one. Specifically, the values in $H$: $DoF_{11}$, $DoF_{12}$, $DoF_{21}$ and $DoF_{22}$ estimate the rotation. Parameters $DoF_{13}$, $DoF_{23}$ estimate the translation. Finally, $DoF_{31}$, $DoF_{32}$ estimate the skewing effect. Parameter $H_{33}$ is used for the scale magnitude. Given the problem setup it can be fixed to a constant value of 1.
To perform estimation
of the $H$ matrix parameters, a set of point have been initially detected and described in both images. These points were then pairwise matched based on the descriptors similarity (see Figure \ref{fig:matched_keypoints} for an illustration)
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{images/sequence_1/matched_visual_keypoints_TUN-99122-lq_avg_frame.png}
\hspace{1cm}
\includegraphics[width=0.45\textwidth]{images/sequence_2/matched_visual_keypoints_TUN-99141-lq_avg_frame.png}
\caption{Sets of images containing the detected and matched keypoints.}
\vspace{-1.5em}
\label{fig:matched_keypoints}
\end{figure}
We formulate the IA as a dynamic optimization problem by considering the loss function $\mathcal{L}_{loss}$ that provides the (partial) sum of distances between pairs of matched keypoints after apply the transformation $H$ on one point from the pair (see Equation (\ref{eq:loss_function})).
\begin{equation}
\begin{split}
& error_k = d_{L_1} \left ( keypoint_{I_2}^k, keypoint_{I_1}^k \cdot H\right ) \\
& \mathcal{L}_{loss} = \sum_{error_k < P_{i}({errors})} error_k
\end{split}
\label{eq:loss_function}
\end{equation}
More specifically, we use $L_1$ distance between pairs of points. When calculating the loss function we take into account lower the distances up to a certain $i^{th}$ percentile $P_{i}$. For the experiments we found empirically the value of $i=80$ to perform well. We use the percentile to smooth the optimization function (\emph{e.g.} filter non-properly matched keypoints).
\section{Introduction}
Image Alignment (IA) has become an important task to be taken into account in real-world applications. IA is the process of overlaying images to be able to further analyze them. It is a crucial step because many systems rely on obtaining the correct formatted data to take a posteriory decision. This concept comes from a more generic one called image registration. IA was initially introduced by \emph{Lukas, Kanade et al.} in \cite{lucas1981iterative}.
In this work, we tackle the Dynamic Optimization Problem (DOP) of IA in a real-world application using a Dynamic Optimization Algorithm (DOA) called Fractal Decomposition Algorithm (FDA), introduced in \cite{nakib2017deterministic} by \emph{Nakib et al.}. We used FDA to perform IA on CCTV camera feed from a tunnel. As the camera viewpoint can change by multiple reasons such as wind, maintenance, etc. the alignment is required to guarantee the correct functioning of video-based traffic security system.
The rest of this paper is organized as follows. First, Section \ref{section:Problem_definition} introduces the problem formulation. Then, Section \ref{section:Methodology} recalls the FDA foundations. Afterwards, \ref{section:Results} introduces the conducted experiments and explains the obtained results. Finally, in Section \ref{section:Conclusions} the conclusions are presented.
\section{Methodology}
\label{section:Methodology}
To solve the stated problem, we propose to use FDA to deal with this dynamic environment.
More specifically, FDA uses a fractal decomposition structure based on hyperspheres (see Figure \ref{fig:FDA_decomposition}) to explore the search space, and an Intensive Local Search (ILS) method to exploit the promising regions.
\begin{figure}[h]
\centering
\includegraphics[width=0.3\textwidth]{images/FDA_decomposition.png}
\hspace{1cm}
\includegraphics[width=0.3\textwidth]{images/FDA_inflation.png}
\caption{\textbf{Left:} 4-level decomposition using the hypersphere fractal structure. \textbf{Right:} Representation of the hypersphere fractal decomposition with inflation coefficient to ensure that the search space is fully covered given a particular dimension.}
\vspace{-1.5em}
\label{fig:FDA_decomposition}
\end{figure}
It was shown in \cite{souquet2020hyperparameters} that this approach is exceptionally beneficial in high dimensional space problems. Herein, our goal is to demonstrate that FDA can also provide accurate results in a low dimension dynamic optimization problems and can be successfully applied to solve a real-world task.
\section{Conclusions}
\label{section:Conclusions}
In this paper, we presented a real-life use case of applying FDA. The obtained results demonstrate the efficiency of the method for solving the problem of image alignment. For the future work, we plan to extend FDA to deal with higher dimensional real-world problems (\emph{e.g.} computer vision) to further validate the FDA's capabilities.
\section{Results}
\label{section:Results}
The Figure \ref{fig:loss_function_optimization} illustrates the optimization process. The red dashed line shows the best fitness error in the current period, and blue line gives the current fitness error obtained at a given evaluation.
\begin{figure}[h]
\centering
\includegraphics[width=0.5\textwidth]{images/camera_regristration_dfda_v1.png}
\caption{Graph providing the dynamic optimization experiments. In blue the fitness error (obtained from the loss function) can be seen. In red the best fitness error obtained so far during this period of time is shown. Particularly, it can be observed that a camera has moved over time 5 times (red peaks).}
\vspace{-1.5em}
\label{fig:loss_function_optimization}
\end{figure}
In Figure \ref{fig:blended_output} a pair of blended images is given in the same coordinate space showing that FDA has provided accurate results.
\begin{figure}[h]
\centering
\includegraphics[width=0.4\textwidth]{images/blend_image_TUN-99122-lq_avg_frame.png}
\hspace{2cm}
\includegraphics[width=0.4\textwidth]{images/blend_image_TUN-99141-lq_avg_frame.png}
\caption{Sets of blended images represented in the same space showing an accurate match.}
\vspace{-1.5em}
\label{fig:blended_output}
\end{figure}
|
\section{Introduction}
Wide-range equations of state (EOS) for matter are demanded in various applications of high energy density physics and physics of plasmas~\cite{Fortov:PhysUspekhi:2009}.
Materials undergoing intense laser \cite{Anisimov:PhysUspekhi:1984,Krasyuk:LaserPhys:2016} or particle beam radiation~\cite{Hoffmann:LaserPartBeams:2005,Gnyusov:LaserPartBeams:2016}, a high- or hypervelocity impact~\cite{Altshuler:PhysUsp:1996}, a high current flow in X or Z-pinch~\cite{Shelkovenko:MRE:2018,McCoy:PRB:2017}, or the confinement fusion~\cite{Betti:NatPhys:2016} may propagate any path on a phase diagram starting from normal conditions. The results of quantitative simulations of such processes via computational fluid dynamics methods are very sensitive to the models used to evaluate thermodynamic and transport properties of matter. Such models have been developing for several decades~\cite{Zeldovich:2002}, however their applicability is restricted to certain regions on the phase diagram. Thus the development of more comprehensive approaches is very important~\cite{Bushman:SovPhysUsp:1983,Lomonosov:LPB:2007}.
The construction of a thermodynamically consistent EOS at extremes is complicated by the presence of free electrons due to the thermal or pressure ionization of atomic shells, which leads to the growth of Coulomb interaction. Its effect may be compared with the energy of thermal (classical) motion of electrons or Fermi energy in case of degenerate electrons using a corresponding coupling parameter $\Gamma = \langle E_{\mathrm{pot}} \rangle / \langle E_{\mathrm{kin}} \rangle = \max(\Gamma_c, \Gamma_q)$, where:
\begin{equation}
\Gamma_c = \frac{e^2 n_e^{1/3}}{k_B T},\quad \Gamma_q = \frac{e^2 n_e^{1/3}}{\varepsilon_F},
\label{eq:gamma_c}
\end{equation}
$k_B$ is the Boltzmann constant,
$\hbar$ is the reduced Planck constant,
$n_e$ is the electron density,
$m_e$ is the electron mass,
$T$ is the temperature,
$e$ is the electron charge,
$\varepsilon_F$ is the Fermi energy
\begin{equation}
\varepsilon_F = (3\pi^2)^{2/3}\frac{\hbar^2}{2m_e}n_e.
\end{equation}
Thus, the states of classical or degenerate matter may vary from an ideal system ($\Gamma_{c,q} \ll 1$) to a strongly coupled one ($\Gamma_{c,q} \gg 1$) and a wide-range model should be valid in all corresponding regions of the phase diagram.
Quantum statistical models have been developed in the framework of continuous electron distribution in a self-consistent field (Thomas--Fermi (TF) model~\cite{Thomas:1927,Fermi:1927} and its extensions~\cite{Feynman:PR:1949,Kirzhnits:SovPhysUsp:1975}) or direct evaluation of discrete spectrum and wave functions (Hartree--Fock--Slater (HFS) model and its extensions~\cite{Slater:PhysRev:1951,Rozsnyai:PRA:1972,Liberman:PRB:1979,Nikiforov_Novikov_Uvarov:2005}). Nowadays, the most advanced methods to evaluate the properties of electrons in matter are based on the Kohn--Sham density functional theory (KS-DFT)~\cite{Hohenberg:PR:1964,KS:1965}, which provides the electronic structure for an arbitrary distribution of ions~\cite{Kresse:PRB:1993}, so that a variety of properties may be derived, including EOS. However, at relatively high temperatures (hundreds of eV) the use of KS-DFT is hampered as it considers only the discrete spectrum of electrons. Orbital free density functional theory (OF-DFT)~\cite{Clerouin:PRA:1992} deals with the continuous spectrum of electrons which allows one to model only high-temperature plasmas, where the contribution of bound electrons can be neglected. The combination of OF and KS-DFT may lead to thermodynamic inconsistency, that is why preferable wide-range models should allow a consistent evaluation of the discrete and continuous spectra contributions. This requirement is satisfied in path-integral Monte Carlo (PIMC) models~\cite{Militzer:PRL:2000,Filinov:PRL:2001}, but the computations may be too expensive compared to KS or OF-DFT.
The implementation of multi-ion DFT algorithms with a proper evaluation of electron transitions from bound to free states is quite complicated. The single-ion limitation used in the average atom (AA) framework is more preferrable at developing various approaches for electron structure predictions, so that one may notice the growing interest to AA models in recent studies~\cite{Shemyakin:ContPlasmaPhys:2012,Ovechkin:HEDP:2016,Kadatskiy:HEDP:2019,Orlov:MRE:2019,Faussurier:2021,Aguiar:2021,CallowHansen:2022}.
In general, an AA model is less computationally expensive, and it may be extended to include the contribution of ions with a pseudo-atom approach~\cite{Starret:PRE:2016,Ovechkin:HEDP:2019}.
AA models have been developing for about a hundred years, so that many improvements have been introduced to the original approach of Thomas and Fermi: quantum and exchange effects~\cite{Dirac:1930,Kirzhnits:JETP:1957,Kirzhnits:JETP:1959,Kalitkin:1960}, shell effects~\cite{Zink:PR:1968,Kirzhnits:JETP:1972,Shpatakovskaya:SovPhysUsp:2012}, binding energy corrections~\cite{Scott:1952} and others.
An enhanced TF model allows one to obtain more realistic thermodynamic properties of electrons, but the approximate treatment of the electronic discrete spectrum as the continuous one limits the applicability of this model significantly~\cite{Dyachkov:PhysPlasmas:2014,Dyachkov:PhysPlasmas:2016}. At the same time the HFS~\cite{Liberman:PRB:1979,Nikiforov_Novikov_Uvarov:2005,Sinko:HEDP:2013,Ovechkin:HEDP:2016} or KS based AA models~\cite{CallowHansen:2022} provide a more realistic spectrum of electons, but for the cost of greater computational complexity. Moreover, the convergence of the self-consistent field cycle is slow (or even not achieved) for large atomic cells (rarefied plasma).
In our study we introduce an AA model, in which semiclassical wave functions for a spherically symmetric atomic cell are used instead of precise ones. The proper joining of discrete and continuous spectra~\cite{Nikiforov_Novikov_Uvarov:2005} provides the thermodynamic consistency, which is required for wide-range EOSs. The use of semiclassical approximation instead of the direct solution of the Shr\"odinger equation makes the model computationally efficient, while keeping a correct electronic structure and robust wide-range convergence during the self-consistent field evaluation. The obtained EOS is compared to the Saha model at the region of rarefied plasma, to KS-DFT in the region of warm dense matter (WDM), and to the TF model at high densities.
\section{Model formulation}
\subsection{Self-consistent field}
The AA approach suggests that electrons are distributed within a spherical cell of a radius $r_0$ with a nucleus at its center. The electrical neutrality of the cell assumes that the number of electrons equals to the nucleus charge $Z$; the total charge density outside the cell is supposed to be zero. Under these limitations, the electrostatic potential $U$ arising inside the cell can be determined from the Poisson equation:
\begin{equation}
\Delta U = 4\pi [n_e (\mathbf{r}) - n_i (\mathbf{r})],
\label{ZTTF:eq:poisson}
\end{equation}
where $n_e$ is the electron density, $n_i = Z\delta(\mathbf{r})$ is the ion density, $\delta(\mathbf{r})$ is the delta function. Another suggestion is the spherical symmetry of the potential and the electron density, so that by applying the appropriate boundary conditions one obtains the following boundary value problem:
\begin{equation}
\left\{
\begin{aligned}
&\frac{1}{r} \frac{\mathrm{d}^2}{\mathrm{d}r^2} \left(r U\right) = n_e(r), \\
&\left.rU(r)\right|_{r = 0} = Z, \quad U(r_0) = 0, \\
&\left.\frac{dU}{dr}\right|_{r = r_0} = 0. \\
\end{aligned}
\right.
\label{eq:boundary_problem}
\end{equation}
The electron-electron correlation and exchange interaction effects may be taken into account using the exchange-correlation functional, which is calculated in the local density approximation based on the formula proposed by Slater~\cite{Slater:1951}:
\begin{equation}
\label{eq:change_cor_func}
\Omega_{xc}[n] = \int_{0}^{r_0} 4 \pi r^2 F_{xc}(n(r)) \mathrm{d}r,
\end{equation}
so that the corresponding term
\begin{equation}
U_{xc}(r) = \frac{\delta \Omega_{xc} }{\delta n(r)} = \frac{\partial F_{xc}(n(r))}{\partial n(r)},
\end{equation}
is added to the potential
\begin{equation}
U_\text{eff}(r) = U(r) - U_{xc}(r).
\end{equation}
The presented general AA model suggests that the potential $U_\text{eff}(r)$ depends on the electron density $n_e(r)$ and vice versa.
Using additional assumptions about $n_e(r)$ one can obtain an explicit dependence $n_e(U)$: the well-known TF
approximation~\cite{Feynman:PR:1949, Kalitkin:1960} allows one to compute a rough solution consistent with the electroneutrality
condition. In a general case, one should find a solution to the problem~\eqref{eq:boundary_problem} iteratively, by applying gradual
changes to the potential and electron density until the desired smallness of energy or chemical potential residual is achieved~\cite{Nikiforov_Novikov_Uvarov:2005}; the TF approximation can be used as a first step of the iterative procedure.
\subsection{Discrete and continuous spectra}
The electron density of an AA can be represented as a sum of discrete $n_{d}(r)$ and continuous $n_{c}(r)$ parts. The discrete part is calculated as a quantum-statistical sum over single-electron states, while the continuous part can be obtained using the TF approximation~\cite{Nikiforov_Novikov_Uvarov:2005}:
\begin{equation}
n_e(r) = n_{d}(r)+n_{c}(r),
\label{eq:electron_density}
\end{equation}
\begin{equation}
n_{d}(r) =\frac{1}{4\pi r^2}\sum_{\varepsilon_{nl} < \varepsilon_0} \frac{2(2l + 1)}
{1 + \exp\left[(\varepsilon_{nl} - \mu)/T)\right]} R^2_{nl}(r),
\label{eq:electron_density_discrete}
\end{equation}
\begin{multline}
n_{c}(r) = \frac{2\sqrt{2}}{\pi^2}\int\limits_{\varepsilon_0}^{\infty} \frac{\sqrt{\varepsilon + U(r)} \mathrm{d}\varepsilon}{1 + \exp\left[(\varepsilon - \mu)/T\right]} = {}\\
\frac{(2T)^{\frac{3}{2}}}{2 \pi^2}\left[ I_{\frac{1}{2}}\left(\frac{U(r)+\mu}{T}\right) - I^{inc}_{\frac{1}{2}}\left(\frac{U(r)+\mu}{T}, y^*\right) \right].
\end{multline}
Here $n$ and $l$ are the principal and orbital quantum numbers, respectively, $R_{nl}(r)$ is the radial wave function of the $\{n, l\}$ state, $I_{\frac{1}{2}}$ and $I^{inc}_{\frac{1}{2}}$ are the complete and incomplete Fermi--Dirac integrals~\cite{Cody:MathComp:1967}, respectively, $y^{*} = \max\left\{(\varepsilon_{0} + U(r))/T,0 \right\}$, $\mu$ is the chemical potential, $\varepsilon_{0}$ is the boundary energy between the discrete and continuous states. It is also convenient to express the occupation numbers explicitly:
\begin{equation}
N_{nl} = \frac{2(2l + 1)}
{1 + \exp\left[(\varepsilon_{nl} - \mu)/T)\right]}.
\end{equation}
The total number of electrons is determined by the sum of the numbers of occupied states at given energies $\varepsilon_{nl}$ for some chemical potential:
\begin{multline}
\label{eq:number_of_states_discrete}
N(\mu) = \int\limits_0^{r_0}4\pi r^2 n(r) \mathrm{d}^3r = \int\limits_0^{r_0}4\pi r^2 [n_d(r) + n_c(r)] \mathrm{d}^3r\\ {} = \sum_{\varepsilon_{nl} < \varepsilon_0}N_{nl}(\mu) + \int\limits_{\varepsilon_0}^{\infty}N_{\varepsilon}(\mu) \mathrm{d}\varepsilon.
\end{multline}
As the spherical cell is electrically neutral, $N(\mu)$ should satisfy the equation $N(\mu) = Z$ which allows to determine an appropriate chemical potential $\mu$.
The boundary between the discrete and continuous spectra $\varepsilon_0$ should be chosen from the condition of the minimum of the grand thermodynamic potential~\cite{Sinko:TVT:1983,Nikiforov_Novikov_Uvarov:2005}:
\begin{multline}
\frac{8\sqrt{2}}{3\pi}\int_0^{r_0}\left[\max\{0,\varepsilon_0+U(r)\}\right]^{\frac{3}{2}}r^2\mathrm{d}r = \\ \sum_{\varepsilon_{nl} <\varepsilon_0} 2(2 l + 1).
\label{eq:boundary_energy}
\end{multline}
This boundary depends on the thermodynamic conditions of matter (density and temperature) which guarantees the thermodynamic consistency of the model.
\subsection{Wave functions}
Wave functions and energy levels in the AA framework are usually calculated via the one-electron Shr\"odinger equation~\cite{Joshua_Izaac:2018}, which is the most time-consuming part in a simulation. Thermodynamic properties are not too demanding on the accuracy of calculating energy levels, so the semiclassical approach by Jeffreys, Wentzel, Kramers, and Brillouin (JWKB method) can be applied. In this approximation, the radial part of a wave function can be expressed through the Bessel functions:
\begin{multline}
R^{(i)}_{nl}(r) = \\
\left\{
\begin{aligned}
& \frac{C_i}{\sqrt{3}}\sqrt{\frac{\xi_i}{|p|}}K_\frac{1}{3}(\xi_i) & (r \leq r_i), \\
& \frac{C_i}{\pi}\sqrt{\frac{\xi_i}{p}}\left[J_{-\frac{1}{3}}(\xi_i) + J_\frac{1}{3}(\xi_i)\right] & (r_i \leq r < r_o),
\end{aligned}
\right.
\label{eq:wave_function_inner}
\end{multline}
\begin{multline}
R^{(o)}_{nl}(r) = \\
\left\{
\begin{aligned}
& \frac{C_o}{\pi}\sqrt{\frac{\xi_o}{p}}\left[J_{-\frac{1}{3}}(\xi_o) + J_\frac{1}{3}(\xi_o)\right] & (r_i < r \leq r_o), \\
& \frac{C_o}{\sqrt{3}}\sqrt{\frac{\xi_o}{|p|}}K_\frac{1}{3}(\xi_o) & (r \geq r_o).
\end{aligned}
\right.
\label{eq:wave_function_outer}
\end{multline}
Here $J_{-1/3}(x)$, $J_{1/3}(x)$ are the Bessel functions of the first kind, $K_{1/3}(x)$ is the modified Bessel function of the second kind, $o$ and $i$ denote the outer and inner turning points, respectively,
\begin{equation}
\xi_i(r) = \left|\int\limits_{r_i}^{r}|p_{nl}(r')|\mathrm{d}r'\right|,\
\xi_o(r) = \left|\int\limits_{r_o}^{r}|p_{nl}(r')|\mathrm{d}r'\right|.
\end{equation}
The semiclassical radial wave functions oscillate in the region of classical motion $(r_i < r < r_o)$, and exponentially decay in the forbidden regions $r < r_i$ and $r > r_o$. The semiclassical momentum is expressed as
\begin{equation}
\label{eq:semiclassical_momentum}
p_{nl}(r) = \sqrt{2\left[\varepsilon_{nl} + U(r) - \frac{(l + 1/2)^2}{2 r^2}\right]}.
\end{equation}
The functions $R_{nl}^{(i)}(r)$ and $R_{nl}^{(o)}(r)$ have singularities at the turning points $r_o$ and $r_i$, respectively, which can be removed using the asymptotic expansion near $r_i$ and $r_o$. The smooth joining of the functions \eqref{eq:wave_function_inner},\eqref{eq:wave_function_outer} in the interval $(r_i, r_o)$ is performed as~\cite{Nikiforov_Novikov_Uvarov:2005}:
\begin{eqnarray}
\label{eq:wave-function-complete}
& R_{nl}(r) = [1 - a(r)]R^{(i)}_{nl}(r) + a(r)R^{(o)}_{nl}(r), \\
& a(r) = \xi_i(r)/\xi_i(r_o).
\end{eqnarray}
It is worth noting that the signs of the outer wave function $R^{(o)}_{nl}(r)$ and the inner one $R^{(i)}_{nl}(r)$ must be consistent:
\begin{equation}
C_i = C_o\,\mathrm{sign}\{J_{-1/3}[\xi_i(r_o)] + J_{1/3}[\xi_i(r_o)]\}.
\end{equation}
The value of the constant $C = |C_i| = |C_o|$ is determined by the wave function normalization:
\begin{equation}
\int_0^{r_0}|R_{nl}(r)|^2\mathrm{d}r = 1.
\end{equation}
\begin{figure*}[t]
\includegraphics[width=0.49\linewidth]{self-consistent-a.pdf}
\includegraphics[width=0.49\linewidth]{self-consistent-b.pdf}
\caption{Modeling of the self-consistent field in the atom of gold ($Z = 79$) at $T = 100\,$eV and $\rho = 19.3\,$g/cm$^3$. (a) Evaluation of the radial electron density distribution using the TF model and the developed semiclassical model. The latter produce the electron shells due to accounting for the discrete spectrum. (b) The deviation of the semiclassical potential from the Thomas--Fermi one. The series of dashed lines illustrate the process of convergence to the self-consistent field.}
\label{fig:image_iterations}
\end{figure*}
\subsection{Energy levels}
The semiclassical energy levels $\varepsilon_{nl}$ are calculated using the Bohr--Sommerfeld quantization condition:
\begin{equation}
S_{nl} = \int_{r_i}^{r_o}p_{nl}(r)\mathrm{d}r = \pi (n - l - 1/2).
\label{eq:bohr_sommerfeld_quantization}
\end{equation}
Here $S_{nl}$ is the action integral, $r_i$, $r_o$ are the inner and outer turning points, respectively, $p_{nl}(r)$ is the semiclassical momentum~\eqref{eq:semiclassical_momentum}.
It is well known that the semiclassical approach is valid only at large enough quantum numbers. Nevertheless, the accuracy of this approximation seems sufficient for describing various physical effects in a wide range of thermodynamics parameters.
Figure~\ref{fig:image_iterations} demonstrates the difference between the self-consistent electron density and potential obtained during the convergent iterative procedure, and the initial TF approximation for the atom of gold at normal density. The semiclassical electron density has notable oscillations corresponding to the contribution of different electron shells. The complicated electronic structure of gold also affects the self-consistent potential which deviates from the TF potential $U_{TF}(r)$.
\subsection{Thermodynamic functions}
\begin{figure*}[t]
\centering
\includegraphics[width=0.49\linewidth]{td_consistency_a.pdf}
~
\includegraphics[width=0.49\linewidth]{td_consistency_b.pdf}
\caption{Verification of the thermodynamic consistency of the developed semiclassical AA model. (a) Verification of identity~\eqref{eq:consT}: the relative deviation does not exceed $5$\% along the curves. (b) Verification of identity~\eqref{eq:consV}: the relative deviation does not exceed $1$\% along the curves.}
\label{fig:consistency}
\end{figure*}
Thermodynamic properties can be obtained by free energy differentiation. Such an approach provides thermodynamic consistency which is an important property for applications. The expression for the free energy $F$ of the AA electron subsystem can be written as a functional of electron density:
\begin{equation}
F[n] = K[n] + U[n] - TS[n] + \Omega_{xc}[n].
\label{eq:free_energy_aa_n}
\end{equation}
Here, $K[n]$ is kinetic energy, $U[n]$ is potential energy, $S[n]$ is entropy, $\Omega_{xc}[n]$ is an exchange-correlation functional. It can be shown that the free energy of the AA model can be rewritten in the following way, convenient for further derivations:
\begin{widetext}
\begin{multline}
\label{eq:free_energy_aa}
F = - T \sum_{n,l} 2(2l + 1)\ln{ \left[1 + \exp \left( \frac{\mu - \varepsilon_{nl}}{T}\right)\right]} + \frac{1}{2} \int_{0}^{r_0} 4 \pi r^2 n_e(r) \left(U(r) - \frac{Z}{r}\right) \mathrm{d}r + \mu N -\\- \int_{0}^{r_0} 4 \pi r^2 n_e(r) U_{xc}(r) \mathrm{d}r + \int_{0}^{r_0} 4 \pi r^2 F_{xc}[n_e(r)]\mathrm{d}r.
\end{multline}
By taking the derivatives, the consistent expressions for thermodynamic functions can be obtained:
\begin{multline}
\label{eq:pressure_qc}
P = - \left(\frac{\partial F}{\partial V}\right)_{T} = -\frac{1}{4 \pi r_0^3} \sum_{\varepsilon < \varepsilon_0} \frac{2l+1}{1+\exp[(\varepsilon_{nl} - \mu)/T]} \Biggr( R_{nl} R^{'}_{nl} + r R_{nl} R^{''}_{nl} - r (R^{'}_{nl})^2 \Biggr) \Biggr |_{r= r_0} +\\ \frac{(2T)^{5/2}}{6 \pi^2} \left[I_{\frac{3}{2}}\left( \dfrac{\mu}{T}\right) - I^{inc}_{\frac{3}{2}}\left( \dfrac{\mu}{T}, y^*\right) \right],
\end{multline}
\begin{multline}
\label{eq:entropy_qc2}
S = -\left(\frac{\partial F}{\partial T}\right)_{V,N} = -\frac{\mu N}{T} + 2 \sum_{\varepsilon < \varepsilon_0} (2l+1) \ln{\left[1 + \exp\left(\frac{ \mu - \varepsilon_{nl}}{T}\right) \right]} + \frac{2}{T}\sum_{\varepsilon < \varepsilon_0} \frac{(2l+1)\varepsilon_{nl}}{1+\exp\left[(\varepsilon_{nl} - \mu)/T\right]} - \\ - \frac{2\sqrt{2}}{3\pi^2 }\ln{\left[1+\exp\left(\frac{ \mu - \varepsilon_0}{T}\right) \right]} \int_{0}^{r_0} 4 \pi r^2 [\varepsilon_0 + U(r)]^{\frac{3}{2}}\mathrm{d}r + \frac{5\sqrt{2}T^{\frac{3}{2}}}{3\pi^2 } \int_{0}^{r_0} 4 \pi r^2 \mathrm{d}r \Biggr[I_{\frac{3}{2}}\left( \dfrac{U(r)+\mu}{T}\right) - \\ I^{inc}_{\frac{3}{2}}\left( \dfrac{U(r)+\mu}{T}, y^*\right) \Biggr] - \frac{1}{T} \int_{0}^{r_0} 4 \pi r^2 U(r) n_c (r) \mathrm{d}r - \int_{0}^{r_0} 4 \pi r^2 \left(\frac{\partial F_{xc}}{\partial T}\right)_{n(r)}\mathrm{d}r,
\end{multline}
\begin{multline}
\label{eq:energy_qc2}
E = F + T S = 2 \sum_{\varepsilon < \varepsilon_0 } \frac{(2l+1)\varepsilon_{nl}}{1+\exp[(\varepsilon_{nl} - \mu)/T]} + \frac{\sqrt{2} T^{5/2}}{ \pi^2}\int_{0}^{r_0} 4 \pi r^2 \mathrm{d}r \left[I_{\frac{3}{2}}\left( \dfrac{U(r)+\mu}{T}\right) - I^{inc}_{\frac{3}{2}}\left( \dfrac{U(r)+\mu}{T}, y^*\right) \right] + \\ + \int_{0}^{r_0}4 \pi r^2 n_d(r) U (r) \mathrm{d}r - \frac{1}{2}\int_{0}^{r_0} 4 \pi r^2 \left[U(r) + U_i(r) + U_{xc}(r) \right] n (r) \mathrm{d}r + \\ +\int_{0}^{r_0} 4 \pi r^2 F_{xc}[n(r)] \mathrm{d}r - T\int_{0}^{r_0} 4 \pi r^2 \left(\frac{\partial F_{xc} [n(r)]}{\partial T}\right)_{n(r)}\mathrm{d}r,
\end{multline}
\begin{equation}
y^{*} = \max\left(\frac{U(r) + \varepsilon_{0}}{T},0 \right).
\end{equation}
\end{widetext}
\subsection{Thermodynamic consistency}
\begin{figure*}[t]
\includegraphics[width=0.49\linewidth]{eos_saha_a.pdf}
\includegraphics[width=0.49\linewidth]{eos_saha_b.pdf}
\caption{Thermodynamic properties of rarefied aluminum plasma evaluated with the chemical model of plasma (dashed line) and the TF (dash-dotted line) and semiclassical AA (solid line) models for electrons. The ideal gas model for ions was used for the TF and AA models, $\rho = 2.7\times10^{-3}$~g/cm$^3$. (a) The compressibility factor of aluminum. (b) The difference between the energy and kinetic part of energy.}
\label{fig:eos_saha}
\end{figure*}
The first law of thermodynamics in the differential form $dE = TdS - PdV$ defines two useful constraints between thermodynamic functions. At $T = \text{const}$:
\begin{equation}
P = T\left(
\frac{\partial S}{\partial V}
\right)_T - \left(
\frac{\partial E}{\partial V}
\right)_T.
\label{eq:consT}
\end{equation}
At $V = \text{const}$:
\begin{equation}
T\left(
\frac{\partial S}{\partial T}
\right)_{V} = \left(
\frac{\partial E}{\partial T}
\right)_{V}.
\label{eq:consV}
\end{equation}
These identities can be used to check the thermodynamic consistency of the AA model.
In Figure~\ref{fig:consistency} we verify Eqs.~\eqref{eq:consT} and \eqref{eq:consV} numerically. The derivatives in \eqref{eq:consT} and \eqref{eq:consV} are calculated with finite differences to demonstrate the convergence of the model. At low temperatures the contribution of the discrete spectrum prevails so the sensibility of results to small changes of temperature or density is higher. Nevertheless, the relative error of identity~\eqref{eq:consT} along isochores does not exceed 5\% in a wide range of temperature, while~\eqref{eq:consV} is satisfied along isotherms even better with the relative error less than 1\%.
\section{Wide-range evaluation of thermodynamic properties}
The described semiclassical AA model can be applied in a wide range of temperatures and pressures. The validity of the model
can be established by a direct comparison with reference models: at low pressure and density of plasma a chemical model is
precise, for WDM DFT simulations show quite accurate results. The basic TF model is also considered to check
the asymptotic behavior of our semiclassical model at high pressures and temperatures.
\subsection{Low density isochores and chemical model of plasma}
In the low-density region, our semiclassical model is compared to the chemical model of plasma~\cite{Zeldovich:2002,Nikiforov_Novikov_Uvarov:2005} and the finite temperature TF model \cite{Feynman:PR:1949}. The chemical model is based on a set of Saha equations for non-degenerate electron gas:
\begin{multline}
\label{eq:saha}
Z_0 \frac{x_{j+1,p}}{g_{j,s}} = {}\\
\frac{2}{3}\sqrt{\frac{2}{\pi}} r_0^3 T^{\frac{3}{2}} \frac{g_{j+1,p}}{g_{j+1,s}}\exp{\left(-\frac{E_{j+1,p}-E_{j,s}}{T}\right)},
\end{multline}
where $Z_0$ is the average ion charge, $x_{j,s}$ is the fraction of ions with charge $j$ in the state $s$, $g_{j,s}$ is the statistical weight of the corresponding ion state, $E_{j,s}$ is the energy level of the corresponding ion state. Equations~\eqref{eq:saha} are combined with the conditions of electroneutrality:
\begin{equation}
\label{eq:saha:neutrality}
\sum_{j,s} x_{j,s} = 1,\
\sum_{j,s} j x_{j,s} = Z_0.
\end{equation}
Eqs.~\eqref{eq:saha},~\eqref{eq:saha:neutrality} together allow determining concentrations~$x_{j,s}$ for further calculations of thermodynamic properties:
\begin{equation}
\label{eq:saha:thermodynamic:P}
P = (Z_0 + 1)\frac{T}{V},
\end{equation}
\begin{equation}
\label{eq:saha:thermodynamic:E}
E = \frac{3}{2}(Z_0 + 1)T + \sum_{j,s} x_{j,s}(E_{j,s} - E_{0,0}).
\end{equation}
According to the chemical model, the pressure growth with temperature shown in Fig.~\ref{fig:eos_saha}a has a clear step-wise view caused by the sequential ionization of electron shells. Being thermally ionized, the free electrons contribute to the total pressure as separate particles. However, the shape of the pressure curve is affected much by the electron density distribution within the atomic cell. The characteristic peaks (shells) on the semiclassical density distribution (multiplied by $4\pi r^2$, see Fig.~\ref{fig:image_iterations}) are responsible for the step-wise ionization with the temperature growth, in agreement with the chemical model calculations. The continuous filling of TF electron states results in a smooth distribution of $4\pi r^2\rho(r)$ which does not show characteristic ``pressure steps'' on a low-density isochor during thermal ionization.
Similiar behavior is presented in Fig.~\ref{fig:eos_saha}b for the difference between energy and its kinetic part ($3PV/2$) for aluminum. There is a notable difference between the semiclassical model and the chemical model of plasma near $10\,$eV. On the other hand, at lower and higher temperatures, the models agree better. The reason for such a discrepancy is probably caused by the semiclassical approximation for evaluating the energy spectrum of bound states.
The energy levels of the outer 3 electrons (3s$^2$ and 3p$^1$), which are ionized first, are calculated with a greater relative error leading to a visible energy deviation compared to the chemical model of plasma.
\subsection{Normal density region with DFT reference}
\begin{figure}[t]
\center{\includegraphics[width=0.99\linewidth]{heat_capacity.pdf}}
\caption{Isochoric electron heat capacity of aluminum at $\rho = 2.7$~g/cm$^3$. The models are: 1) pseudopotential DFT approach (VASP) with 3s and 3p valent electrons~\cite{Levashov_2010}; 2) full potential linear-muffin-tin-orbital (FP-LMTO) DFT approach~\cite{Levashov_2010}; 3) Thomas--Fermi AA model; 4) the developed semiclassical AA model.}
\label{fig:normal_dens}
\end{figure}
WDM, which is usually formed under intense laser irradiation, represents partially ionized matter at near-normal density with strong coupling. The most accurate results in this region can be obtained using DFT. Electronic pressure in a dense medium is insensitive to the excitation of inner electrons while the behavior of isochoric heat capacity is largely determined by this effect. Thus, the results for heat capacity will significantly depend on the method of electronic structure calculation.
Fig.~\ref{fig:normal_dens} shows the isochoric heat capacity of aluminum calculated via the semiclassical model, the TF model, and the DFT approach~\cite{Levashov_2010}. DFT simulations were performed using pseudo-potential and full-electron approaches: the first one takes into account only valence electrons, while other electrons form an invariable core with a softer potential for better convergence; the second one considers all electrons and nuclei during a self-consistent field evaluation. As a result, the pseudopotential approach limits heat capacity growth
while other methods behave differently: the semiclassical and full-electron DFT approaches reproduce the sequential ionization of electron shells, and the TF model provides smooth heat capacity growth. There is a difference between the semiclassical model and the full-electron DFT approach in the ionization temperature of lower-level electrons, but both express very similar tendency.
\subsection{High pressure shock Hugoniot evaluation}
\begin{figure}[t]
\center{\includegraphics[width=0.99\linewidth]{hugoniot.pdf}}
\caption{Shock Hugoniot measurements~\cite{Vladimirov:1984,Rozsnyai:2001} compared to calculations with different models: PIMC model~\cite{Driver:2018} (dashed line), Thomas--Fermi AA model (dotted line), and semiclassical AA model (solid line).}
\label{fig:high_dens}
\end{figure}
Experiments on shock compression are the main source of information for materials at extremes. The Hugoniot equation~\cite{Zeldovich:2002} relates the states of uncompressed and compressed matter and together with an EOS defines the shock Hugoniot curve:
\begin{equation}
E(V, T) - E_0 = \frac{1}{2}\left(P(V,T)+P_0\right)\left(V_0 - V\right).
\label{eq:hugoniot_eq}
\end{equation}
Here, $P_0$, $V_0$, $E_0$ are the pressure, volume, and specific energy of the uncompressed state (before the shock front).
Fig.~\ref{fig:high_dens} shows the calculation of shock Hugoniot curve for aluminum using the semiclassical, TF, and PIMC~\cite{Driver:2018} models in comparison with the experimental data~\cite{Rozsnyai:2001,Vladimirov:1984}. An ideal gas model is used for the ions in the semiclassical and TF models. As can be seen, although the ionization effects are well described by the semiclassical AA model, quantitatively it differs greatly from the PIMC results. First of all, this is due to the choice of a simple model describing the contribution of the ions to the EOS. However, even a simple approximation for the ions allows us to show that the semiclassical approach is capable of describing shell effects in agreement with a more accurate model. Some experimental points available in the considered region of pressures~\cite{Vladimirov:1984} have huge error bars, so all models fit into the inaccuracy of measurements.
\section{Discussion}
\begin{figure}[t]
\center{\includegraphics[width=0.99\linewidth]{internal_energy_wide.pdf}}
\caption{The wide-range evaluation of the internal energy in lead along isotherms at $1$, $10$, and $100\,$eV. Comparison of the developed semiclassical AA model (SC) with the chemical model of plasma at low densities, DFT calculations (VASP), basic Thomas--Fermi AA model (TF), and the Thomas--Fermi AA model with quantum and exchange corrections (TFC).}
\label{fig:pb_pressure}
\end{figure}
An example of a wide-range calculation using the semiclassical AA model, along with other models, is demonstrated in Fig.~\ref{fig:pb_pressure}. Pressure on three isotherms for lead (1, 10 and 100~eV) is shown in the range of density from $10^{-2}$~g/cm$^3$ to $10^3$~g/cm$^3$. At low density, the chemical model of plasma agrees quantitatively with the presented approach at 1 and 100~eV. At 10~eV, however, there is a notable overestimation of energy obtained with the semiclassical AA model. As it was mentioned before, this can be connected with a systematic excess of semiclassical energy levels over the corresponding HFS ones. In the region near the normal density, the semiclassical model predicts the position of the ionization minimum more accurately than the TF one and its modification with quantum and exchange corrections; in the last case the corrections become invalid near the normal density of lead \cite{Dyachkov:PhysPlasmas:2014}. As the temperature rises, the semiclassical model still agrees better with the DFT method than the TF approaches. In the presented diagram TF and TFC models~\cite{Nikiforov_Novikov_Uvarov:2005} can give reasonable results only at high density and high temperatures (in the region of weakly coupled quantum plasma \cite{Dyachkov:PhysPlasmas:2016}, where the continuous electron spectrum determines the values of thermodynamic functions due to ionization. The semiclassical model tends to the TF one in the region of high densities. Thus, the developed semiclassical AA model is able to predict thermodynamic properties of matter with reasonable accuracy at the reduced computational cost.
\section{Conclusion}
In this study the semiclassical AA model is introduced. The model considers the discrete and continuous spectrum of electrons in a thermodynamically consistent way.
The use of semiclassical energy levels and wave functions reduces the computational cost while keeping adequate accuracy and provides stable convergence of the self-consistent potential. The combined treatment of bound and free electrons in a single model is essential for modelling ionization and can be used as an integral part of a wide-range AA model.
The results obtained via the semiclassical AA model demonstrate a reasonable agreement with a chemical model of plasma in the region of rarefied plasma, with DFT in the region of WDM, and with shock Hugoniot data. The main drawback of the semiclassical approach is that the energy levels of valence electrons may slightly differ from the exact ones, but it is usually acceptable for thermodynamical functions evaluation. The presented model may also be used as a part of the pseudo-atom approach~\cite{Starret:PRE:2016} in order to accurately calculate the contribution of ions to plasma properties.
\section*{Acknowledgments}
The authors acknowledge the JIHT RAS Supercomputer Centre, the Joint Supercomputer Centre of the Russian Academy of Sciences,
and the Shared Resource Centre ``Far Eastern Computing Resource''
IACP FEB RAS for providing computing time.
|
\section{Introduction}
Multi-agent systems have attracted considerable attention in many applications due to their natural parallelization, general adaptability, and ability to self-organize \cite{Oh2017}. Such systems have a nonclassical information structure \cite{Malikopoulos2021} and impose several challenges in learning and control \cite{Malikopoulos2022a}.
A recent push in constraint-driven control has brought the idea of long-duration autonomy to the forefront of multi-agent systems research \cite{Egerstedt2018RobotAutonomy}.
As the real-world deployment of robotic systems continues \cite{mahbub2020sae-1,Vasarhelyi2018OptimizedEnvironments} these robotic agents will be left to to interact with their environment on timescales significantly longer than what can be achieved in a laboratory setting.
In addition, constructing a swarm with a large number of robots imposes significant cost constraints on each individual.
For robotic swarms to become viable, we need to develop and employ energy-optimal control techniques under strict cost constraints.
Thus, any approach to long-duration autonomy must emphasize safe energy-minimizing control policies that are driven by interactions with the environment.
Generating emergent flocking behavior has been of particular interest since Reynolds proposed three heuristic rules for multi-agent flocking in computer animation \cite{Reynolds1987}.
In aerial systems, the main energy savings comes from upwash, i.e., trailing regions of upward momentum in the slipstream, which followers exploit to reduce induced drag.
Flocking to minimize energy consumption is known as line flocking in the engineering literature \cite{Beaver2020AnFlockingb}, and its name comes from the linear formation-like flocking behavior of geese, pelicans, etc \cite{Bajec2009OrganizedBirds}.
The simplest method to achieve a V formation may be to generate an optimal set of formation points based on the aerodynamic characteristics of each agent.
This effectively transforms the line flocking problem into a formation reconfiguration problem, where each agent must assign itself to a unique goal and reach it within some fixed time \cite{Nathan2008V-likeBirds}.
The physical effects of V formation flight were explored in recent article \cite{Mirzaeinia2019EnergyReconfiguration}, where the authors demonstrate that the leading and trailing agents consume energy at a significantly higher rate.
This implies that these agents are the limiting factor in the total distance traveled, and the authors propose a formation-reconfiguration scheme based on a load-balancing protocol.
However, formation reconfiguration approaches generally require the formation to be computed offline, and while some articles consider agent heterogeneity (e.g., age, weight, size, and efficiency) \cite{Mirzaeinia2020AnPurposes}, this has not yet been explicitly considered in an engineering context.
Furthermore, formation points must be recalculated online if an agent enters or leaves the system, or if there are significant changes in the ambient environment.
A more flexible approach is to treat line flocking as a data-driven problem, where agents measure the local aerodynamic and hydrodynamic interactions to dynamically position themselves in the upwash field and save energy.
This has been achieved for aerial vehicles in $\mathbb{R}^2$ using a model predictive control approach \cite{Yang2016LoveControl}, where the authors construct a multi-objective optimization problem that minimizes speed differences, maximizes upwash benefit, and minimizes the field of view occlusion between agents.
The authors also demonstrate that solving this multi-objective optimization problem yields emergent V formations, even when the agents are initialized randomly.
A recent review of related optimal flocking techniques is presented in \cite{Beaver2020AnFlockingb}.
Our approach, in contrast to existing work, is constraint-driven.
In our framework, agents seek to travel as efficiently as possible subject to a set of task and safety constraints.
This set-theoretic approach to control is interpretable, i.e., the cause of an agent's action can be deduced by examining which constraints become active during operation.
By examining the conditions that lead to an empty feasible space, our framework also addresses the problem of constraint compatibility, i.e., how each agent ought to behave when its feasible action space is over-constrained.
Our approach is decentralized, and thus it is well-suited to ``open systems," where agents may suddenly enter, leave, or experience failure.
In this article, we describe a new \textit{anseroids} behavior (anserine-oid, meaning `goose-like') that generates dynamic echelon and V formations without any knowledge of the total number of agents in the system, and which are not given any information about the desired formation shape.
To the best of our knowledge, the only results similar to ours are these reported in \cite{Yang2016LoveControl,Roy2020LearningV-formation}.
The former uses particle swarm optimization combined with model predictive control to solve an optimal control problem. However, their approach depends on a multi-objective optimization problem with four components, and they provide no guarantees on the emergence of flocking behavior.
The latter employs reinforcement learning to generate conditions that ensure a V formations does not occur.
The contributions of this article are as follows:
\begin{itemize}
\item The first optimal control algorithm that demonstrates emergent V formations as a result of an energy-minimizing control policy,
\item a physics-based flocking model where agent decisions are coupled through aerodynamic interactions,
\item an interpretable set-theoretic control architecture that intuitively describes the optimal behavior of each agent in the flock,
\item a switching control policy that guarantees a solution to the constrained optimal control problem always exists, and
\item compelling evidence that demonstrate how energy savings is enhanced when heterogeneity is introduced to the flock.
\end{itemize}
The remainder of the article is organized as follows.
In Section \ref{sec:problem}, we discuss our notation and present the dynamics of our problem.
We present our optimal control problem and guarantees on its behavior in Section \ref{sec:control}, and in Section \ref{sec:validation} we validate our results in simulation and experimentally.
In particular, we simulate $2$ UAVs in Section \ref{sec:matlab} to show the aerodynamic interactions between agents, and in Section \ref{sec:experiment} we simulate $11$ UAVs in CrazySwarm and validate our controller experimentally with $5$ drones.
Finally, we summarize our results and directions for future research in Section \ref{sec:conclusion}.
\section{Problem Formulation} \label{sec:problem}
\subsection{Note on Notation}
Most references on optimal control, e.g., \cite{Bryson1975AppliedControl,Ross2015}, consider centralized problems.
Thus, directly adopting their notation may lead to ambiguities about the state space of a decentralized problem.
To relieve this tension, we take the following approach for an agent with index $i$:
endogenous variables, e.g., the position of agent $i$, are written without an explicit dependence on time, while exogenous variables, e.g., the position of agent $j$ as measured by agent $i$, are written with an explicit dependence on time.
This notation is common in the applied mathematics literature \cite{Levine2011OnFlatness}, and makes it explicitly clear how functions evolve with respect to the state (e.g., state dynamics) and how they evolve with respect to time (e.g., external signals measured by the agent).
\subsection{System Dynamics}
We consider a fleet of $N\in\mathbb{N}, N \geq 3$ fixed-wing uncrewed aerial vehicles (UAVs) indexed by the set $\mathcal{A} = \{1, 2, \dots, N\}$.
We denote the state of each UAV $i\in\mathcal{A}$ by
\begin{equation}
\bm{x}_i \coloneqq
\begin{bmatrix}
\bm{p}_i \\ \theta_i
\end{bmatrix},
\end{equation}
where $\bm{p}_i\in\mathbb{R}^2$ is the UAV's position and $\theta_i\in\mathbb{R}$ is the UAV's heading angle.
Each UAV obeys unicycle dynamics,
\begin{equation} \label{eq:dynamics}
\begin{aligned}
\dot{\bm{p}}_i &=
\begin{bmatrix}
v_i \cos{\theta} \\
v_i \sin{\theta}
\end{bmatrix}, \\
\dot{\theta}_i &= \omega_i,
\end{aligned}
\end{equation}
where $v_i\in\mathbb{R}_{>0}$ and $\omega_i\in\mathbb{R}$ are the control variables, which correspond to the linear and angular speed of UAV $i$.
We impose the control bounds,
\begin{equation}
\begin{aligned} \label{eq:bounds}
|\omega_i| &\leq \omega_{\max}, \\
0 < v_{\min} \leq v_i &\leq v_{\max},
\end{aligned}
\end{equation}
where $\omega_{\max}$ is the maximum turning rate and $v_{\min} < v_{\max}$ correspond to the minimum and maximum air speed.
Finally, the total drag force acting on UAV $i$ has the form,
\begin{align} \label{eq:drag}
F_i(\bm{x}_i, t) &= C_1 v_i^2 + \frac{C_2}{v_i^2} - \frac{L}{v_i} W(\bm{p}_i, t),
\end{align}
where $C_1, C_2 \in\mathbb{R}_{>0}$ capture the profile and self-induced drag that include the drag coefficient, air density, and wing area.
The function $W : \mathbb{R}^2\times\mathbb{R} \to \mathbb{R}$ describes the scalar upwash field, which we formally define in Section \ref{sec:control}.
In our modeling framework above, we impose the following assumptions.
\begin{assumption} \label{smp:tracking}
Each UAV is equipped with a low-level flight controller that can track the sequence of control actions that we generate.
\end{assumption}
\begin{assumption} \label{smp:stillAir}
The UAVs are operating in still air with constant aerodynamic properties.
\end{assumption}
We impose Assumptions \ref{smp:tracking} and \ref{smp:stillAir} to determine the behavior of the system in idealized conditions. However, these assumptions are not restrictive on our analysis.
In particular, applying adaptive and robust control techniques, such as control barrier functions \cite{Wang2017SafetySystems} or Gaussian Processes \cite{chalaki2021RobustGP}, can be used to overcome the resulting model mismatch.
\begin{assumption} \label{smp:collision}
Collision avoidance between UAVs in $\mathcal{A}$ can be neglected.
\end{assumption}
Generally, V formations have significant space between individuals without opportunities for collisions between agents \cite{Mirzaeinia2019EnergyReconfiguration,Mirzaeinia2020AnPurposes}.
Thus, we impose Assumption \ref{smp:collision} to focus the scope of our work on the emergence of the V formation.
Furthermore, aerodynamic effects disincentive UAVs from approaching too closely, and collision avoidance can always be guaranteed by introducing pairwise collision avoidance constraints \cite{Beaver2020Energy-OptimalConstraints} or employing a control barrier function as a safety layer \cite{Wang2017SafetySystems}.
\begin{assumption} \label{smp:main}
There exists a global heading angle $\theta_g$ and a small tolerance $\epsilon\in\mathbb{R}_{>0}$ such that $|\theta_i - \theta_g| < \epsilon$ for all UAVs $i\in\mathcal{A}$.
\end{assumption}
We impose Assumption \ref{smp:main} to simplify our analysis of the aerodynamics.
First, it allows us to consider spanwise cuts of the domain, which reduces our analysis from 2D to 1D.
Second, it allows us to model the wake as a scalar field centered on each UAV instead of modeling the wake evolution numerically, i.e., using computational fluid dynamics.
This assumption is common in the multi-UAV literature \cite{Yang2016LoveControl,Mirzaeinia2019EnergyReconfiguration}, although it is usually not stated explicitly.
We impose this assumption as a constraint in our final control algorithm, and it can be interpreted as the ``migratory urge'' proposed by Reynolds \cite{Reynolds1987}; the direction $\theta_g$ could also be computed using consensus, and some agents could simply separate themselves from the flock if Assumption \ref{smp:main} becomes too restrictive.
\subsection{Wake Model} \label{sec:aero}
Under Assumption \ref{smp:main}, we model the wake of each UAV $i\in\mathcal{A}$, as a scalar field centered at $\bm{p}_i$ and aligned with $\theta_i$.
Physically, the upwash field is a consequence of the pressure difference between the top and bottom of the wing \cite{Anderson2017FundamentalsAerodynamics}.
This induces a vortex at the wing tips, which generates an upward velocity (i.e., upwash) far from the wing and downward velocity (i.e., downwash) behind the wing.
Classically the wingtip vortices have been modeled using irrotational flow \cite{Mirzaeinia2019EnergyReconfiguration,Anderson2017FundamentalsAerodynamics,Karamcheti1980PrinciplesAerodynamics}.
However, this model is known to cause nonphysical behavior.
In particular, the vertical air speed approaches infinity at the wing tips.
In this work we augment the irrotational vortex model with a rotational core, which drives the velocity to zero at the wing tips.
Namely, each vortex induces the upwash velocity,
\begin{equation} \label{eq:spanAero}
u_i(r) =
\begin{cases}
\frac{\Gamma}{2\pi r} & \text{ if } |r| \geq r^*, \\
\Omega r & \text{ if } |r| \leq r^*,
\end{cases}
\end{equation}
where $u_i$ is the vertical airspeed, $r$ is the distance to the vortex center, $\Gamma$ is the circulation of the irrotational vortex, $\Omega$ is the angular rotation speed of the rotational core, and
$r^*$ satisfies
\begin{equation}
r^* = \Big(\frac{\Gamma}{2\pi\Omega}\Big)^{\frac{1}{2}}.
\end{equation}
Note that $r^* << b$ for physical systems, and under Assumption \ref{smp:main}, the induced velocity field has the form,
\begin{equation} \label{eq:spanwise}
f(y) = u_i(y-b) - u_i(y+b),
\end{equation}
where $y$ is a relative spanwise position and $2b$ is the wingspan.
The upwash distribution is shown in Fig. \ref{fig:upwash}.
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth]{resources/upwash.png}
\caption{Upwash velocity induced in the spanwise direction due to the wing tip vortices.}
\label{fig:upwash}
\end{figure}
In the streamwise, i.e., longitudinal direction, the wing interactions quickly coalesce into the two strong wingtip vortices.
As the vortices move aft from the wing, their influence approximately doubles.
Modeling the streamwise behavior of the vortices is a challenging problem; we adopt the approach of \cite{Yang2016LoveControl} using a gaussian function,
\begin{equation} \label{eq:streamwise}
g(x) = 2 \exp{-\frac{(x-\mu)^2}{2\sigma^2}},
\end{equation}
where $\mu$ determines the location of maximum upwash benefit, $\sigma$ determines the length of the wake, and $x$ is the relative streamwise position,.
Finally, we define the relative position vector,
\begin{equation}
\bm{s}_{ij}(\bm{p}_i, t) = \bm{p}_i - \bm{p}_j(t),
\end{equation}
and combine \eqref{eq:spanwise} and \eqref{eq:streamwise} to find an expression for the magnitude of UAV $j$'s upwash on $i$,
\begin{equation}
w_i^j(\bm{p}_i, t) \coloneqq f\Big(\bm{s}_{ij}(\bm{p}_i, t)\cdot\hat{y}\Big)\cdot g\Big(\bm{s}_{ij}(\bm{p}_i, t)\cdot\hat{x}\Big),
\end{equation}
where $\hat{x}, \hat{y}$ are unit vectors aligned with and perpendicular to $\theta_{j}(t)$, respectively.
\section{Optimal Feedback Controller} \label{sec:control}
To simplify our analysis of the system-level behavior, we impose the following pairwise assumption on the UAV positions.
\begin{assumption} \label{smp:pairwise}
For each UAV $i\in\mathcal{A}$, there is at most one UAV $j\neq i$ such that the upwash force $w_i^j(\bm{p}_i, t)$ is not negligible.
\end{assumption}
Intuitively, Assumption \ref{smp:pairwise} requires the UAVs to be sufficiently `close' to a V or echelon formation.
This makes our analysis tractable, as each UAV must only consider the influence of its immediate leader.
We extend the results of this section to the general case in Section \ref{sec:experiment}, where the UAVs are simulated in 2D space and the agents' initial conditions do not satisfy Assumption \ref{smp:pairwise}.
Finally, for completeness, we define the necessary conditions for a flock of UAVs to remain stable under Assumption \ref{smp:main}.
\begin{definition} \label{def:stable}
A necessary condition for the flock of UAVs is \emph{stable} at time $t_0$ under Assumption \ref{smp:main} is, for each UAV $i\in\mathcal{A}$,
\begin{equation}
v_i(t) = v_i(t_0), \quad \theta_i(t) = \theta_j(t)
\end{equation}
for all $t \geq t_0$.
\end{definition}
\subsection{Drag Minimization}
As a first step, we seek the control input $v_i$ that minimizes the drag force on UAV $i\in\mathcal{A}$, i.e.,
\begin{equation} \label{eq:fDot}
\frac{\partial F_i}{\partial v_i} = 2C_1 v_i - 2 \frac{C_2}{v_i^3} + \frac{L}{v_i^2} W_i(\bm{p}_i, t) = 0.
\end{equation}
Rearranging terms and multiplying by $v_i^3$ yields,
\begin{equation} \label{eq:vMinDrag}
v_i^4 + \frac{L}{2 C_1} W_i(\bm{p}_i,t) v_i - \frac{C_2}{C_1} = 0.
\end{equation}
\begin{remark}
Note that \eqref{eq:vMinDrag} minimizes the drag experienced by UAV $i\in\mathcal{A}$, which generally maximizes the distance traveled by the UAV per unit of energy expended.
Alternatively, one could minimize the power lost to drag by considering the product of \eqref{eq:drag} and $v_i$. This generally maximizes the flight time of the UAV.
The following analysis holds for both cases.
\end{remark}
\begin{lemma} \label{lma:existance}
There is a unique real positive speed that minimizes the drag experienced by each UAV $i\in\mathcal{A}$.
\end{lemma}
\begin{proof}
The optimal airspeed for UAV $i$ is the solution to \eqref{eq:vMinDrag}, a quartic function of $v_i$ with the discriminant
\begin{equation}
\Delta_4 = - \Bigg(256 \frac{C_2}{C_1} + 27\Big(-\frac{L}{2C_1}W_i(\bm{p}_i, t)\Big)^4 \Bigg) > 0,
\end{equation}
which implies that \eqref{eq:vMinDrag} has two complex conjugate roots and two real roots.
The imaginary roots satisfy,
\begin{equation} \label{eq:imaginaryRoots}
v_i^2 + b v_i + c = 0, \quad b^2 - 4c < 0,
\end{equation}
and thus $c > 0$ .
Next, polynomial long division of \eqref{eq:vMinDrag} on \eqref{eq:imaginaryRoots} yields a quadratic form for the real roots and additional conditions on $a$ and $b$, i.e.,
\begin{align}
v_i^2 - bv + b^2 - c &= 0, \label{eq:realRoots} \\
2bc + \frac{L}{2C_1}W_i(\bm{p}_i, t) - b^3 &= 0, \\
c^2 - b^2 c - \frac{C_2}{C_1} &= 0. \label{eq:lma1C}
\end{align}
Condition \eqref{eq:lma1C} implies,
\begin{equation}
c(c - b^2) = \frac{C_1}{C_2} > 0,
\end{equation}
which, in turn, implies $c > b^2$.
Finally, applying the quadratic equation to \eqref{eq:realRoots} yields,
\begin{equation}
v_i = \frac{b \pm \sqrt{-3b^2 + 4c}}{2}.
\end{equation}
Multiplying the two real roots yields,
\begin{equation}
\frac{1}{4}\Big(4b^2 - 4c \Big) = b^2 - c < 0.
\end{equation}
Thus, the two real solutions to $v_i$ have opposite signs, and \eqref{eq:vMinDrag} has exactly one real positive solution.
\end{proof}
Note that following the proof of Lemma \ref{lma:existance} it is possible to derive the optimal airspeed analytically; however, this is beyond the scope of this article.
Our next result characterizes how the upwash benefit affects the optimal airspeed of each UAV.
\begin{lemma} \label{lma:upwashSpeed}
The optimal airspeed of UAV $i$ decreases when gaining an upwash benefit and increases when experiencing an upwash cost.
\end{lemma}
\begin{proof}
Consider a UAV $i\in\mathcal{A}$ flying in isolation.
In this case $W(\bm{p}_i, t) = 0$, and the optimal airspeed arises when \eqref{eq:fDot} is satisfied, i.e.,
\begin{equation}
v_i^* = \Big(\frac{C_2}{C_1}\Big)^{1/4}.
\end{equation}
Substituting this into \eqref{eq:fDot} and rearranging terms implies,
\begin{equation}
v_i^4 - (v_i^*)^4 = -\frac{L}{2 C_1} v_i W_i(\bm{p}_i, t).
\end{equation}
Thus, as $v_i > 0$ from Lemma \ref{lma:existance}, $W_i < 0$ (upwash cost) implies $v_i > v_i^*$ and $W_i > 0$ (upwash benefit) implies $v_i < v_i^*$.
\end{proof}
\begin{lemma} \label{lma:unstableUpwash}
Under Assumption \ref{smp:main}, flying at the optimal airspeed while receiving an upwash benefit is unstable (Definition \ref{def:stable}), and each UAV will end up with no upwash benefit after finite time.
\end{lemma}
\begin{proof}
For a UAV formation to be stable, it is necessary that $v_i = v_j,$ $\theta_i = \theta_j$ for all UAVs, $i,j\in\mathcal{A}$.
Under our premise, this implies $v_i^* = v_j^*$ for all $i,j\in\mathcal{A}$ and,
\begin{align*}
(v_i^*)^4 + \frac{L}{2 C_1} W_i(\bm{p}_i, t) v_i^* &- \frac{C_2}{C_1} =
0 \\
&= (v_j^*)^4 + \frac{L}{2 C_1} W_j(\bm{p}_j, t) v_j^* - \frac{C_2}{C_1},
\end{align*}
which implies,
\begin{equation} \label{eq:nec1}
W_i(\bm{p}_i, t) = W_j(\bm{p}_j, t).
\end{equation}
The latter is an equivalent necessary condition for stability.
Without loss of generality, let $\theta_i = \theta_j = 0$ for all $i, j \in \mathcal{A}$.
Consider the case when UAV $i$ is ahead of another UAV $j$, i.e.,
\begin{equation} \label{eq:ahead}
(\bm{p}_i - \bm{p}_j) \cdot
\begin{bmatrix}
1 \\ 0
\end{bmatrix}
> 0,
\end{equation}
where $W_i = W_j \neq 0$, and the upwash effect of $i$ upon $j$ is not negligible.
In this case, the streamwise asymmetry of the upwash field requires the existence of another UAV $k$ such that $k$ as a negligible impact on $j$ (Assumption \ref{smp:pairwise}) and a significant impact on $i$; this can only be achieved if UAV $k$ is ahead of $i$ in the streamwise direction.
Repeating this process implies an uncountably infinite number of UAVs, which contradicts our premise that $N$ is finite.
\end{proof}
\begin{theorem} \label{thm:noV}
Under Assumption \ref{smp:main}, flying at the optimal air speed never leads to an emergent V formation.
\end{theorem}
\begin{proof}
Let every UAV $i\in\mathcal{A}$ have a state such that the position $\bm{p}_i$ corresponds to a V formation where no two agents overlap, each UAV receives some upwash benefit, and Assumption \ref{smp:pairwise} is satisfied.
Let $j$ denote the frontmost agent and $i$ satisfy $w_i^j(\bm{p}_i, t) > 0$.
Then, by the asymmetry of the streamwise upwash benefit \eqref{eq:streamwise}, $W_i > W_j$.
Thus, by Lemma \ref{lma:upwashSpeed}, $v_i^* < v_j^*$, and the distance $||\bm{s}_{ij}(\bm{p}_i, t)||$ is increasing.
This implies that the UAVs will travel at different optimal air speeds until the V formation breaks apart.
\end{proof}
Theorem \ref{thm:noV} demonstrates that simply flying at the energy-optimal airspeed can never lead to emergent line flocking!
This highlights a significant shortcoming within applying the robot ecology approach \cite{Egerstedt2021RobotAutonomy} to line flocking.
Specifically, if each UAV $i\in\mathcal{A}$ minimizes its acceleration subject to a constraint that $v_i$ matches $v_i^*$ as closely as possible, then an energy-saving V formation can not occur.
A recent game-theoretic approaches has run into similar challenges \cite{Shi2021AreFormation}.
Therefore, rather than minimizing the ``energy'' spent to actuate by minimizing acceleration, we propose that each agent ought to minimize its ``locomotive power'' expended through the cost function,
\begin{align} \label{eq:costFunction}
J(\bm{x}_i) &= \Big(\frac{v_i - v_i^*}{v_{\max} - v_{\min}}\Big)^2 + \Big(\frac{\omega_i}{\omega_{\max}}\Big)^2,
\end{align}
where $v_i^*$ is the unique optimal airspeed (Lemma \ref{lma:existance}) and both cost components are normalized because they are orthogonal control vectors.
Employing the cost \eqref{eq:costFunction} is a subtle change from the state of the art \cite{Egerstedt2021RobotAutonomy}, but it ends up playing a critical role in the generation and stabilization of emergent V formations.
\subsection{Wake Interactions}
Next, we explore the wake interaction of UAV $j$ on $i \in\mathcal{A}$ using our simplified aerodynamic model.
To simplify our notation, we use the scalars $x_i$ and $y_i$ to denote the relative position of $i$ with respect to $j$ in the streamwise and spanwise directions, respectively.
First, we estimate how UAV $i$ tends to roll due to the local flow field by evaluating the integral,
\begin{equation} \label{eq:moment}
m_i(x_i, y_i, t) = g(x_i) \int_{y_i-b}^{y_i+b} (\xi-y_i) f(\xi, t) d\xi.
\end{equation}
Similarly, we can estimate the lift induced on the wing through momentum transfer,
\begin{equation} \label{eq:force}
w_i(x_i, y_i, t) = g(x_i) \int_{y_i-b}^{y_i+b} f(\xi, t) d\xi.
\end{equation}
Note that both \eqref{eq:moment} and \eqref{eq:force} have analytical closed-form solutions, as the irrotational flow far from the wing-tip is integrable and it transitions to an affine function near the wing tip.
The resulting upwash and tendancy to roll functions are visualized in Fig. \ref{fig:moment}.
Taking the sum of \eqref{eq:force} over all UAVs determines the aggregate upwash effect on $i$,
\begin{equation} \label{eq:fields}
\begin{aligned}
W_i(\bm{p}_i, t) = \sum_{k\in\mathcal{A}\setminus\{i\}} w_k(\bm{p}_i, t), \\
M_i(\bm{p}_i, t) = \sum_{k\in\mathcal{A}\setminus\{i\}} m_k(\bm{p}_i, t),
\end{aligned}
\end{equation}
where $\bm{p}_i$ must be projected onto the streamwise and spanwise components of UAV $k$ to yield $x_i$ and $y_i$.
Note that \eqref{eq:fields} should be multiplied by a constant to compute the upwash force in the correct units.
However, our results only rely on the sign of the derivative of each term, so any positive scaling factor is irrelevant to our analysis.
Furthermore, the sums of \eqref{eq:fields} contain at most one term under Assumption \ref{smp:pairwise}.
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth]{resources/upwash-force-moment.png}
\caption{Upwash force and tendancy to roll calculated by integrating the upwash velocity field along the wingspan at each point in the domain.}
\label{fig:moment}
\end{figure}
Next, we estimate the cost required for UAV $i$ to maintain a constant altitude and orientation in the presence of the upwash field.
In particular, the UAV must expend energy to counter-roll against gradients in the upwash field, and it must pitch upward to counteract a negative upwash (downwash).
This leads to an intuitive physics-based description of the ``cost to flock,''
\begin{equation}
E_i(p_i, t) = \kappa |M_i(p_i, t)| - W_i(p_i, t),
\end{equation}
where $\kappa$ is a system parameter that captures the tradeoff between the cost to roll and the cost to pitch upward.
We require each UAV to satisfy,
\begin{equation}
\dot{E}_i(\bm{p}_i, t) \leq 0,
\end{equation}
where $E_i(\bm{p}_i, t)$ has a finite lower bound.
Thus, each UAV is driven toward an equilibrium point where the energy lost through wake interactions is minimized, or equivalently, the energy saved by flocking is maximized.
\begin{lemma} \label{lma:zeroUpwash}
For two agents $i, j$ such that $\bm{p}_j = \bm{0}$, there exists a unique lateral position $\hat{y}_i > 0$ such that:
\begin{align}
y_i < \hat{y} \implies w_i(\cdot, y_i, \cdot) < 0, \\
y_i = \hat{y} \implies w_i(\cdot, y_i, \cdot) = 0, \\
y_i > \hat{y} \implies w_i(\cdot, y_i, \cdot) > 0.
\end{align}
\end{lemma}
\begin{proof}
By definition $f_i(y_i)$ satisfies,
\begin{equation}
\begin{aligned} \label{eq:yiBehavior}
0\leq y_i < b \implies f_i(y_i) < 0, \\
y_i = b \implies f_i(y_i) = 0, \\
y_i > b \implies f_i(y_i) > 0.
\end{aligned}
\end{equation}
Thus, the equation,
\begin{equation} \label{eq:upwashLemma}
\int_{y_i-b}^{y_i+b} f(\xi) d\xi = 0,
\end{equation}
has a solution on the interval $y_i \in (0, 2b)$.
Let $\hat{y}_i$ denote the smallest solution to \eqref{eq:upwashLemma}, then, for any $\epsilon > 0$, the upwash difference is,
\begin{align*}
D(\epsilon) =& \int_{\hat{y}_i-b}^{\hat{y}_i+b} f(\xi) d\xi - \int_{(\hat{y}_i+\epsilon)-b}^{(\hat{y}_i+\epsilon)+b} f(\xi) d\xi \\
&= \int_{\hat{y}_i-b}^{\hat{y}_i-b+\epsilon} f(\xi) d\xi - \int_{\hat{y}_i+b}^{\hat{y}_i+b+\epsilon} f(\xi) d\xi.
\end{align*}
Note that by \eqref{eq:yiBehavior} the first integral is negative, as $\hat{y}_i \in (0, 2b).$
Meanwhile the second integral is positive, and thus $D(\epsilon)$ is negative for $\epsilon > 0$.
This implies that $\hat{y}_i$ is the unique solution, and $y_i > \hat{y_i} \implies w_i(\cdot, y_i, \cdot) > 0$.
Repeating the process for $\epsilon < 0$ completes the proof.
\end{proof}
\begin{lemma} \label{lma:zeroMoment}
For two agents $i, j$ such that $\bm{p}_j = \bm{0}$, there exists a unique lateral position $\hat{y}_i > b$ such that:
\begin{align}
y_i > \hat{y} \implies m_i(\cdot, y_i, \cdot) < 0, \\
y_i = \hat{y} \implies m_i(\cdot, y_i, \cdot) = 0, \\
y_i < \hat{y} \implies m_i(\cdot, y_i, \cdot) > 0.
\end{align}
\end{lemma}
\begin{proof}
The proof of Lemma \ref{lma:zeroMoment} is similar to that of Lemma \ref{lma:zeroUpwash}, and we omit it for brevity.
\end{proof}
\begin{theorem} \label{thm:equilibrium}
Consider two UAVs $i, j \in \mathcal{A}$ such that $\bm{p}_j = \bm{0}$ and the upwash effect of $j$ on $i$ is not negligible.
Let $[x_i, y_i]^T = \bm{p}_i - \bm{p}_j$ denote the streamwise and lateral position of $i$ relative to $j$. Furthermore, let the constant $\kappa$ be sufficiently small to admit a position $x_i^*, y_i^*$ such that $\nabla E_i(x_i^*, y_i^*, \cdot) = \bm{0}$ and there is a benefit to flocking, i.e., $E_i(x_i^*, y_i^*, \cdot) < 0$.
Then, any equilibrium point that yields $\nabla E_i(x_i^*, y_i^*, \cdot) = 0$ satisfies $x_i^* = -\mu$ $y_i^* > \sqrt{2}\,b$, which corresponds to a V or echelon formation.
\end{theorem}
\begin{proof}
By definition $g(x_i,t)$ is log-convex and has a unique maximum at $x_i = -\mu$.
Next, let $b + r^* \leq y_i \leq 2b - r^*$; the upwash experienced by the UAV is,
\begin{align}
w_i = g(\cdot)\Big(
\int_{y_i-b}^{b-r^*} u_i(\xi-b) d\xi
+ \int_{b-r^*}^{b+r^*} u_i(\xi-b) d\xi \notag \\
+ \int_{b+r^*}^{y_i+b} u_i(\xi-b) dz\xi
- \int_{}^{} u_i(\xi+b) d\xi
\Big),
\end{align}
where the second integral is zero by symmetry of the integrand about $(\xi-b) = 0$.
This implies that
\begin{align}
w_i = g(\cdot)\frac{\Gamma}{2\pi}\Big(
\ln(\frac{r^*}{2b - y_i}) + \ln(\frac{y_i}{r^*}) - \ln(\frac{y_i+2b}{y_i})
\Big) \notag\\
= g(\cdot)\frac{\Gamma}{2\pi}\ln\Big( \frac{y_i^2}{4b^2 - y_i^2} \Big).
\end{align}
Taking a derivative with respect to $y_i$ yields,
\begin{equation}
\frac{\partial w_i}{y_i} = g(\cdot)\frac{\Gamma}{2\pi}\Big( \frac{8b^2}{y(4b^2 - y^2)} \Big),
\end{equation}
which is positive for $y_i < 2b$, thus $w_i$ is strictly increasing for $y_i\in(b+r^*, 2b-r^*)$.
Furthermore, Lemma \ref{lma:zeroUpwash} implies that $w_i = 0$ has a unique solution, which occurs at $y_i = \sqrt{2}\,b$. This constitutes the lower bound on $y_i$ by our premise.
\end{proof}
Finally, based on our analysis of the upwash field properties, we formally define the rules of anseroid behavior to generate emergent V and echelon formations.
\begin{definition} \label{def:anseroid}
Anseroid behavior is characterized by two rules: (1) maximize the energy benefit of the local upwash field, and (2) match the drag-minimizing speed as closely as possible.
\end{definition}
\subsection{Optimal Control Problem}
We employ \emph{constraint-driven control} to implement anseroid behavior for each UAV.
Constraint-driven control is an optimization-based control technique, wherein the rules of interaction are embedded as constraints in an optimization problem.
This technique has been used successfully to control multi-agent constraint-driven systems, particularly in the ecologically-inspired robotics literature \cite{Notomista2019Constraint-DrivenSystems,Wang2017SafetySystems,Beaver2021Constraint-DrivenStudy}.
Our motivation for constraint-driven control is twofold:
first, it enables the UAVs to immediately react to their surroundings without the computational and communication costs associated with decentralized trajectory planning \cite{Beaver2020AnFlockingb,Beaver2022Constraint-DrivenAvoidance}.
This has the added benefit of allowing UAVs to be arbitrarily added to and removed from the domain without a priori knowledge \cite{Beaver2021Constraint-DrivenStudy}.
Second, implementing our desired behavior with constraints can allow for strong guarantees on the system-level behavior, unlike traditional optimization approaches where the objective is composed of performance criteria and soft constraints.
Thus, to implement the anseroid rules of Definition \ref{def:anseroid}, we propose Problem \ref{prb:optCtrl}.
\begin{problem} \label{prb:optCtrl}
Each UAV $i\in\mathcal{A}$ takes the optimal control input that optimizes,
\begin{align*}
\min_{v_i, \omega_i} & \Bigg\{ \Big(\frac{v_i - v_i^*}{v_{\max} - v_{\min}}\Big)^2 + \Big(\frac{\omega_i}{\omega_{\max}}\Big)^2 \Bigg\} \\
\text{subject to: } &
\eqref{eq:dynamics}, \eqref{eq:bounds}, \\
&\dot{E}(\bm{p}_i, t) \leq \rho \\
&|\theta_i - {\theta}_g| \leq \epsilon,
\end{align*}
where $\theta_g$ is the global heading angle (Assumption \ref{smp:main}) and $\rho \leq 0$ determines how quickly the UAVs approach the equilibrium point $\dot{E} = 0$.
Both terms are selected by a designer a priori.
\end{problem}
Note that while our analysis of the system behavior is in continuous-time, in many cases the optimal control problem must be formulated using discrete-time system dynamics.
This mapping is a practical consideration for implementation on a physical robot, and it can be a theoretical concern when the state variables, rather than the control input, appear in the constraints and objective function \cite{Beaver2022Constraint-DrivenAvoidance}.
This is an area of open research in the literature. Potential solutions include allowing the time step to grow arbitrarily small \cite{Xu2022FeasibilityFunctions}, employing control barrier functions to convert state constraints into control constraints \cite{Notomista2022Multi-robotTasks}, or tightening the constraints at each time step to ensure the trajectory remains feasible between time steps \cite{Pant2017SmoothLogic}.
In our next result, we provide a necessary and sufficient condition for the feasible space of Problem \ref{prb:optCtrl} to be empty, and we use this as the foundation for a switching control law.
\begin{theorem} \label{thm:feasible}
For UAV $i\in\mathcal{A}$ satisfying $|\theta_i - \theta_g|\leq \epsilon$, the feasible space of Problem \ref{prb:optCtrl} is empty if and only if the inequality,
\begin{equation} \label{eq:existanceThmStatement}
v_i \, L_{\bm{f}_i}E_i \leq \rho - \frac{\partial E_i}{\partial t},
\end{equation}
does not hold for any
\begin{equation}
v_{\min} \leq v_i \leq v_{\max},
\end{equation}
where $\bm{f} = [\cos(\theta_i), \sin(\theta_i)]^T$ denotes the forward direction of UAV $i$ and $L_{\bm{f}} E$ is the Lie derivative of $E_i$ in the direction of $\bm{f}$.
\end{theorem}
\begin{proof}
Any solution to Problem \ref{prb:optCtrl} must satisfy the inequality constraint $\dot{E}_i(\bm{p}_i, t) \leq \rho$, which expands to,
\begin{equation}
\dot{E}_i = \frac{\partial E_i}{\partial t} + \frac{\partial E_i}{\partial \bm{p}_i} v_i
\begin{bmatrix}
\cos{\theta_i} \\ \sin{\theta_i}
\end{bmatrix} \leq \rho,
\end{equation}
by the definition of the full derivative.
Rearranging terms implies,
\begin{equation} \label{eq:existanceThmProof}
v_i
\frac{\partial E_i}{\partial \bm{p}_i}\begin{bmatrix}
\cos{\theta_i} \\ \sin{\theta_i}
\end{bmatrix}
\leq \rho - \frac{\partial E_i}{\partial t},
\end{equation}
which is equal to \eqref{eq:existanceThmStatement}.
Next, any solution to Problem \ref{prb:optCtrl} must also satisfy the bounds,
\begin{equation} \label{eq:thetaBoundProof}
\quad v_{\min} \leq v_i \leq v_{\max}.
\end{equation}
Thus, if \eqref{eq:existanceThmProof} and \eqref{eq:thetaBoundProof} cannot be jointly satisfied, then the feasible space of Problem \ref{prb:optCtrl} is empty.
Similarly, if any control action $v_i^*$ jointly satisfies \eqref{eq:existanceThmProof} and \eqref{eq:thetaBoundProof}, then $v_i^*$ is a feasible solution of Problem \ref{prb:optCtrl}.
\end{proof}
Intuitively, Theorem \ref{thm:feasible} does not allow UAV $i$ to move in a direction that locally decreases the scalar cost field $E_i$ by less than $|\rho + \frac{\partial E_i}{\partial t}|$.
At each time instant the values of $L_{\bm{f}}E_i$ and $\frac{\partial E_i}{\partial t}$ are fixed by the state of UAV $i$ and the environment, and the UAV must select a value of $v_i$ that $E_i$ is locally decreasing.
In this context, one can interpret $v_i$ as a step size that must overcome the disturbances introduced through $\frac{\partial E_i}{\partial t}$.
Theorem \ref{thm:feasible} describes the conditions where the cost-minimizing constraint is incompatible with the velocity bounds of the agent.
This is closely related to the constraint compatibility problem in the control barrier functions literature, and it is well-studied in the set-theoretic control community \cite{Wang2017SafetySystems,Xiao2022SufficientFunctions,Ibuki2020Optimization-BasedBodies,Beaver2022Constraint-DrivenAvoidance}.
Generally, the problem of constraint incompatibility has been solved in the ecologically-inspired robotics literature by introducing slack variables \cite{Egerstedt2018RobotAutonomy,Ibuki2020Optimization-BasedBodies,Egerstedt2021RobotAutonomy}.
However, this is not fundamentally different from moving the constraint into the objective function to make it ``soft."
A foundational paper in multi-agent control barrier functions proposed operating the system in two modes \cite{Wang2017SafetySystems}: 1) a nominal mode where the agents solve the optimal control problem, and 2) a ``safe mode" where the agents come to a stop when the feasible space is empty.
We take this approach to its logical conclusion -- when the feasible space of Problem \ref{prb:optCtrl} is empty, the controller switches modes and solves a relaxed version of the problem.
This implies an equivalent switched system \cite{Beaver2022Constraint-DrivenAvoidance}, where the UAV uses the premise of Theorem \ref{thm:feasible} to switch between the full and relaxed problem, which is illustrated in Fig. \ref{fig:stateMachine}.
\begin{figure}[ht]
\centering
\includegraphics[width=0.7\linewidth]{resources/flowchart.png}
\caption{The behavior of each UAV visualized as a switched system. Theorem \ref{thm:feasible} determines when the UAV should solve the original or the relaxed optimal control problem.}
\label{fig:stateMachine}
\end{figure}
When Problem \ref{prb:optCtrl} has no feasible solution, it is unreasonable to relax the constraints corresponding to the dynamics or control bounds. The only reasonable options are to relax $\dot{E}_i \leq \rho$ or $|\theta_i - \theta_g| \leq \epsilon$.
We propose that the former should be relaxed to maintain Assumption \ref{smp:main}.
Next, we present the relaxed optimal control problem with Problem \ref{prb:relaxed}, followed by a result that shows such a relaxation only lasts for a finite time interval.
\begin{problem} \label{prb:relaxed}
Each UAV $i\in\mathcal{A}$ takes the optimal control input that optimizes,
\begin{align*}
\min_{v_i, \omega_i} & \Bigg\{ \Big(\frac{v_i - v_i^*}{v_{\max} - v_{\min}}\Big)^2 + \Big(\frac{\omega_i}{\omega_{\max}}\Big)^2 \Bigg\} \\
\text{subject to: } &
\eqref{eq:dynamics}, \eqref{eq:bounds}, \\
&|\theta_i - {\theta}_g| \leq \epsilon,
\end{align*}
where $\theta_g$ is the global heading angle (Assumption \ref{smp:main}).
\end{problem}
If the state UAV $i$ satisfies $|\theta_i - \theta_g| \leq \epsilon$, it is trivial to show that the control action
\begin{equation}
\begin{aligned}
v_i &= \max\big\{\min\big\{ v^*, v_{\max} \big\}, v_{\min} \big\},\\
\omega_i &= 0,
\end{aligned}
\end{equation}
is the optimal solution to Problem \ref{prb:relaxed}.
Note that the solution to Problem \ref{prb:relaxed} corresponds to the unstable case presented in Lemma \ref{lma:unstableUpwash}, where the UAV greedily matches the optimal upwash speed $v^*$.
This mode of operation corresponds to a ``safe mode" that retains some structure of the original problem, i.e., it doesn't require the UAVs to come to a complete stop, without guaranteeing formation flight.
Next, we prove that each UAV will only need to solve Problem \ref{prb:relaxed} for a finite amount of time before the premise of Theorem \ref{thm:equilibrium} is no longer satisfied.
\begin{corollary} \label{cor:finiteTime}
Any ``following'' UAV will only solve Problem \ref{prb:relaxed} for a finite interval of time before the premise of Theorem \ref{thm:feasible} is not satisfied.
\end{corollary}
\begin{proof}
Corollary \ref{cor:finiteTime} follows trivially from the proof Lemmas \ref{lma:upwashSpeed} and \ref{lma:unstableUpwash}; any following UAV $i$ will have a higher upwash benefit, and thus lower optimal airspeed, than the lead UAV.
This implies that the following UAV satisfies $W_i = 0$ after a finite amount of time, which implies that $\dot{E} = 0$ for all control actions.
\end{proof}
Thus, we have introduced Theorem \ref{thm:feasible}, which gives necessary and sufficient conditions for the system to switch between the flocking controller (Problem \ref{prb:optCtrl}) and the safe controller (Problem \ref{prb:relaxed}).
Furthermore, Corollary \ref{cor:finiteTime} demonstrates that the UAVs will spend a finite amount of time solving the relaxed control problem.
\subsection{On Heterogeneity} \label{sec:hetero}
One compelling result of Lemma \ref{lma:upwashSpeed} is the fact that introducing heterogeneity has the potential to further improve the fuel consumption of all UAVs in the flock.
It demonstrates that any UAV gaining an upwash benefit has a lower optimal speed than when flying in isolation, i.e., each $v_i^*$ decreases as $W_i$ increases for each UAV $i\in\mathcal{A}$.
However, each UAV must match the speed of the UAV in front of it to maintain the formation.
This implies that the flock will inevitably fly at the speed of the frontmost agent, with a speed that is significantly higher than $v_i^*$ of the following agent $i$.
However, if heterogeneity is introduced to the flock--through aerodynamic design, payload weight, or fuel consumption--then the slowest agents can be placed at the front of the flock.
This enables the faster agents in the rear, who receive significantly larger upwash benefits, to fly slower and have a steady-state speed closer to their optimal airspeed.
In this way, heterogineity in the optimal airspeed of the UAVs can have benefits across the entire flock.
\section{Experimental Validation} \label{sec:validation}
Next, we present two simulations and a physical experiment using our proposed flocking controllers.
The first simulation is for $N=2$ UAVs in Matlab, and it demonstrates that pairs of agents converge to the equilibrium points described by Theorem \ref{thm:equilibrium}.
Next, we demonstrate a Python-based simulation that uses the CrazySwarm Python API \cite{crazyswarm} for $N=11$ UAVs; this demonstrates that our proposed flocking controller is platform independent, and that the controller scales to larger numbers of agents.
Finally, we run the Python controller in real-time on a fleet of $N=5$ CrazyFlie quadrotors.
The quadrotors are a surrogate for fixed-wing UAVs, and with them we demonstrate that the control algorithm easily runs in real time--including the time it takes to simulate the aerodynamic interactions between agents.
\subsection{Implementation Issues} \label{sec:implementation}
We foresee two major issues when solving Problem \ref{prb:optCtrl} on a real system.
The first is a sensing issue, namely determining the value of $M_i$ and $W_i$ for each UAV $i\in\mathcal{A}$.
We simulate $M_i$ and $W_i$ explicitly in our results, but for a real fixed-wing UAV we proposed a solution that relies on Assumption \ref{smp:tracking}, i.e., that the UAV is equipped with a low-level tracking controller.
At the tracking level, the upwash force and induced roll act as disturbances on the UAV.
By monitoring the roll and pitch signals, we conjecture that the corresponding upwash force and induced tendency to roll can be inferred from the flight controller.
In other words, our proposed anseroid controller is an outer loop that samples the low-level flight controller data.
The second issue is information-theoretic, namely calculating the time derivative of $E_i$ for each UAV $i\in\mathcal{A}$.
The gradient of the upwash field could possibly be predicted using an aerodynamic model, however this does not capture the time-varying component,
\begin{equation}
\dot{E} = \frac{\partial E}{\partial t} + \frac{\partial E}{\partial \bm{p}_i} \dot{\bm{p}}_i,
\end{equation}
where the non-smooth points in $\frac{d E}{d \mathbf{p}}$ can be handled with differential inclusions \cite{Santos2019DecentralizedFunctions}.
The main problem is that UAV $i$ must also have knowledge of $v_j$ to compute $\frac{\partial E_i}{\partial t}$.
Similarly, UAV $j$ must have knowledge of $v_i$ to compute $\frac{\partial E_j}{\partial t}$; this is the fundamental problem of simultaneous actions in decentralized control.
One popular approach is to use a consensus protocol to estimate $\frac{\partial E}{\partial t}$, e.g., ADMM \cite{Summers2012DistributedMultipliers}, which has shown some success in the flocking literature \cite{Lyu2019MultivehicleControl}.
Alternatively, it is possible for the agents to iteratively generate trajectories to converge on a locally optimal control strategy, \cite{Zhan2013FlockingMeasurements,Yuan2017OutdoorControl}.
To minimize the iterative and communication cost, we instead assume that for UAV $i$, the signal $v_j(t)$ is a constant for all $j\in\mathcal{A}\setminus\{i\}$.
As the UAVs form a V shape, their speed will eventually converge to the shape of the formation.
Thus, we expect the error that arises from the constant-speed assumption to asymptotically converge to zero \cite{mahbub2022ACC}.
We demonstrate that the system, using this update scheme, approaches a steady V formation in the following subsections.\footnote{Videos of the experiments and simulation files are available at \url{https://sites.google.com/view/ud-ids-lab/anseroids}.}
Note that $\rho=0$ for all of our results, and this demonstrates that the model mismatch introduced by our scheme is sufficient to ensure that the trivial control policy $v_i = v_j$ for all $i,j\in\mathcal{A}$ does not emerge unless $E_i \approx 0$ or $E_j \approx 0$.
\subsection{Matlab Simulation} \label{sec:matlab}
To demonstrate the behavior of our control algorithm, we apply our proposed constraint-driven control scheme to a system of $N=2$ UAVs Using data for the RQ-11 Raven
\cite{AeroVironment2021RavenRQ-11B}.
The Raven weighs $18.7$ N, with a $1.4$ m wing span and a nominal speed of $12$ m/s.
Approximating the density of air as $\rho = 1.2$ kg/m$^3$ yields the circulation from the Kutta-Joukowski theorem \cite{Anderson2017FundamentalsAerodynamics},
\begin{equation}
\Gamma = \frac{L}{2b \rho u} = \frac{18.7 \text{ N}}{1.4\text{ m } 1.2\text{ kg/m}^3~ 9 \text{ m/s}} = 1.24 \text{ m}^2\text{/s}.
\end{equation}
We expect $r^*$ to be $\approx 5$\% of the span length (e.g., \cite{Yang2016LoveControl} uses 5.4\%); thus we select $r^* = 0.054$ m.
This implies,
\begin{equation}
\Omega = \frac{\Gamma}{2\pi(r^*)^2} \approx 70 \text{ s}^{-2}.
\end{equation}
Next we estimate the drag constants in \eqref{eq:drag}.
Assuming steady flight, the induced drag satisfies \cite{Anderson2017FundamentalsAerodynamics},
\begin{equation}
C_2 = \frac{L^2}{2 \rho \pi b^2 } \approx 95 \text{ N}\cdot\text{m/s}^2.
\end{equation}
Assuming the cruising speed of $12$ m/s is designed to be near-optimal for the Raven,
\begin{equation}
C_1 = \frac{C_2}{(v_i^*)^4} = 5\times10^{-3} \text{ N}\cdot{s/m}^2.
\end{equation}
Note that this implies a drag force of $0.72$ N at cruising speed, and a dimensionless drag coefficient of $C_d \approx 0.04$, which is not unreasonable for a wing.
We initialize the UAVs in a line along the spanwise axis, with an initial orientation of $\theta_G = 0$ and a center-to-center spacing of $2 b = 1.4$ m.
We present a sequence of simulations in Fig. \ref{fig:simSnapshot}, which demonstrates how two agents quickly fall into a leader and follower position, and that the following agent moves to and remains at the point where the flocking cost is minimized.
\begin{figure*}[ht]
\centering
\includegraphics[width=0.32\linewidth]{resources/sim_0001.png}
\includegraphics[width=0.32\linewidth]{resources/sim_0250.png}
\includegraphics[width=0.32\linewidth]{resources/sim_0500.png}
\caption{A sequence of simulation snapshots over $20$ seconds for $N=2$ UAVs initialized in a line.
The annotation shows the instantaneous cost and speed of the top (following) UAV; the contours show the wake of the bottom (lead) UAV.}
\label{fig:simSnapshot}
\end{figure*}
To quantify the impact of the V formation on the pair of UAVs, we selected $\kappa = 0.25$ and calculated the cost functional $E_i$ for each UAV $i\in\mathcal{A}$.
We present the total cost, the maximum cost, minimum cost, and terminal cost for each UAV in Table \ref{tab:cost}.
Note that a cost of zero corresponds to the agent flying in isolation; this demonstrates that the resulting formation yields a significant energy savings for each UAV.
The instantaneous cost is presented in Fig. \ref{fig:costRange}, which shows the cost experienced the agents at each time-instant of the simulation.
Finally, we can see that the cost remains relatively constant for the last $10$ seconds of the simulation, implying the UAVs have reached a steady-state configuration.
\begin{table}[ht]
\centering
\begin{tabular}{rcccc}
& Front UAV & Rear UAV \\\toprule
Total Cost & -35.2 & -155.2 \\
Maximum Cost & -0.3 & -0.3 \\
Minimum Cost & -2.1 & -9.8
\end{tabular}
\caption{The impact of our anseroid controller on the cost of each UAV over the entire simulation; a value of $0$ is equivalent to flying in isolation.}
\label{tab:cost}
\end{table}
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth]{resources/CostGraph.png}
\caption{Cost experienced by the front and rear UAVs over the duration of the Matlab simulation. A cost of zero corresponds to flying in isolation.}
\label{fig:costRange}
\end{figure}
\subsection{Crazy Swarm} \label{sec:experiment}
As a next step toward real-time control of physical UAVs, we developed a decentralized flocking controller using the Crazyswarm library \cite{crazyswarm}.
Crazyswarm is a software library for controlling fleets of Crazyfly quadrotors, and this required us to simulate the aerodynamics of the UAVs to determine the control actions.
Due to the smaller size of the UAVs and the experimental workspace, we scaled the aerodynmic parameters i.e., $\Gamma, \Omega, L$, by $100$ relative to those presented in Section \ref{sec:matlab}, and we present the remaining parameters in Table \ref{tab:parameters}.
To simulate the aerodynamics in real time, we solved the integral \eqref{eq:force} analytically and set $\kappa = 0$.
The simulation result for $N=11$ UAVs is presented in Fig. \ref{fig:crazyswarmSim}.
The front UAV travels almost $35$ m over approximately $60s$, and the formation begins to emerge between $10$ and $15$ m.
This demonstrates that the emergence of a V formation from our controller is independent of the software platform used to implement our algorithm.
\begin{table}[h]
\centering
\begin{tabular}{ccccccc}
$v_{\min}$ & $v_{\max}$ & $w_{\min}$ & $w_{\max}$ & b & $C_1$ & $C_2$ \\ \toprule
0.01 m/s & 2 m/s & -1 rad/s & 1 rad/s & 0.2 m & 160000 & 5000
\end{tabular}
\caption{Parameters used for the Crazyswarm simulation and experiments.}
\label{tab:parameters}
\end{table}
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth]{resources/CrazyswarmSim.png}
\caption{Simulation for $N=11$ UAVs over $60$ seconds in CrazySwarm. The agents are moving from left to right; colored dots and lines correspond to individual UAVs, while the black lines shows the flock shape at different time instants.}
\label{fig:crazyswarmSim}
\end{figure}
We deployed the same code on a fleet of $N=5$ CrazyFlie quadrotors in our experimental testbed \cite{chalaki2021CSM}.
Quadrotors do not have the same aerodynamic interactions as fixed-wing vehicles, and as a consequence we simulated the upwash field $W_i$ for each $i\in\mathcal{A}$ as part of the control loop.
To generate the control actions, we solved Problem \ref{prb:optCtrl} for each UAV using the SciPy optimization package\footnote{\url{ https://docs.scipy.org/doc/scipy/reference/optimize.html}}.
If the resulting solution violates any constraints, we subsequently solve the relaxed Problem \ref{prb:relaxed} to determine the control input.
Calculating the optimal control action of all five agents took an average of $11$ ms, with a maximum time of $22$ ms, on our centralized computer (i7-6950X CPU @ 3.00 GHz); furthermore, state updates were performed using a VICON motion capture system operating at 100 Hz.
Snapshots of the experiment are presented in Fig. \ref{fig:experiment}; this demonstrates that our control algorithm is real-time implementable and leads to an emergent V formation under the strict time and space constraints present in a laboratory environment.
The agents are initialized in an approximately vertical line and spaced $0.5$ m apart.
The agents at the edge of the formation quickly fall back, and a V formation can be seen to emerge when the front agent has moved only $0.5$ m.
The structure of the V formation emerges around $2$ m as the agents approach the boundary of our testbed.
\begin{figure*}[ht]
\centering
\includegraphics[width=0.32\linewidth]{resources/experiment1.png}
\includegraphics[width=0.32\linewidth]{resources/experiment2.png}
\includegraphics[width=0.32\linewidth]{resources/experiment3.png}
\caption{Emergent V formation for $N=5$ Crazyflie drones using the anseroid controller in CrazySwarm over $t=8$ s. The agent positions are shown at the initial time (left), half way through the experiment (center), and the end of the experiment (right).}
\label{fig:experiment}
\end{figure*}
\section{Conclusions} \label{sec:conclusion}
In this article we presented, for the first time, an anseroid behavior that generates emergent V formations for flocks of UAVs.
First, we demonstrated that a greedy energy-minimizing approach is unstable and cannot lead to a stable V formation, and we proposed a constraint-driven approach where agents maximize their upwash benefits.
We proved for a simplified system that this energy-optimizing control leads to emergent flocking, and that the agents are driven to equlibria that resemble a V or echelon formation.
Finally, we demonstrated the performance of our approach through matlab and python simulations, as well as a real-time experiment with five crazyflie drones.
The results of this article present several compelling directions for future work.
First, analyzing the stability of the V formation to disturbances--including the addition and removal of agents, as well as noise, errors and delays--is critical to understand real-world performance of our proposed control algorithm.
Determining the parameter ranges for the wingspan, aerodynamic properties, weight, and energy factor $\kappa$ that make flocking feasible, i.e., satisfy Theorem \ref{thm:equilibrium}, is another interesting research direction.
Relaxing Assumption \ref{smp:main} using a computational fluid dynamics model may yield interesting insights, as could experiments with real fixed-wing aircraft to infer the upwash force and tendancy to troll from the low-level flight controller.
Finally, exploring leader-switching strategies, and incorporating heterogeneity--through payload mass or fuel quantity--is a compelling research direction for more realistic multi-agent and multi-target missions.
|
\section{Introduction}
The current standard model of cosmology was recently confirmed analysing the precise cosmic microwave background (CMB)
radiation data measured by the Planck satellite~\citep{Planck18}.
The 6-parameters cosmological model that best fits the observational data, from CMB radiation and other cosmological observables, is the flat-$\Lambda$CDM, where the 3-dimensional space is Euclidean (i.e., zero spatial curvature or $k = 0$), and the universe is filled in with cold dark matter (CDM) and dark energy –in the form of a cosmological constant $\Lambda$– in addition to the standard baryonic and electromagnetic ingredients~\citep{Planck18,Planck16}.
One distinctive feature of the Planck measurements is the high angular resolution that has allowed to measure the CMB temperature fluctuations at the scales $\sim 4'$ (i.e., $\ell \sim 2500$).
This made it possible to reconstruct, by inverse procedure, the lensing potential map that causes the weak lensing action of matter on the CMB photons path from the last scattering surface to us~\citep{Planck15-XV,Jia15,GAM,Bianchini}.
In fact, the CMB photons detected by Planck are already weakly lensed by the matter they encounter in their paths, and the prediction of this phenomenon in the standard model is a tiny effect of smoothing the CMB acoustic peaks at very small angular scales (i.e., $\ell \stackrel{>}{_{\sim}} 1000$).
Since the observed CMB photons can not be delensed, the solution adopted by the Planck collaboration for the CMB data analyses is to include the weak lensing effect in the model that best fits the CMB angular power spectrum (APS). In practice, this solution contemplates that the amplitude of the weak lensing effect on CMB photons is, in the
flat-$\Lambda$CDM model, $A_{L} = 1$.
The precise measurements of the Planck CMB temperature-temperature (TT) APS at small angles have stimulated
accurate statistical analyses about the suitability of the lensing amplitude parameter $A_{L} = 1$,
as a way to validate the $\Lambda$CDM concordance model.
As a result, some works have reported discordance with this amplitude.
In fact, previous analyses indicate a clear preference for a higher lensing amplitude, i.e., $A_{L} > 1$
\citep{Calabrese,Bianchini,DES22,Ballardini,Valentino20b,Valentino20a,Planck16}\footnote{see the table 9 in page 32, in ref.
https://arxiv.org/pdf/1605.02985.pdf}.
More recently, the Planck collaboration reported:
$A_{L} = 1.180 \,\pm\, 0.065$ (68\%, Planck TT,TE,EE+lowE) and
$A_{L} = 1.243 \,\pm\, 0.096$ (68 \%, Planck TT+lowE), where the lensing amplitude $A_L$ is varied in the
parameter analysis (see section 6.2 in~\cite{Planck18}).
These results motivate us to perform a detailed
examination of the CMB acoustic peaks sensitive to the lens phenomenon, i.e. $\ell > 1000$,
due to a possible excess of lensing power in the TT APS data, a signal that can be accounted for with a
lensing amplitude $A_{L} > 1$, that is, higher than the value adopted by the flat-$\Lambda$CDM Planck's
best-fit model $A_{L} = 1$.
The outline of this work is the following:
In section~\ref{sec2} we describe the Planck data employed in this work.
In section~\ref{sec3} we perform the statistical analyses of the spectra difference (Planck CMB APS minus Planck best-fit
$\Lambda$CDM APS) to investigate:
(i) if it corresponds to white noise or not;
in negative case, (ii) if it has signature and amplitude that can be explained by an extra lensing amplitude $A_{lens} > 0$
interpreted as a possible excess of lensing signature not accounted by the Planck best-fit $\Lambda$CDM APS;
(iii) if this spectrum difference can be explained by modifying some other cosmological parameters in the APS of
the $\Lambda$CDM model.
In section~\ref{sec4} we discuss our results, and also present our conclusions and final remarks.
\section{Data}\label{sec2}
The most accurate current measurements of the CMB temperature fluctuations are part of the third public data release
of the Planck collaboration~\cite{Planck18}, precise data that allow to re-examine many interesting features already reported
with diverse CMB data sets~\citep{WHR1,B-HR,Samal,BR12,Aluri12,BOP,Polastri,Aluri16,Aluri17,Vafaei,Goyal,Saha21,Chiocchetta,Saha22}.
Further, CMB products combined with data sets from other cosmological tracers are being used to study models alternative to the flat-$\Lambda$CDM model (see, e.g.~\cite{WHR2,WHR3,Bessa}).
In particular, the large set of data products released by the Planck team offers the possibility to perform exhaustive analyses
to learn more about features of the matter clustering, investigating the small angular scales where the weak lensing phenomenon
left its signature imprinted. In fact, the Planck CMB data are specially accurate at small angular scales due to their high angular
resolution~\citep{Marques20a,Marques20b,Dong,Avila,Tanseri}.
Thus, we shall analyse in details the small angular scales of the CMB TT APS which is part
of the third data release of the PLANCK collaboration~\cite{Planck18}. Our main goal is to study if the difference between the measured Planck APS and the best-fit $\Lambda$CDM APS (based on the 6-parameters flat-$\Lambda$CDM model)
from the data fit analyses done by the Planck team, is just statistical noise or it has indeed some indicative signature of an incorrect modeling of the weak lensing amplitude (i.e., $A_L \ne 1$), or of other cosmological parameters.
Clearly, the null hypothesis that deserves investigation establishes that the difference between the observed Planck APS
and the $\Lambda$CDM APS is just white noise.
\vspace{0.3cm}
For our analyses some features are important to be considered.
Firstly, the possible signal would be significantly detected only at the angular scales with the smallest
error measurements, that is, for the interval $\ell \sim 1000 - 2000$;
at the same time, one knows that these scales correspond to the regime where the lensing phenomenon has a large impact on
the CMB APS.
Secondly, for these analyses we use the original unbinned data from Planck repository pages, then our analyses shall consider
binnation inspections beside the $\Delta \ell = 30$ case considered by the Planck team.
In our scrutiny, the released Planck CMB TT APS, the Planck APS~\footnote{COM$_{-}$PowerSpect$_{-}$CMB-TT-full$_{-}$R3.01.txt},
covers the range of multipoles $\ell = 2-2508$, where the multipoles of interest here, $\ell \ge 30$, were derived from the
cross-half-mission likelihood Planck, Plik (for details see ref.~\cite{Planck18-V}).
The Planck APS was derived from the Commander component-separation algorithm applied to the combination of Planck 2018 temperature
data between 30 and 857 GHz, including 86\% of the sky (Planck-2018 results VI) (associated $1\sigma$ errors include beam uncertainties).
The analysis was done by optimally combining the spectra in the frequency range $100-217$ GHz, and correcting them for unresolved
foregrounds using the best-fit foreground solution (regions were chosen as to avoid the areas dominated by noise).
For the best-fit $\Lambda$CDM CMB APS, to be subtracted to the Planck APS for the aim of our analyses, we use the 6-parameter
flat-$\Lambda$CDM minimal model, also released by the Planck team and hereafter termed the $\Lambda$CDM
APS~\footnote{COM$_{-}$PowerSpect$_{-}$CMB-base-plikHM-TTTEEE-lowl-lowE-lensing-minimum-theory$_{-}$R3.01.txt}.
This $\Lambda$CDM APS uses the Planck TT, TE, EE+lowE+lensing data.
Both power spectra, the Planck APS and $\Lambda$CDM APS, were publicly released by the Planck collaboration\footnote{
\url{https://archives.esac.esa.int/doi/html/data/astronomy/planck/Cosmology.html}
}.
\section{Methodology and data analyses}\label{sec3}
This section describes the methodological approach we shall follow to study the Planck data APS, specifically concerning the
possibility that the lensing amplitude parameter that best-fits the APS data is larger than 1, $A_L > 1$, where the value
used by the Planck collaboration to obtain the flat-$\Lambda$CDM model is $A_L = 1$.
\subsection{Methodology to investigate the spectrum difference}\label{sec3.1}
We want to investigate if there is an excess of lensing power in the TT APS measured by the Planck Collaboration
$C^{\mbox{\footnotesize Planck}}_{\ell}$, with 1$\sigma$ standard deviation $\sigma^{\mbox{\footnotesize Planck}}_{\ell}$,
with respect to the flat-$\Lambda$CDM APS, $C^{\Lambda\mbox{\footnotesize CDM}}_{\ell}$,
obtained through a best-fit procedure of the 6 parameter flat-$\Lambda$CDM cosmological model to the TT Planck APS data~\citep{Planck18}.
For this, we first define the spectrum difference :
\begin{equation}\label{Dobs}
\delta^{obs}_{\ell} \equiv \frac{C^{\mbox{\footnotesize Planck}}_{\ell} - C^{\Lambda\mbox{\footnotesize CDM}}_{\ell}}
{C^{\Lambda\mbox{\footnotesize CDM}}_{\ell}} \,.
\end{equation}
Notice that the $\Lambda$CDM APS, $C^{\Lambda\mbox{\footnotesize CDM}}_{\ell}$, takes into account the CMB lensing phenomenon with $A_L = 1$, which is the value assumed by the Planck Collaboration in the flat-$\Lambda$CDM model.
We define the symbol $A_{lens} \in [0,1]$
as the parameter that quantifies the excess of lensing amplitude to explain the residual $\delta^{obs}_{\ell}$.
We construct synthetic APS, $C^{syn,L}_{\ell}(A_{lens})$,
with the Planck 2018 flat-$\Lambda$CDM cosmological parameters but for arbitrary lensing amplitude
$A_{lens} \ge 0$.
We also define
\begin{equation}\label{Dsyn}
\delta^{\mbox{\,\footnotesize exc}}_{\ell}(A_{lens}) \,\equiv\, \frac{ C^{\mbox{\,\footnotesize syn},L}_{\ell}(A_{lens})
- C^{\mbox{\,\footnotesize syn},uL}_{\ell} } {C^{\mbox{\,\footnotesize syn},uL}_{\ell}} \,,
\end{equation}
where
\begin{eqnarray}
C^{\Lambda\mbox{\footnotesize CDM}}_{\ell} &=& C^{\mbox{\,\footnotesize syn},L}_{\ell}(A_{lens} = 1) \, , \\
C^{\mbox{\,\footnotesize syn},uL}_{\ell} &=& C^{\mbox{\,\footnotesize syn},L}_{\ell}(\mbox{A}_{lens}=0) \,,
\end{eqnarray}
where the upper letters $L$ and $uL$ mean {\em lensed} and {\em unlensed} APS, respectively.
The quantity $\delta^{\mbox{\,\footnotesize exc}}_{\ell}(A_{lens})$ measures the possible excess of lensing power represented by the relative difference of synthetic lensed and unlensed APS (the difference
that depends on the parameter $A_{lens}$; clearly, the case
$A_{lens} = 0$ implies
that $\delta^{\mbox{\,\footnotesize exc}}_{\ell} = 0$).
Then, a good statistical agreement between the excess quantity $\delta^{\mbox{\,\footnotesize exc}}_{\ell}$ and the observed difference
$\delta^{\mbox{\,\footnotesize obs}}_{\ell}$, measured with the reduced $\chi^2$ best-fit for some $A_{lens} > 0$ value,
will provide an explanation for $\delta^{\mbox{\,\footnotesize obs}}_{\ell}$ as being an excess of lensing power in the Planck APS not accounted for
by the $\Lambda$CDM lensing amplitude $A_{L} = 1$.
Of course, other interpretations for $\delta^{\mbox{\,\footnotesize obs}}_{\ell}$ would be possible, as for instance that it is just statistical noise, and for this reason it is interesting to examine them too. In order to evaluate if the spectrum difference $\delta^{\mbox{\,\footnotesize obs}}_{\ell}$ corresponds to white noise, we shall apply the Ljung-Box test \citep{ljungbox}.
Our analyses include the simulation of APS, from now on termed synthetic APS, which considers the modeling of
the weak lensing phenomenon on CMB photons for cases with $A_{lens} = 0$ (unlensed) and $A_{lens} \ne 0$ (lensed).
The details for the production of these synthetic APS are the following:
\begin{itemize}
\item The Boltzmann code $CLASS$ \citep{class} was modified to allow the code to consider
the lensing amplitude $A_{lens}$ as a parameter.
\item
We use this modified code with the cosmological parameters corresponding to the observed Planck TT APS, as detailed in the
table~\ref{table1}, to produce synthetic lensed and unlensed $\Lambda$CDM TT APS, $C^{\mbox{\,\footnotesize syn},L}_{\ell}$
and $C^{\mbox{\,\footnotesize syn},uL}_{\ell}$, respectively.
\begin{table*
\caption {The cosmological parameters used to generate the synthetic angular power spectra.} \label{table1}
\begin{center}
\begin{tabular}{|l|l|l|l|l|l|l|}
\hline
parameters & $\Omega_b h^2$ & $\Omega_c h^2$ & $100 \theta_{MC}$ & $n_s$ & $\ln(10^{10}A_s)$ & $\tau$\\
\hline \hline
$\Lambda$CDM & 0.022383 & 0.12011 & 1.040909 & 0.96605 & 3.0448 & 0.0543 \\
\hline
\end {tabular}
\end{center}
\end {table*}\label{table1}
\item We then vary the value of the lensing amplitude $A_{lens}$ and construct the quantity $\delta^{\mbox{\,\footnotesize exc}}_{\ell}(A_{lens})$.
\end{itemize}
Moreover the relative error for $\delta^{obs}_{\ell}$, $error(\delta)_{\ell}$ , is calculated by considering the error
from the observed Planck TT APS,
$\sigma^{\mbox{\footnotesize Planck}}_{\ell}$,
and using the standard approach for propagation of uncertainties from equation (\ref{Dobs}).
Our analysis shall consider binnation inspections beside that one adopted by the Planck team.
Thus, after calculating the spectrum difference
for each $A_{lens}$ considered, and the corresponding errors, we have to choose
the bin length $\Delta \ell$, which is the size of the bin that we are dividing our $\delta^{obs}_{\ell}$ array into.
Each bin will be represented by its mean value with error $\sigma_{\ell} \equiv error(\delta)_{\ell} / \sqrt{\,\Delta \ell}$.
In the last step, we compute the $\chi^2$, after binning the spectrum difference and computing the error
\begin{equation}
\chi^2(A_ {lens}) = \sum\frac{[\delta^{\mbox{\,\footnotesize obs}}_{\ell} - \delta^{\mbox{\,\footnotesize exc}}_{\ell}(A_ {lens})]^2}{\sigma^2_{\ell}} \,,
\end{equation}
where $\delta^{obs}_{\ell}$ and $\delta^{exc}_{\ell}(A_{lens})$ are defined in equations (\ref{Dobs}) and (\ref{Dsyn}), respectively.
The sum above is performed over the binned spectra difference data $\{ \delta_{\ell}^{\mbox{\,\footnotesize obs / exc}} \}$.
Finally, the space of parameters is extended to include the
neutrino mass $\sum m_\nu$ and the spatial curvature $\Omega_k$ to explore the possibility that any lensing excess signal can be mimicked by the effects of some of these parameters.
\subsection{Null hypothesis analyses}\label{sec3.2}
In this subsection we shall test if the set of values $\{ \delta^{obs}_{\ell} \}$ is a statistical noise or not,
i.e. we test the randomness of the spectrum difference.
To perform this test we consider the null hypothesis:
\\
\\
\noindent
$H_0$: The spectrum difference, $\{ \delta^{obs}_{\ell} \}$, corresponds to a residual statistical (or white) noise.
\\
\\
That $H_0$ be true means that $A_L= 1$ completely accounts for with the lensing signal in the observed Planck APS.
In order to examine this hypothesis we shall use the Ljung-Box (LB) test~\citep{ljungbox}, which is a modification
of the Box-Pierce Portmanteau {\bf Q} statistic~\citep[][]{boxpierce}.
The LB test is used to look for correlation in a data series,
determining whether there is or not a remaining signature in the residuals after a forecast model has been fitted to the data.
Basically, the LB test is a useful tool to evaluate the autocorrelation between the data in analysis, and to quantify its
statistical significance.
As a first step, it is necessary to compute the autocorrelation in a given data set
$\{ \delta^{obs}_{i} \}$ in an interval $M = [\ell_{min},\ell_{max}]$ with $N$ data points
\begin{eqnarray} \label{auto}
\rho_k =\frac{\sum^{N-k}_{i=1} \left( \delta^{obs}_{i}- \overline{\delta^{obs}_{i}}\right)\left(\delta^{obs}_{i+k}- \overline{\delta^{obs}_{i}}\right)}{\sum^N_{i=1} \left(\delta^{obs}_{i}- \overline{\delta^{obs}_{i}}\right)^2} \,,
\end{eqnarray}
where $\overline{\delta^{obs}_{i}}$ is the average of all $N$ points in the $M$ interval, $k$ is commonly called the \text{lag} and $\rho_k$ is called the lag $k$ autocorrelation.
Since $\rho_k$ measures the correlation between multipoles separated by $k$, the autocorrelation $\rho_k$ can be used, in principle, to detect non-randomness in the data. However, it is more recommended to use tests considering multiple (sometimes called global or total) correlations across all the data interval and for several lags jointly, like the LB test~ \citep{ljungbox}.
The null hypothesis $H_0$ for this test establishes that the first $h$ lags autocorrelations are jointly zero, i.e.
\begin{eqnarray}
H_0: \rho_1 = \rho_2 = \cdots =\rho_k= \cdots =\rho_h = 0 \,,
\end{eqnarray}
where $h$ is the maximum lag considered in the test. In other words, $H_0$ being true implies that all the analyzed data are uncorrelated and correspond to a white noise signal. The LB statistic is defined by
\begin{eqnarray} \label{q}
Q_h = N(N+2)\sum^h_{k=1} \frac{\rho^2_k}{N-k} \,,
\end{eqnarray}
where $\rho_k$ is the estimated correlation using equation (\ref{auto}).
Thus , LB test does not consider just a particular lag $k$ but a set of $h$ estimated correlations.
Since $Q$ asymptotically follows a $\chi^2$ distribution, to determine the statistical significance of the test
it is compared to a $\chi^2$ distribution with $h'=h-q$ degrees of freedom under the condition \citep{ljungbox}
\begin{eqnarray}
Q_h > \chi^2_{1-\alpha,h'} \,,
\end{eqnarray}
where $q$ is the number of parameters used to fit the observed Planck APS and $\alpha$ is the significance level.
Then, small $p$-values ($p<\alpha$) will imply that
significant correlation exists between the data in the set $\{ \delta^{obs}_{i} \}$ and thus, $H_0$ is rejected in the $M$ interval.
A choice of the $h$ parameter in equation (\ref{q}) requires a more detailed discussion.
Several studies has been performed to define which is the optimal value.
For instance, empirically \citet{Ljung1986} suggests $h=5$, \citet{Tsay2010} suggests $h\sim \ln N$, \citet{Hyndman2018}
$h = min(10,N/5)$, \citet{Shumway2011} $h=20$. Furthermore, \citet{Hyndman2014} employed a study with simulations to show that
for very large values of $h$, LB test could lead to not so reliable results.
Recently, \citet{hassani2020} also
used simulations to evaluate the optimal value for the number of lags $h$ involved in the LB test.
Their results have shown that for the order of thousands data, the optimal values are $h=50$ for $\alpha=0.05$ and $h=25$ for $\alpha = 0.01$.
To perform the LB test we first compute $\{ \delta^{obs}_{l} \}$ in different $M$ intervals and apply it for each interval.
Our results are summarized in table~\ref{LBIntervals} where $q=7$, $\alpha =0.01$ and, as recommended by~\citet{hassani2020}, $h =25$ was used.
We rerun the test for different values of $h$ mentioned in the previous paragraph and confirm that the results in the table~\ref{LBIntervals} are robust.
The first interval analyzed is $2\leq \ell \leq 2500$, i.e. the complete range of multipoles measured by Planck.
In such interval we found a rejection of null hypothesis $H_0$ with $p = 0.0$. Then, we applied the test for three different intervals:
$2 \leq \ell \leq 100$, $2\leq \ell \leq 800$ and $2 \leq \ell \leq 1200$, where it was found a no rejection of $H_0$.
These results are indicating that the CMB power spectrum is well fitted by the cosmological parameters found by Planck team, and thus residual $\{ \delta^{obs}_{\ell} \}$ is white noise until, roughly, $\ell \sim 1200$.
We also considered three more intervals,
$1100\leq \ell \leq 2500$, $1600\leq \ell \leq 2500$ and $2000 \leq \ell \leq 2500$.
Results in such intervals are indicating a rejection of the $H_0$ hypothesis at more than $99\%$ CL, i.e.
$\{ \delta^{obs}_{\ell} \}$ is not white noise and some signal could be hidden in such multipole intervals.
One possibility is that rejection of $H_0$ would be due to the not so accurate measurements, represented by the large error bars, observed in the last multipoles (especially for
$\ell \gtrsim 2200$).
In order to evaluate this possibility we consider one more interval.
\begin{table}
\caption {Statistical LB analyses to investigate the rejection or acceptance of the null hypothesis $H_0$}
\begin{center}
\begin{tabular}{|l|l|l|l|}
\hline
Multipole intervals & $p$-value & Reject $H_0$ & $\%$repetition \\
\hline \hline
$\ell = [2, 2500]$ & $0$ & Yes & $100 \%$ \\
\hline
$\ell = [2, 100]$ & $6.0346 \times 10^{-1}$ & No & $95 \%$ \\
\hline
$\ell = [2, 800]$ & $1.1739\times 10^{-1}$ & No & $56 \%$ \\
\hline
$\ell = [2, 1200]$ & $1.709739\times 10^{-2}$ & No & $20 \%$ \\
\hline
$\ell = [1100, 2500]$ & $1.043610\times 10^{-14}$ & Yes & $100 \%$ \\
\hline
$\ell = [1600, 2500]$& $3.70592\times 10^{-8} $ & Yes & $94 \%$ \\
\hline
$\ell = [2000, 2500]$ & $1.567727\times 10^{-3}$ & Yes & $59 \%$ \\
\hline
$\ell = [1100, 2200]$ & $2.6334\times 10^{-13}$ & Yes & $100 \%$ \\
\hline
\end{tabular}
\end{center}\label{LBIntervals}
\end{table}
In fact, we also consider the statistical LB analyses in the interval: $1100 \leq \ell \leq 2200$, and according to the very small $p$-values obtained, we also confirm the rejection of the null hypothesis $H_0$ in this interval at $> 99\%$ confidence level (CL) as we can see in the table \ref{LBIntervals}.
These results confirm a significant correlation among the $\{ \delta^{obs}_{\ell} \}$ and support the hypothesis of the
presence of some structure left in the spectrum difference, even if one does not consider the data with largest errors
for $\ell > 2200$.
Roughly, the null hypothesis is rejected for $\ell \gtrsim 1000 -1200$, precisely at the angular scales where theoretical estimates
indicate that lensing signal starts to be more relevant \citep{Lewis-Challinor}.
So far, we have applied the LB test to the main values of the spectrum difference, i.e. without considering the errors $error(\delta)_{\ell}$. Now, let us investigate the robustness of our results by introducing error measurements in the LB test. In order to continue the analyses we construct artificial sets $\{{\widetilde\delta}^{obs}_{\ell}\}$ defined as $\{ {\widetilde\delta}^{obs}_{\ell}\} \equiv \{ \delta^{obs}_{\ell} + R_{\ell} \}$ and generated as follows: first, for each $\ell$ we generate a $R_{\ell}$ which is a random number obtained from a Gaussian distribution --with zero mean and standard deviation: $error(\delta)_{\ell}$-- that is added to $\{ \delta^{obs}_{\ell}\}$ to get $\{{\widetilde\delta}^{obs}_{\ell}\}$.
These sets $\{{\widetilde\delta}^{obs}_{\ell}\}$ are hypothetical points representing simulated spectrum differences inside 1$\sigma$ error.
We then apply the LB test, in different M intervals, for $100$ simulated spectrum differences $\{{\widetilde\delta}^{obs}_{\ell}\}$ and compute the percentage of repetition of rejection, or not rejection, of the null hypothesis.
The results are shown in the last column in table \ref{LBIntervals}.
For example, in the first interval $2 \leq \ell \leq 2500$, the $100\%$ of repetitions reject $H_0$, while in the second interval the
$95\%$ of repetitions do not reject $H_0$. The percentage of repetitions for $H_0$ no rejection is decreasing for the next two intervals: $2 \leq \ell \leq 800$ and $2 \leq \ell \leq 1200$, which can be an indicative that around $\ell \sim 1000$ there exists the probability of some correlation between the data $\{\delta^{obs}_{\ell}\}$. Nevertheless in all the other cases, percentage of repetitions for rejection, or no rejection, of $H_0$ is sufficiently high to reinforce the results shown in the third column of table~\ref{LBIntervals}.
As a conclusion of this subsection we must point out that there exist a strong evidence to reject the null hypothesis
in the interval $1100 \leq \ell \leq 2500$, where lensing effects on CMB photons are more relevant~\citep{Lewis-Challinor}.
Nevertheless, to be conservative, we can leave out the last multipoles due to their large error bars, i.e., $\ell > 2200$,
and will focus our next analyses in the interval $1100 \leq \ell \leq 2200$.
\subsection{Measuring the excess of lensing power}\label{sec3.3}
The first aim of this work was to investigate if the spectra difference $\delta^{obs}_{\ell}$ defined in
equation~(\ref{Dobs}), corresponds to statistical noise or not.
Results of the previous subsection showed that for several intervals with $\ell \gtrsim 1000-1200$
there seems to be a rejection of the null hypothesis $H_0$.
Now, our second aim is to explore the possibility that the $\delta^{obs}_{\ell}$ data could be well
reproduced through the $\chi^2$ best-fit analyses, described in the subsection~\ref{sec3.1},
by a synthetic APS according to the equation~(\ref{Dsyn}),
$\delta^{\mbox{\,\footnotesize exc}}_{\ell}(A_{lens})$, for some value $A_{lens} > 0$.
As illustrative examples for several $A_{lens}$ cases and diverse binnation schemes $\Delta \ell$,
we show some plots of these best-fit analyses in figures~\ref{fig1},
\ref{fig2}, and~\ref{fig3}.
The fact that one can find a good fit, i.e. $\chi^2 \simeq 1$ as seen in table~\ref{table3}, for the spectrum difference $\delta^{obs}_{\ell}$
for a synthetic APS with $A_{lens} > 0$ suggests that the lensing amplitude parameter in the Planck APS should be
indeed larger than 1, i.e. $1 + A_{lens} > 1$, meaning that this parameter was underestimated
in the analyses of the APS, $C^{\mbox{\footnotesize Planck}}_{\ell}$, done by the Planck collaboration.
Assuming that the spectrum difference $\delta^{obs}_{\ell}$ has the signature of the lensing phenomenon,
as suggested by the plots shown in figures~\ref{fig1}--\ref{fig3},
one can find the parameter value $A_{lens}$ of the synthetic APS that best-fits these data.
In the table~\ref{table3} we display the $\chi^2$ values obtained for different $A_{lens}$
values and diverse $\Delta \ell$ bin lengths.
One notice that in several cases, $\chi^2 \simeq 1$ is fully possible, including the $A_{lens} = 0$ case which
is also considered in our analyses.
This suggests that a more detailed likelihood analyses are in due.
We start considering the $\chi^2$ procedure, of the synthetic APS that best-fits the spectrum difference, $\delta^{obs}_{\ell}$,
for a continuous set of $A_{lens}$ values.
As a result, one obtains a $\chi^2$ value for each $A_{lens}$ considered, that is, $\chi^2$ becomes a function of the $A_{lens}$
parameter: $\chi^2 = \chi^2(A_{lens})$, information that can be plotted as curves $\chi^2$ versus $A_{lens}$, as observed
in the figure~\ref{fig4}, considering moreover different bin lengths $\Delta \ell$.
In the table~\ref{table4} we show the numerical values of the intersections between the curves $\chi^2(A_{lens})$
appearing in figure~\ref{fig4}, obtained for different binnations, and the straight line
$\chi^2 = 1$.
According to these results, we conclude that values $A_{lens} \ne 0$ exist, which imply that the spectrum difference
$\delta^{obs}_{\ell}$ can be well fitted by $\delta^{exc}_{\ell}(A_{lens})$ for some $A_{lens} \ne 0$ values and several $\Delta \ell$ choices.
Furthermore, we perform statistical likelihood analyses to find the best-fit $A_{lens}$ values, as shown in table~\ref{table5}, also for several bin lengths.
Observing the figure~\ref{fig4} one can argue that several binnations provide a best-fit with reduced $\chi^2 = 1$
for some $A_{lens} = A_{lens}(\Delta \ell)$ values, meaning that a solution with $A_{lens} \ne 0$ exists but is not unique.
Complementing this $\chi^2$ studies, the maximum likelihood analyses shown in figure~\ref{fig5} let us to conclude
that $A_{lens} \in [0.10,0.29]$ at 68\%~CL. See also the figure~\ref{fig6} for a detailed study of the illustrative case $\Delta \ell = 17$, where it can be seen that $A_{lens} > 0$ at 2.6 $\!\sigma$.
The same results are found for other binnation schemes.
\begin{figure
\centering
\includegraphics[height=3cm, width=8.5cm]{0.2bin63.png}
\includegraphics[height=3cm, width=8.5cm]{al20_bin63.png}
\includegraphics[height=3cm, width=8.5cm]{0.22bin63.png}
\includegraphics[height=3cm, width=8.5cm]{al22_bin63.png}
\caption{Illustrative examples: plots of the spectrum difference $\delta^{obs}_{\ell}$ together
with $\delta^{exc}_{\ell}(A_{lens})$, for the cases $A_{lens} = 0.20$ (first and second panels) and $A_{lens} = 0.22$
(third and fourth panels), for $\ell \in [1100,2000]$ and $\ell \in [1100,2200]$ as indicated in the plots.
In all these plots we consider $\Delta \ell = 63$.
Although the best-fit curve in the first and third panels seem equals,
they are indeed slightly different.
}\label{fig1}
\end{figure}
\begin{figure
\centering
\includegraphics[height=3cm, width=8.5cm]{0.2test.png}
\includegraphics[height=3cm, width=8.5cm]{al20_bin32.png}
\includegraphics[height=3cm, width=8.5cm]{0.2.png}
\includegraphics[height=3cm, width=8.5cm]{al20_bin40.png}
\caption{Illustrative examples: plots of the spectrum difference $\delta^{obs}_{\ell}$ together with
$\delta^{exc}_{\ell}(A_{lens})$ for $\ell \in [1100,2000]$ and $\ell \in [1100,2200]$ as indicated in the plots.
In all these plots we consider $A_{lens} = 0.20$.
They were obtained for the bin lengths $\Delta \ell = 32$, for the first and the second panels and
$\Delta \ell = 40$ for the third and the fourth panels.
}
\label{fig2}
\end{figure}
\newpage
\begin{figure
\centering
\includegraphics[height=6cm, width=8.5cm]{al20_delta63.png}
\includegraphics[height=6cm, width=8.5cm]{al22_delta63.png}
\caption{Illustrative analyses comparing the spectrum difference $\delta^{obs}_{\ell}$ and
$\delta^{exc}_{\ell}(A_{lens})$ for $\ell = [1100, 2200]$ and $\ell = [1100, 2000]$ with $\Delta \ell = 63$
for $A_{lens} = 0.20$ (upper 2 plots) and $A_{lens} = 0.22$ (lower 2 plots)}
\label{fig3}
\end{figure}
\begin{figure
\centering
\includegraphics[height=6cm,width=8.5cm]{bin.png}
\caption{$\chi ^2$ as a function of $A_{lens}$ and for different number of bins. This plot shows
that, independent of the bin size $\Delta_{\ell}$, the $\chi^2$ function exhibits solutions for the case $\chi^2 = 1$ for
$A_{lens} \ne 0$, which justify our search for the value $A_{lens}$ that best-fits the observed data
$\delta^{obs}_{\ell}$.
}
\label{fig4}
\end{figure}
\begin{table
\caption{Illustrative examples of $\chi^2$ calculations for different values of $A_{lens}$
and several binnation choices $\Delta \ell$. Some of these cases can be seen in figures~\ref{fig1},~\ref{fig2},
and~\ref{fig3}.} \label{table3}
\begin{center}
\begin{tabular}{|l|l|l|l|l|l|l|l|}
\hline
$A_{lens}$ & $\Delta \ell = 17$ & $\Delta \ell = 32$ & $\Delta \ell = 41$ & $\Delta \ell = 51$ & $\Delta \ell = 63$ \\
\hline \hline
0.00 &1.0363 & 1.2809 & 1.02590 & 0.95802 & 1.63446 \\
\hline
0.10 &0.9414 & 1.0747 & 0.7610 & 0.6416 & 1.0798 \\
\hline
0.20 & 0.9329 & 1.0305 & 0.7094 & 0.5851 & 0.8395 \\
\hline
0.24 & 0.9525 & 1.0561 & 0.7459 & 0.6319 & 0.8274 \\
\hline
0.34 & 1.0565 & 1.2238 & 0.9734 & 0.9150 & 0.9980 \\
\hline
\end{tabular}
\end{center}
\end{table}
\vspace{0.5cm}
\subsection{Cosmological parameters dependency}
In the previous section we explore the possibility that the residual structure in CMB lensing signal can be explained by an excess in lensing amplitude $A_{lens} > 0$. However, there is also the possibility that such residual structure could be mimicked by varying in other cosmological parameters. The scrutiny of this dependency is important to confirm if the signal found in the previous section is indeed due to
the lensing effect or it also has contribution of other cosmological parameters.
This analysis is the main objective in this subsection.
As it is well known, the APS, $C^{\Lambda\mbox{\footnotesize CDM}}_{\ell}$, is sensitive to the cosmological parameters \citep{Peebles1968,Doroshkevich1978,Wilson1981}. Such sensitivity is inherited by the quantity $\delta^{obs}_{\ell}$ defined in equation (\ref{Dobs}). So far, the model considered was the simplest $\Lambda$CDM with the six parameters listed in table~\ref{table1} plus lensing amplitude $A_L=1$. Now, the space of parameters is extended to include the neutrino mass $\sum m_\nu$ and the spatial curvature $\Omega_k$. Massive neutrinos slow down the growth of matter perturbations and prevent clustering process \citep{Bond1980,Lesgourgues2006}, resulting, therefore, in an anticorrelation with gravitational lensing amplitude. Moreover, spatial curvature also has correlation with lensing and thus, CMB lensing signal is sensitive to the neutrino mass and the spatial curvature of the universe.
In order to investigate the space of parameters
($A_{lens}$, $\sum m_\nu$, $\Omega_k$) in the light of the spectrum difference $\delta^{obs}_{\ell}$, the other cosmological parameters were fixed by using best-fit values found by Planck 2018 data~\cite{Planck18} (see table \ref{table1}) and the parameters $A_{lens}$, $\sum m_\nu$, and $\Omega_k$ were allowed to vary. Then we use MCMC techniques to find constraints on these three free parameters by considering flat priors and marginalizing to compute the one dimensional posterior distributions. Results are presented in
figure~\ref{fig7} and table~\ref{table7}.
In figure~\ref{fig7} the one dimensional posterior distributions for $\Delta \ell = 17$ and $\Delta \ell =63$ are shown. We can say, in general, that our results are consistent at $68 \%$ CL with an almost flat spatial curvature, with $\sum m_\nu < 0.47$ eV, and with a $\sim 20\%$ lensing amplitude excess quantified by the $A_{lens}$ parameter. This $A_{lens}$ excess is still present at around three standard deviations (see table \ref{table7}), even though these analyses considered a larger number of parameters.
All these results are robust when changing $\Delta \ell$ indicating that the neutrino mass or the spatial curvature weakly impact in the $\delta^{obs}_{\ell}$ intensity signal.
Therefore their effects are not sufficient to explain the signature in the spectrum
difference $\delta^{obs}_{\ell}$, which is more likely explained by a lensing amplitude $A_{lens} \simeq 0.20$, which implies that the lensing amplitude in the Planck's CMB APS should be $\sim 20\%$ larger than the value expected in the flat-$\Lambda$CDM model.
\begin{table
\caption{Table with the numerical values obtained due to the intersection between the horizontal line,
representing the reduced-$\chi^2 = 1$, with the curves produced calculating the reduced-$\chi^2$ for the parameter $A_{L}$,
and considering various bin-length cases $\Delta \ell$.
The first and second points in the table below represent the values of $A_{L}$ where $\chi ^2 = 1$.
As observed in figure~\ref{fig4} there is no intersection for the case $\Delta \ell = 32$.
} \label{tab:title}
\begin{center}
\begin{tabular}{|l|l|l|l|l|l|l|l|}
\hline
bin length & first point & second point \\
\hline \hline
$\Delta \ell = 17$ & $0.028675$ & $0.29479$ \\
\hline
$\Delta \ell =32$ & \hspace{0.4cm}-- & \hspace{0.4cm}-- \\
\hline
$\Delta \ell =41$ & $0.02495$ & $0.3305$ \\
\hline
$\Delta \ell =51$ & $0.0023$ & $0.3477$ \\
\hline
$\Delta \ell =63$ & $0.16053$ & $0.29425$ \\
\hline
\end{tabular}
\end{center}\label{table4}
\end{table}
\begin{table
\caption{Values of $A_{L}$ from the likelihood plot. The last column indicates the confidence level (CL) for having $A_{lens} > 0$ in each binnation case.}~\label{table5}
\begin{center}
\begin{tabular}{|l|c|c|c|l|l|l|l|}
\hline
$\!$bin length$\!$ & $\!\!A_{lens}$ at max. likelihood$\!$ & $\!\!$1$\sigma$ interval$\!$ & $\!\!$CL for $A_{lens}\!>\!0$$\!\!$ \\
\hline \hline
$\Delta \ell = 17$ & $0.16036$ & $[0.0993, 0.2213]$ & 2.6$\sigma$\\
\hline
$\Delta \ell =32$ & $0.177477$ & $[0.1164, 0.2384]$ & 2.8$\sigma$\\
\hline
$\Delta \ell =41$ & $0.17567$ & $[0.1136, 0.2376]$ & 2.8$\sigma$\\
\hline
$\Delta \ell =51$ & $0.172072$ & $[0.1090, 0.2350]$ & 2.7$\sigma$\\
\hline
$\Delta \ell =63$ & $0.22702$ & $[0.1630, 0.2910]$ & 3.5$\sigma$\\
\hline
\end{tabular}
\end{center}\label{likelihoods}
\end{table}
\begin{figure
\centering
\includegraphics[height=7cm, width=8.5cm]{confidence.png}
\caption{Maximum likelihood analyses, complementing the analyses done in figure~\protect\ref{fig4}, showing the 68\% CL of the $A_{lens}$ values that perform the reduced $\chi^2$ best-fit of the spectrum difference $\delta^{obs}_{\ell}$, that is $A_{lens} \in [0.10,0.29]$, for several bin sizes.}
\label{fig5}
\end{figure}
\begin{figure
\centering
\includegraphics[height=7cm, width=8.5cm]{confidence1.png}
\caption{Illustrative example of $\sigma$ intervals for the $\Delta \ell = 17$ case.
As observed, for this binnation case the value $A_{lens} = 0$ (equivalently $A_{L} = 1$)
is excluded with $2.6 \sigma$ (i.e., at 99\% CL).}
\label{fig6}
\end{figure}
\section{Discussions and Final Remarks}\label{sec4}
We now discuss the exhaustive analyses done with the precise Planck APS data and the results obtained investigating the statistical features of the spectrum difference $\delta^{obs}_{\ell}$.
Recent analyses have reported the preference of the lensing amplitude parameter $A_L$ for a value larger than 1, with high statistical significance~\citep{Planck16,Planck18,Bianchini,DES22},
and we want to study if this result could be sourced by a statistical artifact or if it has a physical origin.
Our analyses can be considered complementary, and independent from those done by the Planck collaboration~\cite{Planck18}.
In the previous section we have performed a detailed examination of the unbinned CMB APS.
We have proved that the spectrum difference $\delta^{obs}_{\ell}$ is not statistical noise;
additionally we found several multipole intervals, starting at $\ell \gtrsim 1000$, where this result is true
(see table~\ref{LBIntervals}), supporting the crucial result that the spectrum difference is not statistical noise at small scales.
This result is suggesting that some signal is present in the spectra difference, and the signature shown in
these data must be an indication of the phenomenon that caused it.
However, degeneracy can also happen, that is, the signature present in the data could be reproduced by more than
one source and for this reason we also explore diverse hypotheses to explain the signature and the amplitude of
the spectrum difference data.
Bearing in mind that, at the scales $\ell \gtrsim 1000$, the acoustic peaks are extremely sensitive to the lensing
phenomenon, the first hypothesis examined was that the signature in the spectrum difference corresponds to an underestimated lensing amplitude in the Planck best-fitting procedure.
As shown in the analyses of section~\ref{sec3.3} this hypothesis is indeed verified, where we find a lack of lensing amplitude of around $20\%$ with respect to the Planck APS fitted assuming the flat-$\Lambda$CDM model with $A_L = 1$.
As a matter of fact, we find that there is an excess of signal in $\delta^{obs}_{\ell}$ (see equation~(\ref{Dobs}))
that is well explained by a $\Lambda$CDM APS with a non-null lensing amplitude parameter $A_{lens} > 0$, with values in the interval
$[0.10,0.29]$ at 68\%~CL.
Moreover, we found several scheme binnations that best-fits the spectrum difference $\delta^{obs}_{\ell}$,
with $\chi^2 = 1$, with lensing amplitudes $A_{lens} \simeq 0.2$ (see
tables~\ref{table4} and~\ref{table5}).
According to our likelihood analyses, the synthetic APS produced for these best-fit procedures of $\delta^{obs}_{\ell}$,
with lensing amplitude $A_{lens}$ as a parameter, show that $A_{lens} > 0$, or equivalently $A_L > 1$, with statistical significance of $\sim 3 \sigma$.
Additionally, we have also investigated a possible dependency of the spectrum difference $\delta^{obs}_{\ell}$,
in signature and intensity, on some cosmological parameters associated to lensing amplitude, neutrino mass and spatial curvature. Despite neutrino mass and spatial curvature can have impact in lensing signal, the impact in $\delta^{obs}_{\ell}$ residual is weak and cannot be enough to reproduce the excess in $A_{lens}$, in signature and intensity.
Our likelihood analyses for the examination of the cosmological parameters
$A_{lens}$, $\sum m_{\nu}$, and $\Omega_k$ are displayed in figure~\ref{fig7}.
\begin{figure
\centering
\includegraphics[height=16cm, width=8.5cm]{SeveralParameters.png}
\caption{One dimensional posterior distributions for $A_{lens}$, $\sum m_\nu$ and $\Omega_k$ for $\Delta \ell=17$ and $\Delta \ell=63$, as illustrative examples.
The $\delta^{obs}_{\ell}$ signal is consistent with: $\sum m_\nu < 0.47$eV, an almost flat spatial curvature, and $\sim 20\%$ of excess in $A_{lens}$. These results show the consistency of the different values that these parameters can assume to explain the spectrum difference $\delta^{obs}_{\ell}$.}
\label{fig7}
\end{figure}
\begin{table}
\caption{Values $A_{lens}$, $\sum m_\nu$ and $\Omega_k$ at 1$\sigma$ for $\Delta \ell = 17$ and $\Delta \ell = 63$.
The last column indicates the confidence level (CL) for having $A_{lens} > 0$ in each binnation case.}~\label{table7}
\begin{center}
\begin{tabular}{|l|c|c|c|l|l|l|l|}
\hline
$\!$bin length$\!$ & $\!A_{lens}$ $\!$ & \!$\sum m_\nu $\! & $\!\Omega_K\!$ & \!\!CL for $A_{lens}\!>\!0$\\
\hline \hline
$\Delta \ell = 17$ & $0.18^{+0.063}_{-0.063}$ & $<0.47$ eV & $0.0029^{+0.009}_{-0.009}$ & 3$\sigma$\\
\hline
$\Delta \ell =63$ & $0.22^{+0.065}_{-0.065}$ & $< 0.45$ eV & $0.0013^{+0.0071}_{-0.0071}$ & 2.9$\sigma$\\
\hline
\end{tabular}
\end{center}
\end{table}
\section*{Acknowledgements}
We acknowledge the use of data from the Planck/ESA mission, downloaded from the Planck Legacy Archive.
RM and WHR thank CNPq, TWAS and FAPES (PRONEM No 503/2020) for the fellowships and finantial support under which this work was
carried out.
AB acknowledges a CNPq fellowship.
We acknowledge that our work made use of the CHE cluster, managed and
funded by the COSMO/CBPF/MCTI, with financial support from FINEP and FAPERJ, and operating
at Javier Magnin Computing Center/CBPF.
\bibliographystyle{mnras}
|
\section{Introduction}
\label{section:introduction}
Finite automata, or finite state machines, are the simplest class of computational devices possessing memory.
Their memory is finite and stored exclusively as their \textit{state}, but that alone is enough to make their analysis challenging.
In contrast to simpler devices entirely characterisable by mappings of individual inputs to their corresponding outputs, finite automata may need the entire input history to determine their next state or output.
The learning of finite automata is a classical problem in computational learning theory, and can be phrased as follows: given a language $\mathcal{L}$ and a set of examples $\mathcal{E}$, use $\mathcal{E}$ to find a finite acceptor automaton $\mathcal{D}$ that accepts words in $\mathcal{L}$ and rejects words not in $\mathcal{L}$, or at least find a $\mathcal{D}$ that performs these duties with sufficient accuracy.
For instance, $\mathcal{L}$ can be the language generated by the regular expression $a^*bb^*$.
The set of examples $\mathcal{E}$ could consist of the entire $\mathcal{L}$, with us placing an additional requirement of conciseness on the learning process to arrive at a single, practical solution (say an automaton with at most three states).
Or, $\mathcal{E}$ could contain words such as $aab$ and $bbb$ marked as positive examples (the words $\mathcal{D}$ should accept), and $aaa$ and $aba$ marked as negative examples (words to be rejected by $\mathcal{D}$).
Strictly algorithmic, formal, statistical, and genetic approaches have all been proposed, but this problem remains open mainly because the solutions lack an agreed single measure of merit.
Some have argued for automata that concisely represent $\mathcal{E}$, while others preferred automata that generalised well to $\mathcal{L}$ in scenarios where $\mathcal{E}$ contains only a few examples representative of the original language.
Another degree of freedom arises in the constraints that are placed on $\mathcal{E}$.
One can insist that $\mathcal{E}$ consists of only positive examples (i.e. $\mathcal{E} \subseteq \mathcal{L}$), that $\mathcal{E}$ contains enough examples to sanction good generalisation to $\mathcal{L}$ (e.g. all repetitions of a pattern beyond a certain count signify that the pattern may be repeated indefinitely), or that $\mathcal{E}$ contains both positive and negative examples (i.e. $\mathcal{E} \cap \mathcal{L} \neq \emptyset \neq \mathcal{E} \cap \mathcal{L}^c$, where $\mathcal{L}^c$ denotes words not in $\mathcal{L}$).
We give a detailed overview of related work in \Cref{section:related_work}.
This paper proposes a neural model that learns a finite acceptor automaton for target language $\mathcal{L}$ from a given $\mathcal{E}$ consisting of positive and negative examples.
We insist that the logic of the resulting automaton is fully explainable -- fully comprehensible by humans, that the resulting automata generalise well to $\mathcal{L}$ despite being based on only $\mathcal{E}$, and that the representations are concise, or at least such that they do not contain too much redundancy of computational logic.
Internally, our model learns a regular grammar rather than a finite automaton.
Briefly, a grammar $\mathcal{G}$ is a triplet of $\mathcal{A}_T,\mathcal{A}_N,\mathcal{P}$ -- the terminal alphabet, non-terminal alphabet, and set of productions, respectively, where one of the letters of $\mathcal{A}_N$ is marked as the start symbol (the ``root'' of $\mathcal{G}$).
A left-regular grammar is a grammar in which all productions of $\mathcal{P}$ are of the form $A \to c$, $A \to \epsilon$, or $A \to Bc$ for $A,B \in \mathcal{A}_N$, $c \in \mathcal{A}_T$, $\epsilon$ the empty string.
A right-regular grammar uses productions of the form $A \to cB$ instead of $A \to Bc$.
Regular grammars are equivalent to finite automata in their language-generating power and admit straightforward conversions into each other.
We do not consider empty languages or grammars that give empty languages.
In our setting, there is a target language $\mathcal{L} = \left< \mathcal{G} \right>$ generated by the ground-truth grammar $\mathcal{G}$.
We are shown examples $\mathcal{E}$ taken from both $\mathcal{L}$ and $\mathcal{L}^c$, and our model internally learns a hypothesis grammar $\mathcal{G}'$.
In other words, we do not train neural automata to become accurate acceptors for given languages $\mathcal{L}$, but work instead with a fully explainable regular language learning parser under an acceptor training setup.
We use less information than many previous methods for grammatical inference and let the grammar emerge under single-bit supervision.
Our contributions are:
\begin{itemize}
\item We introduce of a novel neural model tailored specifically to the learning of regular grammars from positive and negative acceptor examples.
There are no limits on the complexity of the regular grammars it can learn.
\item We describe in detail a procedure for the recovery of the learned grammars and parse trees from the internals of our model, warranting explainability.
\item We systematically evaluate our model across the dimensions of grammar size, grammar complexity, and training data quantity.
\end{itemize}
\section{Related Work}
\label{section:related_work}
From the perspective of algorithmic learning, we divide the literature on learning of regular languages into two groups: one concerning the learning of finite automata and the associated regular expressions, the other addressing the problem of grammar induction.
\textit{Learning finite automata.} Early work on the subject leveraged Hidden Markov Models \cite{rabiner1986introduction} and probabilistic finite state machines \cite{carrasco1994learning} as the models for learning.
Later approaches readily recognized the utility of regular expressions as a form for description and tended to be deterministic.
Polynomial-time algorithms were proposed for learning of regular expressions without union operation from chosen classes of positive examples \cite{brazma1993efficient} and for learning of unambiguous regular expressions of maximum loop depth 2 \cite{kinber2010learning}. A genetic programming approach leveraging both positive and negative examples was used in \cite{svingen1998learning}, and a further polynomial-time method for $1$-unambiguous regular expressions aiming for simplicity of the resulting expressions was presented in \cite{fernau2009algorithms}.
Most of the recent work on regular expression learning takes the line of natural language processing (NLP) and proposes methods for particular uses in real-world datasets.
A genetic programming method for text extraction from XML documents is outlined in \cite{bartoli2014automatic}.
\cite{prasse2012learning} focuses on identifying email campaigns with high precision while phrasing it as an optimisation problem, and \cite{bui2014learning} gives a method based on Support-Vector Machines tailored to clinical texts.
Note that all of the above work either fences out a particular sub-class of regular expressions that are to be learned (e.g. union-less or $1$-unambiguous) or makes additional assumptions based on the nature of the particular real-world problem it is designed to solve.
\textit{Grammar induction.}
Pioneering work on the learning of grammars from examples often attempted to construct probabilistic context-free grammars that generated the target language $\mathcal{L}$ \cite{fu1975grammatical}.
Subsequent attempts employed a wider variety of tactics including Bayesian methods \cite{chen1995bayesian} and genetic programming \cite{wyard1993context}.
The learning was done chiefly from positive examples, though methods for automatic negative example generation from positive examples were later introduced \cite{smith2005guiding}.
Similarly to above, contemporary work on grammar induction is almost exclusively guided by the desired applications in NLP rather than interest in theory of computation.
As such, black-box recurrent neural automata \cite{choi2018learning}, reinforcement learning models \cite{yogatama2016learning}, and partially interpretable structured attention \cite{kim2017structured} and transformer \cite{goertzel2020guiding} models dominate the sub-field at present.
Our approach spans both of these groups.
We train on accept/reject information just like an acceptor automaton but learn a grammar.
In contrast to research on learning finite automata and regular expressions, there are no inherent limits on the complexity of grammars our model can learn -- it can learn any regular grammar, not a sub-class characterised by a particular type of regular expressions.
Further, in contrast with the recent efforts in grammar induction, the neural architecture we present learns a grammar in a fully explainable fashion.
Explainability is present in the training procedure, grammar extraction, and the use of our model for direct parsing by the learned grammar.
Given the above, our approach is more akin to research in program synthesis and fits better under the paradigm of ``algorithm learning''.
This is because we can view the language examples together with the annotations that mark them as positive and negative as input-output pairs for the broader \textit{programming by example (PbE)} task \cite{smith2000programming}.
In PbE, a program in a given programming language is to be synthesised based on a set of input-output pairs.
In our case, the programming language is the rewriting scheme for regular grammars and the program is the particular regular grammar that generates $\mathcal{L}$ or a substantial overlap therewith.
\section{Model}
\label{section:architecture}
Our network consists of three components: \textit{Terminal Grammar Unit}, \textit{Non-terminal Grammar Unit}, and \textit{Start Selector Unit}.
These are engaged sequentially following a simplified CYK algorithm to perform the function of an acceptor for the target language $\mathcal{L}$.
After training, the internals of each unit can be directly inspected to recover the learned regular grammar $\mathcal{G}' = \left(\mathcal{A}_T, \mathcal{A}_N', \mathcal{P}'\right)$.
Note that while the terminal alphabet $\mathcal{A}_T$ is considered known (from $\mathcal{L}, \mathcal{E}$) and shared with our model, the non-terminal alphabet $\mathcal{A}_N'$ is learned indirectly through use in learned productions $\mathcal{P}'$.
We present our method as for the learning of left-regular grammars, but the entire setup can be inverted to produce right-regular grammars.
\begin{figure*}[t!]
\centering
\includegraphics[width=0.95\textwidth]{figures/Blocks.png}
\caption{
An overview of the model's neural structure.
\textit{From left.} The Terminal Grammar Unit is applied to the first letter of the input word $w$.
The non-terminal grammar unit is then applied recurrently (i.e. the trained weights are shared between individual instances) to the remaining letters.
The final parse belief vector $u_{\ell}$ is fed into the Start Selector Unit to yield the accept/reject verdict.
}
\label{figure:blocks}
\end{figure*}
\subsection{Application Algorithm}
\label{section:application_algorithm}
For each example word $w{:}a_1a_2a_3\dots a_{\ell}$ we one-hot encode letters $a_i$ into $t$-dimensional vectors $v_i$, where $t = \left|\mathcal{A}_T\right|$ is the size of the terminal alphabet.
If $w$ is a positive example we associate it with label $y=1$, otherwise $y=0$.
$v_1$ is then fed into the Terminal Grammar Unit to produce an intermediate parse $n'$-dimensional (belief) vector $u_1$.
$n'$ is a parameter of the training, giving an upper bound on the number of non-terminals that may appear in $\mathcal{P}'$.
The ground-truth non-terminal alphabet $\mathcal{A}_N$ is not known to us during training and so neither is its size $n = \left|\mathcal{A}_N\right|$.
$n'$ is therefore just a guess for $n$, made in a thought process similar to that for the number of latent dimensions to be used in a disentangling variational autoencoder.
For $1 \leq i < \ell$, $v_{i+1}$ and $u_i$ are then provided as the input for the Non-terminal Grammar Unit, which in turn produces the intermediate parse vector $u_{i+1}$.
Finally, $u_{\ell}$ is processed by the Start Selector Unit, which outputs a number $o(w)$ between $0$ and $1$ representing the belief that $w \in \mathcal{L}$.
The whole procedure is illustrated in \Cref{figure:blocks} on the unit level, and in \Cref{figure:internals} with internals and on a simple example.
In training, $o(w)$ is assigned a loss value computed as the binary cross-entropy between $y$ (true label) and $o(w)$ (predicted label).
The model is trained to minimise total loss combining the cross-entropy and two other losses reflecting the quality of the hypothesis grammar.
\subsection{Terminal Grammar Unit (TGU)}
The TGU serves as the hypothesis grammar parser for potential terminal productions.
It takes a $t$-dimensional vector $v_1$ -- the one-hot encoding of the first letter -- as input, uses it to query the trained hypothesis terminal production $n'$-by-$t$ matrix $P_T$, and clamps the output between $0$ and $1$:
$
\text{TGU}\left( v_1 \right) := \text{clamp}\left(\sigma\left(P_T\right) v_1\right),
$
where $\text{clamp}(v) := \max\left(\min\left(1, v\right), 0\right)$ and $\sigma$ is the logistic sigmoid, each applied element-wise.
If $A$ and $b$ are one-hot-encoded as the $i$-th and $j$-th euclidean basis vectors, then the $i,j$-th entry of $\sigma\left(P_T\right)$ represents the belief of the model (the strength of TGU's hypothesis) that the production $A \to b$ is a part of the grammar $\mathcal{G}'$ being learned from $\mathcal{E}$.
\subsection{Non-terminal Grammar Unit (NGU)}
\label{section:ngu}
The NGU is the recurrent unit of our architecture and is the parser for the hypothesis grammar's non-terminal productions.
It holds $n'$ $n'$-by-$t$ trained matrices, each representing the hypothesised non-terminal productions with one of the $n'$ potential non-terminals $\mathcal{A}_N'$ on the left-hand side.
To perform its parse, the NGU takes the previous parse vector $u_i$, current terminal vector $v_{i+1}$, and yields $u_{i+1}$ where the $k$-th entry of $u_{i+1}$ for $1 \leq k \leq n'$ is given by
$
\text{NGU}\left(u_i, v_{i+1} \right)_k := \text{clamp}\left(\left(\sigma\left(P_N^k\right) v_{i+1}\right)^{\text{T}}u_i\right).
$
If $A,B$ and $c$ are one-hot-encoded as the $k$-th, $i$-th, and $j$-th euclidean basis vectors respectively, then the $i,j$-th entry of $\sigma\left(P_N^k\right)$ represents the strength of NGU's hypothesis that the production $A \to Bc$ belongs to $\mathcal{P}'$.
\subsection{Start Selector Unit (SSU)}
The final parsing belief vector $u_\ell$ describes the model confidence about each of the $n'$ non-terminals being the root non-terminal.
In order for a parse to be successful, the parsing must terminate by reaching a particular non-terminal letter having the role of the start symbol.
We allow for the start symbol to emerge in an unsupervised fashion by letting the model learn which of the non-terminals should be considered to have the function of the root of $\mathcal{G}'$.
We achieved this by making a constant softmax choice (paying ``constant attention'') in the Start Selector Unit:
$
o(w), \text{SSU}\left( u_\ell \right) := \text{softmax}\left(s\right)^\text{T} u_\ell,
$
where $s$ is a trained $n'$-dimensional vector.
We also experimented with using a two- and three-layer multi-layer perceptron (MLP) networks and found no difference in performance but observed a tendency of the model to encode the productions of the grammar's start symbol in the MLP, hindering explainability.
We also considered fixing one of the entries of $u_\ell$ but observed a decrease in recall scores.
\begin{figure*}[t!]
\centering
\includegraphics[width=0.90\textwidth]{figures/Units4.png}
\caption{
An illustration of the model's internals.
The productions $\mathcal{P}'$ of the model hypothesis grammar $\mathcal{G}' = \left(\mathcal{A}_T, \mathcal{A}_N', \mathcal{P}'\right)$ generate the language given by the regular expression $\left(a|b\right)^*cc^*$.
The $\times$ symbol represents matrix multiplication, $\cdot$ represents vector inner product, and the colour coding of productions and matrix entries marks equivalences by the encoding-decoding procedures of \Cref{section:ngu} and \Cref{section:grammar_extraction}.
We present the model with a string $abb$.
\textit{Left.} The Terminal Grammar Unit matches the initial $a$ encoded as $v_1 = (1\,0\,0)^{\text{T}}$ with the production $A \to a$ and produces the belief vector $u_1 = (1\,0)^\text{T}$.
\textit{Middle}. The Non-terminal Grammar Unit has its production matrices queried by $v_2 = (0\,1\,0)^{\text{T}}$ and finds a match in the production $A \to Ab$.
This is then dotted with the prior belief $u_1$ to produce the next parse belief $u_{2}$.
The same is repeated for $v_3,u_2 = v_2,u_1$ giving $u_\ell = u_3 = (1\,0)^\text{T}$.
\textit{Right.} The model has been trained to recognize $C$ as the root symbol, but the terminal parse $u_\ell$ on the given negative example is $(1\,0)^\text{T}$, leading to $0$ as the output of the acceptor.
}
\label{figure:internals}
\end{figure*}
\subsection{The Role of clamp$\left( \cdot \right)$}
If the hypothesis grammar is or nears being ambiguous, the vector inner product in NGU often results in belief values in $u_{i+1}$ being greater than $1$ (e.g. as in \Cref{figure:tree}).
To keep the model trainable and explainable at the same time, we limit all grammar unit outputs to $1$.
The values in our model never go below $0$, but we keep the lower bound in the definition of clamp$\left( \cdot \right)$ for consistency with similar work in neural networks.
\subsection{Training Losses}
\label{section:losses}
The training loss for our model consists of three components, with relative contributions to the total loss being controlled by hyperparameter factors.
\subsubsection{Prediction loss} The binary cross-entropy between the true labels ($1,0$ for positive,negative examples $w$) and predicted labels $o(w)$.
The prediction loss guides the model towards learning a grammar that generates exactly $\mathcal{L}$.
\subsubsection{Sharpening loss} For every entry $e$ in $\sigma\left(P_T\right),\sigma\left(P_N^k\right)$ we compute the sharpening contribution $
\mathcal{S}\left( e \right) := 1 - \left(2e - 1\right)^2
$ and then compute the total sharpening loss $
\mathcal{S} := \frac{1}{n'(n' + 1)t}\sum_e \mathcal{S}\left( e \right).
$
The sharpening loss helps to ensure that the values of $\sigma\left(P_T\right),\sigma\left(P_N^k\right)$ are eventually clearly interpretable as productions of $\mathcal{G}'$.
\subsubsection{Production use loss} Simply the mean of all entries $e$ of $\sigma\left(P_T\right),\sigma\left(P_N^k\right)$, i.e.
$
\mathcal{U} := \frac{1}{n'(n' + 1)t} \sum_e e \,.
$
Intuitively, this loss encourages the use of smaller grammars in contrast to larger ones.
\subsubsection{Total loss}
The total loss for a batch $\mathcal{B}$ is then
$
\ell(\mathcal{B}) := \text{BCE}\left(\mathcal{B}\right) + \beta \mathcal{S} + \gamma \mathcal{U},
$
where $\beta, \gamma \geq 0$ are hyperparameters.
\begin{figure}[t!]
\centering
\includegraphics[width=0.45\textwidth]{figures/Tree3.png}
\caption{
An example of parse tree construction from parsing input vectors $v_i,u_i$.
Expanding on example from \Cref{figure:internals} we consider an instance of our model employing grammar $\mathcal{G}^* := \left(\mathcal{A}_T, \mathcal{A}_N', \mathcal{P}^*\right)$ to parse the word $bbacc$.
$\mathcal{P}^*$ is an extension of $\mathcal{P}'$, with the additional productions introducing ambiguity leading to two more potential parse trees.
The ambiguity leads to values exceeding $1$ in the dot product in the NGU and clamping is engaged to keep the values between $0$ and $1$.
\textit{Table from the top.} The initial terminal $b$ can be parsed by either of $A$ or $C$.
For $i=2$, each of $A \to Ab, C \to Ab, C \to Cb$ can be engaged producing partial parses rooted at $A$ or $C$.
For $i=3$, either $A \to Aa$ or $C \to Ca$ can be used, conditional on the root of the previous intermediate parse.
For $i=4,5$, $C \to Cc$ applies.
}
\label{figure:tree}
\end{figure}
\subsection{Grammar Extraction}
\label{section:grammar_extraction}
We set a confidence threshold $\tau$ for when an entry of a production matrix $P_\bullet$ is to be interpreted as signifying the presence of the production in the grammar.
In our experimentation, $\tau = 0.95$ proved to be a reliable choice, though any threshold strictly below $1$ is eventually achievable owing to the sharpening loss.
Denote $\left(M\right)_{ij}$ the $i,j$-th entry of a matrix $M$, $a_k$ the $k$-th terminal in $\mathcal{A}_T$, and $A_k$ the $k$-th non-terminal in $\mathcal{A}_N'$.
Then the productions $\mathcal{P}'$ of the induced grammar $\mathcal{G}'$ can be extracted from the TGU and NGU by the following procedure:
\begin{itemize}
\item If $\left(\sigma(P_T)\right)_{ij} \geq \tau$, add $A_i \to t_j$ to $\mathcal{P}'$.
\item For each $1 \leq k \leq n$, if $\left(\sigma(P_N^k)\right)_{ij} \geq \tau$ then add $A_k \to A_i a_j$ to $\mathcal{P}'$.
\item Let $\mu := \argmax_k \text{softmax}\left( s \right)_k$. Add $S \to A_\mu$ to $\mathcal{P}'$.
\end{itemize}
Note that by constructing $\mathcal{P'}$ in this manner, $\mathcal{A}_N'$ may contain non-terminals that are never used or that can never be reached from the start symbol $S$ in a derivation.
\subsection{Parse Tree Construction}
While the extracted induced grammar forms a basis for parser construction, the TGU and NGU can be used to construct the parse trees directly.
For all $1 \leq i \leq \ell$, $u_\ell$ multi-hot-encodes all non-terminals that can be used to generate the prefix sub-word $w_{\leq i} = a_1 \dots a_i$ of $w$.
This can be done inductively in $i$ as follows: for $1 \leq j \leq n$, if $\left(\sigma(P^k) v_i\right)_j \geq \tau$ and $u_i \geq \tau$, $N_k$ is a parse tree node for $u_{i+1}$ with one child being the root of the parse tree for $u_i$ and the other being the leaf terminal $a_i$.
This is illustrated in \Cref{figure:tree}.
Such a tree may exist for every $j$, giving all the at most $n'$ possible parse tree roots for any prefix sub-word $w_{\leq i}$.
\section{Experiments}
\label{section:experiments}
To systematically evaluate the grammar induction performance of our model, we generate left-regular grammars of varying complexity using a randomized procedure, and then use the said grammars to produce training datasets consisting of positive and negative examples.
Once trained, the instances of our model are inspected for their induced grammars $\mathcal{G}'$ and tested for how closely they match the ground-truth $\mathcal{G}$.
Due to the lack of related work addressing the problem of general regular grammar induction from positive and negative examples (cf. \Cref{section:related_work}), the objective of our experimentation is to investigate the robustness of our method to increases in grammar complexity and potential brevity in training examples (i.e. the cases when the training examples may be plenty but are not long enough to be wholly confident about the recurrence in a production).
This is in line with the example evaluation approach taken in other work \cite{svingen1998learning,bartoli2014automatic,prasse2012learning}.
\subsection{Grammar Generation}
During the ground-truth grammar generation phase, we vary the number of terminals $t$, the number of non-terminals $n$, and the average number of productions per non-terminal $\mathfrak{p}$, thus controlling the complexity of the generated grammar.
For each non-terminal $A_i$, the number of productions is subsequently sampled from a geometric distribution with parameter $\pi = \mathfrak{p}^{-1}$.
Each production for $A_i$ is either a terminal production of the form $A_i \to a$ with probability $p_T = 0.4$ or a non-terminal production of the form $A_i \to A_j a$ with probability $p_N = 1 - p_T$.
$A_j$ is chosen uniformly at random from the set $\mathcal{A}_N$ of candidate non-terminals and $a$ is chosen uniformly at random from the set $\mathcal{A}_T$ of terminals.
We designate one symbol in $\mathcal{A}_N$ as the start symbol $S$ without preference.
In order for the language $\mathcal{L} = \left<\mathcal{G}\right>$ to be non-empty, we add a randomly sampled terminal production to the set $\mathcal{P}$ of productions if there are none, and furthermore ensure the derivation reachability of all non-terminals in $\mathcal{A}_N$ from the start non-terminal $S$ by adding a single production $A_i \to A_j a$ for each unreachable non-terminal $A_j$ and some reachable non-terminal $A_i$. $A_i,a$ are again taken at random.
\begin{figure}[t!]
\centering
\includegraphics[width=0.49\textwidth]{figures/progression.jpg}
\caption{
A visualisation of the training progression of a single run from the experiment described in \Cref{section:experiments}.
The bright and dark tiles represent hypothesis grammar values close to $1$ and $0$ respectively.
From left, we see random hypothesis grammar space becoming more orderly as the training progresses.
The final induced grammar after the 15th epoch can be recovered following \Cref{section:grammar_extraction}.
}
\label{figure:progression}
\end{figure}
\subsection{Generation of Training Examples}
\label{section:generation_of_examples}
Let $\mathcal{G}$ be a ground-truth grammar generated as above, and let $\mathcal{L} = \left<\mathcal{G}\right>$ be its corresponding language.
The training data for a single instance of our model is a dataset $\mathcal{E}$ of words formed from the terminal alphabet $\mathcal{A}_T$, consisting of both negative and positive examples for $\mathcal{L}$.
To generate the examples for $\mathcal{L}$, we construct a minimal deterministic finite automaton (DFA) $\mathcal{D}$ equivalent to $\mathcal{G}$ in the sense that $\mathcal{D}$ accepts a word $w$ over $\mathcal{A}_T$ if and only if $w \in \mathcal{L}$.
Constructing finite automata to generate $\mathcal{E}$ instead of using the ground-truth $\mathcal{G}$ directly helps us avoid introducing unintentional bias into $\mathcal{E}$ that would favour $\mathcal{G}$ or grammars very similar to $\mathcal{G}$.
In other words, we avoid information leakage by constructing equivalent automata and optimising them prior to generating $\mathcal{E}$.
\subsubsection{Positive examples}
We perform a breadth-first search of depth $d$ on $\mathcal{D}$, memorising the path taken on each branch of the search.
Whenever an accepting state is encountered, the word consisting of transition symbols of the given path read out in sequence is returned.
The search then continues to explore the path as before until the depth $d$ has been reached.
Observe that the length $\ell$ of the word is $\leq d$.
\subsubsection{Negative examples from non-accepting paths}
Paths (including intermediate paths, i.e. paths of length $\leq d$) of the above breadth-first search that do not end at an accepting state are added to $\mathcal{E}$ as negative examples.
\subsubsection{Negative examples with invalid postfix}
Let $\sigma$ be a state of $\mathcal{D}$ reached by the above search after $k$ steps.
Let $\mathcal{A}\left(\sigma\right)$ be the set of letters $a \in \mathcal{A}_T$ that label out-transitions of $\sigma$.
For all $a \in \mathcal{A}_T \backslash \mathcal{A}\left(\sigma\right)$, let $w$ be the word formed by appending $a$ to the $a_1\dots a_k$ labels of the transitions on the path traversed to reach $\sigma$.
Such $w$ are also negative examples.
\subsubsection{Negative examples with invalid infix}
Let $a_1\dots a_k a$ be a word for an invalid postfix negative example, and let $b_1\dots b_l$ be the symbols of some valid path between two states of $\mathcal{D}$ such that the latter state is accepting.
We consider any word formed as $a_1\dots a_k a b_1 \dots b_l$ a negative example.
\subsubsection{Random negative examples}
To further enrich $\mathcal{E}$, we generated a number of words of uniformly random length up to $d$ with each letter drawn from $\mathcal{A}_T$ uniformly at random such that $\mathcal{D}$ would not accept them, and added them to $\mathcal{E}$ as negatives.
Intuitively, the positive examples guide the model towards learning a grammar $\mathcal{G}'$ that recalls all words of $\mathcal{L}$ (i.e. $\mathcal{L} \subseteq \left<\mathcal{G}'\right>$),
the negative examples with invalid prefix or infix increase precision (i.e. minimise $\mathcal{L}^c \cap \left<\mathcal{G}'\right>$),
and the negative examples form non-accepting paths further contribute to increases in precision by discouraging spurious links between the hypothesis non-terminals $\mathcal{A}_N'$ of $\mathcal{G}'$ while also indirectly reducing redundancy of computational logic in $\mathcal{G}'$.
We experimented with various ratios of positive to negative examples and ended up settling for $1{:}1$, with random negative examples providing more negative examples wherever needed to reach this ratio.
\subsection{Evaluation}
\label{section:evaluation}
Given a trained model, we extract the induced grammar $\mathcal{G}'$ as per \Cref{section:grammar_extraction}.
We then convert $\mathcal{G}'$ into an equivalent DFA, from whom we compute the \emph{minimal} DFA $\mathcal{D}'$.
The minimality is in the number of states as arrived at by the Hopcroft's algorithm.
For each regular language $\mathcal{L}$, there exists a unique (up to a re-labelling isomorphism) minimal recognizer DFA $\mathcal{D}$ \cite[p.~159-164]{hopcroft2001introduction}.
Given the minimal DFA $\mathcal{D}$ of the ground-truth grammar and the minimal DFA $\mathcal{D}'$ of the induced grammar, we canonically re-label their states and follow the algorithm of \cite{almeida2009testing} to test whether $\mathcal{D},\mathcal{D}'$ are isomorphic.
Isomorphism of $\mathcal{D},\mathcal{D}'$ means that $\mathcal{G},\mathcal{G}'$ are fully equivalent, implying that our model has achieved the maximum recall and precision and that we do not need to evaluate further.
In the case that $\mathcal{G},\mathcal{G}'$ are not equivalent, we assess similarity of languages by comparing the finite subsets $\mathcal{L}_d,\mathcal{L}_d'$ of each.
These consist of all words of $\mathcal{L}, \mathcal{L}' = \left<\mathcal{G}\right>$ up to length $d$.
We then measure the recall $\frac{\mathcal{L}_d \cap \mathcal{L}_d'}{\mathcal{L}_d}$, precision $\frac{\mathcal{L}_d \cap \mathcal{L}_d'}{\mathcal{L}_d'}$, and accuracy $\frac{\mathcal{L}_d \cap \mathcal{L}_d'}{\mathcal{L}_d \cup \mathcal{L}_d'}$ of our model.
\subsection{Results}
\label{section:results}
We generated grammars with $t = 4$, $n \in \left\{ 2, 3, 4 \right\}$, and $\mathfrak{p} \in \left\{2, 3, 4, 5 \right\}$, resulting in 12 different grammar configurations.
For each grammar, we generated the full dataset of positive and negative examples of strings up to length 16, training models on strings of length up to 6, 8, 10, 12, 14, and 16, with 5 runs per grammar per length.
During training, we used a batch size of 80, Adam optimizer with learning rate $0.005$, $\beta$ of $0.05$ after 60\% of epochs and $0$ before then.
There was a maximum of 60 epochs and $n' = 5$.
We extracted the grammar from the model using a confidence threshold $\tau = 0.95$, and evaluated it as described above.
Overall, the model learned the grammar from which the data was generated \textit{exactly} (i.e. the automata $\mathcal{D},\mathcal{D}'$ were isomorphic) in 85\% of all runs.
For more complex grammars, we observed an increase in accuracy (+2-5\%) when increasing length of examples while keeping the grammar complexity fixed.
Generally, grammars with smaller numbers of non-terminals $n$ were easier to learn for our model, and for all $n$, higher number of productions per non-terminals also led to better results.
This is somewhat counter-intuitive but perhaps due to our methodology for negative example generation: more complex grammars had more negative examples from non-accepting paths and fewer random negative examples.
As pointed out in \Cref{section:related_work}, no previous method addressed the problem of learning the entire class of regular grammars from positive and negative examples, and so no direct comparison can be made at present.
\section{Conclusion}
\label{section:conclusion}
We have introduced a purely neural explainable model for the induction of regular grammars from positive and negative acceptor examples, and demonstrated its ability to induce grammars to high levels of accuracy.
Our model can be used both for grammar induction and as a regular language parser.
This duality of purpose arises from it simultaneously inducing a grammar and attempting partial parses throughout its training as an acceptor automaton.
The ultimate but distant goal of grammatical inference in the context of algorithm learning is the induction of general grammars (\textsc{Type}-0 in the Chomsky hierarchy) which are equivalent in power to Turing machines.
We see regular grammars as a simple but rich class of grammars for induction from examples under the Programming by Example paradigm and hope that our work will help to facilitate further advancements in explainable neural grammar inference.
\bibliographystyle{plain}
|
\section{Introduction and statements of the main results}
In this paper, we study the solvability of the second boundary value problem of the following fourth order Monge-Amp\`ere type equation on a bounded, smooth, uniformly convex domain $\Omega\subset\mathbb R^n$ ($n\geq 2$):
\begin{equation}
\label{Abreu-sin}
\left\{
\begin{alignedat}{2}\sum_{i, j=1}^{n}U^{ij}D_{ij}w~& =-\gamma\div (|Du|^{q-2} Du)+ \text{\bf{b}}\cdot Du+c(x, u):=f(x, u, Du, D^2u)~&&\text{\ in} ~\ \ \Omega, \\\
w~&= (\det D^2 u)^{-1}~&&\text{\ in}~\ \ \Omega,\\\
u ~&=\varphi~&&\text{\ on}~\ \ \partial \Omega,\\\
w ~&= \psi~&&\text{\ on}~\ \ \partial \Omega.
\end{alignedat}
\right.
\end{equation}
Here $\gamma\geq 0$, $q>1$, $U=(U^{ij})_{1\leq i, j\leq n}$ is the cofactor matrix of the Hessian matrix $$D^2 u=(D_{ij}u)_{1\leq i, j\leq n}\equiv \left(\frac{\partial^2 u}{\partial x_i \partial x_j}\right)_{1\leq i, j\leq n}$$
of an unknown uniformly convex function $u\in C^2(\overline{\Omega})$; $\varphi\in C^{3,1}(\overline\Omega)$, $\psi\in C^{1,1}(\overline\Omega)$, $\text{\bf{b}}:\overline{\Omega}\rightarrow\mathbb R^n$ is a vector field on $\overline{\Omega}$, and $c(x, z)$ is a function on ${\overline\Om}\times\mathbb R$. When the right-hand side $f$ depends only on the independent variable, that is $f=f(x)$, (\ref{Abreu-sin}) is the {\it Abreu equation} arising from the problem of finding extremal metrics on toric manifolds in K\"ahler geometry \cite{Ab}, and it is equivalent to
$$\sum_{i,j=1}^n\frac{\partial^2 u^{ij}}{\partial x_i\partial x_j}=f(x),$$
where $(u^{ij})$ is the inverse matrix of $D^2u$. The general form in \eqref{Abreu-sin} was introduced by the second author in \cite{LeCPAM, LeJAM, LePRS} in the study of convex functionals with a convexity constraint related to the Rochet-Chon\'e model \cite{RC} for the monopolist's problem in economics, whose Lagrangian depends on the gradient variable; see also Carlier-Radice \cite{CR} for the case where the Lagrangian does not depend on the gradient variable.
When $\gamma>0$, we call \eqref{Abreu-sin} a {\it singular Abreu equation} because its right-hand side depends on $D^2u$ which can be just a matrix-valued measure for a merely convex function $u$.
Our focus in this paper will be on the case $\gamma>0$. For simplicity, we will take $\gamma=1$.
The Abreu type equations can be included in a class of fourth order Monge-Amp\`ere type equations of the form
\begin{equation}\label{4-eq-g}
U^{ij}D_{ij}[g(\det D^2 u)]=f
\end{equation}
where $g:(0,\infty)\rightarrow (0,\infty)$ is an invertible function. In particular, when $g(t)=t^{\theta}$, one can take
$\theta=-1$ and $\theta=-\frac{n+1}{n+2}$ to get the Abreu type equation and the {\it affine mean curvature} type equation \cite{Ch}, respectively.
It is convenient to write (\ref{4-eq-g}) as a system of two equations for $u$ and $w=g(\det D^2 u)$. One is a Monge-Amp\`ere equation for the convex function $u$ in the form of
\begin{equation}\label{MA1}
\det D^2 u=
g^{-1}(w)
\end{equation}
and other is the following linearized Monge-Amp\`ere equation for $w$:
\begin{equation}\label{LMA1}
U^{ij} D_{ij }w=f.
\end{equation}
The second-order linear operator $\sum_{i, j=1}^nU^{ij}D_{ij}$ is the linearized Monge-Amp\`ere operator associated with the convex function $u$ because its coefficient matrix comes from linearizing the Monge-Amp\`ere operator:
\[U= \frac{\partial\det D^2 u }{\partial (D^2 u)}.\]
When $u$ is sufficiently smooth, such as $u\in W^{4, s}_{\text{loc}}(\Omega)$ where $s>n$, the expression $\displaystyle\sum_{i, j=1}^{n}U^{ij}D_{ij}w$ can be written as $\displaystyle\sum_{i, j=1}^{n}D_i(U^{ij}D_{j}w)$, since the cofactor matrix $(U^{ij})$ is divergence-free, that is, \[\displaystyle\sum_{i=1}^n D_i U^{ij}=0\] for all $j$.
The regularity and solvability of equation \eqref{4-eq-g}, under suitable boundary conditions, are closely related to the regularity theory of the linearized Monge-Amp\`ere equation, initiated in the fundamental work of Caffarelli-Guti\'errez \cite{CG}. In the past two decades, there have been many progresses on the study of these equations and related geometric problems,
including \cite{D1, D2, D3, D4, TW1, TW2, TW3, Z1, Z2, CW, CHLS, Le1, Le2}, to name a few.
According to the decomposition \eqref{MA1} and \eqref{LMA1}, a very natural boundary value problem for the class of fourth order equation (\ref{4-eq-g}) is the second boundary value problem where one describes the values of $u$ and $w$ on the boundary $\partial\Omega$ as in \eqref{Abreu-sin}.
\subsection{Previous results and difficulties}
\label{pre_res}
A summary of solvability results for (\ref{Abreu-sin}), or more generally, the second boundary value problem for (\ref{4-eq-g}), for the case $f\equiv f(x)$ is as follows. For the second boundary value problem of the affine mean curvature equation, that is, (\ref{4-eq-g}) with $g(t) = t^{-\frac{n+1}{n+2}}$, Trudinger-Wang \cite{TW2, TW3} proved the existence of a unique $C^{4,\alpha}(\overline\Omega)$ solution when $f\in C^\alpha(\overline\Omega)$ with $f\leq 0$, and
a unique $W^{4,p}(\Omega)$ solution when $f\in L^{\infty}(\Omega)$ with $f\leq 0$. The analogous result for the Abreu equation (\ref{Abreu-sin}) was then obtained by the fourth author \cite{Z2}.
For the $W^{4,p}(\Omega)$ solution, the second author \cite{Le1} solved (\ref{Abreu-sin}) for $f\in L^p(\Omega)$ with $p>n$ and $f\leq 0$. The sign on $f$ was removed by Chau-Weinkove \cite{CW} under the assumption that $f\in L^{p}(\Omega)$ with $p>n$ and $f^+:=\max\{f, 0\}\in L^{q}(\Omega)$ with $q>n+2$ for the affine mean curvature equation. Finally, in \cite{Le2}, the second author showed that the $W^{4,p}(\Omega)$ solution exists under the weakest assumption
$f\in L^{p}(\Omega)$ with $p>n$ for a broad class of equations like \eqref{4-eq-g}, including both the affine mean curvature equation and the Abreu equation.
We will concentrate on the singular Abreu equation \eqref{Abreu-sin}, and its solvability in $C^{4,\alpha}$ and $W^{4, s}$ $(s>n)$ in this paper. We obtain solvability by establishing a priori higher order derivative estimates and then using the degree theory.
Essentially, establishing a priori estimates requires establishing the Hessian determinant estimates for $u$, and H\"older estimates for $w$.
For the singular Abreu equation, the dependence of the right-hand side on $D^2u$ creates two new difficulties in applying the regularity theory of the linearized Monge-Amp\`ere equation. The first difficulty lies in obtaining the \textit{a priori} lower and upper bounds for $\det D^2 u$, which is a critical step in
applying the regularity results of the linearized Monge-Amp\`ere equation. The appearance of $D^2u$ has very subtle effects on the Hessian determinant estimates. The second author \cite{LeCPAM} obtained the Hessian determinant estimates for the case $f=-\div (|Du|^{q-2} Du)$ in two dimensions with $q\geq 2$ by using a special algebraic structure of the equation. In a recent work
of the second and the fourth authors \cite{LZ},
the Hessian determinant estimates for the case $1<q<2$ were established by using partial Legendre transform.
The second difficulty, granted that the bounds $0<\lambda\leq\det D^2 u\leq\Lambda<\infty$ have been established, consists in obtaining H\"older estimates for $w$ in the linearized Monge-Amp\`ere equation (\ref{LMA1}), which has no lower order terms on the left-hand side. This requires certain integrability condition for the right-hand side, as can be seen from the simple equation $\Delta w=f$. In previous works \cite{CG, GN1, GN2}, classical regularity estimates for linearized Monge-Amp\`ere equation were obtained for $L^n$ right-hand side. This integrability breaks down even in the case $f=- \Delta u$ (where $q=2$, $\text{\bf{b}}=0$ and $c=0$), which is a priori at most $L^{1+\varepsilon}$ for some small constant $\varepsilon(\lambda,\Lambda, n)>0$ (see \cite{DFS, F, Sc}). With the H\"older estimates for the linearized Monge-Amp\`ere equation with $L^{n/2+\varepsilon}$ right-hand side in \cite{LN2}, the second author \cite{LeCPAM} established the solvability of \eqref{Abreu-sin} for the case $f=-\div (|Du|^{q-2} Du)$ in two dimensions with $q\geq 2$.
When $1<q<2$, $f=-\div (|Du|^{q-2} Du)$ becomes more singular in $D^2 u$ and hence it has lower integrability (if any).
However, in two dimensions, the second and the fourth authors \cite{LZ} solved the second boundary value problem \eqref{Abreu-sin} for
$f=-\div (|Du|^{q-2} Du)+c(x,u)$ for any $q>1$ under suitable assumptions on $c$ and the boundary data. The proof was based on the interior and global H\"older estimates for linearized Monge-Amp\`ere equation with the right-hand being the divergence of a bounded vector field which were established in \cite{LeCMP, LeJMAA}. The solvability of the singular Abreu equations (\ref{Abreu-sin}) in higher dimensions, even the simplest case $f=-\triangle u$, has been widely open. Only some partial results were obtained in \cite{LeJAM} under either a smallness condition (such as replacing $f=-\Delta u$ by $f=-\delta \Delta u$ for a suitably small constant $\delta>0$) or a radial symmetry condition.
\subsection{Statements of the main results}
The purpose of this paper is to solve the higher dimensional case of (\ref{Abreu-sin}).
We will first consider the case that the right-hand side has no drift term $\text{\bf{b}}\cdot Du$. This case answers in the affirmative the question raised in \cite[Page 6]{LZ}. In fact, we can establish the solvability for singular Abreu equations that are slightly more general than \eqref{Abreu-sin} where $\div (|Du|^{q-2} Du)$ is now replaced by
$\div (DF(Du))$ for a suitable convex function $F$. Our first main theorem states as follows.
\begin{thm}[Solvability of the second boundary value problem for singular Abreu equations in higher dimensions]
\label{thm:SBV-p-lap}
Let $\Omega\subset\mathbb R^n$ be an open, smooth, bounded and uniformly convex domain. Let $r>n$. Let $F\in W^{2, r}_{\text{loc}}(\mathbb R^n)$ be a convex function.
Assume that $\varphi\in C^{5}(\overline{\Omega})$ and $\psi\in C^{3}(\overline{\Omega})$ with $\min_{\partial \Omega}\psi>0$.
Consider the following second boundary value problem for a uniformly convex function $u$:
\begin{equation}
\label{eq:Abreu-plap}
\left\{
\begin{alignedat}{2}\sum_{i, j=1}^{n}U^{ij}D_{ij}w~& =-\div(DF(Du))+c(x, u)~&&\text{\ in} ~\ \ \Omega, \\\
w~&= (\det D^2 u)^{-1}~&&\text{\ in}~\ \ \Omega,\\\
u ~&=\varphi~&&\text{\ on}~\ \ \partial \Omega,\\\
w ~&= \psi~&&\text{\ on}~\ \ \partial \Omega.
\end{alignedat}
\right.
\end{equation}
Here $(U^{ij})= (\det D^2 u)(D^2 u)^{-1}$, and $c(x, z)\leq 0$.
\begin{enumerate}
\item[(i)] Assume $c\in C^\alpha(\overline\Omega\times \mathbb R)$ where $\alpha\in (0, 1)$. Then, there exists a uniformly convex solution $u\in W^{4, r}(\Omega)$ to \eqref{eq:Abreu-plap} with
$$\|u\|_{W^{4,r}(\Omega)}\leq C$$
for some $C>0$ depending on $\Omega$, $n$, $\alpha$, $F$, $r$, $c$, $\varphi$ and $\psi$.
Moreover, if $F\in C^{2,\alpha_0}(\mathbb R^n)$ where $\alpha_0\in (0, 1)$, then there exists a uniformly convex solution $u\in C^{4, \beta}(\overline{\Omega})$ to \eqref{eq:Abreu-plap} where $\beta=\min\{\alpha,\alpha_0\}$ with
$$\|u\|_{C^{4,\beta}(\overline{\Omega})}\leq C$$
for some $C>0$ depending on $\Omega$, $n$, $\alpha$, $\alpha_0$, $F$, $c$, $\varphi$ and $\psi$.
\item[(ii)] Assume $c(x,z)\equiv c(x)\in L^p(\Omega)$ with $p>n$ where $c(x)\leq 0$. Then, for $s= \min\{r,p\}$,
there exists a uniformly convex solution
$u\in W^{4, s}(\Omega)$ to \eqref{eq:Abreu-plap} with
$$\|u\|_{W^{4,s}(\Omega)}\leq C$$
for some $C>0$ depending on $\Omega$, $n$, $p$, $F$, $r, s$, $\|c\|_{L^p(\Omega)}$, $\varphi$ and $\psi$.
\end{enumerate}
\end{thm}
We will prove Theorem \ref{thm:SBV-p-lap} in Section \ref{SBV-plap-pf}.
We also discuss the solvability and regularity estimates of (\ref{Abreu-sin}) in the case that the right-hand side has more general lower order terms
and no sign restriction on $c$. We mainly focus on the most typical case that the right-hand side has a Laplace term:
\begin{equation}
\label{Abreu-lap}
\left\{
\begin{alignedat}{2}\sum_{i, j=1}^{n}U^{ij}D_{ij}w~& =-\triangle u+ \text{\bf{b}}\cdot Du+c(x,u)~&&\text{\ in} ~\ \ \Omega, \\\
w~&= (\det D^2 u)^{-1}~&&\text{\ in}~\ \ \Omega,\\\
u ~&=\varphi~&&\text{\ on}~\ \ \partial \Omega,\\\
w ~&= \psi~&&\text{\ on}~\ \ \partial \Omega.
\end{alignedat}
\right.
\end{equation}
Here, $(U^{ij})= (\det D^2 u)(D^2 u)^{-1}$.
Our second main result is the following theorem.
\begin{thm}[Solvability of the second boundary value problem for singular Abreu equations with lower order terms in high dimensions]
\label{SBV-1}
Let $\Omega\subset\mathbb R^n$($n\geq 3$) be an open, smooth, bounded and uniformly convex domain.
Assume that $\varphi\in C^{5}(\overline{\Omega})$ and $\psi\in C^{3}(\overline{\Omega})$ with $\min_{\partial \Omega}\psi>0$.
Consider the
second boundary value problem \eqref{Abreu-lap} with $c(x,z)\equiv c(x)$.
\begin{enumerate}
\item[(i)] If $\text{\bf{b}}\in C^\alpha(\overline\Omega;\mathbb R^n)$ and $c\in C^\alpha(\overline\Omega)$ where $\alpha\in (0, 1)$, then there exists a uniformly convex solution $u\in C^{4, \alpha}(\overline{\Omega})$ to \eqref{Abreu-lap} with
$$\|u\|_{C^{4,\alpha}(\overline{\Omega})}\leq C$$
for some $C>0$ depending on $\Omega$, $n$, $\alpha$, $\|\text{\bf{b}}\|_{C^\alpha(\overline\Omega)}$, $\|c\|_{C^\alpha(\overline\Omega)}$, $\varphi$ and $\psi$.
\item[(ii)] If $\text{\bf{b}}\in L^\infty(\Omega;\mathbb R^n)$ and $c\in L^p(\Omega)$ with $p>2n$, then there exists a uniformly convex solution $u\in W^{4, p}(\Omega)$ to \eqref{Abreu-lap} with
$$\|u\|_{W^{4,p}(\Omega)}\leq C$$
for some $C>0$ depending on $\Omega$, $n$, $p$, $\|\text{\bf{b}}\|_{L^\infty(\Omega)}$, $\|c\|_{L^p(\Omega)}$, $\varphi$ and $\psi$.
\end{enumerate}
\end{thm}
We will prove Theorem \ref{SBV-1} in Section
\ref{SBV-1-pf}. Furthermore, in two dimensions, when $\|\text{\bf{b}}\|_{L^\infty(\Omega)}$ is small, depending on $\Omega, \psi$ and $\psi$, the conclusions of Theorem \ref{SBV-1} still hold; see Remark \ref{SBV-1-rem}.
The lack of non-positivity of $c$ in (\ref{Abreu-lap}) can raise more difficulties in the $L^\infty$-estimate and the use of Legendre transform in the Hessian determinant estimates. Compared to the weakest assumption $c\in L^{p}(\Omega)$ with $p>n$ in \cite{Le2},
we need $p>2n$ in Theorem \ref{SBV-1}$(ii)$. However, in two dimensions, this assumption can be weakened provided stronger conditions on $\text{\bf{b}}$ are imposed, but $\|\text{\bf{b}}\|_{L^\infty(\Omega)}$ can be arbitrarily large. This is the content of our final main result.
\begin{thm} [Solvability of the second boundary value problem for singular Abreu equations with lower order terms in two dimensions]
\label{SBV-2}
Let $\Omega\subset\mathbb R^2$ be an open, smooth, bounded and uniformly convex domain.
Assume that $\varphi\in C^{5}(\overline{\Omega})$ and $\psi\in C^{3}(\overline{\Omega})$ with $\min_{\partial \Omega}\psi>0$.
Consider the
second boundary value problem \eqref{Abreu-lap}.
Assume that $\text{\bf{b}}\in C^1(\overline{\Omega};\mathbb R^n)$ with $\div(\text{\bf{b}})\leq \frac{32}{diam(\Omega)^2}$, and $c(x,z)\equiv c(x)\in L^p(\Omega)$ with $p>2$. Then there exists a uniformly convex solution $u\in W^{4, p}(\Omega)$ to \eqref{Abreu-lap} with
$$\|u\|_{W^{4,p}(\Omega)}\leq C$$
for some $C>0$ depending on $\Omega$, $p$, $\text{\bf{b}}$, $\|c\|_{L^p(\Omega)}$, $\varphi$ and $\psi$.
\end{thm}
The proof of Theorem \ref{SBV-2} will be given in Section
\ref{SBV-2-pf}.
\begin{rem} Some remarks are in order.
\begin{enumerate}
\item[(1)] Theorem \ref{thm:SBV-p-lap} applies to all convex functions $F(x)=|x|^q/q$ $(q>1)$ on $\mathbb R^n$ for which (\ref{eq:Abreu-plap}) becomes (\ref{Abreu-sin}) when $\text{\bf{b}}=0$. Note that, if $1<q<2$, then $|x|^q\in W^{2, r}_{\text{loc}}(\mathbb R^n)$ for all $n<r<n/(2-q)$, while if $q\geq 2$, we have $|x|^q\in W^{2, r}_{\text{loc}}(\mathbb R^n)$ for all $r>n$.
\item[(2)]
By the Sobolev embedding theorem, the solutions $u$ obtained in our main results at least belong to $C^{3,\beta}(\overline{\Omega})$ for some $\beta>0$.
\item [(3)] The condition $\div(\text{\bf{b}})\leq \frac{32}{diam(\Omega)^2}$ in Theorem \ref{SBV-2} is due to the method of its proof in obtaining a priori $L^{\infty}$ estimates that uses a Poincar\'e type inequality on planar convex domains in Lemma \ref{lem:Poincare}.
\end{enumerate}
\end{rem}
\subsection{On the proofs of the main results} Let us now say a few words about the proofs of our main results using a priori estimates and degree theory. We focus on the most crucial point that overcomes the obstacles encountered in previous works: obtaining the a priori H\"older estimate for $w=(\det D^2 u)^{-1}$ in higher dimensions, once the Hessian determinant bounds on $u$ have been obtained.
In this case, global H\"older estimates for $Du$ follow. Here, we use a new equivalent form (see Lemma \ref{low_lem}) for the singular Abreu equation to deal with the difficulties mentioned in Section \ref{pre_res}.
In particular, in Theorem \ref{thm:SBV-p-lap}, instead of establishing the H\"older estimate for $w$, we establish the H\"older estimate for $\eta= w e^{F(Du)}$. The key observation is that $\eta$ solves a linearized Monge-Amp\`ere equation with a drift term in which the very singular term \[\div (DF(Du))=\text{trace}(D^2 F(Du)D^2 u)\] no longer appears.
Thus, the proof of Theorem \ref{thm:SBV-p-lap} reduces the global higher order derivative estimates for (\ref{eq:Abreu-plap}) to the global H\"older estimates of linearized Monge-Amp\`ere equations with drift terms. To the best of the authors' knowledge, these global H\"older estimates with full generality are not available in the literature. In the case of Theorems \ref{SBV-1} and \ref{SBV-2}, the drift terms are also H\"older continuous. However, they do not vanish on the boundary and this seems to be difficult to prove H\"older estimates for $\eta$ at the boundary, not to
mention global H\"older estimates. We overcome this difficulty by observing that {\it each of our singular Abreu equation is in fact equivalent to a family of linearized Monge-Amp\`ere equations with drifts}. In particular, at each boundary point $x_0$,
\[\eta^{x_0}(x)= w(x) e^{F(Du(x))-DF(Du(x_0))\cdot (Du(x)-Du(x_0))-F(Du(x_0))}\]
solves a linearized Monge-Amp\`ere equation with a H\"older continuous drift that vanishes at $x_0$. This gives pointwise H\"older estimates for $\eta^{x_0}$ (and hence for $\eta$) at $x_0$. Combining this with interior H\"older estimates for linearized Monge-Amp\`ere equations with bounded drifts, we obtain the global H\"older estimates for $\eta$ and hence for $w$. Section \ref{LOT} will discuss all these in detail.
For reader's convenience, we recall the following notion of pointwise H\"older continuity.
\begin{defi}[Pointwise H\"older continuity] A continuous function $v\in C(\overline{\Omega})$
is said to be pointwise $C^\alpha$ ($0<\alpha<1$) at a boundary point $x_0\in\partial\Omega$, if there exist constants $\delta, M>0$ such that
\[|v(x)- v(x_0)|\leq M|x-x_0|^\alpha\quad\text{for all } x\in \Omega\cap B_\delta(x_0).\]
\end{defi}
Throughout, we use the convention that repeated indices are summed.
\vskip 15pt
The rest of the paper is organized as follows. In Section \ref{diff}, we establish a new equivalent form for the singular Abreu equations which transform them into linearized Monge-Amp\`ere equations with drift terms, and the dual equations under Legendre transform. The global H\"older
estimates for the linearized Monge-Amp\`ere equation with drift terms, under suitable hypotheses, will be addressed in Section \ref{LOT}. With these estimates, we can prove Theorem \ref{thm:SBV-p-lap} in Section \ref{SBV-plap-pf}. The proofs of Theorems \ref{SBV-1} and \ref{SBV-2} will be given in Sections \ref{SBV-1-pf}, and \ref{SBV-2-pf}, respectively. In the final Section \ref{rem_sect}, we discuss (\ref{Abreu-sin}) with more general lower order terms,
and present a proof of Theorem \ref{LMA-G} on global H\"older estimates for solutions to the linearized Monge-Amp\`ere equation with a drift term that are pointwise H\"older continuous at the boundary.
\vskip 30pt
\section[Equivalent forms of singular Abreu equations]{Equivalent forms of the singular Abreu equations}
\label{diff}
In this section, we derive some equivalent forms for the following general singular Abreu equations:
\begin{equation}\label{Abreu-plaplace}
\left\{
\begin{alignedat}{2}
U^{ij} D_{ij} w&=- \div (DF(Du)) + Q(x, u, Du), ~&&\text{\ in} ~\ \ \Omega, \\\
w~&= (\det D^2 u)^{-1}~&&\text{\ in}~\ \ \Omega,
\end{alignedat}
\right.
\end{equation}
where $U=(U^{ij})= (\det D^2 u) (D^2 u)^{-1}$, $F\in W^{2, n}_{\text{loc}}(\mathbb R^n)$, and $Q$ is a function on $\mathbb R^n\times \mathbb R\times \mathbb R^n$.
\subsection{Singular Abreu equations and linearized Monge-Amp\`ere equations with drifts}
Our key observation is the following lemma.
\begin{lem}[Equivalence of singular Abreu equations and linearized Monge-Amp\`ere equations with drifts]
\label{low_lem} Assume that a locally uniformly convex function $u\in W^{4, s}_{\text{loc}}(\Omega)$ ($s>n$) solves (\ref{Abreu-plaplace}).
Then
\[\eta = w e^{F(Du)}\]
satisfies
\begin{equation}\label{eq:F(p)-eq-n}
U^{ij}D_{ij} \eta -(\det D^2u) DF(Du)\cdot D \eta =e^{F(Du)}Q(x, u, Du).
\end{equation}
\end{lem}
\begin{proof} Let $(u^{ij})= (D^2 u)^{-1} = wU$.
By computations using
$D_j U^{ij}= 0$ and $w=(\det D^2u)^{-1}$, we have
\[
U^{ij}D_{ij} w=D_j(U^{ij}D_i w)=D_j(u^{ij}D_i(\log w))=-D_j(u^{ij}D_i(\log\det D^2u))
\]
and
\begin{equation}\label{q-la}
D_j\left[u^{ij} D_i \left(F(Du)\right)\right]=\div (DF(Du)).
\end{equation}
It follows that equation \eqref{Abreu-plaplace} can be written as
\begin{equation}
\label{zetaQ}
D_j(u^{ij}D_i\zeta)=-Q(x, u, Du)
\end{equation}
where
\begin{equation*}
\zeta=\log\det D^2u-F(Du).
\end{equation*}
In other words, in \eqref{Abreu-plaplace}, the singular term \[\div (DF(Du)) =\text{trace} (D^2 F(Du)D^2 u)\] can be absorbed into the left-hand side to turn it into a divergence form equation.
Next, observe that $\zeta= -\log\eta$, and
\[D_i \zeta = -D_i \eta/\eta = -D_i \eta \det D^2u e^{-F(Du)}.\]
Thus (\ref{zetaQ}) becomes
\begin{eqnarray*}
Q(x, u, Du) &=& - D_j(u^{ij}D_i \zeta) \\&=& D_j\Big(u^{ij} \det D^2u e^{-F(Du)} D_i\eta\Big) \\&=& D_j\Big(U^{ij} e^{-F(Du)} D_i \eta\Big) \\&=& U^{ij} D_j \Big( e^{-F(Du)} D_i\eta\Big) \,\,(\text{using the divergence free property of } (U^{ij})) \\
&=&U^{ij}D_{ij} \eta e^{-F(Du)} -U^{ij} D_i\eta e^{-F(Du)} D_kF (Du)D_{kj}u\\
&=& U^{ij} D_{ij}\eta e^{-F(Du)} -(\det D^2u) e^{-F(Du)}DF (Du)\cdot D \eta.
\end{eqnarray*}
Therefore, (\ref{eq:F(p)-eq-n}) holds, and the lemma is proved.
\end{proof}
\begin{rem} In general, \eqref{Abreu-plaplace} is not the Euler-Lagrange equation of any functional. However, the introduction of \[\eta =w e^{F(Du)} = (\det D^2 u)^{-1} e^{F(Du)}\] in Lemma \ref{low_lem} has its root in an energy functional. Indeed, when $Q\equiv 0$, \eqref{Abreu-plaplace} becomes
\[D_{ij} (U^{ij} (\det D^2 u)^{-1})+ \div (DF(Du))=0, \]
and this is the the Euler-Lagrange equation of the Monge-Amp\`ere type functional
\[\int_{\Omega} \Big(F(Du) -\log \det D^2 u\Big) dx = \int_\Omega \log \Big((\det D^2 u)^{-1} e^{F(Du)}\Big)\,dx.\]
\end{rem}
\begin{rem}
Taking $F(x)=|x|^{q}/q$ with $q>1$ in Lemma \ref{low_lem} where $x\in\mathbb R^n$, we find that an equivalent form of \[U^{ij}D_{ij} w=-\div (|Du|^{q-2} Du)+Q(x, u, Du), \quad w=(\det D^2u)^{-1}\] is
\begin{equation}\label{p-eq-n}
U^{ij}D_{ij}\eta-(\det D^2u)|Du|^{q-2} Du \cdot D\eta=Q(x, u, Du)e^{\frac{|Du|^q}{q}},
\end{equation}
where \[\eta = w e^{\frac{|Du|^q}{q}}.\]
\end{rem}
Lemma \ref{low_lem} shows that $\eta=w e^{F(Du)}$, where $u$ is a solution of (\ref{Abreu-plaplace}), satisfies a linearized Monge-Amp\`ere equation with a drift term. This fact plays a crucial role in the study of
singular Abreu equations in higher dimensions in latter sections.
Once we have the determinant estimates for $\det D^2u$ for the second boundary value problem of (\ref{Abreu-plaplace}), we can estimate $u$ in $C^{1,\alpha}(\overline{\Omega})$ provided the boundary data is smooth. This gives nice regularity properties for the right-hand side of (\ref{eq:F(p)-eq-n}) (and particularly, \eqref{p-eq-n}) and the drift
on the left-hand side. Then the higher regularity estimates for \eqref{Abreu-plaplace} can be reduced to global H\"older estimates for the following linearized Monge-Amp\`ere equation with a drift term:
\begin{equation}\label{LMA-lower}
U^{ij} D_{ij}\eta+\text{\bf{b}}\cdot D\eta+f(x)=0.
\end{equation}
This is the content of Section \ref{LOT}.
\subsection{Singular Abreu equations under the Legendre transform}
In this section, we derive the dual equation of \eqref{Abreu-plaplace} under the Legendre transform in any dimension.
After the Legendre transform, the dual equation is still a linearized Monge-Amp\`ere equation.
Denote the Legendre transform $u^{\ast}$ of $u$ by $$u^{\ast}(y)=x \cdot Du-u, \quad \text{where }y=Du(x)\in \Omega^\ast =Du(\Omega). $$
Then
\[x= Du^* (y), \quad\text{and } u(x) = y\cdot Du^* (y) -u^*(y).\]
\begin{prop}[Dual equations for singular Abreu equations]
\label{new2}
Let $u\in W^{4, s}_{\text{loc}}(\Omega)$ ($s>n$) be a uniformly convex solution to (\ref{Abreu-plaplace}) in $\Omega$. Then in $\Omega^*=Du(\Omega)$, its Legendre transform $u^{\ast}$ satisfies the following dual equation
\begin{eqnarray}\label{new-eq-leg}
u^{\ast ij}D_{ij} \left(w^*+F(y)\right)=Q\left(D u^{*},y\cdot D u^{*}-u^*,y\right).
\end{eqnarray}
Here $( u^{\ast ij})$ is the inverse matrix of $D^2u^{\ast}$, and $w^*= \log\det D^2 u^{\ast}$.
\end{prop}
\begin{proof}
When (\ref{Abreu-plaplace}) is a Euler-Lagrange equation of a Monge-Amp\`ere type functional, we can derive its dual
equation from the dual functional as in \cite[Proposition 2.1]{LZ}. Here for the general case,
we prove it by direct calculations. Note that for the case that the right-hand side has no singular term, the dual equation
has been obtained in \cite[Lemma 2.7]{Le2}. We include a complete proof here for reader's convenience.
For simplicity, let $d=\operatorname{det} D^{2} u$ and $d^{*}=\operatorname{det} D^{2} u^{*}$. Then $d(x)=d^{*-1}(y)$ where $y= Du(x)$. We will simply write $d= d^{*-1}$ with this understanding.
We denote by $\left(u^{i j}\right)$ and $\left(u^{* i j}\right)$ the inverses of the Hessian matrices $D^{2} u=\left(D_{i j} u\right)=\left(\frac{\partial^{2} u}{\partial x_{i} \partial x_{j}}\right)$ and $D^{2} u^{*}=\left(D_{i j}u^{*}\right)=\left(\frac{\partial^{2} u^{*}}{\partial y_{i} \partial y_{j}}\right)$, respectively.
Let $(U^{\ast ij})=(\det D^2 u^*) (u^{\ast ij})$ be the cofactor matrix of $D^2 u^\ast$.
Note that $w={d}^{-1}=d^{*}$. Thus
\[
D_{j}w=\frac{\partial w}{\partial x_{j}}=\frac{\partial d^{*}}{\partial y_{k}} \frac{\partial y_{k}}{\partial x_{j}}=\frac{\partial d^{*}}{\partial y_{k}} D_{k j}u=\frac{\partial d^{*}}{\partial y_{k}} u^{* k j} .
\]
Clearly,
\[
d^{*-1} \frac{\partial d^{*}}{\partial y_{k}}=\frac{\partial}{\partial y_{k}}\left(\log d^*\right)=D_{y_k} w^{*},
\]
from which it follows that \[D_{x_j}w=D_{y_k}w^{*}\left(U^{*}\right)^{k j}.\] Similarly, \[D_{i j} w=(\frac{\partial}{\partial y_{l}} D_{j} w) u^{* l i}.\] Hence, using
\[
U^{i j}=\operatorname{det} D^{2} u\cdot u^{i j}=\left(d^{*}\right)^{-1} D_{y_i y_j} u^{*},
\]
and the fact that $U^{*}=\left(U^{* i j}\right)$ is divergence-free, we obtain
\begin{eqnarray}
\label{Abu*}
U^{i j} D_{i j} w
&=&\left(d^{*}\right)^{-1} D_{y_i y_j} u^{*} u^{* l i} \frac{\partial}{\partial y_{l}} D_jw\nonumber\\
&=&\left(d^{*}\right)^{-1}(\frac{\partial}{\partial y_{j}} D_{j} w) \nonumber\\
&=&\left(d^{*}\right)^{-1} \frac{\partial}{\partial y_{j}}\left(D_{y_k} w^{*} U^{* k j}\right)\nonumber\\
&=&\left(d^{*}\right)^{-1} U^{* k j} D_{y_k y_j}w^{*}\nonumber\\
&=&u^{*ij}D_{ij}w^*.
\end{eqnarray}
On the other hand, by \eqref{q-la}, we have
\begin{eqnarray}
\label{Qu*}
\div (DF(Du))&=&D_{x_j}\left[u^{ij}D_{x_i}\left(F(Du)\right)\right]\nonumber\\
&=&u^{* l j} \frac{\partial}{\partial y_{l}}\left[u^*_{ij}u^{* ki} \frac{\partial}{\partial y_{k}}\left(F(y)\right)\right]\nonumber\\
&=&u^{\ast ij}D_{y_iy_j} \left(F(y)\right).
\end{eqnarray}
Combining (\ref{Abu*}) with (\ref{Qu*}) and recalling \eqref{Abreu-plaplace}, we obtain
\[ u^{\ast ij}D_{ij} \left(w^*+F(y)\right) = Q(x, u(x), Du(x)) = Q\left(D u^{*},y\cdot D u^{*}-u^*,y\right), \]
which is
(\ref{new-eq-leg}).
The lemma is proved.
\end{proof}
\vskip 30pt
\section[Linearized Monge-Amp\`ere equations with drifts]{H\"older estimates for linearized Monge-Amp\`ere equation with drifts}
\label{LOT}
In this section, we study global H\"older estimates for the linearized Monge-Amp\`ere equation with drift
\begin{equation}
\label{LMA-eq}
\left\{
\begin{alignedat}{2}
U^{ij}D_{ij} v +\text{\bf{b}}\cdot Dv~& = f ~&&\text{ in } ~ \Omega, \\\
v&= \varphi~&&\text{ on }~\partial \Omega,
\end{alignedat}
\right.
\end{equation}
where $U=(U^{ij})= (\det D^2 u)(D^2 u)^{-1}$ and $\text{\bf{b}}:\Omega\rightarrow\mathbb R^n$ is a vector field.
When there is no drift term, that is $\text{\bf{b}}\equiv 0$, global H\"older estimates for (\ref{LMA-eq}) were established under suitable assumptions on the bounds $0<\lambda\leq \det D^2 u\leq\Lambda$ on the Hessian determinant of $u$, and the data. In particular, the case $f\in L^n(\Omega)$ was treated in \cite[Theorem 1.4]{Le1} (see also \cite[Theorem 4.1]{LN1} for a more localized version) and the case $f\in L^{n/2+\varepsilon}(\Omega)$ was treated in \cite[Theorem 1.7]{LN2}.
We would like to extend the above global H\"older estimates to the case with bounded drift. In this case, the interior H\"older estimates for (\ref{LMA-eq}) were obtained as a consequence of the interior Harnack inequality proved in \cite[Theorem 1.1]{LeCCM}.
Note that Maldonado \cite{M} also proved a Harnack's inequality for (\ref{LMA-eq}) with different and stronger conditions on $\text{\bf{b}}$.
Therefore, to obtain global H\"older estimates for (\ref{LMA-eq}) with a bounded drift $\text{\bf{b}}$, it remains to prove the H\"older estimates at the boundary. Without further assumptions on $\text{\bf{b}}$, this seems to be difficult with current techniques. However, when $\text{\bf{b}}$ is pointwise H\"older continuous, and vanishes at a boundary point $x_0$, we can obtain the pointwise H\"older continuity of $v$ at $x_0$.
This can be deduced from the following result, which is a drift version of \cite[Proposition 2.1]{Le1}.
\begin{prop}[Pointwise H\"older estimate at the boundary for solutions to non-uniformly elliptic, linear equations with pointwise H\"older continuous drift]\label{ptwH}
Assume that $\Omega\subset\mathbb R^n$ is a bounded, uniformly convex domain. Let
$\varphi\in C^{\alpha}(\partial\Omega)$ for some $\alpha\in (0, 1)$, and $g\in L^n(\Omega)$. Assume that
the matrix $(a^{ij})$ is measurable, positive definite and satisfies $\det (a^{ij})\geq \lambda$ in $\Omega$. Let $\text{\bf{b}}\in L^{\infty}(\Omega;\mathbb R^n)$.
Let $v\in C(\overline{\Omega})\cap W^{2, n}_{loc}(\Omega)$ be the
solution to
$$a^{ij} D_{ij}v + \text{\bf{b}}\cdot Dv~=g ~\text{in} ~\Omega, \quad
v= \varphi ~\text{on}~ \partial\Omega.
$$
Suppose there are constants $\mu,\tau\in (0, 1)$, and $M>0$ such that at some $x_0\in\partial\Omega$, we have
\begin{equation}
\label{bbvanish}
|\text{\bf{b}} (x)|\leq M|x-x_0|^\mu\quad \text{for all } x\in\Omega\cap B_{\tau}(x_0).\end{equation}
Then, there exist $\delta, C$ depending only on $\lambda, n, \alpha,\mu, \tau, M$, $\|\text{\bf{b}}\|_{L^{\infty}(\Omega)}$, and $\Omega$ such that
\begin{equation*}|v(x)-v(x_{0})|\leq C|x-x_{0}|^{\frac{\min\{\alpha,\mu\}}{\min\{\alpha,\mu\} +4}}\left(\|\varphi\|_{C^{\alpha}(\partial\Omega)} + \|g\|_{L^{n}(\Omega)}\right)~\text{for all}~ x\in \Omega\cap B_{\delta}(x_{0}). \end{equation*}
\end{prop}
We will prove Proposition \ref{ptwH} in Section \ref{ptwH_pf}.
Once we have the pointwise H\"older estimates at the boundary, global H\"older estimates for (\ref{LMA-eq}) follow. This is the content of the following theorem.
\begin{thm}[Global H\"older estimates for solutions to the linearized Monge-Amp\`ere equation with a drift term that are pointwise H\"older continuous at the boundary]
\label{LMA-G}
Assume that $\Omega\subset\mathbb R^n$ is a uniformly convex domain with boundary $\partial\Omega\in C^3$.
Let $u \in C(\overline
\Omega)
\cap
C^2(\Omega)$ be a convex function satisfying
\[\lambda \leq \det D^2 u \leq \Lambda \quad \text{in}\quad \Omega\]
for some positive constants $\lambda$ and $\Lambda$. Moreover, assume that $u|_{\partial\Omega}\in C^3$. Let $(U^{ij})= (\det D^2 u)(D^2 u)^{-1}$.
Let $\text{\bf{b}}\in L^{\infty}(\Omega;\mathbb R^n)$ with $\|\text{\bf{b}}\|_{L^{\infty}(\Omega)}\leq M$, $f\in L^{n}(\Omega)$ and $\varphi\in C^{\alpha}(\partial\Omega)$ for some $\alpha\in (0,1)$.
Assume that $v \in C(\overline{\Omega}) \cap W^{2,n}_{loc}( \Omega)$ is a solution to the following linearized Monge-Amp\`ere equation with a drift term
\begin{equation*}
\left\{
\begin{alignedat}{2}
U^{ij}D_{ij} v +\text{\bf{b}}\cdot Dv ~& = f ~&&\text{ in } ~ \Omega, \\\
v&= \varphi~&&\text{ on }~\partial \Omega.
\end{alignedat}
\right.
\end{equation*}
Suppose that there exist $\gamma\in (0, \alpha]$, $\delta>0$ and $K>0$ such that
\begin{equation}
\label{Bdr_assumed}
|v(x)-v(x_{0})|\leq K|x-x_{0}|^{\gamma} \quad \text{for all}~ x_0\in\partial\Omega, \text{ and } x\in \Omega\cap B_{\delta}(x_{0}). \end{equation}
Then, there exist a constant
$\beta\in (0, 1)$ depending on $n$, $\lambda$, $\Lambda$, $\gamma$ and $M$, and a constant $C >0 $ depending only on $\Omega$, $u|_{\partial\Omega}$, $\lambda$, $\Lambda$, $n$, $\alpha$, $\gamma$, $\delta$, $K$ and $M$ such that
$$|v(x)-v(y)|\leq C|x-y|^{\beta}\Big( \|\varphi\|_{C^\alpha(\partial\Omega)} + \|f\|_{L^{n}(\Omega)} \Big),~\forall x, y\in \Omega. $$
\end{thm}
The proof of Theorem \ref{LMA-G} is similar to that of \cite[Theorem 1.4]{Le1} for the case without a drift. For completeness and for reader's covenience, we present its proof at the end of the paper in Section \ref{rem_sect}.
\begin{rem}
It would be interesting to prove the global H\"older estimates in Theorem \ref{LMA-G} without the assumption (\ref{Bdr_assumed}).
\end{rem}
In Section \ref{wH_sec}, we will apply Theorem \ref{LMA-G} to establish the global H\"older estimates for Hessian determinants of singular Abreu equations provided that the Hessian determinants are bounded between two positive constants; see Theorem \ref{wHolder_thm}.
\subsection{Pointwise H\"older estimates at the boundary}
\label{ptwH_pf}
In this section, we prove Proposition \ref{ptwH}.
\begin{proof}[Proof of Proposition \ref{ptwH}] The proof is similar to that of \cite[Proposition 2.1]{Le1}. Due to the appearance of the drift $\text{\bf{b}}$ and the pointwise H\"older continuity condition (\ref{bbvanish}), we include the proof for reader's convenience.
Let \[K= \|\text{\bf{b}}\|_{L^{\infty}(\Omega)}, \quad\text{and }L=\text{diam}(\Omega).\] In this proof, we fix the exponent \[\gamma=\min\{\alpha,\mu\}/2.\]
However, the proof works for any exponent $\gamma\in (0, 1)$ such that $\gamma<\min\{\alpha,\mu\}$, and in this case, we replace the exponent $\frac{\min\{\alpha,\mu\}}{\min\{\alpha,\mu\} +4}$ in the proposition by $\frac{\gamma}{\gamma + 2}$.
Clearly $\varphi\in C^{\gamma}(\partial\Omega)$ with $\|\varphi\|_{C^{\gamma}(\partial\Omega)} \leq C(\alpha,\mu, L) \|\varphi\|_{C^{\alpha}(\partial\Omega)}$.
By considering the equation satisfied by $(\|\varphi\|_{C^{\gamma}(\partial\Omega)} + \|g\|_{L^{n}(\Omega)})^{-1}v$,
we can assume that
$$\|\varphi\|_{C^{\gamma}(\partial\Omega)} + \|g\|_{L^{n}(\Omega)}=1,$$
and it suffices to prove that, for some $\delta=\delta(n, \lambda, \alpha,\tau, K, M,\mu,\Omega)>0$, we have
$$|v(x)-v(x_{0})|\leq C(n, \lambda, \alpha,\tau, K, M,\mu,\Omega)|x-x_{0}|^{\frac{\gamma}{\gamma +2}}~\text{for all}~ x\in \Omega\cap B_{\delta}(x_{0}). $$
Moreover, without loss of generality, we assume that
$$\Omega\subset \mathbb R^{n}\cap \{x_{n}>0\},~x_0=0\in\partial\Omega.$$
Since $\det (a^{ij})\geq \lambda$, by the Aleksandrov-Bakelman-Pucci (ABP) estimate for elliptic, linear equations with drifts (see \cite[inequality (9.14)]{GT}), we have
\begin{eqnarray}\label{h-from-above}
\|v\|_{L^{\infty}(\Omega)}&\leq& \|\varphi\|_{L^{\infty}(\partial\Omega)} \nonumber\\&&+ \text{diam}(\Omega)\Bigg\{\exp\Big[\frac{2^{n-2}}{n^n \omega_n} \int_\Omega \Big( 1+ \frac{|\text{\bf{b}}|^n}{\det (a^{ij})}\Big)dx \Big]-1\Bigg\}^{1/n}\Big \|\frac{g}{(\det (a^{ij}))^{1/n}}\Big\|_{L^n(\Omega)}
\nonumber\\&\leq& C_0
\end{eqnarray}
for a constant $C_0(n,\lambda, K, L)>1$. Here we used $\omega_n=|B_1(0)|$, and $\|\varphi\|_{C^{\gamma}(\partial\Omega)} + \|g\|_{L^{n}(\Omega)}=1$.
Hence, for any $\varepsilon \in (0,\tau^{\gamma})$
\begin{equation}|v(x)-v(0)\pm \varepsilon|\leq 3C_{0}:= C_{1}.
\label{3C0C1}
\end{equation}
Consider now the functions
$$\psi_{\pm}(x) := v(x)- v(0)\pm \varepsilon\pm C_{1} \kappa(\delta_2) x_{n}$$
where
\[\kappa (\delta_2):= (\inf \{y_{n}: y\in \overline{\Omega}\cap\partial B_{\delta_{2}}(0)\})^{-1}\]
in the region \[A:= \Omega\cap B_{\delta_{2}}(0)\] where $\delta_{2}<1$ is small to be chosen later.
The uniform convexity of $\Omega$ gives
\begin{equation}
\inf \{y_{n}: y\in \overline{\Omega}\cap\partial B_{\delta_{2}}(0)\} \geq C_{2}^{-1}\delta^2_{2}
\end{equation}
where $C_2$ depends on the uniform convexity of $\Omega$. Thus,
\[\kappa(\delta_2) \leq C_2 \delta^{-2}_2.\]
Note that, if $x\in\partial \Omega$ with $|x|\leq \delta_{1}(\varepsilon):= \varepsilon^{1/\gamma}(\leq\tau)$ then, we have from $\|\varphi\|_{C^{\gamma}(\partial\Omega)}\leq 1$ that
\begin{equation}
\label{bdr_vineq0}|v(x)-v(0)| =|\varphi(x)-\varphi(0)| \leq |x|^{\gamma} \leq \varepsilon.
\end{equation}
It follows that, if we choose $\delta_{2}\leq \delta_{1}$, then from (\ref{3C0C1}) and (\ref{bdr_vineq0}), we have
$$\psi_{-}\leq 0, \psi_{+}\geq 0~\text{on}~\partial A.$$
From (\ref{bbvanish}), we have
\[|\text{\bf{b}}| \leq M\delta_2^\mu \quad\text{in } A,\]
and therefore
\[a^{ij}D_{ij}\psi_{-} + \text{\bf{b}}\cdot D\psi_{-}= g - C_1 \kappa(\delta_2)\text{\bf{b}}\cdot e_n\geq -|g|-C_1 C_2 M \delta_2^{\mu-2}\quad\text{in } A, \]
where $e_n=(0, \cdots, 0, 1)\in\mathbb R^n$.
Similarly,
\[a^{ij}D_{ij}\psi_{+} + \text{\bf{b}}\cdot D\psi_{+}= g + C_1 \kappa(\delta_2)\text{\bf{b}}\cdot e_n\leq |g| + C_1 C_2 M \delta_2^{\mu-2}\quad\text{in } A. \]
Again, applying the ABP estimate for elliptic, linear equations with drifts, we obtain
$$\psi_{-}\leq C(n,\lambda, K, L)\text{diam} (A) \|g + C_1 C_2 M \delta_2^{\mu-2}\|_{L^{n}(A)}\leq C_3(n, \lambda, K, M,\Omega, \tau, \mu)\delta^\mu_{2}~\text{in}~ A.$$
In the above inequality, we used $\|g\|_{L^{n}(A)} \leq 1$ and
\[\|g + C_1 C_2 M \delta^{\mu-2}\|_{L^{n}(A)} \leq \|g\|_{L^{n}(A)} +C_1 C_2 M \delta_2^{\mu-2} |A|^{1/n} \leq C(n, \lambda, K, M,\Omega, \tau, \mu)\delta_2^{\mu-1}. \]
Similarly, we have $$
\psi_{+}\geq - C(n,\lambda, K, L)\text{diam} (A) \|g+ C_1 C_2 M \delta_2^{\mu-2}\|_{L^{n}(A)}\geq -C_3(n, \lambda, K, M,\Omega, \tau, \mu)\delta^\mu_{2}~\text{in}~ A.$$
We now restrict $\varepsilon\leq C_3^{\frac{-\gamma}{\mu-\gamma}}$ so that
$$\delta_{1} = \varepsilon^{1/\gamma}\leq [\varepsilon/C_3]^{1/\mu}.$$
Then, for $\delta_{2}\leq \delta_{1}$, we have $C_3\delta_{2}^\mu\leq \varepsilon$, and thus,
$$|v(x)-v(0)|\leq 2\varepsilon + C_{1} \kappa(\delta_2) x_{n}\quad\text{in } A.$$
Therefore, choosing $\delta_{2}= \delta_{1}$, we find
\begin{equation*}|v(x)-v(0)|\leq 2\varepsilon + C_{1} \kappa(\delta_2) x_{n}\leq 2\varepsilon + \frac{2C_{1}C_{2}}{\delta_{2}^2}x_{n}~\text{in}~ A.\end{equation*}
Summarizing, we obtain the following inequality
\begin{equation}
\label{op-ineq0}
|v(x)-v(0)|\leq 2\varepsilon + \frac{2C_{1}C_{2}}{\delta_{2}^2}|x| \leq 2\varepsilon + 2C_{1}C_{2}\varepsilon^{-2/\gamma}|x|
\end{equation}
for all $x,\varepsilon$ satisfying the following conditions
\begin{equation}
\label{xe-ineq}
|x|\leq \delta_{1}(\varepsilon):= \varepsilon^{1/\gamma}, \quad \varepsilon\leq C_3^{\frac{-\gamma}{\mu-\gamma}}: = c_{1}.
\end{equation}
Let us now choose
$\varepsilon = |x|^{\frac{\gamma}{\gamma + 2}}.$
Then the conditions in (\ref{xe-ineq}) are satisfied as long as
$$|x|\leq \min\{c_{1}^{\frac{\gamma +2}{\gamma}}, 1\}:=\delta.$$
With this choice of $\delta$, and recalling (\ref{op-ineq0}), we have
$$|v(x)-v(0)| \leq (2+ 2C_1 C_2)|x|^{\frac{\gamma}{\gamma + 2}}\quad \text{for all }x\in \Omega\cap B_{\delta}(0).$$
The proposition is proved.
\end{proof}
\subsection{Singular Abreu equations with Hessian determinant bounds}
\label{wH_sec}
In this section, we apply Theorem \ref{LMA-G} to establish the global H\"older estimates for Hessian determinants of singular Abreu equations provided that the Hessian determinants are bounded between two positive constants. This is the content of the following theorem.
\begin{thm}[H\"older continuity of Hessian determinant of singular Abreu equations under Hessian determinant bounds]
\label{wHolder_thm}
Assume that $\Omega\subset\mathbb R^n$ is a uniformly convex domain with boundary $\partial\Omega\in C^3$. Let $F\in W^{2, r}_{\text{loc}}(\mathbb R^n)$ for some $r>n$, and let $g\in L^s(\Omega)$ where $s>n$.
Let $\varphi\in C^4(\overline{\Omega})$ and $\psi\in C^2(\overline{\Omega})$ with $\min_{\partial\Omega} \psi>0$. Assume that $u\in W^{4, s}(\Omega)$ is a uniformly convex solution to the singular Abreu equation:
\begin{equation*}
\left\{
\begin{alignedat}{2}
U^{ij} D_{ij} w&=- \div (DF(Du))+ g(x), ~&&\text{\ in} ~\ \ \Omega, \\\
w~&= (\det D^2 u)^{-1}~&&\text{\ in}~\ \ \Omega,\\\
u~&= \varphi~&&\text{\ on}~\ \ \partial\Omega,\\\
w~&= \psi~&&\text{\ on}~\ \ \partial\Omega,
\end{alignedat}
\right.
\end{equation*}
where $U=(U^{ij})= (\det D^2 u) (D^2 u)^{-1}$. Suppose that, for some positive constants $\lambda$ and $\Lambda$, we have
\[\lambda \leq \det D^2 u \leq \Lambda \quad \text{in}\quad \Omega.\]
Then, there exist
constants $\beta, C >0 $ depending only on $\Omega$, $\varphi, \psi$, $\lambda$, $\Lambda$, $n$, $r$, $F$, and $\|g\|_{L^{n}(\Omega)}$, such that
\[\|w\|_{C^{\beta}(\overline{\Omega})}\leq C. \]
\end{thm}
\begin{proof} Since $F\in W^{2, r}_{\text{loc}}(\mathbb R^n)$, by the Sobolev embedding theorem, we have $F\in C^{1, \alpha}(\mathbb R^n)$ where $\alpha=1-n/r\in (0, 1)$. From the Hessian determinant bounds on $u$, and $u=\varphi$ on $\partial\Omega$ where $\varphi\in C^{4}(\overline{\Omega})$,
by \cite[Proposition 2.6]{LS}, we have
\begin{equation}
\label{uc1a0}
\|u\|_{C^{1,\alpha_0}(\overline\Omega)}\leq C_1,\end{equation}
where $\alpha_0\in (0, 1)$ depends on $\lambda,\Lambda$, and $n$. The constant $C_1$ depends on $\Omega, n,\lambda,\Lambda$ and $\varphi$.
By Lemma \ref{low_lem},
the function
\[\eta(x)= w(x) e^{ F(Du(x))}\] satisfies
\begin{equation}
\label{etaeq}
U^{ij} D_{ij}\eta -(\det D^2 u) DF(Du(x)) \cdot D\eta= g(x)e^{ F(Du(x))}\equiv f(x). \end{equation}
From (\ref{uc1a0}), we deduce that $\eta|_{\partial\Omega}\in C^{\alpha_0}$ with estimate
\begin{equation}
\label{etabdr} \|\eta\|_{C^{\alpha_0}(\partial\Omega)} \leq C_\ast(\psi, C_1, F).
\end{equation}
{\it Step 1: Pointwise H\"older continuity of $\eta$ at the boundary.} Fix $x_0\in\partial\Omega$.
Let us denote
\[\tilde F(y):=F(y)-F(Du(x_0))-D F(Du(x_0))\cdot (y-Du(x_0))\quad\text{for } y\in\mathbb R^n.\]
Then, we have
\begin{equation*}U^{ij} D_{ij} w(x)=- \div (D\tilde F(Du(x))) + g(x)\quad\text{in }\Omega.\end{equation*}
By Lemma \ref{low_lem}, the function
\[\eta^{x_0}(x)= w(x) e^{\tilde F(Du(x))}\] satisfies
\begin{equation}
\label{etax0}
U^{ij} D_{ij}\eta^{x_0} -(\det D^2 u) (DF (Du(x))- DF (Du(x_0))) \cdot D\eta^{x_0}= g(x)e^{\tilde F(Du(x))}\equiv f^{x_0}(x). \end{equation}
Clearly,
\begin{equation}
\label{fx0Ln}
\|f^{x_0}\|_{L^{n}(\Omega)} \leq C_2,
\end{equation}
where $C_2$ depends on $\|F\|_{C^1(B_{C_1}(0))}$ and $\|g\|_{L^{n}(\Omega)}$.
The vector field
\[\text{\bf{b}}(x)=(\det D^2 u)\cdot (DF(Du(x))- DF (Du(x_0))) \]
satisfies in $\Omega$ the estimate
\begin{equation}
\label{bbetax0}
|\text{\bf{b}}(x)| \leq \Lambda \|DF\|_{C^\alpha(B_{C_1}(0))} |Du(x)-Du(x_0)|^\alpha \leq \Lambda C_1\|DF\|_{C^\alpha(B_{C_1}(0))} |x-x_0|^{\alpha_1},\end{equation}
where
\[\alpha_1=\alpha\alpha_0.\]
We also have $\eta^{x_0}\mid_{\partial\Omega}\in C^{\alpha_1}(\partial\Omega)$ with
\begin{equation}
\label{etax0bdr}
\|\eta^{x_0}\|_{C^{\alpha_1}(\partial\Omega)} \leq C_3(\alpha, \alpha_0, C_1, \psi, \|DF\|_{C^\alpha(B_{C_1}(0))}).\end{equation}
Note that
\[\det (U^{ij})=(\det D^2 u)^{n-1}\geq \lambda^{n-1}.\]
Hence, from (\ref{etax0}), (\ref{bbetax0}) and (\ref{etax0bdr}), we can apply Proposition \ref{ptwH} and find constants \[\gamma=\alpha_1/(\alpha_1 +4)\in (0, 1),\] and $\delta, C_4>0$ depending only on $n,\lambda, \Lambda, \alpha, F,\varphi, \psi $, and $\Omega$ such that, for all $x\in \Omega\cap B_{\delta}(x_{0})$,
\begin{equation}
\label{etax0x0}
|\eta^{x_0}(x)-\eta^{x_0}(x_{0})|\leq C_4|x-x_{0}|^{\gamma}\left( \|\eta^{x_0}\|_{C^{\alpha_1}(\partial\Omega)} + \|f^{x_0}\|_{L^{n}(\Omega)}\right)\leq C_5|x-x_{0}|^{\gamma},\end{equation}
where $C_5= C_4(C_2 + C_3)$.
Due to
\[\eta(x) = \eta^{x_0}(x) e^{F(Du(x_0))+DF(Du(x_0))\cdot (Du(x)-Du(x_0))},\]
and (\ref{uc1a0}),
(\ref{etax0x0}) implies the pointwise $C^\gamma$ continuity of $\eta$ at $x_0$ with estimate
\begin{equation}
\label{etaptw}
|\eta(x)-\eta(x_0)|\leq C_6|x-x_0|^\gamma \quad \text{for all } x\in \Omega\cap B_{\delta} (x_0),
\end{equation}
where $C_6$ depends on $\Omega$, $\varphi$, $\psi$, $\lambda$, $\Lambda$, $n$, $\alpha$, $F$ and $\|g\|_{L^n(\Omega)}$.
{\it Step 2: Global H\"older continuity of $\eta$ and $w$.} From (\ref{etaptw}), we can apply Theorem \ref{LMA-G} to (\ref{etaeq}) to conclude the global H\"older continuity of $\eta$. Since $w= \eta e^{-F(Du)}$, $w$ is also globally H\"older continuous. In other words,
there exist a constant
$\beta\in (0, 1)$ depending on $n,\lambda,\Lambda,\alpha$ and $F$, and a constant $C >0 $ depending only on $\Omega$, $\varphi$, $\psi$, $\lambda$, $\Lambda$, $n$, $r$, $F$ and $\|g\|_{L^n(\Omega)}$ such that
$$\|w\|_{C^{\beta}(\overline{\Omega})}\leq C. $$
The theorem is proved.
\end{proof}
\vskip 30pt
\section{Proof of Theorem \ref{thm:SBV-p-lap}}
\label{SBV-plap-pf}
In this section, we prove Theorem \ref{thm:SBV-p-lap} using a priori estimates and degree theory. With Theorem \ref{wHolder_thm} at hand, a key step is to establish a priori Hessian determinant estimates for uniformly convex solutions $u\in W^{4, s}(\Omega)$ ($s>n$) of (\ref{eq:Abreu-plap}).
For the Hessian determinant estimates, we will use the maximum principle and the Legendre transform; see also \cite[Theorem 1.2]{LZ} with a slightly different proof for the case of $F(x)=|x|^q/q$ $(q>1)$ and $c(x, z)$ being smooth.
\begin{lem}[Hessian determinant estimates]\label{lem:det-est-plap}
Let $\Omega\subset\mathbb R^n$ be an open, smooth, bounded and uniformly convex domain.
Assume that $\varphi\in C^{5}(\overline{\Omega})$ and $\psi\in C^{3}(\overline{\Omega})$ with $\min_{\partial \Omega}\psi>0$. Let $r, s>n$. Let $F\in W^{2, r}_{\text{loc}}(\mathbb R^n)$ be a convex function, and $c(x,z)$ be a function on $\overline\Omega\times\mathbb R$. Suppose $c(x,z)\leq 0$ with
$c\in C^\alpha(\overline\Omega\times \mathbb R)$ where $\alpha\in (0, 1)$ or
$c(x,z)\equiv c(x)\in L^s(\Omega)$.
Assume that $u\in W^{4, s}(\Omega)$ is a uniformly convex solution to the second boundary value problem
\begin{equation*}
\left\{
\begin{alignedat}{2}U^{ij}D_{ij}w~& =-\div (DF(Du))+c(x, u)~&&\text{\ in} ~\ \ \Omega, \\\
w~&= (\det D^2 u)^{-1}~&&\text{\ in}~\ \ \Omega,\\\
u ~&=\varphi~&&\text{\ on}~\ \ \partial \Omega,\\\
w ~&= \psi~&&\text{\ on}~\ \ \partial \Omega,
\end{alignedat}
\right.
\end{equation*}
where $(U^{ij})=(\det D^2 u)(D^2 u)^{-1}$.
Then
\[
C^{-1}\leq \det D^2u\leq ( \min_{\partial\Omega}\psi)^{-1}\quad\text{in }\Omega,
\]
where $C>0$ is a constant depending on $\Omega$, $n$, $\varphi$, $\psi$, $F$ and $c$. In the case of $c(x,z)\equiv c(x)\in L^s(\Omega)$, the dependence of $C$ on $c$ is via $\|c\|_{L^{n}(\Omega)}$.
\end{lem}
\begin{proof}
From the convexity of $F$ and $u$, we have
\[
-\div (DF(Du))=-\text{trace} (D^2 F(Du) D^2 u)\leq 0.
\]
This combined with $c(x,u)\leq 0$ yields \[U^{ij}D_{ij}w=-\div (DF(Du))+c(x,u)\leq 0\quad \text{in }\Omega.\]
Hence, by the maximum principle,
$w$ attains its minimum value in $\overline{\Omega}$ on the boundary. Thus \[w\geq\min_{\partial\Omega}w= \min_{\partial\Omega}\psi>0\quad \text{in }\Omega.\] This together with $\det D^2u=w^{-1}$ gives the upper bound for the Hessian determinant:
\[
\det D^2u\leq C_1:= ( \min_{\partial\Omega}\psi)^{-1}\quad\text{in }\Omega.
\]
From the above upper bound, by using $u=\varphi$ on $\partial\Omega$ together with $\Omega$ being smooth and uniformly convex,
we can construct suitable barrier functions to deduce that
\begin{equation}\label{eq:Du-bound-plap}
\sup_{\Omega}|u| + \|Du\|_{L^{\infty}(\Omega)}\leq C_2,
\end{equation}
where $C_2$ depends on $n, \varphi, \psi$ and $\Omega$.
We now proceed to establish a positive lower bound for the Hessian determinant.\\
Let \[u^*(y)= x\cdot Du(x)-u(x)\] be the Legendre transform of $u(x)$ where \[y = Du(x) \in \Omega^*:= Du(\Omega).\]
Then, (\ref{eq:Du-bound-plap}) implies
\begin{equation}\label{eq:u^*-bound-plap}
\text{diam}(\Omega^\ast) + \|u^*\|_{L^{\infty}(\Omega^\ast)}\leq C_3(n, \varphi, \psi, \Omega).
\end{equation}
In view of Proposition \ref{new2}, $u^*$ satisfies
\begin{equation}
\label{dualu}
u^{\ast ij}D_{ij} \left(w^*+F(y)\right)=c(Du^*,y\cdot Du^*-u^*)\quad \text{in }\Omega^\ast,
\end{equation}
where \[(u^{\ast ij})= (D^2 u^{\ast})^{-1},\quad \text{and } w^*=\log\det D^2u^*.\]
Note that, for $y= Du(x)\in \partial\Omega^\ast$ where $x\in\partial\Omega$, we have
\[w^\ast(y) =\log (\det D^2 u(x))^{-1} =\log \psi(x). \]
By the ABP maximum principle applied to (\ref{dualu}), and recalling (\ref{eq:u^*-bound-plap}), we find
\begin{eqnarray*}
\sup_{\Omega^*}(w^*+F(y)) &\leq& \sup_{\partial\Omega^*}(w^*+F(y))+C(n, \text{diam}(\Omega^\ast))\left\|\frac{c(Du^*,y\cdot Du^*-u^*)}{{(\det D^2u^*)}^{-1/n}}\right\|_{L^n(\Omega^*)}\\
&=& \sup_{\partial\Omega^*}(w^*+F(y))+C(n, \text{diam}(\Omega^\ast))\Bigg(\int_{\Omega}|c(x,u)|^n dx\Bigg)^{1/n}\\
&\leq& C_4
\end{eqnarray*}
where $C_4$ depends on $\Omega$, $n$, $\varphi$, $\psi$, $F$ and $c$. Clearly, in the case of $c(x,z)\equiv c(x)\in L^s(\Omega)$, the dependence of $C_4$ on $c$ is via $\|c\|_{L^{n}(\Omega)}$. In the above estimates, we used
\begin{eqnarray*}
\left\|\frac{c(Du^*,y\cdot Du^*-u^*)}{{(\det D^2u^*)}^{-1/n}}\right\|_{L^n(\Omega^*)} &=&\Bigg(\int_{\Omega^*}|c(Du^*,y\cdot Du^*-u^*)|^n \det D^2u^*\,dy\Bigg)^{1/n}\\
&=&\Bigg(\int_{\Omega}|c(x,u)|^n \det D^2u^* \det D^2u\,dx\Bigg)^{1/n}\\
&=& \Bigg(\int_{\Omega}|c(x,u)|^n dx\Bigg)^{1/n}.
\end{eqnarray*}
It follows that
\[\sup_{\Omega^*}w^*(y)=\sup_{\Omega^*} \log\det D^2 u^\ast \leq C_5\]
which implies
\[
\det D^2u\geq e^{-C_5}>0 \quad\text{in }\Omega,
\]
where $C_5$ depends on $\Omega$, $n$, $\varphi$, $\psi$, $F$ and $c$. This is the desired positive lower bound for the Hessian determinant, and
the proof of the lemma is completed.
\end{proof}
Now, we can give the proof of Theorem \ref{thm:SBV-p-lap}.
\begin{proof}[Proof of Theorem \ref{thm:SBV-p-lap}] We divide the proof, using a priori estimates and degree theory, into three steps. Steps 1 and 2 establish higher order derivative estimates for $u\in W^{4, s}(\Omega)$ ($s>n$) solutions. Step 3 confirms the existence of $W^{4, s}(\Omega)$ or $C^{4,\beta}(\overline{\Omega})$ solutions via degree theory.
In the following, we fix $s>n$ with the additional requirement that
\begin{equation*}
\left\{
\begin{alignedat}{2}
s&= r~&&\text{\ in case }~ (i), \\\
s~&= \min\{r, p\} ~&&\text{\ in case }~(ii).
\end{alignedat}
\right.
\end{equation*}
{\it Step 1: Determinant estimates and second order derivative estimates for uniformly convex $u\in W^{4, s}(\Omega)$ ($s>n$) solutions $u$ of (\ref{eq:Abreu-plap}).}
By Lemma \ref{lem:det-est-plap}, we have
\begin{equation}\label{eq:d-boud-plap}
0<\lambda\leq \det D^2u\leq \Lambda:= ( \min_{\partial\Omega}\psi)^{-1} \quad\text{in }\Omega,
\end{equation}
where $\lambda$ depends on $\Omega$, $n$, $F$, $\varphi$, $\psi$, and on either $c$ in case $(i)$, or $\|c\|_{L^{n}(\Omega)}$ in case $(ii)$.
From (\ref{eq:d-boud-plap}) and $u=\varphi$ on $\partial\Omega$ where $\varphi\in C^{5}(\overline{\Omega})$,
by \cite[Proposition 2.6]{LS}, we have
\begin{equation}\label{C1beta}\|u\|_{C^{1,\alpha_0}(\overline\Omega)}\leq C_1,\end{equation}
where $\alpha_0\in (0, 1)$ depends on $\lambda,\Lambda$, and $n$. The constant $C_1$ depends on $\Omega, n,\lambda,\Lambda$ and $\varphi$.
With (\ref{eq:d-boud-plap}) and $F\in W^{2, r}_{\text{loc}}(\mathbb R^n)$, we can use Theorem \ref{wHolder_thm} to find $\beta_0\in (0,1)$, and $C_5>0$ depending on $\Omega, n, F, r, \varphi, \psi, c$, such that
\[
\|w\|_{C^{\beta_0}(\overline{\Omega})}\leq C_2(\Omega, n, F, r, \varphi, \psi, c).
\]
Hence
$
\det D^2u=w^{-1}\in C^{\beta_0}(\overline{\Omega})$.
By the global Schauder estimates for the Monge-Amp\`ere equation in \cite{S2,TW3}, we have
\begin{equation}
\label{ue2beta}
\|u\|_{C^{2,\beta_0}(\overline{\Omega})}\leq C_3(\Omega, n, F, r, \varphi, \psi, c).
\end{equation}
Combining this with \eqref{eq:d-boud-plap}, we find \[C_4^{-1} I_n\leq D^2u\leq C_4 I_n \quad\text{in }\Omega\] for some $C_4(\Omega, n, F, r, \varphi, \psi, c)>0$. Here $I_n$ denotes the identity $n\times n$ matrix. In other words, the linear operator $U^{ij}D_{ij}$ is uniformly elliptic with coefficients $U^{ij}$ bounded in $C^{\beta_0}(\overline{\Omega})$.
\vskip 7pt
{\it Step 2: Global higher order derivative estimates for uniformly convex $W^{4, s}(\Omega)$ $(s>n)$ solutions $u$ of (\ref{eq:Abreu-plap}).} Denote the right-hand side of \eqref{eq:Abreu-plap} by
\begin{equation}\label{eq:f_e}
f:=-\div (DF(Du)) + c(x, u)=-\text{trace} (D^2 F(Du) D^2 u) +c(x,u).
\end{equation}
Observe that,
one has
the following estimate
\begin{equation}
\label{GLr}
\|\text{trace} (D^2 F(Du) D^2 u)\|_{L^{r}(\Omega)} \leq C(\Omega, n, F, r, \varphi, \psi, c).\end{equation}
Indeed, we have
\begin{eqnarray*}
\|\text{trace} (D^2 F(Du) D^2 u)\|^r_{L^r(\Omega)} &\leq& n^2 \|D^2 u\|^r_{L^{\infty}(\Omega)}\| D^2 F(Du) \|^r_{L^r(\Omega)} \\
&\leq& n^2C_3^r \int_{\Omega} |D^2 F(Du(x))|^{r} dx\quad (\text{using } (\ref{ue2beta}))\\
&=& n^2C_3^r \int_{Du(\Omega)} |D^2 F(y)|^{r} \frac{1}{\det D^2 u((Du)^{-1}(y))}dy \\
&\leq& n^2C_3^r \lambda^{-1} \int_{B_{C_1}(0)} |D^2 F(y)|^{r} dy \quad (\text{using } (\ref{eq:d-boud-plap}) \text{and } (\ref{C1beta}))\\
&\leq& C_3^r \lambda^{-1} C(n, C_1, F, r).
\end{eqnarray*}
We consider cases $(i)$ and $(ii)$ separately.
\vskip 7pt
\begin{enumerate}
\item[(i)]
{\it The case of $ c\in C^\alpha(\overline\Omega\times\mathbb R)$.} Recall that $s=r$ in this case.
We have from (\ref{GLr}) that $f=-\text{trace} (D^2 F(Du) D^2 u) + c(x, u)\in L^s(\Omega)$ with estimate
\[\|f\|_{L^{s}(\Omega)} \leq C(\Omega, n, F, r, s, \varphi, \psi, c).\]
By {\it Step 1}, \begin{equation*}
U^{ij}D_{ij}w =f~\text{\ in} ~ \Omega,
w ~=\psi~\text{\ on}~\partial \Omega,
\end{equation*}
is a uniformly elliptic equation in $w$ with $C^{\beta_0}(\overline{\Omega})$ coefficients. Thus, from the standard $W^{2,p}$ theory for uniformly elliptic linear equations (see \cite[Chapter 9]{GT}),
we obtain the following $W^{2, s}(\Omega)$ estimate:
\[
\|w\|_{W^{2,s}(\Omega)}\leq C(\Omega, n, q, s, \varphi, \psi, c).
\]
This combined with $\det D^2 u=w^{-1}$ in $\Omega$ with $u=\varphi$ on $\Omega$ yields the following
$W^{4,s}$ estimate of $u$:
$$\|u\|_{W^{4,s}(\Omega)}\leq C(\Omega, n, F, r, s, \varphi, \psi, c).$$
Moreover, in the particular case that $F\in C^{2,\alpha_0}(\mathbb R^n)$, we find that $f\in C^{\gamma}(\overline{\Omega})$ where $\gamma\in(0, 1)$ depends only on $\alpha$, $F, \alpha_0$, and $\beta_0$ with estimate
\begin{equation}
\label{fcgamma}
\|f\|_{C^{\gamma}(\overline{\Omega})}\leq C(\Omega, n, \alpha, q, c, \varphi, \psi).\end{equation}
Thus, we can apply the classical Schauder theory (see \cite[Chapter 6]{GT} for example) to \eqref{eq:Abreu-plap} which, by {\it Step 1}, is a uniformly elliptic equation in $w$ with $C^{\beta_0}(\overline{\Omega})$ coefficients. We conclude that $w \in C^{2, \beta}(\overline{\Omega})$, where $\beta\in (0, 1)$ depends only on $n,\gamma$ and $\beta_0$, with estimate
\[\|w\|_{C^{2, \beta}(\overline{\Omega})}\leq C(\Omega, n, \alpha, \alpha_0, F, c, \varphi, \psi).\]
Due to \[\det D^2 u=w^{-1} \quad \text{in }\Omega, \quad u=\varphi \quad
\text{on }\Omega,\] this implies that $u\in C^{4, \beta}(\overline{\Omega})$ with estimate
\begin{equation}
\label{uc4b}
\|u\|_{C^{4, \beta}(\overline{\Omega})}\leq C(\Omega, n, \alpha, \alpha_0, F, c, \varphi, \psi).
\end{equation}
With this estimate, we go back to $f=-\text{trace} (D^2 F(Du) D^2 u) + c(x, u)$ and find that one can actually take $\gamma=\min\{\alpha,\alpha_0\}$ in (\ref{fcgamma}). Repeating the above process, one find that (\ref{uc4b}) holds for $\beta=\min\{\alpha,\alpha_0\}$.
\item[(ii)]{\it The case of $c(x,z)\equiv c(x) \in L^p(\Omega)$ with $p>n$.} Recall that in this case $s=\min\{r, p\}$. Then, we have from \eqref{eq:f_e} and (\ref{GLr}) that
\begin{equation*}\|f\|_{L^{s}(\Omega)} \leq (\Omega, n, p, F, r, s, \varphi, \psi, \|c\|_{L^p(\Omega)}).\end{equation*}
Arguing as in the case $(i)$ above, we obtain the following $W^{4,s}$ estimate of $u$:
$$\|u\|_{W^{4,s}(\Omega)}\leq C(\Omega, n, p, F, r, s,\varphi, \psi, \|c\|_{L^p(\Omega)}).$$
\end{enumerate}
{\it Step 3: Existence of solutions via degree theory.} From the $C^{4,\beta}(\overline{\Omega})$ or $W^{4, s}(\Omega)$ estimates for uniformly convex $W^{4, s}(\Omega)$ solutions $u$ of (\ref{eq:Abreu-plap}) in {\it Step 2}, we can use the Leray-Schauder degree theory as in \cite{CW,TW2,LeCPAM}
to prove the existence of $C^{4,\beta}(\overline{\Omega})$ or $W^{4, s}(\Omega)$ solutions to (\ref{eq:Abreu-plap}) as stated in the theorem. We omit details here.
\end{proof}
\vskip 10pt
\section{Proof of Theorem \ref{SBV-1}}
\label{SBV-1-pf}
In this section, we prove Theorem \ref{SBV-1}.
As in the proof of Theorem \ref{thm:SBV-p-lap} in Section \ref{SBV-plap-pf}, we focus on a priori estimates for smooth, uniformly convex solutions. The most crucial ones are the Hessian determinant estimates.
Without the sign of $c$, we first need to obtain the \textit{a priori} $L^\infty$-bound for $u$.
\begin{lem}[A priori $L^\infty$-bound for uniformly convex $W^{4, n}$ solutions]
\label{lem:bound-u-high}
Let $\Omega\subset\mathbb R^n$($n\geq 3$) be an open, smooth, bounded and uniformly convex domain.
Assume that $\varphi\in C^{5}(\overline{\Omega})$ and $\psi\in C^{3}(\overline{\Omega})$ with $\min_{\partial \Omega}\psi>0$.
Assume $\text{\bf{b}}\in L^\infty(\Omega;\mathbb R^n)$. Suppose that there exist functions $g_1, g_2\in L^1(\Omega)$ and a constant $0\leq m<n-1$ such that
\begin{equation}\label{c-cond}
|c(x,z)|\leq |g_1(x)|+|g_2(x)|\cdot |z|^{m} \quad\text{in }\Omega\times\mathbb R.
\end{equation}
Assume that $u\in W^{4, n}(\Omega)$ is a uniformly convex solution to \eqref{Abreu-lap}. Then there exists a constant $C>0$ depending on $\Omega$, $n$, $\varphi$, $\psi$, $\|\text{\bf{b}}\|_{L^\infty(\Omega)}$, $\|g_1\|_{L^1(\Omega)}$, $\|g_2\|_{L^1(\Omega)}$ and $m$ such that
\[
\|u\|_{L^\infty(\Omega)}\leq C.
\]
\end{lem}
\begin{proof} From $u\in W^{4, n}(\Omega)$ and the Sobolev embedding theorem, we have $u\in C^2(\overline{\Omega})$.
For a convex function $u\in C^2(\Omega)$ with $u = \varphi$ on $\partial\Omega$, we have (see, e.g., \cite[inequality (2.7)]{Le2})
\begin{equation}\label{eq:u-bound-high}
\|u\|_{L^{\infty}(\Omega)} \leq \|\varphi\|_{L^{\infty}(\Omega)}+C_1\left(n, \Omega,\|\varphi\|_{C^{2}(\Omega)}\right)\left(\int_{\partial \Omega}\left(u_{\nu}^{+}\right)^{n}\,dS\right)^{1 / n},
\end{equation}
where $u_{\nu}^{+}=\max \left(0, u_{\nu}\right)$, $\nu$ is the unit outer normal of $\partial\Omega$ and $dS$ is the boundary measure.
Thus, to prove the lemma, it suffices to prove
\[
\int_{\partial \Omega}\left(u_{\nu}^{+}\right)^{n} d S \leq C(\Omega, n, \varphi, \psi, \|\text{\bf{b}}\|_{L^\infty(\Omega)}, \|g_1\|_{L^1(\Omega)}, \|g_2\|_{L^1(\Omega)}, m).
\]
For this, we use the arguments as in the proof of \cite[Lemma 4.2]{LeCPAM}.
Observe that,
since $u$ is convex with boundary value $\varphi$ on $\partial\Omega$, we have
$u_\nu\geq -\|D\varphi\|_{L^{\infty}(\Omega)}$ and hence
\begin{equation}
\label{unuab}
|u_\nu|\leq u_\nu^{+}+ \|D\varphi\|_{L^{\infty}(\Omega)},\quad \text{and }
(u_\nu^{+})^n\leq u^n_\nu + \|D\varphi\|^n_{L^{\infty}(\Omega)} \quad\text{on }\partial\Omega.
\end{equation}
Let $\rho$ be a strictly convex defining function of $\Omega$, i.e.
\[
\Omega:=\left\{x \in \mathbb R^{n}: \rho(x)<0\right\}, \rho=0 \text { on } \partial \Omega \text { and } D \rho \neq 0 \text { on } \partial \Omega.
\]
Let
\[
\tilde{u}=\varphi+\mu\left(e^{\rho}-1\right) .
\]
Then, for $\mu$ large, depending on $n$, $\Omega$ and $\|\varphi\|_{C^{2}(\overline{\Omega})}$, the function $\tilde{u}$ is uniformly convex, belongs to $C^{5}(\overline{\Omega})$. Furthermore, as in \cite[Lemma 2.1]{Le2}, there exists a constant $C_2>0$ depending only on $n$, $\Omega$, and $\|\varphi\|_{C^{4}(\overline{\Omega})}$ such that the following facts hold:
(i) $\|\tilde{u}\|_{C^{4}(\overline{\Omega})} \leq C_2, \quad$ and $\operatorname{det} D^{2} \tilde{u} \geq C_2^{-1}>0$ in $\Omega$,
(ii) letting $\tilde{w}=\left[\operatorname{det} D^{2} \tilde{u}\right]^{-1}$, and denoting by $\left(\tilde{U}^{i j}\right)$ the cofactor matrix of $D^{2} \tilde{u}$, we have
\[
\big\|\tilde{U}^{i j} D_{i j} \tilde{w}\big\|_{L^{\infty}(\Omega)} \leq C_2.
\]
Let $K(x)$ be the Gauss curvature at $x \in \partial \Omega$. Then, since $\Omega$ is uniformly convex, we have \begin{equation}\label{Kom} 0<C^{-1}(\Omega)\leq K(x)\leq C(\Omega) \quad\text{on }\partial\Omega.\end{equation}
From the estimate (4.10) in the proof of \cite[Lemma 4.2]{LeCPAM} with $\theta=0$ and $f_\delta:= -\Delta u +\textbf{b}\cdot Du+ c$ which uses (i) and (ii), we obtain
\begin{equation}\label{eq:K-u_nu-high}
\begin{aligned}
\int_{\partial \Omega} K \psi u_{\nu}^{n} d S \leq & \int_{\Omega}\left(\Delta u-\textbf{b}\cdot Du-c\right)(u-\tilde{u}) d x+C_3\left(\int_{\partial \Omega}\left(u_{\nu}^{+}\right)^{n} d S\right)^{(n-1)/ n}+C_3,
\end{aligned}
\end{equation}
where $C_3$ depends on $C_2$, $\Omega$ and $\varphi$.
We will estimate the first term on the right-hand side of \eqref{eq:K-u_nu-high} by splitting it into three terms. Firstly, using $u\Delta u=\div (u Du)-|Du|^2$ and integrating by parts, we have
\begin{align}
\int_{\Omega}\Delta u(u-\tilde{u})\,dx&\leq\int_{\Omega}u\Delta u\,dx+C_2\int_{\Omega}\Delta u\,dx\nonumber\\
&=\int_{\partial\Omega}\varphi u_{\nu}\,dS-\int_{\Omega}|Du|^2\,dx+C_2\int_{\partial\Omega}u_{\nu}\,dS\nonumber\\
&\leq C(\varphi, C_2)\int_{\partial\Omega}|u_{\nu}|\,dS-\int_{\Omega}|Du|^2\,dx\nonumber\\
&\leq C_4(n,\varphi, C_2)\left(\int_{\partial\Omega}(u^{+}_{\nu})^n\,dS\right)^{\frac{1}{n}} + C_4(n,\varphi, C_2)\quad(\text{recalling } (\ref{unuab})).\label{eq:1st-RHS}
\end{align}
Secondly, by integration by parts, we find
\begin{align}
\int_{\Omega}|Du|^2\,dx&=\int_{\Omega}(\div(uDu)-u\Delta u)\,dx\nonumber\\
&=\int_{\partial\Omega}\varphi u_{\nu}\,dS-\int_{\Omega}u\Delta u\,dx\nonumber\\
&\leq C_5(\varphi)\int_{\partial\Omega}u_{\nu}^+\,dS+\|u\|_{L^\infty(\Omega)}\int_{\Omega}\Delta u\,dx + C_5(\varphi)\nonumber\\
&\leq(C_5+\|u\|_{L^\infty(\Omega)})\int_{\partial\Omega}u_{\nu}^+\,dS + C_5.\label{eq:Du-L2bound-high}
\end{align}
In view of \eqref{eq:Du-L2bound-high} with \eqref{eq:u-bound-high}, we can estimate
\begin{align}
\int_{\Omega} \text{\bf{b}}\cdot Du(\tilde{u}-u)\,dx &\leq |\Omega|^{1/2}\|\text{\bf{b}}\|_{L^\infty(\Omega)}(\|\tilde{u}\|_{L^\infty(\Omega)}+\|u\|_{L^\infty(\Omega)})\left(\int_{\Omega}|Du|^2\,dx\right)^{\frac{1}{2}}\nonumber\\
&\leq C_6+C_6\left(\int_{\partial \Omega}\left(u_{\nu}^{+}\right)^{n} d S\right)^{\frac{2}{n}}\label{eq:2nd-RHS}
\end{align}
where $C_6$ depends on $\Omega, n$, $\varphi$ and $\|\text{\bf{b}}\|_{L^\infty(\Omega)}$.
Moreover, $C_6$ depends linearly on $\|\text{\bf{b}}\|_{L^\infty(\Omega)}$.
Finally, using \eqref{c-cond} and (\ref{eq:u-bound-high}), we have
\begin{align}
\int_{\Omega}-c(u-\tilde{u})\,dx
&\leq(\|u\|_{L^\infty(\Omega)}+\|\tilde{u}\|_{L^\infty(\Omega)})\int_{\Omega}|g_1|+|g_2||u|^{m}\,dx\nonumber\\
&\leq C+C\|u\|_{L^\infty(\Omega)}^{m+1}\nonumber\\
&\leq C_7+C_7\left(\int_{\partial\Omega}(u_{\nu}^+)^n\,dS\right)^{\frac{m+1}{n}}.\label{eq:3rd-RHS}
\end{align}
Here $C_7$ depends on $\Omega$, $n$, $\varphi$, $\|g_1\|_{L^1(\Omega)}$, $\|g_2\|_{L^1(\Omega)}$ and $m$.
It follows from (\ref{unuab}) that
\begin{equation}
\label{unu+}
\int_{\partial\Omega}K\psi (u_{\nu}^+)^n\,dS\leq C_8(\Omega,\varphi,\psi) + \int_{\partial\Omega}K \psi u_{\nu}^{n} \,d S.
\end{equation}
Combining (\ref{Kom})--\eqref{eq:1st-RHS}, \eqref{eq:2nd-RHS}--(\ref{unu+}) while recalling that $0\leq m<n-1$ and $n\geq 3$, we obtain
\begin{align*}
C^{-1}(\Omega)\min_{\partial\Omega}\psi\int_{\partial\Omega} (u_{\nu}^+)^n\,dS&\leq C_8 + \int_{\partial\Omega}K \psi u_{\nu}^{n} \,d S\\
&\leq C_9\left[1+\left(\int_{\partial\Omega}(u_{\nu}^+)^n\,dS\right)^{\frac{n-1}{n}}+\left(\int_{\partial\Omega}(u_{\nu}^+)^n\,dS\right)^{\frac{m+1}{n}}\right],
\end{align*}
where $C_9$ depends on $C_3, C_4$, $C_6$, $C_7$ and $C_8$.
It follows that
\[
\int_{\partial\Omega}(u_{\nu}^+)^n\,dS\leq C
\]
where $C$ depends on $\Omega$, $n$, $\varphi$, $\psi$, $\|\text{\bf{b}}\|_{L^\infty(\Omega)}$, $\|g_1\|_{L^1(\Omega)}$, $\|g_2\|_{L^1(\Omega)}$ and $m$.
The proof of the lemma is completed.
\end{proof}
\begin{rem} We have the following observations regarding the two dimensional version of Lemma \ref{lem:bound-u-high}.
\label{lem51_rem}
\begin{enumerate}
\item[(i)] The above proof fails in two dimensions. This is because, in two dimensions, the right-hand side of \eqref{eq:2nd-RHS} is of the same order of magnitude as the left-hand side of (\ref{eq:K-u_nu-high}). Therefore, when $C_6$ is large, plugging \eqref{eq:2nd-RHS} into (\ref{eq:K-u_nu-high}) does not give any new information.
\item[(ii)] On the other hand, since $C_6$ depends linearly on $\|\text{\bf{b}}\|_{L^\infty(\Omega)}$, in two dimensions, one can still absorb the right-hand side of \eqref{eq:2nd-RHS} into the left-hand side of (\ref{eq:K-u_nu-high}) as long as $\|\text{\bf{b}}\|_{L^\infty(\Omega)}$ is small, depending on $\Omega,\varphi$ and $\psi$. In this case, we still have the $L^\infty$ estimate.
\item[(iii)]In Section \ref{SBV-2-pf}, we will establish the $L^\infty$ estimate in two dimensions under a stronger condition on
$\text{\bf{b}}$ but $\|\text{\bf{b}}\|_{L^\infty(\Omega)}$ can be arbitrarily large.
\end{enumerate}
\end{rem}
Next, we establish the Hessian determinant estimates.
\begin{lem}[Hessian determinant estimates]
\label{upper_lower}
Let $u\in W^{4,p}(\Omega)$ be a uniformly convex solution to the fourth order equation
\begin{equation}
\label{eq_lower}
\left\{
\begin{alignedat}{2}\sum_{i, j=1}^{n}U^{ij}D_{ij}w~& =-\Delta u+\text{\bf{b}}\cdot D u+c(x)~&&\text{\ in} ~\ \ \Omega, \\\
w~&= (\det D^2 u)^{-1}~&&\text{\ in}~\ \ \Omega,\\\
u ~&=\varphi~&&\text{\ on}~\ \ \partial \Omega,\\\
w ~&= \psi~&&\text{\ on}~\ \ \partial \Omega,
\end{alignedat}
\right.
\end{equation}
where $(U^{ij})=(\det D^2 u)(D^2 u)^{-1}$, $\min_{\partial\Omega}\psi>0$, $\text{\bf{b}}\in L^\infty(\Omega;\mathbb R^n)$ and $c\in L^{p}(\Omega)$ with $p>2n$. Then there exists a constant $C>0$ depending on $\Omega$, $n$, $p$, $\varphi$, $\psi$, $\|\text{\bf{b}}\|_{L^\infty(\Omega)}$ and $\|c\|_{L^p(\Omega)}$ such that
$$0<C^{-1}\leq \det D^2u\leq C\quad\text{in }\Omega.$$
\end{lem}
\begin{proof}
The proof uses a trick in Chau-Weinkove \cite{CW}. For simplicity,
denote \[d:=\det D^2u\quad\text{and } (u^{ij})= (D^2 u)^{-1}.\] Let \[G=d e^{Mu^2},\] where $M>0$ is a large constant to be determined later. By Lemma \ref{lem:bound-u-high},
we have \[\|u\|_{L^\infty(\Omega)}\leq C_0\]
where $C_0>0$ depends on $\Omega$, $n$, $\varphi$, $\psi$, $\|\text{\bf{b}}\|_{L^\infty(\Omega)}$, and $\|c\|_{L^1(\Omega)}$.
Since $w=d^{-1}$, we have
$w=G^{-1}e^{Mu^2}$. Direct calculations yield
\begin{eqnarray*}
D_iw &=& -G^{-2}D_iGe^{Mu^2}+2MuD_iuG^{-1}e^{Mu^2}, \\[5pt]
D_{ij} w&=& 2G^{-3}D_iGD_jGe^{Mu^2}-G^{-2}D_{ij}Ge^{Mu^2}\\[5pt] && -2MuD_juD_iGG^{-2}e^{Mu^2}-2MuD_iuD_jGG^{-2}e^{Mu^2}\\[5pt]
&&+2MD_iuD_juG^{-1}e^{Mu^2}+2MuD_{ij}uG^{-1}e^{Mu^2}+4M^2u^2D_iuD_juG^{-1}e^{Mu^2}.
\end{eqnarray*}
Then, using $U^{ij} G^{-1} e^{Mu^2}= u^{ij}$, we have
\begin{eqnarray*}
U^{ij}D_{ij} w&=&2G^{-2}u^{ij}D_iGD_jG-G^{-1}u^{ij}D_{ij}G-4MuG^{-1}u^{ij}D_iuD_jG\\[5pt]
&&+2Mu^{ij}D_iuD_ju+2Mnu+4M^2u^2u^{ij}D_iuD_ju\\[5pt]
&=&G^{-2}u^{ij}D_iGD_jG+u^{ij}(2MuD_iu-G^{-1}D_iG)(2MuD_ju-G^{-1}D_jG)\\[5pt]
&&-G^{-1}u^{ij}D_{ij}G+2Mu^{ij}D_iuD_ju+2Mnu\\[5pt]
&\geq&-G^{-1}u^{ij}D_{ij}G+2Mu^{ij}D_iuD_ju+2Mnu.
\end{eqnarray*}
Thus, from the first equation in \eqref{eq_lower}, we obtain
\[G^{-1}u^{ij}D_{ij}G\geq 2Mu^{ij}D_iuD_ju+2Mnu+\Delta u-\text{\bf{b}}\cdot Du-c.\]
Using the following matrix inequality (see, for example, \cite[Lemma 2.8(c)]{LeCCM})
\[u^{ij} D_i v D_j v \geq \frac{|Dv|^2}{\mbox{trace}\,(D^2u)} =\frac{|Dv|^2}{\Delta u}\]
together with $\Delta u\geq n d^{1/n}$, we find that
\begin{eqnarray}
\label{CWABP}
G^{-1}u^{ij}D_{ij}G
&\geq& 2M\frac{|Du|^2}{\Delta u}+\frac{1}{2}\Delta u-\text{\bf{b}}\cdot Du+\frac{1}{2}\Delta u+2Mnu-c\nonumber\\[5pt]
&\geq&2\sqrt{M}|Du|-|\text{\bf{b}}|\cdot |Du|+\frac{n}{2}d^{\frac{1}{n}}+2Mnu-c\nonumber\\[5pt]
&\geq&-(c-2Mnu-\frac{n}{2}d^{\frac{1}{n}})^+ \quad\text{in }\Omega,
\end{eqnarray}
provided \[M\geq \frac{1}{4}\|\text{\bf{b}}\|_{L^\infty(\Omega)}^2.\] Hence, by the ABP estimate applied to (\ref{CWABP}) in $\Omega$ where $G= \psi^{-1}e^{M\varphi^2}$ on $\partial\Omega$, we have
\begin{eqnarray}
\sup_{\Omega}G&\leq&\sup_{\partial\Omega}(\psi^{-1}e^{M\varphi^2})
+C(n,\Omega)\left\|\frac{(c-2Mnu-\frac{n}{2}d^{\frac{1}{n}})^+}{\big[\det (G^{-1} (D^2 u)^{-1})\big]^{\frac{1}{n}}}\right\|_{L^n(\Omega)}\nonumber\\[5pt]
&=&\sup_{\partial\Omega}(\psi^{-1}e^{M\varphi^2})
+C(n,\Omega)\left\|\frac{de^{Mu^2}(c-2Mnu-\frac{n}{2}d^{\frac{1}{n}})^+}{d^{-\frac{1}{n}}}\right\|_{L^n(\Omega)}\nonumber\\[5pt]
&\leq&\sup_{\partial\Omega}(\psi^{-1}e^{M\varphi^2})
+C_1\left\|d^{1+\frac{1}{n}}(c-2Mnu-\frac{n}{2}d^{\frac{1}{n}})^+\right\|_{L^n(\Omega)}.\label{ABP_lower}
\end{eqnarray}
Here $C_1$ depends on $n$, $\Omega$ and $C_0$ (via $\|u\|_{L^\infty(\Omega)}$).
Note that, for any $p_0>n$, we have
\begin{multline}
\label{CWineq}
\left\|d^{1+\frac{1}{n}}(c-2Mnu-\frac{n}{2}d^{\frac{1}{n}})^+\right\|_{L^n(\Omega)}\leq\left(\int_{\big\{c-2Mnu\geq (n/2)d^{\frac{1}{n}}\big\}}d^{n+1}(c-2Mnu)^n\,dx\right)^{\frac{1}{n}}\\
\leq\left(\int_{\big\{ c-2Mnu\geq (n/2)d^{\frac{1}{n}}\big\}}d^{n+1}(c-2Mnu)^n\frac{(c-2Mnu)^{p_0-n}}{\big[(n/2)d^{1/n}\big]^{p_0-n}}\,dx\right)^{\frac{1}{n}}\\
=(n/2)^{-\frac{p_0-n}{n}}\left(\int_{\big\{c-2Mnu\geq (n/2)d^{\frac{1}{n}}\big\}}d^{n-\frac{p_0}{n}+2}(c-2Mnu)^{p_0}\,dx\right)^{\frac{1}{n}}.
\end{multline}
We now choose $p_0$ such that \[2n<p_0<\min\{n(n+2), p\}.\]
Let $\gamma=1-\frac{p_0}{n^2}+\frac{2}{n}$.
Then $0<\gamma<1$. Moreover, from (\ref{ABP_lower}) and (\ref{CWineq}), we have
\begin{eqnarray*}
\sup_{\Omega}G&\leq&C+C\left(\int_\Omega d^{n\gamma} (|c|+|u|)^{p_0} \,dx\right)^\frac{1}{n}\\[4pt]
&\leq &C+C\left(\int_\Omega (de^{Mu^2})^{n\gamma} (|c|^p+1) \,dx\right)^\frac{1}{n}\\[4pt]
&\leq &C_2+C_2(\sup_{\Omega}G)^\gamma \cdot \left(\int_\Omega (|c|^p+1) \,dx\right)^\frac{1}{n}.
\end{eqnarray*}
Here $C_2$ depends on $\Omega, M,\varphi,\psi$, $C_1$, $\gamma$ and $p$.
It follows that
\[\sup_{\Omega}G\leq C_3(C_2, \gamma, \|c\|_{L^{p}(\Omega)}).\]
Since $G= d e^{Mu^2}$, we also get an upper bound for $d=\det D^2 u$:
\[\det D^2 u\leq C_3\quad\text{in }\Omega.\]
It remains to establish a positive lower bound for $\det D^2u$.
Once we have the upper bound of the Hessian determinant of $u$, using $u=\varphi$ on $\partial\Omega$ and a suitable barrier, we obtain
\[\sup_{\Omega}|u| + \sup_\Omega|Du|\leq C_4(C_3,\varphi,\Omega).\]
Then we can apply the Legendre transform to get the lower bound of the determinant.
According to Proposition \ref{new2}, the Legendre transform $u^*$ of $u$ satisfies
\[
u^{\ast ij}D_{ij} \left(w^*+\frac{|y|^2}{2}\right)=\text{\bf{b}}(D u^{*})\cdot y+ c(Du^{*}) \quad\text{in }\Omega^*=Du(\Omega),
\]
where $u^{\ast ij}= (D^2 u^{\ast})^{-1}$ and $w^*=\log\det D^2u^*$. Applying the ABP estimate to $w^*+\frac{|y|^2}{2}$ on $\Omega^*$, and then changing of variables $y=Du(x)$ with $dy=\det D^2u \,dx$, we obtain
\begin{eqnarray*}
\sup_{\Omega^*}\Big(w^*+\frac{|y|^2}{2}\Big)&\leq& \sup_{\partial\Omega^*}\Big(w^*+\frac{|y|^2}{2}\Big)+C(n) \text{diam}(\Omega^*)
\left\|\frac{\text{\bf{b}}(D u^{*}) \cdot y+ c(Du^{*})}{(\det u^{*ij})^{\frac{1}{n}}}\right\|_{L^n(\Omega*)}\\
&\leq&C(\psi, C_4)+C(n, C_4)\left(\int_{\Omega^*}\frac{|\text{\bf{b}}(D u^{*}) \cdot y+ c(Du^{*})|^n}{(\det D^2u^*)^{-1}}\,dy\right)^{\frac{1}{n}}\\
&=&C(\psi, C_4)+C(n, C_4) \left(\int_{\Omega}|\text{\bf{b}} \cdot Du+ c(x)|^n\,dy\right)^{\frac{1}{n}}\\[5pt]
&\leq& C(\psi, C_4)+C(n, C_4)(\|\text{\bf{b}}\|_{L^n(\Omega)}\sup_\Omega |Du|+\|c\|_{L^n(\Omega)}).
\end{eqnarray*}
In particular, we have
\[\sup_{\Omega^*}w^* \leq C_5\]
where $C_5>0$ depending on $\Omega$, $n$, $\varphi$, $\psi$, $\|\text{\bf{b}}\|_{L^\infty(\Omega)}$ and $\|c\|_{L^p(\Omega)}$.
Since $w^*=\log\det D^2u^*$, the above estimate gives the lower bound for $\det D^2u$:
\[\det D^2 u\geq e^{-C_5} \quad\text{in }\Omega,\]
completing the proof of the lemma.
\end{proof}
\begin{rem}
If there is no first order term, $\textbf{b}\cdot Du$ on the right-hand of \eqref{Abreu-lap}, we can directly obtain Hessian determinant bounds by the same trick used in the proof of Lemma \ref{upper_lower} without getting \textit{a priori} $L^\infty$-bound of $u$. Moreover, these bounds are valid for all dimensions.
\end{rem}
We are now ready to prove Theorem \ref{SBV-1}.
\begin{proof}[Proof of Theorem \ref{SBV-1}] The proof uses a priori estimates and degree theory as in that of Theorem \ref{thm:SBV-p-lap}.
We obtain the existence of a uniformly convex solution in $C^{4,\alpha}(\overline{\Omega})$ in case $(i)$, and in $W^{4, p}(\Omega)$ in case $(ii)$, with stated estimates provided that we can establish these estimates for $W^{4,p}(\Omega)$ solutions. Thus, it remains to establish these a priori estimates.
Assume now $u\in W^{4,p}(\Omega)$ is a uniformly convex smooth solution to \eqref{Abreu-lap}. By Lemma \ref{lem:bound-u-high} and the assumption on $c$ in either $(i)$ or $(ii)$, we can obtain the Hessian determinant estimates for $u$ by Lemma \ref{upper_lower}.
Once we have the Hessian determinant estimates, Theorem \ref{wHolder_thm} applies
with
\[F(x)=|x|^2/2,\quad\text{and }g(x)= b(x)\cdot Du(x) + c(x).\]
This gives the H\"older estimates for $w$. The rest of the proof of Theorem \ref{SBV-1}, which is concerned with global higher order derivative estimates, is similar to {\it Step 2} in the proof of Theorem \ref{thm:SBV-p-lap}$(i)$ and $(ii)$.
We omit the details.
\end{proof}
\begin{rem}
\label{SBV-1-rem}
In two dimensions, when $\|\text{\bf{b}}\|_{L^\infty(\Omega)}$ is small, depending on $\Omega, \psi$ and $\psi$, the conclusions of Theorem \ref{SBV-1} still hold. Indeed, in this case, by Remark \ref{lem51_rem}, we still have the $L^\infty$ estimate in Lemma \ref{lem:bound-u-high}. The proof of Theorem \ref{SBV-1} then follows.
\end{rem}
\section{Proof of Theorem \ref{SBV-2}}
\label{SBV-2-pf}
In this section, we will prove Theorem \ref{SBV-2}.
As in the proof of Theorem \ref{thm:SBV-p-lap}, it suffices to derive the \textit{a priori} estimates for $W^{4,p}(\Omega)$ solutions. Here, we recall that
\[p>2.\]
Theorem \ref{SBV-2} can be deduced from the following Theorem \ref{thm:W4p-est}.
\begin{thm}[A priori $W^{4,p}(\Omega)$ estimates for $W^{4,p}(\Omega)$ solutions]
\label{thm:W4p-est}
Let $\Omega\subset\mathbb R^2$, $\varphi$, $\psi$, $\text{\bf{b}}$ and $c$ be as in Theorem \ref{SBV-2}. Assume that $u\in W^{4,p}(\Omega)$ is a uniformly convex solution to \eqref{Abreu-lap}. Then
\[
\|u\|_{W^{4,p}(\Omega)}\leq C,
\]
where $C>0$ is a constant depending on $\Omega$, $p$, $\varphi$, $\psi$, $\text{\bf{b}}$ and $c$.
\end{thm}
The rest of this section is devoted to the proof of Theorem \ref{thm:W4p-est}.
We will first obtain the $L^\infty$-bound of $u$ and $L^2$-bound of $Du$. For this, the following Poincar\'e type inequality will be useful.
\begin{lem}[Poincar\'e type inequality on planar convex domains]
\label{lem:Poincare}
Let $\Omega\subset\mathbb R^2$ be an open, smooth, bounded and uniformly convex domain. Assume that $u\in C^1(\Omega)\cap C(\overline{\Omega})$ and $u|_{\partial\Omega}=\varphi$. Then
\[
\int_{\Omega}|u|^2\,dx\leq C(\varphi,\operatorname{diam}(\Omega))\|u\|_{L^\infty(\Omega)}+\frac{\operatorname{diam}(\Omega)^2}{16}\int_{\Omega}|Du|^2\,dx.
\]
\end{lem}
\begin{proof}
Note that for any one-variable function $f\in C^1(a,b)\cap C^0[a,b]$ where $a< b$, one has
\begin{equation}\label{eq:one-f}
\int_{a}^b |f(x)|^2\,dx\leq(b-a)(|f(a)|+|f(b)|)\|f\|_{L^\infty(a, b)}+\frac{(b-a)^2}{8}\int_{a}^b |f'(x)|^2\,dx.
\end{equation}
Indeed, denoting $c:=\frac{a+b}{2}$, then using H\"older's inequality and Fubini's theorem, one obtains
\begin{align}
\int_a^c|f(x)|^2\,dx&=\int_a^cf(a)(2f(x)-f(a))\,dx+\int_a^c\left(\int_a^x f'(t)\,dt\right)^2\,dx\nonumber\\
&\leq 2(c-a)|f(a)|\cdot\|f\|_{L^\infty(a, b)}-(c-a)f(a)^2+\int_a^c(x-a)\int_{a}^x |f'(t)|^2\,dtdx\nonumber\\
&=2(c-a)|f(a)|\cdot\|f\|_{L^\infty(a, b)}-(c-a)f(a)^2+\int_a^c|f'(t)|^2\int_t^c(x-a)\,dxdt\nonumber\\
&\leq 2(c-a)|f(a)|\cdot\|f\|_{L^\infty(a, b)}-(c-a)f(a)^2+\frac{(c-a)^2}{2}\int_{a}^c|f'(x)|^2\,dx\nonumber\\
&\leq (b-a)|f(a)|\cdot\|f\|_{L^\infty(a, b)}+\frac{(b-a)^2}{8}\int_{a}^c|f'(x)|^2\,dx.\label{eq:f-ac}
\end{align}
Similarly, we have
\begin{equation}\label{eq:f-cb}
\int_{c}^{b}|f(x)|^2\,dx\leq (b-a)|f(b)|\cdot\|f\|_{L^\infty(a, b)}+\frac{(b-a)^2}{8}\int_{c}^{b}|f'(x)|^2\,dx.
\end{equation}
Combining \eqref{eq:f-ac} with \eqref{eq:f-cb}, we obtain \eqref{eq:one-f}.
Next, by the convexity of $\Omega$, we can assume that there are $c, d\in\mathbb R$, and one-variable functions $a(x_1)$, $b(x_1)$,
such that
$$\Omega=\{(x_1,x_2):c<x_1<d, a(x_1)<x_2<b(x_1)\}.$$ It is clear that
$d-c\leq\operatorname{diam}(\Omega)$ and $b(x_1)-a(x_1)\leq\operatorname{diam}({\Omega})$. Then, by \eqref{eq:one-f} and $u=\varphi$ on $\partial\Omega$, we have
\begin{multline*}
\int_{a(x_1)}^{b(x_1)}|u(x_1,x_2)|^2\,dx_2\leq 2\operatorname{diam}({\Omega})\|\varphi\|_{L^\infty(\Omega)}\|u\|_{L^\infty(\Omega)}\\+\frac{\operatorname{diam}({\Omega})^2}{8}\int_{a(x_1)}^{b(x_1)}|D_{x_2}u(x_1,x_2)|^2\,dx_2.
\end{multline*}
Integrating the above inequality over $c<x_1<d$ yields
\begin{equation}\label{eq:int-D2u}
\int_{\Omega}|u|^2\,dx\leq 2\operatorname{diam}({\Omega})^2\|\varphi\|_{L^\infty(\Omega)}\|u\|_{L^\infty(\Omega)}+\frac{\operatorname{diam}({\Omega})^2}{8}\int_{\Omega}|D_{x_2}u|^2\,dx.
\end{equation}
Similarly,
\begin{equation}\label{eq:int-D1u}
\int_{\Omega}|u|^2\,dx\leq 2\operatorname{diam}({\Omega})^2\|\varphi\|_{L^\infty(\Omega)}\|u\|_{L^\infty(\Omega)}+\frac{\operatorname{diam}({\Omega})^2}{8}\int_{\Omega}|D_{x_1}u|^2\,dx.
\end{equation}
Combining \eqref{eq:int-D2u} and \eqref{eq:int-D1u}, we obtain
\[
\int_{\Omega}|u|^2\,dx\leq 2\operatorname{diam}({\Omega})^2\|\varphi\|_{L^\infty(\Omega)}\|u\|_{L^\infty(\Omega)}+\frac{\operatorname{diam}({\Omega})^2}{16}\int_{\Omega}|Du|^2\,dx,
\]
completing the proof of the lemma.
\end{proof}
\subsection{Estimates for $\sup_\Omega |u|$ and $\|Du\|_{L^2(\Omega)}$}
Now we derive bounds for $u$ and $\|Du\|_{L^2(\Omega)}$.
\begin{lem}[$L^\infty$ and $W^{1,2}$ estimates]\label{lem:bound-Du}
Let $\Omega\subset\mathbb R^2$, $\varphi$, $\psi$, $\textbf{b}$ and $c$ be as in Theorem \ref{SBV-2}. Assume that $u\in W^{4,p}(\Omega)$ is a uniformly convex solution to \eqref{Abreu-lap}. Then there exists a constant $C>0$ depending on $\Omega$, $\varphi$, $\psi$, $\text{\bf{b}}$ and $\|c\|_{L^1(\Omega)}$ such that
\[
\|u\|_{L^\infty(\Omega)}\leq C\quad\text{and}\quad\|Du\|_{L^2(\Omega)}\leq C.
\]
\end{lem}
\begin{proof}
To prove the lemma where $n=2$, by \eqref{eq:u-bound-high} and \eqref{eq:Du-L2bound-high}, it suffices to prove
\begin{equation}
\label{unuL2}
\int_{\partial \Omega}u_{\nu}^{2} d S \leq C(\Omega, \varphi, \psi, \text{\bf{b}}, \|c\|_{L^1(\Omega)}),
\end{equation}
where $\nu$ is the unit outer normal of $\partial\Omega$.
Let $\tilde u$ be as in the proof of Lemma \ref{lem:bound-u-high} so that $(i)$ and $(ii)$ there are satisfied. Let $K(x)$ be the Gauss curvature at $x \in \partial \Omega$. Then, as in (\ref{eq:K-u_nu-high}), we have, for some $C_1(\Omega,\varphi)>0$
\begin{equation}\label{eq:K-u_nu}
\begin{aligned}
\int_{\partial \Omega} K \psi u_{\nu}^{2} d S \leq & \int_{\Omega}\left(\Delta u-\text{\bf{b}}\cdot Du-c\right)(u-\tilde{u}) d x+C_1\Big(\int_{\partial \Omega}u_{\nu}^{2} d S\Big)^{1 / 2}+C_1
\end{aligned}
\end{equation}
Next, we will estimate the RHS of \eqref{eq:K-u_nu} term by term. First, from the inequality before last in \eqref{eq:1st-RHS}, we have
\begin{equation}
\int_{\Omega}\Delta u(u-\tilde{u})\,dx\leq
C(\Omega,\varphi)\left(\int_{\partial\Omega}u_{\nu}^2\,dS\right)^{\frac{1}{2}}-\int_{\Omega}|Du|^2\,dx.\label{eq:RHS-1st}
\end{equation}
Using $u=\varphi$ on $\partial\Omega$, and integrating by parts, we get
\begin{align}
\int_{\Omega}(\text{\bf{b}}\cdot Du)\tilde{u}\,dx&= \int_{\Omega}(\text{\bf{b}} \tilde u)\cdot Du\,dx\nonumber\\
&=\int_{\partial\Omega}u\tilde{u}\text{\bf{b}}\cdot \nu\,dS-\int_{\Omega}\div(\text{\bf{b}}\tilde{u})u\,dx\nonumber\\
&=\int_{\partial\Omega}\varphi\tilde{u}(\text{\bf{b}}\cdot\nu)\,dS-\int_{\Omega}(\text{\bf{b}}\cdot D\tilde{u}+\tilde{u}\,\div \text{\bf{b}}) u\,dx\nonumber\\&\leq C(1+ \|u\|_{L^{\infty}(\Omega)})\leq C_3 + C_3 \Big(\int_{\partial \Omega}u_{\nu}^{2} d S\Big)^{1 / 2},\label{eq:RHS-2nd-1}
\end{align}
where $C_3$ depends on $\Omega$, $\varphi$, $\sup_{\partial\Omega}|\text{\bf{b}}|$, $\|\text{\bf{b}}\|_{L^{\infty}(\Omega)}$ and $\|\div\text{\bf{b}}\|_{L^{\infty}(\Omega)}$.
Moreover,
\begin{align*}
\int_{\Omega}-(\text{\bf{b}}\cdot Du)u\,dx =\frac{1}{2}\int_{\Omega}-\text{\bf{b}}\cdot D(u^2)\,dx&=\frac{1}{2}\left[\int_{\Omega}(\div \text{\bf{b}})u^2\,dx-\int_{\partial\Omega}u^2\text{\bf{b}}\cdot\nu\,dS\right].
\end{align*}
Note that $\div \text{\bf{b}}\leq \frac{32}{\operatorname{diam}(\Omega)^2}$. Then by Lemma \ref{lem:Poincare} and \eqref{eq:u-bound-high}, we have
\begin{align*}
\frac{1}{2}\int_{\Omega}(\div \text{\bf{b}})u^2\,dx&\leq C(\varphi,\operatorname{diam}(\Omega))\|u\|_{L^\infty(\Omega)}+\int_{\Omega}|Du|^2\,dx\\
&\leq C(\Omega,\varphi)+C(\Omega,\varphi)\left(\int_{\partial\Omega}u_{\nu}^2\,dS\right)^{\frac{1}{2}}+\int_{\Omega}|Du|^2\,dx.
\end{align*}
Hence
\begin{align}
\int_{\Omega}-(\text{\bf{b}}\cdot Du)u\,dx &=\frac{1}{2}\left[\int_{\Omega}(\div \text{\bf{b}})u^2\,dx-\int_{\partial\Omega}\varphi^2\text{\bf{b}}\cdot\nu\,dS\right]\nonumber\\
&\leq C_4+C_4\left(\int_{\partial\Omega}u_{\nu}^2\,dS\right)^{\frac{1}{2}}+\int_{\Omega}|Du|^2\,dx,\label{eq:RHS-2nd-2}
\end{align}
where $C_4$ depends on $\Omega$, $\varphi$, and $\sup_{\partial\Omega}|\text{\bf{b}}|$.
Finally, as in \eqref{eq:3rd-RHS}, we get
\begin{equation}
\int_{\Omega}-c(u-\tilde{u})\,dx\leq C_5+C_5\left(\int_{\partial\Omega}u_{\nu}^2\,dS\right)^{\frac{1}{2}},\label{eq:RHS-3rd}
\end{equation}
where $C_5$ depends on $\Omega,\varphi$ and $\|c\|_{L^1(\Omega)}$.
Combining \eqref{eq:K-u_nu}--\eqref{eq:RHS-3rd}, we obtain
\begin{align*}
C^{-1}(\Omega)\inf_{\partial\Omega}\psi\int_{\partial\Omega}u_{\nu}^2\,dS&\leq\int_{\partial\Omega}K \psi u_{\nu}^{2} \,d S\leq C_6\left[1+\left(\int_{\partial\Omega}u_{\nu}^2\,dS\right)^{\frac{1}{2}}\right],
\end{align*}
where $C_6>0$ depends on $\Omega$, $\varphi$, $\psi$, $\text{\bf{b}}$ and $\|c\|_{L^1(\Omega)}$. From this, we deduce (\ref{unuL2}), completing
the proof of the lemma.
\end{proof}
\subsection{Hessian determinant estimates for $u$}
\begin{lem}[Hessian determinant estimates]\label{det-est-dim2}
Let $\Omega\subset\mathbb R^2$, $\varphi$, $\psi$, $\text{\bf{b}}$ and $c$ be as in Theorem \ref{SBV-2}. Assume that $u\in W^{4,p}(\Omega)$ is a uniformly convex solution to \eqref{Abreu-lap}. Then
\[
0<C^{-1}\leq \det D^2u\leq C \quad\text{in }\Omega,
\]
where $C>0$ is a constant depending on $\Omega$, $\varphi$, $\psi$, $\text{\bf{b}}$ and $\|c\|_{L^2(\Omega)}$.
\end{lem}
\begin{proof}
We first prove the lower bound of $\det D^2u$. Note that in two dimensions, we have $\text{trace } U=\Delta u$. Hence we can rewrite the first equation in \eqref{Abreu-lap} as
\begin{equation}\label{eq:re-Abreu}
U^{ij}D_{ij}\Big(w+|x|^2/2\Big)=\text{\bf{b}}(x)\cdot Du(x)+c(x):=Q(x)\quad\text{in }\Omega.
\end{equation}
By Lemma \ref{lem:bound-Du}, we have \[\|Q\|_{L^2(\Omega)}\leq C_0\]
where $C_0$ depends on $\Omega$, $\varphi$, $\psi$, $\text{\bf{b}}$ and $\|c\|_{L^2(\Omega)}$.
Applying the ABP estimate to \eqref{eq:re-Abreu} and using $\det U= \det D^2 u$, we have
\begin{eqnarray*}
\sup_{\Omega}\left(w+|x|^2/2\right)&\leq& \sup_{\partial\Omega}\psi+C(\Omega)+C(\Omega)\left\|\frac{Q}{(\det U)^{1/2}}\right\|_{L^2(\Omega)}\\
&\leq& C(\Omega,\psi)+C(\Omega)\|Q\|_{L^2(\Omega)}\cdot\sup_{\Omega}(\det D^2u)^{-\frac{1}{2}}\\
&\leq& C(\Omega,\psi)+C(\Omega)(\sup_{\Omega}w)^{\frac{1}{2}}.
\end{eqnarray*}
Therefore
$\sup_{\Omega}w\leq C_1$,
where $C_1$ depends on $\Omega$, $\varphi$, $\psi$, $\text{\bf{b}}$ and $\|c\|_{L^2(\Omega)}$. Consequently,
\begin{equation}\label{eq:lower-bound-d}
\det D^2u\geq C_1^{-1}>0\quad\text{in }\Omega.
\end{equation}
Hence by the boundary H\"older estimate for solutions of non-uniformly elliptic equations \cite[Proposition 2.1]{Le1}, we know from (\ref{eq:re-Abreu}) that $w$ is H\"older continuous on $\partial\Omega$ with estimates depending only on $C_1$, $\Omega$ and $\psi$. Then by constructing a suitable barrier near the boundary as in \cite[Lemma 2.5]{Le2}, we can obtain
\[\|Du\|_{L^\infty(\Omega)}\leq C_2,\]
where $C_2$ depends on $C_1$, $\Omega$, $\varphi$ and $\psi$.
The upper bound of the Hessian determinant can be obtained similar as in Lemma \ref{upper_lower}. Let $u^\ast(y)$ be the Legendre transform of $u(x)$ where \[y= Du(x)\in Du(\Omega):=\Omega^*.\] Then
\[\text{diam}(\Omega^*)\leq C_2.\]
By Proposition \ref{new2} (with $F(x)=|x|^2/2$), $u^\ast$ satisfies
\begin{equation}\label{eq:dual-eq}
U^{*ij}D_{ij}\Big(-w^{*}-|y|^2/2\Big)=-Q(Du^*)\det D^2u^*\quad\text{in }\Omega^\ast
\end{equation}
where $(U^{\ast ij})= (\det D^2 u^*) (D^2 u^*)^{-1}$, and $w^*=\log\det D^2u^*$.
Applying the ABP maximum principle to \eqref{eq:dual-eq}, and recalling that \[w^\ast(y) = \log (\det D^2 u(x))^{-1}=\log w(x) =\log \psi(x)\quad\text{on } \partial\Omega^\ast,\] we obtain
\begin{align*}
\sup_{\Omega^*}(-w^*-|y|^2/2)&\leq \sup_{\partial\Omega^*}(-w^*-|y|^2/2) +C(\text{diam}(\Omega^*))\|Q(Du^*)(\det D^2u^*)^{1/2}\|_{L^2(\Omega^*)}\\
&\leq-\log \min_{\partial\Omega}\psi+C(C_2)\|Q\|_{L^2(\Omega)},
\end{align*}
where we used
\[
\int_{\Omega^*}[Q(Du^*)]^2\det D^2u^*\,dy=\int_{\Omega}[Q(x)]^2\det D^2u^*\det D^2u\,dx=\int_{\Omega}[Q(x)]^2\,dx= \|Q\|_{L^2(\Omega)}^2.
\]
Therefore, we have
\[\sup_{\Omega^*}(-w^*)\leq C_3\]
where $C_3$ depends on $C_0$, $C_2$ and $ \min_{\partial\Omega}\psi$.
This implies $w^*\geq -C_3$ in $\Omega^*$, and hence
\begin{equation}\label{eq:upper-bound-d}
\det D^2u\leq e^{C_3}\quad\text{in }\Omega.
\end{equation}
The lemma follows from
(\ref{eq:lower-bound-d}) and (\ref{eq:upper-bound-d}).
\end{proof}
\subsection{Proof of Theorem \ref{thm:W4p-est}} Finally, we can prove Theorem \ref{thm:W4p-est} which implies Theorem \ref{SBV-2}.
\begin{proof}[Proof of Theorem \ref{thm:W4p-est}]
Once we have the determinant estimates, we can establish the higher estimates by using the regularity of the linearized Monge-Amp\`ere equation with drift terms as in Section \ref{SBV-plap-pf} and Section \ref{SBV-1-pf}.
However, in two dimensions, we can also establish these estimates as in \cite{LeCPAM}.
By Lemma \ref{det-est-dim2}, we have
\begin{equation}\label{det22}
0<\lambda\leq \det D^2u\leq \Lambda \quad\text{in }\Omega
\end{equation}
for $\lambda$, $\Lambda$ depending on $\Omega$, $\varphi$, $\psi$, $\textbf{b}$, $p$ and $\|c\|_{L^p(\Omega)}$. By the interior $W^{2,1+\varepsilon}$ estimates for Monge-Amp\`ere equation \cite{DFS,F, Sc}, we have $D^2 u\in L^{1+\varepsilon}_{loc}(\Omega)$ for some constant $\varepsilon(\lambda,\Lambda)>0$.
By the global $W^{2,1+\varepsilon}$ estimates for the Monge-Amp\`ere equation \cite{S3}, there exists a constant $C_0>0$ depending on $\Omega$, $\varphi$, $\psi$, $\textbf{b}$, $p$ and $\|c\|_{L^p(\Omega)}$ such that
\[
\|u\|_{W^{2,1+\varepsilon(\lambda,\Lambda)}(\Omega)}\leq C_0.
\]
Let $q:=\min\{p,1+\varepsilon(\lambda,\Lambda)\}>1$. Then \[G:= -\Delta u+\text{\bf{b}}\cdot Du+c\] satisfies
\[\|G\|_{L^q(\Omega)} \leq C_1\]
where $C_1>0$ depending on $\Omega$, $p$, $\varphi$, $\psi$, $\textbf{b}$ and $\|c\|_{L^p(\Omega)}$.
Recall that
\[
U^{ij}D_{ij} w=G\quad\text{on }\Omega,\quad\text{and } w=\psi \quad\text{on }\partial\Omega.
\]
By the global H\"older estimate for the linearized Monge-Amp\`ere equation \cite{LN2} with $L^{q}$ right-hand side where $q>n/2$, we deduce
\[
\|w\|_{C^\alpha(\overline{\Omega})}\leq C(\Omega, \varphi, \psi, p, \text{\bf{b}}, c)
\]
where $\alpha\in (0, 1)$ depends on $\Omega, \varphi, \psi, p, \text{\bf{b}}, c$. The proof of $W^{4, p}(\Omega)$ estimate for $u$ is now the same as that of
Theorem \ref{thm:SBV-p-lap}$(ii)$. Hence, the theorem is proved.
\end{proof}
\section{Extensions and the proof of Theorem \ref{LMA-G}}
\label{rem_sect}
In this section, we discuss (\ref{Abreu-sin}) with more general lower order terms,
and present a proof of Theorem \ref{LMA-G} for completeness.
\subsection{Possible extensions of the main results}
The following remarks indicate some possible extensions of our main results.
\begin{rem}
From the proofs in Sections \ref{SBV-plap-pf}-\ref{SBV-2-pf} and the $L^\infty$-estimates in Lemma \ref{lem:bound-u-high}, it can be seen that some conclusions of Theorems \ref{thm:SBV-p-lap}, \ref{SBV-1} and \ref{SBV-2}
also hold for more general cases of $c=c(x,z)$. Consider, for example,
\[c(x,z)=g_1(x)+g_2(x) h(z).\]
Then the following facts hold:
\begin{enumerate}
\item[(1)] The conclusions in Theorem \ref{thm:SBV-p-lap}(ii) hold when $g_1\leq 0, g_2\leq 0; g_1, g_2\in L^p(\Omega)$ with $p>n$, and $h\geq 0$ with $h\in C^\alpha(\mathbb R)$.
\item[(2)] The conclusions in Theorem \ref{SBV-1}(i) hold when $g_1, g_2\in C^\alpha(\overline\Omega)$, and $h\in C^\alpha(\mathbb R)$ with $|h(z)|\leq C |z|^{m}$ for $0\leq m<n-1$.
\item[(3)] The conclusions in Theorem \ref{SBV-1}(ii) hold when $g_1, g_2\in L^p(\Omega)$ with $p>2n$, and $h\in C^\alpha(\mathbb R)$ with $|h(z)|\leq C |z|^{m}$ for $0\leq m<n-1$.
\item[(4)] The conclusions in Theorem \ref{SBV-2} hold when $g_1, g_2\in L^p(\Omega)$ with $p>2$, and $h\in C^\alpha(\mathbb R)$ with $|h(z)|\leq C |z|^{m}$ for $0\leq m<1$.
\end{enumerate}
\end{rem}
\begin{rem} Since we use the trace of $\text{\bf{b}}$ on $\partial\Omega$ in
(\ref{eq:RHS-2nd-1}), it is natural to have $\text{\bf{b}}\in C(\overline{\Omega};\mathbb R^n)$. It would be interesting to obtain the conclusion of Theorem \ref{SBV-2} for $\text{\bf{b}}\in C(\overline{\Omega};\mathbb R^n)$ instead of $\text{\bf{b}}\in C^{1}(\overline{\Omega};\mathbb R^n)$.
\end{rem}
\subsection{Global H\"older estimates for pointwise H\"older continuous solutions at the boundary}
\label{sect_LOT_proof}
In this section, we prove Theorem \ref{LMA-G}.
The proof is similar to that of \cite[Theorem 1.4]{Le1} for the case without a drift. For completeness, we include the proof which includes the following ingredients: interior H\"older estimates for linearized Monge-Amp\`ere equations with bounded drifts, and rescalings using a consequence of the boundary Localization Theorem for the Monge-Amp\`ere equation which we will recall below.
Under the assumption $\lambda\leq\det D^2 u\leq \Lambda$, the linearized Monge-Amp\`ere operator $U^{ij}D_{ij}$ is elliptic, but it can be degenerate and singular in the sense that the eigenvalues of $U=(U^{ij})$ can tend to zero or infinity. To prove estimates for the linearized Monge-Amp\`ere equation that are independent of the bounds on the eigenvalues of $U$, as in \cite{CG} and subsequent works, we work with {\it sections} of $u$
instead with Euclidean balls. For a convex function $u\in C^{1}({\overline\Om})$ defined on the closure of a convex, bounded domain $\Omega\subset \mathbb R^n$,
the section of $u$ centered at $x\in \overline \Omega$ with height $h>0$ is defined by
\begin{equation*}
S_{u} (x, h) :=\Big\{y\in \overline \Omega: \quad u(y) < u(x) + Du(x) \cdot (y- x) +h\Big\}.
\end{equation*}
Before proving the global H\"older estimate, we recall the interior H\"older estimate.
The following interior H\"older estimate for the
nonhomogeneous linearized Monge-Amp\`ere equation with drift terms
is a simple consequence of the interior Harnack inequality proved in \cite[Theorem 1.1]{LeCCM}. In \cite{M}, Maldonado proved a similar Harnack's inequality for linearized Monge-Amp\`ere equation with drift terms with different and stronger conditions on $\text{\bf{b}}$.
\begin{thm}[Interior H\"older estimate for the
nonhomogeneous linearized Monge-Amp\`ere equation with drift terms, {\cite{LeCCM}}]
\label{inter-H}
Suppose that $u\in C^2(\Omega)$ is a strictly convex function in a bounded domain $\Omega\subset \mathbb R^n$ with
section $S_u(0, 1)$ satisfying
\[B_{r_1}(0)\subset S_{u}(0, 1)\subset B_{r_2}(0)\]
for some positive constants $r_1\leq r_2$, and
its Hessian determinant satisfying
\[ \lambda \leq \det D^2 u\leq\Lambda\quad\text{in }\Omega\]
where $\lambda$ and $\Lambda$ are positive constants. Let $(U^{ij})= (\det D^2 u)(D^2 u)^{-1}$. Let $\text{\bf{b}}: S_u(0, 1)\rightarrow\mathbb R^n$ be a vector field such that
$\|\text{\bf{b}}\|_{L^{\infty}(S_{u}(0, 1))} \leq M$.
Let $v \in W^{2,n}_{loc}(
S_{u}(0, 1))$ be a solution to
\[
U^{ij}D_{ij} v + \text{\bf{b}}\cdot Dv = f\text{ in }S_{u}(0, 1).\]
Then, there exist
constants $\beta_0, C >0 $ depending only $\lambda, \Lambda$, $n$, $r_1$, $r_2$, and $M$
such that
$$|v(x)-v(y)|\leq C|x-y|^{\beta_0}\Big(
\|v\|_{L^{\infty}( S_{u}(0, 1))} + \|f\|_{L^{n}( S_{u}(0, 1))} \Big),~\quad\text{for all }x, y\in S_{u}(0, 1/2). $$
\end{thm}
To bridge the interior H\"older estimates in Theorem \ref{inter-H} and the boundary H\"older estimates in (\ref{Bdr_assumed}), we need to control the shape of sections of the convex function $u$ that are tangent to the boundary $\partial \Omega$. The following proposition, proved by Savin in \cite{S3} (see also \cite[Proposition 3.2]{LS}), provides such a tool.
It is a consequence of the boundary Localization Theorem for the Monge-Amp\`ere equation, proved by Savin in \cite[Theorem 2.1]{S1} and \cite[Theorem 3.1]{S2}.
\begin{prop}[Shape of sections tangent to the boundary, {\cite{S3}}]\label{tan_sec}
Assume that $\Omega\subset\mathbb R^n$ is a uniformly convex domain with boundary $\partial\Omega\in C^3$.
Let $u \in C(\overline
\Omega)
\cap
C^2(\Omega)$ be a convex function satisfying
\[\lambda \leq \det D^2 u \leq \Lambda \quad \text{in}\quad \Omega\]
for some positive constants $\lambda$ and $\Lambda$. Moreover, assume that $u|_{\partial\Omega}\in C^3$.
Assume that for some $y \in \Omega$ the section $S_{u}(y, h) \subset \Omega$
is tangent to $\partial \Omega$ at some point $x_0\in\partial\Omega$, that is, $\partial S_{u}(y, h)\cap
\partial\Omega =x_0$, for some $h \le h_0(\lambda, \Lambda, \Omega, u|_{\partial\Omega}, n)$. Then there exists a small positive
constant $k_0$ depending on $\lambda$, $\Lambda$, $\Omega, u|_{\partial\Omega} $ and $n$ such that
$$k_0 E_h \subset S_{u}(y, h) -y\subset k_0^{-1} E_h, \quad \quad k_0 h^{1/2} \leq \text{dist}(y,\partial \Omega) \leq k_0^{-1} h^{1/2}, \quad $$
where $E_h= h^{1/2}A_{h}^{-1}B_1(0)$ is an ellipsoid with $A_h$ being a linear transformation with the following properties
\begin{equation*} \|A_{h}\|, \|A_{h}^{-1} \| \le k_0^{-1} |\log h|;~ \det A_{h}=1.
\end{equation*}
\end{prop}
Now, we are ready to prove Theorem \ref{LMA-G}.
\begin{proof}[Proof of Theorem \ref{LMA-G}]
By considering the equation satisfied by $(\|\varphi\|_{C^{\alpha}(\partial\Omega)} + \|f\|_{L^{n}(\Omega)})^{-1}v$,
we can assume that
$$\|\varphi\|_{C^{\alpha}(\partial\Omega)} + \|f\|_{L^{n}(\Omega)}=1,$$
and we need to show that
$$\|v\|_{C^{\beta}(\overline{\Omega})}\leq C(\lambda, \Lambda, n, \alpha, \Omega, u|_{\partial\Omega}, \gamma,\delta, K, M),$$
for some $\beta\in (0, 1)$ depending on $n,\lambda,\Lambda$, $\Omega, u|_{\partial\Omega}$, $\gamma$, and $M$.
{\it Step 1: H\"older estimates in the interior of a section tangent to the boundary.}
Let $y\in \Omega $ with \[r=r_y:=\text{dist} (y,\partial\Omega) \le c_1(n,\lambda,\Lambda,\Omega, u|_{\partial\Omega}),\] and consider the maximal interior section $S_u(y, h)$ centered at $y$, that is
\[h=h_y:=\sup\{t\,|~ S_u(y,t)\subset \Omega\}.\]
By Proposition \ref{tan_sec} applied at the point $x_0\in \partial S_u(y, h) \cap \partial \Omega,$ we can find a constant $k_0(n,\lambda,\Lambda,\Omega, u|_{\partial\Omega})>0$ such that
\begin{equation}
\label{hrk0}
k_0 h^{1/2} \leq r \leq k_0^{-1} h^{1/2},
\end{equation}
and $S_u(y,h)$ is equivalent to an ellipsoid $E_h$, that is,
$$k_0E_h \subset S_u(y, h)-y \subset k_0^{-1}E_h,$$
where
\begin{equation}E_h:= h^{1/2}A_{h}^{-1}B_1(0),~ \text{ with } \|A_{h}\|, \|A_{h}^{-1} \| \le k_0^{-1} |\log h|;~ \det A_{h}=1.
\label{Eh0}
\end{equation}
Let
$$T\tilde x:=y+h^{1/2}A_{h}^{-1}\tilde x.$$
We rescale $u$ by
$$\tilde u(\tilde x):=\frac {1}{ h} [u(T\tilde x)-u(y)-D u(y)\cdot (T\tilde x-y)].$$
Then
\[\lambda\leq \det D^2\tilde u(\tilde x)\leq\Lambda,\]
and
\begin{equation}
\label{normalsect0}
B_{k_0}(0) \subset \tilde S_1 \subset B_{k_0^{-1}}(0), \quad \quad \tilde S_1:= S_{\tilde u}(0, 1)=h^{-1/2} A_{ h}(S_u(y, h)- y).
\end{equation}
Define the rescalings $\tilde v$ for $v$, $\tilde \text{\bf{b}} $ for $\text{\bf{b}}$, and $\tilde g$ for $g$ by
$$\tilde v(\tilde x):= v(T\tilde x)- v(x_{0}),\quad \tilde \text{\bf{b}} (\tilde x) = h^{1/2} A_{ h} \text{\bf{b}}(T\tilde x), \quad \tilde g(\tilde x):= h g(T\tilde x),\quad \tilde x\in \tilde S_{1}.$$
Simple computations give
\[D \tilde v(\tilde x)= h^{1/2}(A_{h}^{-1})^{t} D v(T\tilde x), \]
\begin{equation*}
D^2 \tilde u(\tilde x)= (A_{h}^{-1})^{t} D^2 u(T\tilde x) A_{h}^{-1},\quad D^2 \tilde v(\tilde x)= h(A_{h}^{-1})^{t} D^2 v(T\tilde x) A_{h}^{-1},
\end{equation*}
and the cofactor matrix $\tilde U=(\tilde U^{ij})$ of $D^2 \tilde u$ satisfies
$$\tilde U(\tilde x): = (\det D^2 \tilde u) (D^2 \tilde u)^{-1} = (\det D^2 u) A_{h} (D^2 u)^{-1} (A_{h})^{t} = A_h U(T\tilde x) (A_h)^{t}.$$
Therefore, we find that
\begin{equation*}\tilde U^{ij} D_{ij} \tilde v= \mbox{trace}\, (\tilde U D^2 \tilde v)=h (U^{ij} D_{ij} v)(T\tilde x)~\text{ in }~\tilde S_1.
\end{equation*}
It is now easy to see that
$\tilde v$ solves
\[\tilde U^{ij} D_{ij}\tilde v + \tilde \text{\bf{b}} \cdot D\tilde v= \tilde g\quad \text{in }\tilde S_{1}.\]
Due to (\ref{Eh0}), and the smallness of $h$ (see (\ref{hrk0})), we have the following bound
\begin{equation*}
\|\tilde \text{\bf{b}}\|_{L^{\infty}(\tilde S_1)} \leq k_0^{-1}h^{1/2}|\log h|\cdot\|\text{\bf{b}}\|_{L^{\infty}(S_u(y, h))} \leq k_0^{-1}h^{1/2}|\log h| M \leq M.
\end{equation*}
Now, we apply the interior H\"older estimates in Theorem \ref{inter-H} to $\tilde v $ to obtain a small constant $\beta\in (0,1)$ depending only on $n, \lambda, \Lambda$, $k_0$, and $M$, such that
\begin{multline*}\abs{\tilde v (\tilde z_{1})-\tilde v(\tilde z_{2})}\leq C_1(n, \lambda, \Lambda, M)\abs{\tilde z_{1}-\tilde z_{2}}^{\beta}\Big \{\|\tilde v \|_{L^{\infty}(\tilde S_{1})} + \|\tilde g\|_{L^{n}(\tilde S_{1})}\Big\},\\ \text{for all } \tilde z_{1}, \tilde z_{2}\in \tilde S_{1/2}:= S_{\tilde u}(0, 1/2).\end{multline*}
By (\ref{normalsect0}), we can decrease $\beta$ in the above inequality if necessary, and thus assume that
$$2\beta\leq \gamma.$$
A simple computation using (\ref{Eh0}) gives
$$ \|\tilde g\|_{L^{n}(\tilde S_{1})} = h^{1/2}\|g\|_{L^{n}(S_u(y, h))}.$$
Moreover, from (\ref{hrk0}) and (\ref{Eh0}), we infer the following inclusions regarding sections and balls
\begin{equation}
\label{ballsecth2}
B_{c_2 \frac{r}{\abs{\log r}}}(y)\subset S_u(y,h/2) \subset S_u(y,h) \subset B_{C_2 r\abs{\log r}}(y),
\end{equation}
for some $c_2\in (0, 1)$ and $C_2>0$ depending on $n,\lambda,\Lambda,\Omega, u|_{\partial\Omega}$. We also deduce that
$$\text{diam} (S_u(y,h))\leq C(n,\lambda,\Lambda,\Omega, u|_{\partial\Omega})r\abs{\log r}\leq \delta$$
if
\[r\leq c_3(n,\lambda,\Lambda, \Omega, u|_{\partial\Omega}, \delta).\]
We now consider $r$ satisfying the above inequality.
By (\ref{Bdr_assumed}), we have
$$\norm{\tilde v }_{L^{\infty}(\tilde S_{1})} \leq K \text{diam} (S_u(y, h))^{\gamma} \leq C_3 (r\abs{\log r})^{\gamma},$$
where $C_3= C_3(n,\lambda,\Lambda,\Omega, u|_{\partial\Omega},\gamma, K)$.
Hence
\begin{equation*}\abs{\tilde v (\tilde z_{1})-\tilde v(\tilde z_{2})}\leq C_4\abs{\tilde z_{1}-\tilde z_{2}}^{\beta}\Big\{(r\abs{\log r})^{\gamma} + h^{1/2}\|g\|_{L^{n}(S_u
(y, h))}\Big\} \quad \text{for all }\tilde z_{1}, \tilde z_{2}\in \tilde S_{1/2}\end{equation*}
where $C_4=C_4(n,\lambda,\Lambda,\Omega, u|_{\partial\Omega},\delta, \gamma, K, M)$
Each $z\in S_u(y, h/2)$ corresponds to a unique $\tilde z= T^{-1}z\in \tilde S_{1/2}$.
Rescaling back, recalling $2\beta\leq\gamma$, and using
$\tilde z_1-\tilde z_2= h^{-1/2}A_{h}(z_1-z_2)$,
and the fact that
\begin{eqnarray*}\abs{\tilde z_1-\tilde z_2}&\leq& \| h^{-1/2}A_{h}\|\abs{z_1-z_2} \\&\leq& k_0^{-1} h^{-1/2}\abs{\log h}\abs{z_1-z_2}\leq
C_5(n,\lambda,\Lambda, \Omega, u|_{\partial\Omega}) r^{-1}\abs{\log r}\abs{z_1-z_2},\end{eqnarray*}
we find
\begin{equation}
\label{vbh2}
|v(z_1)-v( z_2)| \le |z_1-z_2|^{\beta} \quad \text{for all } z_1, z_2 \in S_u(y, h/2),
\end{equation}
provided that $r=r_y\leq c_3<1$ is small.
{\it Step 2: Global H\"older estimates.}
We now combine (\ref{vbh2}) with (\ref{Bdr_assumed}) and (\ref{ballsecth2}) to prove $$\|v\|_{C^\beta(\bar \Omega)} \leq C(n,\lambda,\Lambda,\Omega, u|_{\partial\Omega},\alpha,\delta, \gamma, K, M).$$
Indeed, as in (\ref{h-from-above}), there exists a constant $C_\ast(n,\lambda, M,\text{diam}(\Omega))$ such that
\begin{equation}
\label{vmax}
\|v\|_{L^{\infty}(\Omega)}\leq C_\ast.\end{equation}
It remains to estimate $\abs{v(x)-v (y)}/\abs{x-y}^{\beta}$ for $x$ and $y$ in $\Omega$. Let $r_{x} = \text{dist}(x, \partial\Omega)$ and $r_{y}= \text{dist} (y, \partial\Omega).$ Assume, without loss of generality, that
$r_{y}\leq r_{x}$. Take $x_{0}\in\partial\Omega$ and $ y_{0}\in \partial\Omega$ such that \[r_{x}= \abs{x- x_{0}} \quad\text{and }r_{y} = \abs{y-y_{0}}.\]
From (\ref{vmax}) and the interior H\"older estimates in Theorem \ref{inter-H}, we only need to consider the case $r_{y}\leq r_{x}\leq c_3<1$. Consider the following cases.\\
{\it Case 1:
$\abs{x-y}\leq c_2 \frac{r_{x}}{\abs{\log r_{x}}}.$}
In this case, by (\ref{ballsecth2}), we have $$y\in B_{c_2 \frac{r_{x}}{\abs{\log r_{x}}}}(x)\subset S_u(x,h_x/2),$$
where
\[h_x=\sup\{t\,|~ S_u(x,t)\subset \Omega\}.\]
In view of (\ref{vbh2}), we have
$$\frac{\abs{v(x)-v (y)}}{\abs{x-y}^{\beta}}\leq 1.$$
{\it Case 2:
$
\abs{x-y}\geq c_2 \frac{r_{x}}{\abs{\log r_{x}}}.$}
In this case, we have
\begin{equation}
\label{rxlogx}
r_{x}\leq c_2^{-1}\abs{x-y}\abs{\log \abs{x-y}}.
\end{equation}
Indeed, if
$$1>r_{x}\geq \abs{x-y}\abs{\log \abs{x-y}}\geq \abs{x-y} $$
then
$$r_{x}\leq \frac{1}{c_2}\abs{x-y}\abs{\log r_{x}}\leq \frac{1}{c_2}\abs{x-y}\abs{\log \abs{x-y}}.$$
Due to (\ref{rxlogx}), we have
$$\abs{x_0-y_0}\leq r_{x} + \abs{x-y} + r_{y}\leq C_6(n,\lambda,\Lambda,\Omega, u|_{\partial\Omega}) \abs{x-y}\abs{\log \abs{x-y}}.$$
Therefore, by (\ref{Bdr_assumed}), $\|\varphi\|_{C^{\alpha}(\partial\Omega)} \leq 1$, and $2\beta\leq \gamma\leq \alpha$, we obtain
\begin{eqnarray*}\abs{v(x)-v (y)}&\leq& \abs{v(x)- v(x_{0})} +
\abs{v(x_{0})- v(y_{0})} + \abs{v(y_{0})- v(y)}\\ & \leq& C \left(r_{x}^{\gamma} + \abs{x_{0}- y_{0}}^{\alpha} + r_{y}^{\gamma}\right) \\ &\leq& C\left(\abs{x-y}\abs{\log \abs{x-y}}\right)^{\gamma}\leq C \abs{x-y}^{\beta},
\end{eqnarray*}
where $C=C(n,\lambda,\Lambda,\alpha, \Omega, u|_{\partial\Omega},\delta, \gamma, K, M)$. This gives an estimate for $\abs{v(x)-v (y)}/\abs{x-y}^{\beta}$ in {\it Case 2}.
The proof of the theorem is complete.
\end{proof}
|
\section{Introduction}
\label{sec:intro}
\par The axion, initially proposed as a solution to the strong CP problem \cite{Peccei,Weinberg}, is now one of the most well-motivated candidates of dark matter \cite{Preskill,Sikivie, Willy}. The axions are stable bosons, with large occupation numbers and can rethermalize through their gravitational interactions forming a Bose-Einstein condensate (BEC) \cite{BEC1,BEC2}. Owing to the large occupation number of the ground state, the BEC condensate has been treated classically as a localized, coherently oscillating clump called an axion star, if the kinetic pressure is balanced by gravity and axiton or oscillon, if it is balanced by self-interactions. \cite{axion_star1,axion_star2,axion_star3,axion_star4,axion_star5,Chavanis,Wilczek}.
\par The interaction of axions with electromagnetic field has been proposed as a venue for their detection \cite{axion_detection}. Of particular interest it the Primakoff effect \cite{Primakoff}, which is the interaction of an axion with a virtual photon to produce a real photon.
One possible "laboratory" that exploits this effect for the detection of the axion is that of neutron stars, where some of the strongest magnetic fields in nature can be found \cite{axion_neutron}.
\par The interaction between a dilute axion star (AS) and a neutron star (NS) has also been proposed to explain Fast Radio Bursts \cite{FRB1,FRB2,FRB3,FRB4}, where the magnetic field of the neutron star induces an electric field in the axion star which then accelerates electrons in the atmosphere of the neutron star \cite{FRB_axion_neutronstar_electron} or induces a dipole moment in neutrons on the upper upper core of the neutron star \cite{FRB_axion_neutronstar_neutron}. However, due to tidal forces, the AS will break apart in a time scale of the order of seconds, much longer than the millisecond duration of FRBs \cite{tidal}. It has been pointed out though that the axion to photon conversion in the collision of an \textit{dense} AS with a NS may still be able to explain FRBs \cite{FRB_dense,Bai}.
\par Recently, the electromagnetic power emitted by an AS-NS binary system was derived analytically \cite{ANS_binaries}. It was assumed that the AS is described by an exponential profile, that the axis of rotation of the NS is perpendicular to the AS-NS plane and that plasma effects were negligible due to the low frequency of the plasma. For circular orbits, it was found that the spectral flux density of the emitted radiation is modulated mainly by the high frequency spin of the NS. There is also significant enhancement of emitted radiation for elliptical orbits due to the periodic approach of the AS to the NS.
\par In this work, we include the plasma co-rotating with the NS through the equation $k=\sqrt{m^{a}-\omega^{2}_{p}}$, where $\omega_{p}$ is the plasma frequency. We also assume that the rotation axis of the neutron star has a tilt with respect to the the normal to the NS-AS plane, use a profile for the AS that is more accurate than the exponential in the case of attractive self-interactions \cite{Eby} and allows us to compare with previous works where a constant magnetic field was assumed \cite{dipole}, and finally explore the relevance of plasma effects when we consider an axion mass $m_{a} \sim \omega_{p}$.
\par The paper is organized as follows: In section 2, we review the interactions between the axion and electromagnetism. In section 3, we establish the profile of the axion star that we will be using and the magnetic fields of the pulsar. In section 4, we calculate the power emitted by the axion condensate and provide analytic formulae. In section 5, we condlude with some remarks for future research.
\section{Axion Electrodynamics}
The Lagrangian of the axion field couple to electromagnetism is:
\begin{equation}
\label{eqn:Lagrangian}
\mathcal{L}=\frac{1}{2}\partial_{\mu}\partial^{\mu}\varphi-\frac{1}{2}m^{2}_{\varphi}\varphi^{2}-\frac{1}{4} F_{\mu \nu}F^{\mu \nu}-\frac{g_{a\gamma}}{4}\varphi F_{\mu \nu}\tilde{F}^{\mu \nu}
\end{equation}
where $\phi$ is the scalar field of the axion and:
\begin{equation}
\label{eqn:Tensor}
F_{\mu \nu}=\partial_{\mu}A_{\nu}-\partial_{\nu}A_{\mu} \,,
\qquad
\Tilde{F^{\mu \nu}}=\frac{1}{2}\epsilon^{\mu \nu \rho \sigma} F_{\rho \sigma}
\end{equation}
\par Note that in the above Lagrangian we have neglected gravity and assumed that there are no free currents and charges. Since the occupation number of the axion field is huge, we may treat its field as a classical field. The equations of motion are:
\begin{subequations}
\label{eqn: Maxwell}
\begin{align}
(\Box + m^{2}_{a}) \varphi = & g_{a\gamma} \bf{E} \cdot \bf{B}
\\
\nabla\cdot \bf{E} & =\rho
\\
\nabla \times \bf{B} - \pdv{\bf{E}}{t} & = \bf{J}
\\
\nabla \cdot \bf{B} & =0
\\
\nabla \times \bf{E}+\pdv{\bf{B}}{t} & = 0
\end{align}
\end{subequations}
where the charge density $\rho$ and current density $\bf{J}$ are:
\begin{subequations}
\begin{align}
\rho &= -g_{a\gamma}\nabla\varphi \cdot \textbf{B} \\
\textbf{J} & = g_{a \gamma} (\pdv{\varphi}{t} \textbf{B}+\nabla \varphi \times \textbf{E}) \label{eqn: current}
\end{align}
\end{subequations}
\par This is a set of coupled, partial differential equations. To make progress, we follow \cite{dipole} and expand the electromagnetic fields, charge density and current density in the small parameter $g_{a\gamma}\varphi_{0}$:
\begin{subequations}
\begin{align}
\bf{E}=\bf{E}^{(0)}+\bf{E}^{(1)}+...,
\qquad
\bf{B}=\bf{B}^{(0)}+\bf{B}^{(1)}+...
\end{align}
\end{subequations}
where $\bf{E}^{(0)}$ and $\bf{B}^{(0)}$ are the background fields that mix with the axion field. The 2 dynamical Maxwell equations for the first order electromagnetic fields are:
\begin{subequations}
\begin{align}
\nabla \cdot \textbf{E}^{(1)} & = \rho^{(1)},
\\
\nabla \times \textbf{B}^{(1)}-\pdv{\textbf{E}^{(1)}}{t} & = \textbf{J}^{(1)}
\end{align}
\end{subequations}
\par Introducing now the vector and scalar potential through the standard equations $\textbf{B}^{(1)}=\nabla \times \textbf{A}$, $\textbf{E}^{(1)}=-\nabla A^{0} - \pdv{\bf{A}}{t}$, our problem is reduced to calculating $A^{0}$ and $\textbf{A}$ from the equations:
\begin{subequations}
\begin{equation}
\label{eqn:wave equation potential}
(-\nabla^{2} + \frac{\partial^{2}}{\partial t^{2}} )A^{0} = \rho^{(1)} \\
\end{equation}
\begin{equation}
\label{eqn:wave equation vector}
(-\nabla^{2} + \frac{\partial^{2}}{ \partial t^{2}} )\textbf{A} = \textbf{J}^{(1)}
\end{equation}
\end{subequations}
\par We now turn to the field of the AS and the background fields $\textbf{E}^{(0)}$ and $\textbf{B}^{(0)}$ of the NS.
\section{Axions stars and Neutron stars}
\subsection{Axion stars}
\par For the axion star, we use the Ansatz proposed in \cite{Schiappacasse_2018} for a spherical symmetric star, oscillating coherently with frequency $\omega$:
\begin{equation}
\label{eqn:profile 1}
\varphi=\varphi_{0}f(r) cos(\omega t)
\end{equation}
\par For the non-relativistic case that we examine here, we may set $\omega= m_{a}$.
The spatial profile $f(r)$ must satisfy the boundary condition $ f(r) \xrightarrow[]{} 0$ when $r \xrightarrow[]{} \infty$ and have an inverted parabola at $r=0$. We will consider the profile:
\begin{equation}
\label{eqn:profile 2}
f(r) = \text{sech}(r/R)
\end{equation}
\par Although we should include a polynomial that captures the behavior of the field at $r \xrightarrow[]{} 0$, these profiles give us the correct qualitative behavior of the field.
\\
It was found in \cite{Wilczek}, where the potential considered was that of the QCD axion:
\begin{equation}
\label{eqn:potential}
V(\varphi)=\Lambda^{4} (1-cos(\varphi/f_{a}))
\end{equation}
that if $|\varphi_{0}| \ll f_{a}\, (\frac{10^{-6} eV}{m_{a}})$, where $f_{a}$ is the axion decay constant, we are in the diluted, stable branch where $R \omega \gg 1$ and we can ignore the self-interactions. In this regime, the axion star is supported by its kinetic pressure acting against gravity. Its typical mass and radius is \cite{Chavanis,Wilczek}:
\begin{align}
\label{eqn: dilute AS}
M_{dilute} & \leq 1.0 \times 10^{-9} M_{\odot} \left( \frac{f_{a}}{10^{13} \,\text{GeV}} \right) \left( \frac{10^{-6}\, \text{eV}}{m_{a}} \right) \\
R_{dilute} & = 27 km \left( \frac{10^{13} \, \text{GeV}}{f_{a}} \right) \left( \frac{10^{-6} \, \text{eV} }{m_{a}} \right)
\end{align}
If $|\varphi_{0}|=\mathcal{O}(1) f_{a} $, we are in the dense branch where $R \geq$few$ \times \omega^{-1}$ and we have to include the self-interations, while the non-relativistic approximation breaks down and our Ansatz \ref{eqn:profile 1} does not capture the correct profile of the oscillon. In addition, its lifetime was determined to be short, of the order $ \mathcal{O}(\frac{10^{3}}{m})$ and therefore not relevant to cosmological investigations. The radius of a dense axion star is given by \cite{axion_star1}:
\begin{equation}
\label{eqn: dense AS}
R_{dense} = 3 m \times \left( \frac{10^{13} \, \text{GeV}}{f_{a}} \right)^{1/2} \left( \frac{10^{-6} \, \text{eV}}{m_{a}} \right)^{1/2} \left( \frac{M_{dense}}{10^{-12} M_{\odot}} \right)^{0.3}
\end{equation}
\par In \cite{decay, Oscilon_dm}, in addition to \ref{eqn:potential}, a number of other potentials were studied and the lifetime of oscillons was found to be up to 5 orders of magnitude larger. However, they arrived to similar results regarding the single frequency approximation i.e. we need a more sophisticated profile than \ref{eqn:profile 1} to study dense oscillons-axion stars. In this work, we will assume that the profile in \ref{eqn:profile 1} is also correct for dense axion stars and leave a multiple frequency profile for future work.
\subsection{Neutron Stars}
\par We are modelling the neutron star as a rotating magnetic dipole of angular frequency $\Omega$ surrounded by co-rotating plasma. In the near zone $r \ll \frac{1}{\Omega}$ the magnetic field is given by:
\begin{equation}
\textbf{B}_{NS}=\frac{B_{0}}{2} \left(\frac{R_{NS}}{r} \right)^{3} (3(\hat{m} \cdot \hat{r}) \hat{r} - \hat{m})
\end{equation}
where $R_{NS}$ is the neutron star's radius, $B_{0}$ is the intensity of the magnetic field on the neutron star's surface and $\hat{m}$ is the unit vector in the direction of the magnetic moment. The number density of the electrons was derived by Goldreich and Julian assuming perfect conductivity of the plasma \cite{Goldreich}:
\begin{equation}
n_{c}=\frac{2 \bf{\Omega} \cdot \textbf{B}}{e} \frac{1}{1-\Omega^{2} r^{2} \sin^{2} \theta}
\end{equation}
while the plasma frequency is given by:
\begin{equation}
\omega_{p}=\sqrt{\frac{4 \pi \alpha n_{c}}{m_{e}}} \approx 10^{-5} \text{eV} \sqrt{\frac{\hat{\Omega} \cdot \textbf{B}}{10^{12} G} \frac{1 s}{T_{s}}}
\label{eqn: plasma freq}
\end{equation}
where $T_{s}$ is the period of the NS.
\begin{figure}
\centering
\includegraphics{drawing.png}
\caption{Schematic of the NS showing the different angles. The rotation axis makes a $\tilde{\theta}$ angle with the z axis and the magnetic dipole moment makes a constant $\gamma$ angle with the rotation axis, while precessing around it with $\Omega$ angular frequency}
\label{fig:xi=0}
\end{figure}
\par We now turn to the rotation axis of the NS. For simplicity, we will assume that it lies on the y-z plane, at an angle $\tilde{\theta}$ from the z axis:
\begin{equation}
\label{eqn: rotation axis}
\hat{\Omega}=
\begin{pmatrix}
0 \\ \sin\tilde{\theta} \\ \cos \tilde{\theta} \\
\end{pmatrix}
\end{equation}
We can now perform two successive rotations using SO(3) matrices to obtain the magnetic moment unit vector $\hat{m}$ that precesses around the rotation axis. One is around an axis perpendicular to $\hat{\Omega}$ by an angle $\gamma$ and the next one is around $\hat{\Omega}$ by an angle $\Omega t$. The details are given in Appendix \ref{appendix A}. Here, we simply write down the final answer:
\begin{equation}
\label{eqn:magnetic moment}
\hat{m}=
\begin{pmatrix}
\cos(\Omega t) \sin \gamma \\ \cos \gamma \sin \tilde{\theta} + \sin \gamma \cos \tilde{\theta} \sin(\Omega t) \\ \cos \gamma \cos \tilde{\theta} -\sin \gamma \sin \tilde{\theta} \sin (\Omega t)
\end{pmatrix}
\end{equation}
\par Furthermore, we argue that we can ignore the background electric field in \ref{eqn: current}. Its value at the surface of the star is $E_{0}$=$B_{0} \Omega a$ and in order to ignore the second term in the right hand side of \ref{eqn: current}, it must hold:
\begin{equation}
\nabla \varphi \times \textbf{E} \ll \pdv{\varphi}{t} \textbf{B} \Rightarrow \frac{\textbf{E}_{0}}{R} \ll m_{a} \textbf{B}_{0} \Rightarrow{} \frac{\Omega R_{NS}}{c} \ll m_{a}R
\end{equation}
\par For typical values of $R_{NS}$=10 km, $\Omega=10^{-3}-1 s$, this inequality is true for the range of $m_{a}R$ that we are interested in.
\subsection{AS-NS binary}
Assuming that $M_{NS} \gg M_{AS}$, we can place the NS in the origin of our coordinate system and the AS will be orbiting around it. The AS's motion is described by the familiar equation for elliptical orbits:
\begin{equation}
r_{a}=\frac{(1-e^{2})r_{0}}{1+e \cos \phi_{a}} ,\ r_{0}=\frac{l^{2}_{0}}{G M_{NS}(1-e^{2})}
\end{equation}
where $r_{0}$ is the semi-major axis, $l_{0}$ is the angular momentum per unit mass, $\phi_{a}$ is the azimuthal angle of the AS and $e$ is the eccentricity. We will restrict our analysis to the cases where $e=0$ and $0 < e < 1$, to circular and elliptical orbits respectively. Since the AS lies on the $x-y$ plane, its position vector is:
\begin{equation}
\textbf{r}_{a}=r_{a}(\cos \phi_{a}, \sin \phi_{a},0)
\end{equation}
Later on, we will also need the time dependence of the angle $\phi_{a}$ which is given by the parametric equations \cite{Goldstein}:
\begin{equation}
\omega t = \psi - e \sin \psi ,\hspace{0.2 cm} \cos \phi_{a} = \frac{\cos \psi - e}{1 - e \cos \psi}
\end{equation}
where $0 \leq \psi \leq 2 \pi$ and $\omega$ is the angular frequency of the AS's rotation around the NS. The period of rotation is $T_{AS}=\frac{2 \pi }{\omega}$, which we will assume to be 10s.
\par. We will also assume for simplicity that Earth's position vector lies on the z axis.
\begin{table}
\begin{center}
\begin{tabular} {| c | c | c | c |}
\hline
$M_{NS} (M_{\odot})$ & $M_{dense} (M_{\odot})$& $R_{NS} (km)$ & $r_{0} (km)$ \\
\hline
1.4 & $10^{-12}$ & 15 & $7.8 \times 10^{3}$ \\
\hline \hline
$\gamma (rad)$ & $T_{AS} (s)$ & $T_{s} (s)$ & $m_{a} (\text{eV})$\\
\hline
0.3 & 10 & 1 & $10^{-6}$ \\
\hline \hline
$f_{a} (\text{GeV})$ & $c_{\gamma}$ & $r_{e} (kpc)$ & $B_{0} (G)$\\
\hline
$10^{14}$ & $10^{5}$ & 1 & $10^{12}$ \\
\hline
\end{tabular}
\caption{A table with the relevant constants of our system}
\label{tab: table 1 }
\end{center}
\end{table}
\par Table \ref{tab: table 1 } summarizes the constants relevant to our system, which we take to be the same as in \cite{ANS_binaries}, except for $f_{a}$ which we take to be $f_{a}=10^{13} \text{GeV}$, so that we are consistent with the equation $m_{a} f_{a}= m_{\pi} f_{\pi} = (10^{8} \text{eV})^{2}$ that must be true for the QCD axion.
\section{Analytic calculation of emitted power}
\par The radiating fields are given by the known equations:
\begin{align}
\label{eqn:radiating fields}
\textbf{B} & =\nabla \times \textbf{A}, \\
\textbf{E} & =i \omega \textbf{A}-\nabla A^{0}=\frac{i}{\omega} \left(\omega^{2} \textbf{A} - \textbf{k} (\textbf{k} \cdot \textbf{A}) \right)
\end{align}
where we have taken the Fourier transform in time and used the Lorentz gauge $\nabla \cdot \textbf{A}+\partial_{0}A^{0}=0$.It is therefore sufficient to solve \ref{eqn:wave equation vector}. Using the retarded Green's function, the solution is:
\begin{equation}
\textbf{A} (\textbf{r}_{e}) =\frac{g_{a \gamma} \varphi_{0} \omega \textbf{B}_{NS}}{4 \pi r_{e}}\int d^{3}r' e^{ik |\textbf{r}_{e} -\textbf{r}'|} \, f(|\textbf{r}'-\textbf{r}_{a}|)
\label{eqn: vector integral}
\end{equation}
The magnetic field varies negligibly within the AS, so we can take it out of the integral. We will only write the final result of this integration here and leave the details for Appendix \ref{Appendix B}.
\begin{equation}
\textbf{A} (\textbf{r}_{e}) \approx \frac{g_{a \gamma} \varphi_{0} \omega (\pi R)^{2} \textbf{B}_{NS}(\textbf{r}_{a}) }{4 k r_{e}} \frac{\tanh(\pi k R/2)}{\cosh(\pi k R/2)} e^{i k|\textbf{r}_{e}-\textbf{r}_{a}|}
\label{eqn: vector solution}
\end{equation}
The time averaged Poynting vector is given by:
\begin{equation}
\label{eqn:Poynting}
\langle \textbf{S} \rangle = \frac{1}{2} \Re(\textbf{E} \times \textbf{B}^{\star})=\frac{1}{2 \omega} \Re \left( \textbf{k}(\omega^{2}|\textbf{A}|^{2}-|\textbf{k} \cdot \textbf{A}|^{2}) - \omega^{2}_{p} (\textbf{k} \cdot \textbf{A}) \textbf{A}^{\star} \right)
\end{equation}
Inserting in \ref{eqn:Poynting} the expression in \ref{eqn: vector solution}, we find for the Poynting vector and the power per solid angle :
\begin{equation}
\label{eqn: poynting solution}
\langle \textbf{S} \rangle = \frac{\pi^4 \omega}{32 r^{2}_{e}} \left( \frac{g_{a\gamma}\varphi_{0} R^{2}}{k} \right)^{2} \left( \frac{\tanh(\pi k R/2)}{\cosh(\pi k R/2)}\right)^{2} \left(\textbf{k}(\omega^{2}|\textbf{B}_{NS}|^{2} - (\textbf{k} \cdot \textbf{B}_{NS})^{2}) - \omega^{2}_{p} \textbf{B}_{NS}(\textbf{k} \cdot \textbf{B}_{NS}) \right)
\end{equation}
\begin{equation}
\label{eqn: solid angle }
\frac{dP}{d \Omega} = r^{2}_{e} \hat{k} \cdot \langle \textbf{S} \rangle = \frac{\pi^{4} (g_{a \gamma} \varphi_{0} \omega ^{2} R^{2})^{2}}{32 k \omega} \left( \frac{\tanh (\pi k R /2)}{\cosh (\pi k R /2)} \right)^{2} |\textbf{B}_{NS}|^{2} \sin^{2} \beta
\end{equation}
where $\beta$ is the angle between the magnetic field at the AS's position and Earth's position vector from the AS. It comes as no surpirse that this is the same expression as the one derived in \cite{dipole}, where a constant magnetic field was assumed.
\begin{figure}
\label{fig: power}
\centering
\includegraphics{power.png}
\caption{Plot of the function $\left( \frac{\tanh (\pi \omega R /2)}{\cosh (\pi \omega R /2)} \right)^{2}$ versus $\omega R$. The function has a peak at $\omega R$=0.56 and is exponentially suppressed for $\omega R \gg 1$}
\end{figure}
.
\subsection{AS-NS binary in vacuum}
\par For the benchmark values considered in Table \ref{tab: table 1 }, $\omega_{p} \sim 10^{-9} \text{eV}$, so we can neglect plasma effects for the moment, since $m_{a} \sim 10^{-6} \text{eV}$. The power per solid angle, equation \ref{eqn: solid angle }, is exponentially suppressed with respect to $\omega R$, which means that only dense axion stars can radiate efficiently in vacuum. We can understand this by considering energy-momentum conservation in the axion to photon conversion process \cite{Review}. Let us assume that the magnetic field has only spatial variation, $\textbf{B}_{0}(\textbf{x})$. Energy conservation is satisfied because $\textbf{B}_{0}(\textbf{x})$ is time-indepedent, while its fourier transform guarantees momentum conservation, since the axion and the photon have different dispersion relations. In our case, we assumed that the magnetic field is constant, therefore momentum conservation is not guaranteed and the power is suppressed. \par We may define the spectral flux density as $\mathcal{S}=\frac{dP}{d\Omega}\left(r^{2}_{e} \mathcal{B}\right)^{-1}$, where $\mathcal{B}=\frac{0.1 m_{a}}{2 \pi}$ is an estimate for the Doppler width. Inserting the benchmark values we are considering, the result is:
\begin{align}
\label{eqn: spectral dense}
S & = 2.36 \times 10^{-5} \text{Jy} \left( \frac{c_{
\gamma}}{10^{5}} \right)^{2} \left( \frac{10^{13} \text{GeV}}{f_{a}} \right)^{2} \left( \frac{10^{-6} \text{eV}}{m_{a}} \right) \left( \frac{M_{dense}}{10^{-12} M_{\odot}}\right)^{1.2} \left( \frac{B_{0}}{10^{12} \text{GeV}} \right)^{2} \times \nonumber \\ & \times \left( \frac{1 kpc}{r_{e}} \right)^{2} \sin^{2} \beta \left( \frac{1+e\cos \phi_{a}}{1-e^{2}} \right)^{6} |3(\hat{m} \cdot \hat{r})\hat{r} - \hat{m}|^{2}
\end{align}
\begin{figure}[h]
\centering
\includegraphics{dense_e_0.png}
\caption{Spectral flux density versus time for e=0. The blue color represents the rotation axis on the z axis, while orange and green represent the rotation axis at $\tilde{\theta}= \frac{\pi}{4}$ and $\tilde{\theta}= \frac{\pi}{2}$ respectively}
\label{fig: e=0}
\end{figure}
In Figure \ref{fig: e=0}, we plot the spectral flux density versus time for three different $\tilde{\theta}$ angles. For $\tilde{\theta}=0$, our results are similar to what was found in \cite{ANS_binaries} i.e., the signal is modulated by the rotation of the AS around the NS and by the spinning of the NS around its axis. The envelope at the top is approximately flat because the separation between the AS and NS does not change with time. This picture changes if we allow the NS's rotation axis to be tilted. The position of the AS now plays a more important role in modulating the signal, as we can see from the peaks and troughs in the plot that have a period of 5 seconds. The locations of these maxima and minima can be explained if we remember that the AS starts its orbit on the x axis, where it receives the minimum amount of magnetic field lines from the NS, since the latter's rotation axis is now on the y-z plane. At 2.5 seconds, the AS has made one quarter of the orbit and it is found on the y-axis, where it now receives the maximum number of magnetic field lines. At 5 seconds, it is on the $x$ axis again, where it receives the minimum amount and so on. We would expect that $S$ would be larger the closer the NS rotation axis is to the x-y plane. This also confirmed by Figure \ref{fig: e=0}, where we see that as $\tilde{\theta}$ increases, $S$ also increases.
\begin{figure}[h]
\centering
\includegraphics{dense_e_0.5.png}
\caption{Spectral flux density versus time for e=0. The blue color represents the rotation axis on the z axis, while orange and green represent the rotation axis at $\tilde{\theta}= \frac{\pi}{4}$ and $\tilde{\theta}= \frac{\pi}{2}$ respectively}
\label{fig: e=0.5}
\end{figure}
In the case where $e=0.5$, Figure \ref{fig: e=0.5}, the situation is different, because now the separation between the NS and the AS changes periodically with time. The orbit stars from the perihelion, where the signal is maximum. At 5 seconds, it is on the aphelion, where it is minimum and so on. As it was pointed out in \cite{ANS_binaries}, this is easily explained if we remember that $|\textbf{B}_{NS}|^{2} \sim \frac{1}{r^{6}}$, so there is significant enhancement to the signal if the AS is close to the NS. Hence, the modulation due to the NS spinning is not as important. This is confirmed by our results when we allow the NS's rotation axis to be tilted. In Figure \ref{fig: e=0.5}, it is evident that $S$ is slightly enhanced for the cases $\tilde{\theta}= \frac{\pi}{4}, \frac{\pi}{2}$, but not by much compared to the $\tilde{\theta}=0$ case.
\subsection{AS-NS binary in plasma}
We would like to investigate how the plasma surrounding the NS affects the emitted power. To that end, we consider a lower axion mass, of the order $m_{a} = 10^{-9} - 10^{-8} \text{eV}$. This is outside of the QCD axion range, but it is valid for ALPs, where $f_{a}$ and $m_{a}$ are independent. We will keep the value of $f_{a}=10^{13} \text{GeV}$ for the axion decay constant.
\par It was shown in \cite{dipole} that when $\omega \rightarrow \omega_{p}$, the emitted power per solid angle becomes:
\begin{equation}
\frac{dP}{d\Omega}({\omega \rightarrow \omega_{p}}) \approx \frac{(g_{a \gamma} \varphi_{0})^{2}}{128} \frac{(\pi \omega R)^{6}}{\omega^{2}} \left(1-\frac{\omega_{p}^{2}}{\omega^{2}}\right)^{1/2} |\textbf{B}_{NS}|^{2} \sin^{2} \beta
\end{equation}
There is significant enhancement in power now if $\omega R \gg 1$,due to the factor $(\omega R)^{6}$. Therefore dilute axion stars can radiate much more efficiently than dense axion stars. Inserting in \ref{eqn: dilute AS} $m_{a}=10^{-9} \text{eV}$, we find for the radius $R_{AS,dilute}=27000 \text{km}$ and $M_{AS,dilute} \approx 10^{-6} M_{\odot}$. The Roche limit for this AS is then $R_{Roche}=R_{AS,dilute} \left(\frac{2 M_{NS}}{M_{AS.dilute}} \right)^{1/3} \approx 38 \times 10^{5} \text{km} $, much larger than the semi-major axis $r_{0}=7.8 \times 10^{3} \text{km}$. Hence, even though dilute axion stars could theoretically radiate in the presence of plasma, they will be pulled apart by tidal forces and disintegrate. We tried to avoid this problem by increasing $r_{0}$. However, if we increase $r_{0}$, the plasma frequency, equation \ref{eqn: plasma freq}, becomes smaller. This means that we have to lower the axion mass too, increasing the AS radius, via equation \ref{eqn: dilute AS}, leading again to a Roche limit larger than $r_{0}$. It is an open question what happens to the remnants of this disintegrated AS which we leave for future work.
\section{Conclusions and Outlook}
In this paper, we have derived analytical expressions for the spectral flux density emitted by an AS that is gravitationally bound to a NS. We have used a $\text{sech}$ profile to describe the AS that captures its essential characteristics and took into account plasma effects using the Goldreich-Julian model for the NS. We have also allowed the rotation axis of the NS to be tilted towards the plane of the AS-NS, placing it on the y-z plane. If the orbit is circular, we find that there is an order of magnitude increase to the spectral flux density whenever the AS is on the y axis, because of the increased number of magnetic field lines that now pass through the neutron star. If, on the other hand, the orbit is an ellipse, there is no significant enhancement to the flux, since now the main modulation comes from the AS's orbit.
\par
We also considered the effect that the plasma may have on the flux. Given that in the presence of plasma only dilute stars can radiate efficiently, we lowered the axion mass to $m_{a} = 10^{-9} \text{eV}$ to investigate this possibility. We found that by lowering the mass, we increase the AS radius, which in turn increases the Roche limit by several orders of magnitude more than the AS-NS distance, meaning that the dilute AS will be torn apart by tidal forces before it interacts with the NS's magnetic field.
\par
Some interesting venues for future research would be to investigate the effect of back-reaction to the AS profile. This would probably require solving numerically the set of equations \ref{eqn: Maxwell}. Back-reaction could also shed some light to the time scale of disintegration of a dilute AS in the gravitational field of a neutron star or a black hole. Another interesting direction of research would be to include a multi-frequency profile for the dense AS. As it was mentioned in the introduction, given that there have been attempts to explain FRBs with dense axion stars, a more accurate profile for the AS would be a step in that direction.
\acknowledgments
I would like to thank David Sadek, Jose Alberto Ruiz Cembranos, Dennis Maseizik and especially Pierre Sikivie for useful discussions.
|
\section{Introduction}
Online advertising platforms serve personalized advertisements based on users' potential interests. For example, Baidu Search Ads (a.k.a. ``Phoenix Nest'') has been successfully using ultra-high dimensional input data and ultra-large-scale deep neural networks for training CTR (Click-Through Rate) models since 2013~\cite{fan2019mobius,zhao2019aibox,fei2021gemnn}. In an advertising platform based on eCPM (effective Cost Per Mille), the advertisements fed to a specific user are ranked by the product of the bid price offered by the advertisers and the user's predicted CTR from the CTR prediction model. Intuitively, it tends to put the advertisements with high predicted CTR and bid price at the top of the rank list to attract potential customers for advertisers and achieve high profit for the online advertising platform.
Since the predicted CTR has a substantial impact on the rank of the displayed ads, the deviation of the predicted CTR from the actual CTR has a significant influence on the revenue of advertisers and the advertising platform. If the predicted CTR of an ad is lower than the actual CTR, the ad might not get exposed to customers. In this case, the advertisers will not gain the expected revenues, and the advertising platform also loses the charges which it should have attained. On the other hand, if the predicted CTR of an ad is higher than the actual CTR, the ad will be misplaced at the top of the rank list but does not lead to the expected amount of clicks from customers. Then the advertisers will be overcharged by the advertising platform and cannot gain a reasonable profit. Thus, an effective CTR prediction model is a critical component for an eCPM advertising platform to achieve satisfactory advertising effects for the advertisers and abundant revenues for the advertising platform.
With the prompt progress achieved in machine learning in the past decade, we have witnessed the rapid evolution in the architecture of the CTR prediction model. The earliest works are mainly based on linear logistic regression (LR) model~\cite{lee2012estimating}, non-linear gradient boosting decision trees (GBDT)~\cite{friedman2001greedy,li2007mcrank,zheng2007regression,li2010robust,he2014practical,li2022package}, Bayesian models~\cite{graepel2010web} or factorization machines (FM)~\cite{rendle2010factorization,ta2015factorization,hu2018collaborative}. These models take a shallow structure and thus cannot effectively describe high-order latent patterns in the user-ad behavior. Inspired by the great success of deep learning in computer vision and natural language processing, researchers attempt to build deep neural networks (DNN) for CTR prediction. Factorization-machine supported neural network (FNN)~\cite{zhang2016deep} feeds the output of FM into a deep fully-connected~neural network. Convolutional Click Prediction Model (CCPM)~\cite{liu2015convolutional} predicts the CTR by a deep convolutional neural network. Zhang~{et al.}~\cite{zhang2014sequential} utilize the recurrent neural network to model the sequence of user behaviors for CTR prediction. The following works~\cite{cheng2016wide,qu2016product,guo2017deepfm,wang2017deep,wang2018dkn,lian2018xdeepfm,song2019autoint,zhou2019deep,huang2019fibinet,fan2019mobius,feng2019deep,ma2019hierarchical,zhao2019aibox,pi2020search,rendle2020neural,fei2021gemnn,xu2021agile} explore more advanced neural networks for more effectively modeling higher-order patterns. Although extensive efforts have been devoted to improving the CTR prediction on the model side, the data side in the CTR prediction has been relatively less exploited. The lack of training data is commonly encountered for many machine learning tasks, including CTR prediction model training.
The existing mainstream CTR prediction model is based on the behavior ID features based on the historical interactions between users and ads. Each ad, as well as each user, is assigned a unique behavior ID. Each behavior ID of a user/ad is mapped to a vector, termed as embedding vector. The ad embedding vectors and user embedding vectors are learned jointly based on the historical behaviors of users on ads. Since the user/ad embedding vectors are learned based on historical behaviors, the quality of the learned features is heavily dependent on the richness of users' past behaviors on ads. When a new ad is added to the advertising system, we have no access to the user-ad behavior on this new ad. It leads to the lack of the training data issue, and the behavior ID embedding for this new ad is not reliable at all for predicting the CTR of any user on the ad. The lack of training data for new ads is normally termed the cold-start problem.
A straightforward solution to solving the cold-start problem is using content-based features by understanding the visual content of ads~\cite{mo2015image,ge2018image}.
Specifically, Mo~\emph{et al.}~\cite{mo2015image} learn image features of display ads directly from raw pixels and user feedback in the target task. Ge~\emph{et al.}~\cite{ge2018image} exploits both ad image features and user behavior image features. Nevertheless, due to the high computational cost of image feature extraction, Mo~\emph{et al.}~\cite{mo2015image} and Ge~\emph{et al.}~\cite{ge2018image} adopt a pre-trained image feature extractor to get the image feature in the offline phase and then use the extracted image feature for predicting CVR/CTR in the online phase. Since the feature extractor is not related to CVR/CTR prediction in the training phase, the extracted image features might not be effective for CVR/CTR prediction. Zhao~\emph{et al.}~\cite{zhao2019you} devise a pre-ranking model. When an advertiser uploads ads, the offline pre-ranking model determines inferior and superior ads based on their visual content. Only the superior ads with attractive visual content will be fed into the online ranker for further ad display. Since the pre-ranking model is conducted in the offline phase, the inefficiency caused by complex feature extraction is no longer an issue, and it makes the end-to-end training of feature extractor feasible. The offline pre-ranking model is trained by the accumulated historical data from all users in a learn-to-rank manner. Since the training data is collected from all users, it reveals the preference of the majority of users and might not reveal the preference of a specific user for personalized advertising.
In this work, we seek to model the visual content in ads for boosting the performance CTR prediction effectively. We map the visual content of an ad into a set of visual indices (IDs) through supervised clustering.
After we obtain the visual IDs for ads, we learn a visual embedding vector for each visual ID based on the user's clicks on ads. The learned visual ID embedding vector for each ad can effectively describe the visual content in an ad and can well generalize to ads newly added to the advertising system. Meanwhile, the visual content encoded in the visual ID embedding vector is complementary to the ad behavior ID embedding vector, which is beneficial to CTR prediction for not only new ads but also the ads which have accumulated rich user behaviors.
In this case, as visualized in Figure~\ref{structure}, both ad behavior ID embedding and visual ID embedding are the input of the CTR prediction model, and they work together to achieve a more accurate CTR prediction.
\begin{figure}[h]
\centering
\includegraphics[width=3.4in]{figures/f1.pdf}
\caption{The structure of the proposed model. Different from the traditional CTR prediction model using only ad behavior ID embedding and user behavior ID embedding, our model additionally incorporates the visual content by devising a visual ID embedding for each ad (the purple rectangular). Since the visual ID embedding is only dependent on the visual content of the ad, it generalizes well for the new ads.}
\label{structure}
\end{figure}
To generate the visual IDs for ads, a straightforward solution is using k-means to cluster the visual features from an off-the-shelf image feature extractor, \emph{e.g.}, a convolutional neural network.
Nevertheless, the image features from the off-the-shelf image extractor such as ResNet-50~\cite{he2016deep} pre-trained on ImageNet~\cite{deng2009imagenet} might not be effective for discriminating the commercial images. To bridge the domain gap between the pre-trained dataset and the commercial application, we can fine-tune the pre-trained image extractor on commercial images in ads. Traditionally, fine-tuning the ResNet-50 requires human labours to annotate the label of each image, which is prohibitively expensive for the huge-scale commercial image corpus. Inspired by the recent success achieved by CLIP~\cite{radford2021learning} in self-supervised learning, we exploit the textual description of the image in an ad as the supervision and construct a contrastive learning task to optimize the image feature extractor. After contrastive learning through supervision from textual descriptions of images, we can obtain the visual features. But it still deviates from our goal of obtaining the visual IDs. Thus, we devise a supervised clustering module based on the learned codebooks when fine-tuning the image feature extractor. Since the hard assignment in clustering is not differentiable, we soften the cluster assignment for achieving end-to-end training.
Meanwhile, to enrich the codebooks for partitioning the visual feature space into finer cells, we adapt residual quantization and product quantization.
After training the image feature extractor as well as the codebooks for supervised clustering using the textual description through contrastive learning, we use them to generate the visual IDs.
The visual IDs are further mapped into visual ID embeddings, which will be the input of the CTR prediction model.
\vspace{0.05in}
In summary, the \textbf{contributions} of this work are four-fold:
\begin{itemize}
\item We propose to incorporate visual content in the CTR prediction model through learning the visual ID embedding for ads. The learned visual ID embedding generalizes well to new ads and complements the ad behavior ID embedding. \vspace{0.0in}
\item We devise an effective approach to learn the visual ID using the supervision from textual descriptions in a contrast-learning manner. We formulate the visual ID learning into a supervised clustering problem based on learned codebooks. To support end-to-end learning, we soften the cluster assignment and make it differentiable.\vspace{0.05in}
\item To enrich the codebooks for partitioning the visual feature space into finer cells, we adopt both residual quantization and product quantization. \vspace{0.05in}
\item We launched the proposed CTR prediction model incorporating the visual content in Baidu's online advertising system. After launching, we achieved a $1.46\%$ improvement in CTR and a $1.10\%$ increase~in~total~charges.
\end{itemize}
\section{Related Work}
\textbf{CTR prediction model.} CTR prediction is a long-standing problem in online advertising and recommender system. The focus of CTR prediction is on learning an effective embedding for users as well as products (ads). Early works on embedding learning are mainly based on linear regression (LR)~\cite{jahrer2010combining,mcmahan2013ad,chapelle2014modeling,wang2021efficient} or factorization machine (FM)~\cite{koren2009matrix,rendle2010factorization,rendle2012factorization,juan2016field,hu2018collaborative}.
Specifically, LR-based methods use a linear projection to map the features into an embedding vector. The weights of the linear projection are optimized based on the binary cross-entropy loss. In parallel, FM-based methods map features into a latent space and model the interactions between users and products (ads) through the inner product of their
embedding vectors. Inspired by the great success achieved by deep learning in computer vision~\cite{he2016deep} and natural language processing~\cite{vaswani2017attention}, several methods
exploit deep neural networks to learn embedding vectors. Specifically, CNN-based CTR prediction model~\cite{liu2015convolutional} exploits the interactions between neighbors in the feature space for enhancing the discriminating power of embedding features. Zhang \emph{et al.}~\cite{zhang2016deep} utilizes deep neural networks to enhance the features from traditional embedding methods such as FMs, restricted Boltzmann machines (RBMs), and denoising auto-encoders (DAEs) for generating more effective embedding vectors. Wide$\&$Deep learning~\cite{cheng2016wide} jointly trains wide linear models and deep neural networks to combine their benefits for the recommender system. Product-based Neural Networks (PNN)~\cite{qu2016product} investigates the interactions between diffident features through inner-product and outer-product operations in neural networks for CTR prediction. Deep crossing~\cite{shan2016deep} builds a deep neural network that automatically combines features to produce superior models. DeepFM~\cite{guo2017deepfm} combines a factorization machine (FM) and a deep neural network (DNN). In DeepFM, the FM extracts low-order features, whereas the DNN generates the high-order features. DKN~\cite{wang2018dkn} develops a word-entity-aligned and knowledge-aware convolutional neural network to fuse s semantic-level and knowledge-level representations of news. Deep Interest Network (DIN)~\cite{zhou2018deep} models the user's rich historical behaviors through a sequence model. Deep Interest Evolution Network (DIEN)~\cite{zhou2019deep} improves DIN by using a more advanced sequence model, GRU. Deep Session Interest Network~(DSIN)~\cite{feng2019deep} also focuses on modeling users' behaviors in a sequence of sessions. It adopts Transformer~\cite{vaswani2017attention}, which has shown better performance than GRU in modeling sentences and documents. Search-based Interest Model
(SIM)~\cite{pi2020search} models the user's lifelong sequential behavior data through Transformer to exploit richer user behaviors. In these aforementioned methods, the CTR prediction is mainly based on the user's ID embedding and the item's ID embedding learned from the historical user behaviors. Thus, they might not perform well when the users' historical behaviors are not sufficient. Different from above works, we utilize the visual content to complement the user's ID embedding and the item's ID embedding to achieve good CTR prediction performance for new ads and new users.
\vspace{0.1in}
\noindent \textbf{Visual features in CTR prediction.} Mo \emph{et al.}~\cite{mo2015image} feed the raw images in products (ads) into a convolutional neural network (CNN) and train the CNN through user-click supervision. It addresses the cold start problem when the ID features are not reliable when the historical behaviors are not rich enough. Ge \emph{et al.}~\cite{ge2018image} extract image features not only on the products (ads) side but also on the user side. Zhao~\emph{et al.}~\cite{zhao2019you} observes the efficiency problem encountered in~\cite{mo2015image,ge2018image} and moves the image feature extractor to the offline phase. Specifically, before ranking the products (ads) through the CTR prediction model, they pre-rank the products (ads) based on their visual features. Category-specific CNN (CSCNN)~\cite{liu2020category} early fuses the product category with the visual image when modeling the product visual content and meanwhile devise an efficient architecture to make the online deployment of CNN model feasible.
Wang~\emph{et al.}~\cite{wang2021hybrid} proposes a hybrid bandit model using visual content as priors for creative ranking. Different from the above-mentioned methods, we map each ad into a visual ID. We train our image encoder and the visual ID generator based on the text-image pairs in an end-to-end manner. Then we learn the visual ID embedding using the objective of optimizing CTR prediction accuracy. Thus, our visual ID embedding not only encodes the visual content but also effectively describes the user-item interactions. Recently, the teams in Baidu Research and Baidu Search Ads developed a series of works~\cite{yu2020combo,yu2021multi,yu2022tree} to exploit the use of vision BERT for boosting the text-visual relevance for video ads.
\vspace{0.05in}
\section{Visual ID Generation}
\vspace{0.05in}
\textbf{Supervision signal.}
We denote the image in the ad by $I$ and denote the visual index of the image by $\mathrm{ID}_I$. Straightforwardly, $\mathrm{ID}_I$ can be obtained from k-means clustering on ad image features extracted from an off-the-shelf image feature extractor, \emph{e.g.}, ResNet pre-trained on ImageNet dataset. Nevertheless, there is a domain gap between the natural images in ImageNet and the commercial images in the ads. A solution to adapting the target domain is fine-tuning the pre-trained ResNet on the commercial images in the ads. But there are no class labels available for supervising the network fine-tuning, and it is prohibitively expensive to annotate the image labels manually. Inspired by the great success achieved by CLIP~\cite{radford2021learning} in self-supervised learning, we exploit the textual description of the images in the ad as the supervision to fine-tune the network through contrastive learning. We formulate the process of learning visual IDs for images in ads into a deep supervised clustering problem. Below we introduce the details.
\vspace{0.08in}
\noindent \textbf{Deep supervised clustering.} We denote the textual description of the image $I$ by $T$. We denote the image feature extractor by $\mathrm{Encoder}_{\mathrm{img}}$ and that for encoding the textual description by $\mathrm{Encoder}_{\mathrm{txt}}$. We denote the image feature of $I$ generated from $\mathrm{Encoder}_{\mathrm{img}}$ by $\mathbf{x} \in \mathbb{R}^d$ and the text feature of $T$ from $\mathrm{Encoder}_{\mathrm{txt}}$ by $\mathbf{y} \in \mathbb{R}^d$. That is,
\begin{equation}
\mathbf{x} = \mathrm{Encoder}_{\mathrm{img}}(I), \mathbf{y} = \mathrm{Encoder}_{\mathrm{txt}}(T).
\end{equation}
To make the clustering feasible, we devise a dictionary denoted by $\mathcal{C} = \{\mathbf{c}_i\}_{i=1}^N$. $\mathcal{C}$ are weights of the network which are randomly initialized and optimized through contrastive learning.
Straightforwardly, we can map the visual feature $\mathbf{x}$ from the image extractor to its closest codeword in $\mathcal{C}$. In this case, the visual ID of the image $I$ is just the index of its closest codeword:
\begin{equation}
\label{hard}
\mathrm{ID}_I = \mathrm{argmax}_{i\in[1,N]} -\|\mathbf{c}_i - \mathbf{x}\|_2,
\end{equation}
where $\|.\|_2$ denotes the $L_2$ norm. The visual feature $\mathbf{i}$ is mapped to $\mathbf{c}_{\mathrm{ID}_I}$ to learn $\mathbf{C}$ through contrastive learning. Nevertheless, the above formulation can only optimize $\mathbf{C}$ but fail to back-propagate the gradient to update the image encoder due to the fact that the hard assignment in Eq.~(\ref{hard}) is not differentiable. Thus, we soften the hard assignment and map the image feature $\mathbf{x}$ into a weighted summation of codewords:
\begin{equation}
\label{soft}
\hat{\mathbf{x}} = \sum_{i=1}^N s_i \mathbf{c}_i,
\end{equation}
where $s_i$ is the weight computed by
\begin{equation}
\label{eq:soft2}
s_i = \frac{e^{-\beta \|\mathbf{x} - \mathbf{c}_i\|_2}}{\sum_{j=1}^Ne^{-\beta \|\mathbf{x} - \mathbf{c}_j\|_2}},
\end{equation}
where $\beta$ is a pre-defined positive constant controlling the softness of the assignment. When $\beta \to +\infty$, the soft-assignment operation in Eq.~\eqref{soft} will degenerate to the hard assignment. To approximate the hard assignment well, $\beta$ cannot be set to be a small value. On the other hand, if $\beta$ is too large, the softmax function in Eq.~\eqref{soft} tends to fall into the saturation region and leads to gradient vanishing. In implementation, we set $\beta$ to be a small value in the first epoch and gradually increases it to a large value in the training process.
\begin{figure*}[t]
\centering
\includegraphics[width=7in]{figures/train.pdf}
\caption{The pipeline of learning visual IDs. Given a batch of $N$ text-image pairs. An image is only relevant with the text in the pair and is irrelevant with other $N-1$ images. The text extractor generates the text features $\{\mathbf{y}_i\}_{i=1}^N$ and the image extractor generates the image features $\{\mathbf{x}_i\}_{i=1}^N$. An image feature $\mathbf{x}_i$ is mapped into a weighted summation of its neighboring codewords $\hat{\mathbf{x}}_i$ through the soft assignment. Then the similarity between the text features $\{\mathbf{y}_i\}_{i=1}^N$ and the mapped image codewords $\{\hat{\mathbf{x}}_i\}_{i=1}^N$ are optimized through the contrastive loss. }
\label{visualID}\vspace{0in}
\end{figure*}
To partition the feature space into a fine level, we need to set the number of codewords in the codebook $\mathbf{C}$ large enough. Nevertheless, the increase in the number of codewords will inevitably lead to an increase in memory and computation costs. To achieve high efficiency, we conduct a two-level quantization by exploiting residual quantization.
In the first level, we map the feature $\mathbf{x}$ into codewords from the coarse-level dictionary $\mathcal{C}_0 = \{\mathbf{c}_{0,i}\}_{i=1}^{N_0}$ in a soft-assignment manner:
\begin{equation}
\label{eq:x}
\bar{\mathbf{x}} = \sum_{i=1}^{N_0} s_{0,i} \mathbf{c}_{0,i},
\end{equation}
where
\begin{equation}
s_{0,i} = \frac{e^{-\beta \|\mathbf{x} - \mathbf{c}_{0,i}\|_2}}{\sum_{j=1}^Ne^{-\beta \|\mathbf{x} - \mathbf{c}_{0,j}\|_2}},
\end{equation}
Then we compute the residual vector $\mathbf{r}$:
\begin{equation}
\mathbf{r} = \mathbf{x} - \hat{\mathbf{x}} .
\end{equation}
After that, we conduct product quantization (PQ)~\cite{jegou2011product,ge2013optimized} to split the image residual features $\mathbf{r} \in \mathbb{R}^{d}$ into $K$ segments:
\begin{equation}
\mathbf{r} \to [\mathbf{r}_1,\cdots,\mathbf{r}_K],
\end{equation}
and each segment of the residual vector $\mathbf{r}_k \in \mathbb{R}^{\frac{d}{K}}$, $\forall k \in [1,K]$.
We conduct clustering on each split segment of the residual vector $\mathbf{r}_k$ based on the split-specific codebook $\mathcal{C}_k = \{\mathbf{c}_{k,i}\}_{i=1}^{N_k}$, where $\mathbf{c}_{k,i} \in \mathbb{R}^{\frac{d}{K}}$.
To be specific, we conduct
\begin{equation}
\label{soft2}
\bar{\mathbf{r}}_k = \sum_{i=1}^{N_k} s_{k,i} \mathbf{c}_{k,i},
\end{equation}
where
\begin{equation}
s_{k,i} = \frac{e^{-\beta \|\mathbf{r}_k - \mathbf{c}_{k,i}\|_2}}{\sum_{j=1}^Ne^{-\beta \|\mathbf{r}_k - \mathbf{c}_{k,j}\|_2}}.
\end{equation}
After that, $\{\hat{\mathbf{r}}_k\}_{k=1}^K$ are concatenated into the recovered residual feature:
\begin{equation}
\hat{\mathbf{r}} \gets [\hat{\mathbf{r}}_1, \cdots, \hat{\mathbf{r}}_K].
\end{equation}
The recovered feature vector is obtained by summing up the recovered residual feature $\hat{\mathbf{r}}$ and coarse-quantize vector $\bar{\mathbf{x}}$ in Eq.~(\ref{eq:x}):
\begin{equation}
\label{eq:last}
\hat{\mathbf{x}} = \hat{\mathbf{r}} + \bar{\mathbf{x}}.
\end{equation}
Straightforwardly, the coarse-level codebook $\mathcal{C}_0$ and $K$ fine-level sub-codebooks $\{\mathbf{C}_k\}_{k=1}^K$ can partition the whole feature space into $\prod_{k=0}^K N_k$ cells. But the total memory complexity of storing the codebooks $\{\mathbf{C}_k\}_{k=0}^K$ is only $\mathcal{O}(dN)$.
\vspace{0.15in}
\noindent \textbf{Contrastive learning.} %
The contrastive learning is conducted within each mini-batch.
Given a mini-batch of text-image pairs $\{(I_i,T_i)\}_{i=1}^B$, the image $I_i$ is only relevant with the text $T_i$ and is irrelevant with other texts, $T_j$ ($j\neq i$).
Thus, after we obtain the text features $\{\mathbf{y}_i\}_{i=1}^B$ and the mapped image features $\{\hat{\mathbf{x}}_i\}_{i=1}^B$ from the codebooks $\{\mathcal{C}\}_{k=0}^K$ based on Eq.~(\ref{soft})-(\ref{eq:last}), we construct the contrastive loss defined as
\begin{align}
&\mathcal{L} =\\\notag
&- \frac{1}{B} \sum_{i=1}^B\left \{ \mathrm{log}(\frac{e^{\mathrm{sim}(\hat{\mathbf{x}}_i,\mathbf{y}_i)}}{\sum_{j=1}^B e^{\mathrm{sim}(\hat{\mathbf{x}}_j,\mathbf{y}_i)}}) + \mathrm{log}(\frac{e^{\mathrm{sim}(\hat{\mathbf{x}}_i,\mathbf{y}_i)}}{\sum_{j=1}^B e^{\mathrm{sim}(\hat{\mathbf{x}}_i,\mathbf{y}_j)}}) \right\},
\end{align}
where $\mathrm{sim}(\cdot,\cdot)$ is the function measuring the cosine similarity between two vectors defined as
\begin{equation}
\mathrm{sim}(\mathbf{x}, \mathbf{y}) = \frac{\mathbf{x}^{\top} \mathbf{y}}{\|\mathbf{x}\|_2 \|\mathbf{y}\|_2}.
\end{equation}
The contrastive learning loss aims to enlarge the gap between similarities of positive pairs and those of negative pairs. The process of learning image encoder, text encoder and dictionaries are visualized in Figure~\ref{visualID} and summarized in Algorithm~\ref{alg:basic}.
\begin{algorithm} [h]
\caption{Image encoder and dictionary learning.}
\label{alg:basic}
\textbf{Input}: {The image encoder $\mathrm{Encoder}_{\mathrm{img}}$, the text encoder $\mathrm{Encoder}_{\mathrm{txt}}$, $K+1$ codebooks $\{\mathcal{C}_k\}_{k=0}^K$, and text-image pairs $\{I_i, T_i\}_{i=1}^B$}, the positive constant $\beta$.
\textbf{for} {$i\ \in [1,B]$}{
~~~$ \mathbf{x}_i \gets \mathrm{Encoder}_{\mathrm{img}}(I_i)$
~~~$\mathbf{y}_i \gets \mathrm{Encoder}_{\mathrm{txt}}(T_i)$
~~~$\bar{\mathbf{x}}_{i} = \sum_{j=1}^N \mathbf{c}_{0,j} \frac{e^{-\beta\|\mathbf{x} - \mathbf{c}_{0,j}\|_2}}{\sum_{l=1}^N e^{-\beta\|\mathbf{x} - \mathbf{c}_{0,l}\|_2}}$
~~~ $ \mathbf{r}_i \gets \mathbf{x}_i - \bar{\mathbf{x}}_{i}$
~~~$[\mathbf{r}_{i,1}, \cdots, \mathbf{r}_{i,K}] \gets \mathbf{r}_i$\\
\textbf{for} {$k\ \in [1,K]$}{
~~~$\hat{\mathbf{r}}_{i,k} = \sum_{j=1}^N \mathbf{c}_{1,j} \frac{e^{-\beta\|\mathbf{r}_{i,k} - \mathbf{c}_{k,j}\|_2}}{\sum_{l=1}^N e^{-\beta\|\mathbf{r}_{i,k} - \mathbf{c}_{k,l}\|_2}}$
}
~~~$\hat{\mathbf{r}}_i \gets [\hat{\mathbf{r}}_{i,1}, \cdots, \hat{\mathbf{r}}_{i,K} ]$
~~~$\hat{\mathbf{x}}_i \gets \hat{\mathbf{r}}_i + \bar{\mathbf{x}}_{i}$ \\
}
$ \mathcal{L} = \frac{-1}{B}\sum_{i=1}^B \mathrm{log} \left( \frac{e^{ \frac{\hat{\mathbf{x}}_i^{\top}\mathbf{y}_i} {\|\mathbf{x}_i\|_2 \|\mathbf{y}_i\|_2} }}{\sum_{j=1}^B e^{\frac{\hat{\mathbf{x}}_i^{\top}\mathbf{y}_j} {\|\mathbf{x}_i\|_2 \|\mathbf{y}_j\|_2}}} \right) $\;
Update $\mathrm{Encoder}_{\mathrm{img}}$, $\mathrm{Encoder}_{\mathrm{txt}}$, the codebooks $\{\mathcal{C}_k\}_{k=0}^K$ based on $\mathcal{L}$ using stochastic gradient descent.
\end{algorithm}
\vspace{0.15in}
\noindent \textbf{Visual ID generation.} After we train the models using the contrastive learning, we use the image encoder $\mathrm{Encoder}_{\mathrm{img}}$ to extract the image feature $\mathbf{x}$. Then $\mathbf{x}$ is mapped to its closest codewords in the learned coarse-level dictionary $\mathcal{C}_0 = \{ \mathbf{c}_{0,i}\}_{i=1}^{N_0}$. We define the index of the mapped codeword in $\mathcal{C}_0$ as the visual ID $\mathrm{ID}_0^v$. Then, we compute the residual feature vector $\mathbf{r} = \mathbf{x} - \mathbf{c}_{0,\mathrm{ID}_0^v}$. After that, $\mathbf{r}$ is equally split into $K$ segments $\{r_k\}_{k=1}^{K}$, and each segment is mapped to its closest codeword in the codebook $\mathbf{C}_k = \{ \mathbf{c}_{k,i}\}_{i=1}^{N_k}$. We define the index of the closest codeword in each dictionary $\mathbf{C}_k$ as the visual ID $\mathrm{ID}_k^v$, that is,
\begin{equation}
\begin{split}
\mathrm{ID}_{k}^v = \mathrm{argmin}_{i}{\|\mathbf{r}_k - \mathbf{c}_{k,i}\|}. \\
\end{split}
\end{equation}
We visualize the process of generating visual ID in Figure~\ref{visualIDgen}.
In total, for each image, we generate $K+1$ visual IDs ($\{\mathrm{ID}_{k}^v\}_{k=0}^K$), which will be used as the input of the CTR model to encode the visual content of the ad. It is worth noting that, the image extractor $\mathrm{Encoder}_{\mathrm{img}}$ and $K+1$ codebooks $\{\mathcal{C}_k\}_{k=0}^K$ are only used in generating visual ID. They will not be involved when learning the visual ID embedding for CTR prediction. An alternative choice is to plug the image extractor $\mathrm{Encoder}_{\mathrm{img}}$ and the codebooks $\{\mathcal{C}_k\}_{k=0}^K$ into the CTR prediction model to achieve end-to-end training. But it is prohibitively expensive due to the fact that the image feature extractor is normally heavy.
\begin{figure} [t!]
\includegraphics[width=3.5in]{figures/ctr_2.pdf}
\vspace{0.1in}
\caption{ The pipeline of generating visual IDs. An image is fed into the image encoder to generate the image feature $\mathbf{x}$.
It first finds its closest codeword in the coarse-level codebook $\mathcal{C}_0 = \{\mathbf{c}_{0,i}\}_{i=1}^{N_0}$ and the index of the losest codeword is the first visual ID, $\mathrm{ID}_0$.
Meanwhile, we obtain the residual vector $\mathbf{r} = \mathbf{x} - \mathbf{c}_{0,\mathrm{ID}_0}$, which is equally split into $K$ segments $[\mathbf{r}_1,\cdots,\mathbf{r}_K]$.
Each segment $\mathbf{r}_{k}$ is mapped to its closest codeword in the dictionary $\mathbf{C}_k$ and the index of the closest codeword in $\mathbf{C}_k$ corresponds to the visual ID, $\mathrm{ID}_K$. In total, we obtain $K+1$ visual IDs for an image.}
\label{visualIDgen}
\end{figure}
\section{CTR Prediction Model Learning}
\vspace{0.08in}
The CTR prediction model in our advertising platform takes the user ID, the ad ID and the ad visual ID as input to predict the CTR. We denote a user by $u$ and an ad by $a$.
The CTR prediction model consists of embedding layers, the ad tower, the user tower and the prediction head.
To be specific, the user embedding layer maps the ID of the user $u$ to a vector:
\begin{equation}
\mathbf{u} = \mathrm{EMB}_{\mathrm{user}}(\mathrm{ID}_u),
\end{equation}
the ad embedding layer maps the ID of the ad $a$ to a vector:
\begin{equation}
\mathbf{a} = \mathrm{EMB}_{\mathrm{ad}}(\mathrm{ID}_a).
\end{equation}
The visual embedding layer maps the ad's each visual IDs $\mathrm{ID}^v_k$ ($k\in[0,K]$) to a feature vector in the following manner
\begin{equation}
\begin{split}
{\mathbf{x}} &= \mathrm{EMB}_{\mathrm{visual}}^{(0)}(\mathrm{ID}^v_{0})\\
\mathbf{r}_k &= \mathrm{EMB}_{\mathrm{visual}}^{(k)}(\mathrm{ID}^v_{k}),~\forall k\in[1,K]. \\
\end{split}
\end{equation}
Then the visual embedding of the ad $\mathbf{v}$ is obtained by concatenating $\{\mathbf{r}_k\}$ and summing up the concatenated vector with ${\mathbf{x}}$
\begin{equation}
\begin{split}
\mathbf{r} &= [\mathbf{r}_1,\cdots, \mathbf{r}_K],\\
\mathbf{v} &= \mathbf{x} + \mathbf{r} .\\
\end{split}
\end{equation}
The user tower takes $\mathbf{u}$ as input, and generates the user feature:
\begin{equation}
\hat{\mathbf{u}} = \mathrm{Tower}_{\mathrm{user}}(\mathbf{u}).
\end{equation}
In parallel, the ad tower takes $\mathbf{a}$ and $\mathbf{v}$ as input, and generates the ad feature:
\begin{equation}
\hat{\mathbf{a}} = \mathrm{Tower}_{\mathrm{ad}}([\mathbf{a},\mathbf{v}]).
\end{equation}
In practice, both $\mathrm{Tower}_{\mathrm{ad}}$ and $ \mathrm{Tower}_{\mathrm{user}}$ are implemented by several fully-connected layers.
The prediction head takes the user feature $\hat{\mathbf{u}} \in \mathbb{R}^D $ and the ad feature $\hat{\mathbf{a}} \in \mathbb{R}^D$ as input and generates the predicted CTR:
\begin{equation}
\hat{y} = \mathrm{sigmoid}([\hat{\mathbf{u}},\hat{\mathbf{a}}] \mathbf{w} + b),
\end{equation}
where $\mathbf{w} \in \mathbb{R}^{2D}$ and $b \in \mathbb{R}$ are weights of the prediction head, and $\mathrm{sigmoid}$ is the function defined as $\mathrm{sigmoid}(x) = \frac{1}{1+e^{-x}}$.
In the training phase, we construct a binary cross-entropy loss to update the weights CTR prediction model. To be specific, given the predicted CTR $\hat{y} \in (0,1)$ and the ground-truth CTR by $y \in \{0,1\}$, the binary cross-entropy loss is constructed by
\begin{equation}
\mathcal{L}_{\mathrm{BCE}} = -[(1-y)\mathrm{log}(\hat{y}) + y\mathrm{log}(1-\hat{y})].
\end{equation}
\begin{algorithm}[h]
\caption{CTR prediction model learning.}
\label{alg:basic2}
\textbf{Input}: {The user-ad pairs $\{a_i,u_i\}_{i=1}^M$ and the ground-truth labels $\{y_i\}_{i=1}^M$, user embedding layer $\mathrm{EMB}_{\mathrm{user}}$, ad embedding layer $\mathrm{EMB}_{\mathrm{ad}}$, and visual embedding layers $\mathrm{EMB}_{\mathrm{visual}}^{(1)}$ and $\mathrm{EMB}_{\mathrm{visual}}^{(2)}$, user tower $\mathrm{Tower}_{\mathrm{user}}$, ad tower $\mathrm{Tower}_{\mathrm{ad}}$, the prediction head with weights $\{\mathbf{w},b\}$}\;
\textbf{for}~~{$i\ \in [1,M]$}{
~~~fetch the ad ID for $a_i$, $\mathrm{ID}_{a}$\;
~~~$\mathbf{a} = \mathrm{EMB}_{\mathrm{ad}}(\mathrm{ID}_a)$ \;
~~~fetch the visual IDs of $a_i$, $\{\mathrm{ID}^v_{k}\}_{k=0}^K$\;
~~~$\mathbf{x} = \mathrm{EMB}_{\mathrm{visual}}^{(0)}(\mathrm{ID}_{0}^v)$\;
\textbf{for}~~{$k\ \in [1,K]$}{
~~~$\mathbf{r}_k = \mathrm{EMB}_{\mathrm{visual}}^{(k)}(\mathrm{ID}_{k}^v)$\;
}
~~~$\mathbf{r} = [\mathbf{r}_1,\cdots, \mathbf{r}_K]$\;
~~~$\mathbf{v} = \mathbf{r} + \mathbf{x}$\;
~~~$\hat{\mathbf{a}} = \mathrm{Tower}_{\mathrm{ad}}([\mathbf{a},\mathbf{v}]$ \;
~~fetch the user ID for $u_i$, $\mathrm{ID}_{u}$\;
~~~$\mathbf{u} = \mathrm{EMB}_{\mathrm{user}}(\mathrm{ID}_u)$\;
~~~$\hat{\mathbf{u}} = \mathrm{Tower}_{\mathrm{user}}(\mathbf{u})$\;
~~~$\hat{y}_i = \mathrm{sigmoid}([\hat{\mathbf{u}},\hat{\mathbf{a}}] \mathbf{w} + b),$
}
$ \mathcal{L} = - \frac{1}{M}\sum_{i=1}^M (1-y_i)\mathrm{log}(\hat{y}_i) + y_i\mathrm{log}(1-\hat{y}_i) $\;
Update $\mathrm{EMB}_{\mathrm{user}}$, $\mathrm{EMB}_{\mathrm{ad}}$, $\mathrm{EMB}_{\mathrm{visual}}^{(1)}$, $\mathrm{EMB}_{\mathrm{visual}}^{(2)}$, $\mathrm{Tower}_{\mathrm{user}}$, $\mathrm{Tower}_{\mathrm{ad}}$, $\mathbf{w}$, and $b$ using $\mathcal{L}$ by stochastic gradient descent.
\end{algorithm}
\noindent Algorithm~\ref{alg:basic2} summarizes the CTR prediction model learning.
\vspace{0.1in}
\section{Experiments}
\vspace{0.1in}
\noindent\textbf{Implementation.} We implement the image encoder $\mathrm{Encoder}_{\mathrm{img}}$ by ResNet50~\cite{he2016deep} with $50$ layers and the text encoder $\mathrm{Encoder}_{\mathrm{txt}}$ by the BERT-base model~\cite{devlin2019bert} with $12$ Transformer layers and $768$ hidden size. ResNet50 is pre-trained on ImageNet dataset and the BERT-base model is pre-trained on sentences from the BooksCorpus dataset with 800 Million words and English Wikipedia with 2,500 million words. We use the hidden feature before the last classification layer of ResNet50 as the image feature $\mathbf{x}$ and use the hidden state of the [cls] token in the last layer of the BERT-base as the text feature $\mathbf{y}$. To make the dimension of the image feature identical to that of the text feature, we add a fully-connected layer to project the feature obtained from the ResNet to the $768$-dimensional vector.
By default, we set the number of codebooks for the residual vector, $K$, as $4$. Meanwhile we set the default number of codewords in each codebook $\mathcal{C}_k$, $N_k$, as $256$. In the training process, we set the initial value of $\beta$ as $1$ and gradually increase it to $10$.
\vspace{0.08in}
\noindent \textbf{Training data.} To train the image extractor $\mathrm{Encoder}_{\mathrm{img}}$, the text extractor $\mathrm{Encoder}_{\mathrm{txt}}$ and the codebooks $\{\mathcal{C}_k\}_{k=0}^K$, we use $50$ million text-image pairs collected from our advertising system. To train the CTR prediction model, we use $70$ million user-click data per day collected from our advertising system and the model is fine-tuned for around 1 month.
\subsection{Offline Experiments}
We evaluate the influence of our model on the CTR prediction in the offline phase. We first define the CTR prediction model using only ad behavior ID features and user behavior ID features as the baseline. We use the AUC improvement of CTR prediction over the baseline model as the evaluation metrics. Meanwhile, we use an observation window of 40-day length to demonstrate more details.
\vspace{0.1in}
\noindent \textbf{Residual quantization.} By default, we conduct a two-level residual quantization. In the first level, we conduct a coarse level quantization using $\mathcal{C}_0$ and then conduct the second-level quantization on the residual. Here, we compare the two-level residual quantization with a single-level quantization by removing the first-level coarse level quantization.
We report the AUC difference between the model with the baseline model along $40$ days in Figure~\ref{fig:res}.
\begin{figure}[h]
\vspace{-0.1in}
\centering
\includegraphics[width=2.65in]{fig/res.eps}
\vspace{-0.1in}
\caption{Comparisons with the setting without residual quantization.}
\label{fig:res
\end{figure}
As shown in the figure, using the residual quantization, the model can get more improvement compared with its counterpart without residual quantization. To be specific, at the $40$-th day, using residual quantization, the model achieves a $0.1418\%$ AUC improvement, whereas the model without residual quantization only achieves a $0.1242\%$ AUC increase.
\vspace{0.1in} \noindent \textbf{Comparisons with alternative configurations.} We soften the quantization operation as Eq.~\eqref{soft} and Eq.~\eqref{eq:soft2} to make it differentiable. Thus, the codeword assignment can be incorporated into a neural network, and the codewords can be learned in an end-to-end manner. An alternative solution is a two-stage learning process. In the first stage, we can learn the raw feature vectors without quantization through the contrastive loss based on the text-image pairs. After that, we use the unsupervised k-means clustering to conduct residual quantization and product quantization for generating the codebooks and the visual IDs. We term this configuration as the two-stage baseline. Meanwhile, we also compare with the baseline directly using the raw feature vectors learned without quantization through the contrastive loss based on text-image pairs as the visual representation for CTR prediction. We term it as w/o quantization baseline. As shown in Figure~\ref{fig:quan}, ours consistently outperforms two compared baselines, demonstrating the effectiveness of learning codebooks in an end-to-end manner.
\begin{figure}[t]
\centering
\includegraphics[width=2.8in]{fig/quan-eps-converted-to.pdf}
\vspace{-0.1in}
\caption{Comparisons with the two-stage setting and the setting without quantization.}
\label{fig:quan}\vspace{-0.2in}
\end{figure}
\begin{figure}[b!]
\vspace{-0.25in}
\centering
\subfigure[The influence of $N$]{\includegraphics[width=2.65in]{fig/fig6a-eps-converted-to.pdf}}\\
\vspace{-0.1in}
\subfigure[The influence of $K$]{\includegraphics[width=2.65in]{fig/fig6b-eps-converted-to.pdf}}
\vspace{-0.1in}
\caption{The influence of parameters on the performance.}
\label{fig:hyper}\vspace{-0.1in}
\end{figure}
\begin{figure*}[t]
\centering
\subfigure[$\mathrm{ID}^v_0=1388$, $\mathrm{ID}^v_1=145$, $\mathrm{ID}^v_2=251$, $\mathrm{ID}^v_3=86$, $\mathrm{ID}^v_4=87$]{
\includegraphics[width=1\textwidth, height=0.17\textwidth ]{figures/visual1.pdf}}
\subfigure[$\mathrm{ID}^v_0=1511$, $\mathrm{ID}^v_1=201$, $\mathrm{ID}^v_2=106$, $\mathrm{ID}^v_3=42$, $\mathrm{ID}^v_4=149$]{
\includegraphics[width=1\textwidth, height=0.17\textwidth]{figures/visual2.pdf}}
\subfigure[$\mathrm{ID}^v_0=84$, $\mathrm{ID}^v_1=104$, $\mathrm{ID}^v_2=61$, $\mathrm{ID}^v_3=22$, $\mathrm{ID}^v_4=202$]{
\includegraphics[width=1\textwidth, height=0.17\textwidth]{figures/visual3.pdf}}
\caption{Visualization of ads with the same visual IDs.}
\label{fig:visual}
\end{figure*}
\vspace{0.05in}
\noindent \textbf{Influence of the number of codewords.} A larger number of codewords can partition the feature space into finer cells and leads to smaller quantization errors.
We evaluate the influence of the number of codewords, $N$, on the AUC of CTR prediction. To this end, we show the results when $N$ varies among $\{64,256,1024\}$. As shown in Figure~\ref{fig:hyper} (a), when $N$ increases from $64$ to $256$, the AUC of CTR prediction gets improved considerably. But when $N$ further increases to $1024$, the AUC gets saturated. Since the increase of $N$ will bring more computation cost, considering both efficiency and effectiveness, we set $N=256$ by default.
\vspace{0.1in}
\noindent \textbf{Influence of the number of segments.} We split the residual vector into $K$ segments and conduct product quantization based on split segments. A larger $K$ also leads to a
finer partition of the feature space and smaller quantization errors. We evaluate the influence of the number of segments, $K$, on the AUC of CTR prediction. To this end, we show the results when $K$ is chosen from $\{1,2,4\}$. As shown in Figure~\ref{fig:hyper} (b), when $K$ increases from $1$ to $4$, the AUC of CTR prediction gets boosted significantly.
To be specific, when $K=1$, at the $40$-th day, the AUC improvement over baseline is only $0.0903\%$, whereas the AUC improvement is $0.1418\%$ when $K=4$.
But the increase of $K$ also brings more computational cost. Thus, to balance the efficiency and effectiveness, we do not further increase $K$ beyond $4$ and set $K=4$, by default.
\vspace{0.1in}
\noindent \textbf{Influence on new ads and old ads.} As we know, incorporating the visual content cannot bring more benefits to new ads since their ID features are not reliable. Here, we demonstrate the influence of our model in new ads and the old ones. As shown in Table~\ref{tab:newold}, using the visual ID embedding, it has a more influential impact on new ads than old ads.
To be specific, for new ads, it brings $+0.6098\%$ increase in AUC of the CTR prediction, whereas it only brings $+0.1238\%$ increase for old ads.
\begin{table}[t]
\caption{The influence of visual ID embedding on new ads and old ads.}
\centering
\begin{tabular}{cc|cc} \hline
\multicolumn{2}{c|}{old ads} & \multicolumn{2}{c}{new ads} \\ \hline
\# of instances & AUC diff & \# of instances & AUC diff \\ \hline
$61,982,597$ & $+0.1238\%$ & $4,935,208$ & $+0.6098\%$ \\\hline
\end{tabular}
\label{tab:newold}\vspace{-0.1in}
\end{table}
\vspace{0.1in}
\noindent \textbf{Visualization.} In Figure~\ref{fig:visual}, we visualize the ads with the same visual IDs, $\{\mathrm{ID}_{k}^v\}_{k=0}^4$. It is not difficult to observe from the figure that the ads with the same visual IDs have a similar visual appearance, demonstrating the effectiveness of the learned visual IDs.
\subsection{Online Experiments}
\begin{table}[h]
\centering
\caption{The influence of launching the proposed model in our advertising platform. Note that ``$+\cdot\%$'' is relative improvement for online experiments.}
\begin{tabular}{cc}\\ \hline
Charge & CTR \\ \hline
$+1.10\%$ & $+1.46\%$\\ \hline
\end{tabular}
\label{online}
\end{table}
In Table~\ref{online}, we show the influence of launching the proposed model in Baidu online advertising platform. Before launching our model, the CTR prediction model in our advertising platform uses only the ad ID features and user ID features based on the users' historical behaviors. As shown in the table, after launching our model, the charge of our advertising platform gets increased relatively by $+1.10\%$, and the CTR of the ads gets improved relatively by $+1.46\%$. That is, the launching of the proposed model brings significant revenue boosts for the advertisers as well as our ads platform.
\vspace{0.1in}
\section{Conclusion}
\vspace{0.1in}
The traditional CTR prediction model relies on the user behavior ID features and product (ad) behavior ID features. But the effectiveness of these ID features requires rich historical user-product (ad) behaviors, and thus they do not generalize well to the new products (ads). In this work, we investigate exploiting visual content in ads to boost the CTR prediction accuracy. We adopt the textual description of images as the supervision to learn the image feature extractor and map the visual content in each ad to visual IDs. Then we learn the embedding for visual IDs based on the user's clicking behaviors. We feed the learned visual ID embedding besides the ad behavior ID embedding and the user behavior ID embedding into a CTR prediction model to generate a more reliable CTR prediction. The systematic offline experiments demonstrate the improved AUC of the CTR prediction by incorporating the visual content. After launching the proposed model in Baidu advertising platform, we achieve a $ 1.10 \%$ charge increase and a $1.46\%$ CTR increase.
\vspace{0.5in}
\bibliographystyle{plain}
|
\section{Conclusion}
We propose an approach to leverage the \textit{coarsely-aligned} datasets for image-to-image translation. Our proposed approach, Coarsely-Aligned Paired Image Translation (\method{CAPIT}\xspace), is built upon several key insights of the coarsely-aligned data.
These include foreground masking, local re-alignment to overcome pixel shifts, and a ranking-based loss to address stochastic variation.
\method{CAPIT}\xspace improves both the generated image quality and the performance of downstream tasks in autonomous driving.
Furthermore, our approach is general and can be extended to other applications where coarsely-aligned data is easily available.
\section{Experiments}\label{s_exp}
First, we evaluate our \textit{coarsely}-paired image translation approach on the Ithaca365 dataset~\cite{diaz2022Ithaca365} and compare our results to several baseline methods, as shown in \autoref{ss_mainR}. We evaluate the quality of image translation using the FID metric \cite{heusel2017gans} and several downstream tasks. In \autoref{sec:ablation}, we perform an ablation study, demonstrating the key components of our approach in leveraging \emph{coarsely}-paired dataset. In \autoref{sec:exp-foreground}, we provide an analysis on the effect of mask accuracy in foreground-masked loss computation. In \autoref{sec:exp-misalignment}, we provide an analysis on the effect window size in misalignment-tolerating loss computation. Finally, in \autoref{sec:exp-GANloss} we do an additional experiment on another datasets (Facades \cite{Tylecek13} and map to aerial photos~\cite{isola2017image} datasets) to justify the design choice in our GAN loss.
\subsection{Datasets}
Our main experiment uses the Ithaca365 dataset \cite{diaz2022Ithaca365} which contains diverse scenes (\textit{e.g.}\xspace, urban, rural, and campus area) along a repeated 15km route. We use one traversal per weather condition (night and snowy as source domains and sunny as target domain). Each condition is split into $5208/1138$ images for training and testing, where the two splits do not have overlapping location. The images are paired between different conditions based on closest GPS poses. We use pretrained Mask R-CNN~\cite{wu2019detectron2,he2017mask} instance segmentation to obtain the foreground masks for our loss computation in \autoref{eq:fgmask}.
For additional analysis on GAN loss (\autoref{sec:exp-GANloss}), we also do experiments on additional datasets, Facades \cite{Tylecek13} and map to aerial photos~\cite{isola2017image} datasets.
\subsection{Main Results}\label{ss_mainR}
We compare our approach to several SOTA unpaired translation models (CUT \cite{park2020contrastive}, ACL-GAN \cite{zhao2020aclgan}, MUNIT \cite{huang2018munit}, CycleGAN \cite{CycleGAN2017}, ForkGAN~\cite{zheng2020forkgan}, and ToDayGAN~\cite{anoosheh2019night}) to show the benefit of leveraging \emph{coarsely}-paired images for supervision.
\autoref{fig:gen} shows some qualitative results of translated images, which generally demonstrate that our approach can translate an image closer to the reference image.
\autoref{table:comparison-all} shows the performance of our approach compared with other baselines evaluated through different metrics and dowstream tasks (\textit{i.e.}\xspace, semantic segmentation, monocular depth estimation, and visual localization)
\subsubsection{FID} We assess the quality of the translated images through Frechet Inception Distance (FID)~\cite{bynagari2019gans} to indicate distance between the distribution of translated images (fake sunny images) and of target images (real sunny images). \autoref{table:comparison-all} shows that our approach has lower FID than the baselines, meaning that our translated images are more similar to the real sunny images.
We also compute the FID between real sunny images and original adverse weather images, \textit{i.e.}\xspace, snowy and night, to show the performance gap in adverse conditions without any domain adaptation. Finally, the FID between two different sunny traversals is computed (last row in \autoref{table:comparison-all}) as a lower bound (best case) indication.
\subsubsection{Semantic Segmentation}
We evaluate the performance of image translation model through a downstream task of semantic segmentation using pixel accuracy (PixAcc) as the evaluation metric.
\emph{All} semantic segmentation experiments are evaluated using the PSPNet~\cite{zhao2017pyramid} pretrained on the MIT ADE20k dataset~\cite{zhou2017scene}. First, we apply the PSPNet model for semantic segmentation directly on night and snowy images. Then, to assess the quality of our image translation model, we apply the same model on the fake sunny images translated from night and snowy using our image translation model, as well as other baseline models. \autoref{table:comparison-all} shows that our image translation method improves the performance of semantic segmentation in adverse conditions more than all baselines.
As the Ithaca365 dataset~\cite{diaz2022Ithaca365} does not provide semantic segmentation labels, we generate \emph{pseudo} ground-truth label by applying the pretrained PSPNet on the sunny \emph{coarse} correspondence images. Additionally, when computing pixel accuracy, if a pixel has a moving foreground label in either the original adverse image or the reference sunny image, it is not included in the accuracy computation. As the pseudo ground-truth reference is approximate, we run semantic segmentation on \emph{another} set of \textit{coarsely-aligned} sunny images, comparing against the pseudo ground-truth to give a sense of achievable upper bound accuracy (last row in \autoref{table:comparison-all}).
\subsubsection{Monocular Depth Estimation}
We also evaluate the performance of image translation through monocular depth estimation downstream task. We use a standard inlier metric $\delta_\text{threshold}$: the percentage of pixels whose inaccuracy level $\max\left(\frac{\hat{d}_{i,j}}{d_i},\frac{d_{i,j}}{\hat{d}_{i,j}}\right)$ are below a threshold (we use 1.25); $d_{i,j}$ and $\hat{d}_{i,j}$ are the ground-truth and predicted depth for pixel $i,j$~\cite{lee2019big}.
We train BTS~\cite{lee2019big} monocular depth estimation model on the sunny data using projected LiDAR as the sparse ground truth. We test the same depth model (trained on sunny condition) on the adverse weather condition images without domain adaptation, as well as on the fake sunny images translated using our approach and several baseline models. Finally, we evaluate the depth model on the in-domain real sunny dataset to give us an upper bound (best case) accuracy (last row in \autoref{table:comparison-all}). We show (\autoref{table:comparison-all}) the benefit of our image translation model compared to the baselines by the domain-induced performance gap that it closes.
\input{tables/table-comparison}
\subsubsection{Image-based Localization}
\label{ssec:labeling5}
Our final downstream task evaluation is on image-based localization, formulated as an image-retrieval problem. Specifically, given a \textit{query} image, we find the \textit{closest} image using SIFT matching~\cite{lowe1999object} from a set of images in \textit{library}. We evaluate based on localization error (loc-err), \textit{i.e.}\xspace Euclidean distance from query image location to its match.
We use images from a sunny traversal as the \textit{library}, while the query images come from the adverse weather images (night and snowy). First, we directly use the adverse weather image as the \textit{query} without any adaptation; this serves as a bottom baseline. Then, we employ image translation models to first convert the night/ snowy images to fake sunny images for query.
The result in (\autoref{table:comparison-all}) shows that our image translation approach improves the localization error (loc-err) more than the baselines.
\begin{figure*}[th]
\centering
\includegraphics[width=1\linewidth]{figures/viz.png}\vskip-5pt
\caption{\small Visualization results of image translation of our approach compared to several baselines.}
\label{fig:gen}
\vskip -15pt
\end{figure*}
\subsection{Ablation Study}\label{sec:ablation}
We perform an ablation study to assess the gains from each of our proposed components and loss terms (shown in \autoref{table:ablation}).
We begin with a baseline \emph{paired} translation algorithm (Row-1) that employs conditional-GAN (cGAN) and L1 loss~\cite{isola2017image}.
Comparing Row-1 and Row-2, we see that applying \emph{foreground masking} consistently improves the performance. Comparing Row-2 and Row-3, we see that \emph{unpaired-GAN (uGAN) loss} outperforms the original paired conditional-GAN (cGAN) loss.
Therefore, we employ the uGAN loss instead of cGAN loss.
Comparing Row-3 and Row-4, we see that using \emph{masked patchNCE loss} between the generated image and the coarsely-aligned target image can largely overcome the stochastic misalignment and improve the performance. Finally, comparing Row-4 and Row-5 shows that accounting for the pixel misalignment in the \emph{misalignment-tolerating L1 loss} computation can further boost the model.
\input{tables/table-ablation}
In short, while we have shown (\autoref{ss_mainR}) the benefit of leveraging the coarsely-aligned data, \emph{the gains are by no means trivial.} Indeed, by comparing Row-1 in \autoref{table:ablation} (baseline paired translation) to existing unpaired algorithms in \autoref{table:comparison-all}, we hardly see the benefit of coarsely-aligned data. However, with our proposed losses, which are dedicated to tackling the misalignment due to foreground objects, camera pose shifts, and stochastic variations, our \method{CAPIT}\xspace can leverage the valuable contextual information provided by the coarsely-paired data to excel in image translation.
\subsection{Effect of Mask Accuracy on Foreground-Masked Loss}\label{sec:exp-foreground}
Masks are used to mask out foregrounds (which are unlikely static) in the reconstruction loss computation during \textit{training} only; they can be pre-generated prior to training. As masks are only used for training, \emph{ground-truth masks can be used} if available. If the dataset does not provide such labels (such as in the Ithaca365 dataset), we can can apply pre-trained Mask R-CNN to each \textit{coarsely}-aligned image pair and take the \textit{union} as masks (\autoref{eq:fgmask}).
\autoref{table:maskthreshold} shows the FID evaluation of image translation different \emph{confidence thresholds} are used for Mask R-CNN mask proposals acceptance. Higher threshold produces lower FP but higher FN. False positive (FP) errors mask out loss computation on unnecessary area, reducing the training signal amount. False negative (FN) errors compute reconstruction loss on foreground areas that may not match in the two images, causing \textit{noise} in the training signal. As shown, Mask R-CNN \emph{default} threshold (0.5) produces the most accurate masks.
\input{tables/maskthreshold}
\subsection{Effect of Window Size on Misalignment-Tolerating Loss}\label{sec:exp-misalignment}
Window size (in \autoref{eq:k}) is studied in \autoref{table:windowsize}.
If images are perfectly aligned, we can directly compare pixels of two images at the same location. However with misalignment, the window acts a search space of correspondence when comparing the reconstruction loss.
A window too small cannot cover misaligned pixels, while too large over-simplifies the translation problem (\textit{i.e.}\xspace, find similar but irrelevant pixels). We use $k_w\!=\!k_h\!=\!k\!=\!3$ in \autoref{eq:k}.
\input{tables/windowsize}
\subsection{Additional Analysis on GAN Loss}\label{sec:exp-GANloss}
In the previous ablation study, we find that using unpaired-GAN (uGAN) loss results in better performance than conditional-GAN (cGAN) for our \emph{coarsely-paired} training. Inspired by this finding, we do further analysis (\autoref{table:additional}) on different ways of using GAN loss on well-aligned datasets: Facades \cite{Tylecek13} and map-to-aerial photos~\cite{isola2017image}. Specifically, we train the Pix2pix~\cite{isola2017image} paired translation model using three different GAN losses.
Let (realA, realB) be a \textit{paired} batch of images from domain A \& B.
First, in the \emph{conditional-GAN} training, we use a conditional-discriminator following the original Pix2Pix setup \cite{isola2017image}, where every image fed into the discriminator is conditioned on the input image (realA).
Second, in the \emph{paired-GAN} training, we feed $G$(realA) and realB into the discriminator $D$.
Third, in the \emph{unpaired-GAN} training, we sample \textit{another} batch of images realB' from domain B, and feed $G$(realA) and realB' into $D$.
Interestingly, we find that the unpaired-GAN loss combined with paired L1 results in the best performance for all three experiments. This result not only aligns with our previous finding and supports our GAN loss design, but also suggests that we should reexamine the discriminator design even with perfectly-aligned image pairs.
Note that while the overall training objectives (\textit{i.e.}\xspace, expectation over all images) of paired and unpaired GAN are the same, they lead to different training dynamics of $D$ in mini-batch training. We find this important since GAN training iterates between generator and discriminator training and is known to be less stable.
Specifically, we find that \textit{unpaired} training can better capture the overall difference between real and fake images, while \textit{paired} training overly focuses on each paired batch.
\input{tables/GANloss}
\section{Introduction}
Although the development of autonomous driving perception has rapidly advanced in recent years, work has typically focused on ideal weather conditions. However, to ensure safe operation, these perception systems must also operate robustly under various adverse weather conditions (such as snowy and night)~\cite{pitropov2021canadian}.
One possible solution is to train joint or specialized models to cover all conditions.
However, this requires having more models or increased model size, as well as costly labeling and retraining for all tasks and conditions.
Moreover, sensor signals (\textit{e.g.}\xspace, images) in adverse weather are likely of poor quality, making some labeling tasks harder.
In this paper, we investigate an alternative idea of turning images captured in adverse weather into a benign one (\textit{i.e.}\xspace, sunny), upon which the downstream tasks are more likely to attain high accuracy, even without additional retraining.
This problem generally can be cast as an image-to-image translation task~\cite{wang2018high}, training neural networks to transfer images from a source domain (\textit{e.g.}\xspace, night) to another target domain (\textit{e.g.}\xspace, sunny).
Image-to-image translation training can typically be categorized into paired \cite{isola2017image} or unpaired options \cite{CycleGAN2017,zhao2020aclgan,huang2018munit}.
Paired translations are trained with perfectly-aligned source and target image pairs, learning to map one to another. While this approach is straightforward, collecting such image pairs is challenging, if not impossible, in many application scenarios. For example, it is infeasible to obtain two driving images with the exact same semantic layouts (\textit{e.g.}\xspace, the same set of cars and pedestrians and their placements) under different weather conditions. Thus, for many problem setups (\textit{e.g.}\xspace, autonomous driving), existing works usually drop the notion of paired training and utilize the more challenging unpaired training, learning to map between two image distributions, usually accompanied by additional constraints (\textit{e.g.}\xspace, cycle consistency~\cite{CycleGAN2017}).
\input{figurestext/introfigure}
In this paper, we propose Coarsely-Aligned Paired Image Translation (\method{CAPIT}\xspace), a new fashion for learning the image-to-image translation model by leveraging \emph{coarsely-aligned} image pairs (see \autoref{fig:intro}).
We argue that \emph{coarsely-aligned} image pairs are much easier to obtain. For instance, many of us drive the same routes daily (\textit{e.g.}\xspace, to and from work and school), even under different weather conditions.
We can form coarsely-aligned pairs by picking images that are closest in GPS locations across traversals. While such data may not be perfectly aligned due to the shifts in camera poses or mobile objects at different times, they offer a wealth of training signals.
For example, they reveal consistency and appearance changes of background objects between image domains. Moreover, as we will show, these provide sufficient (weak) supervision to train state-of-the-art image-to-image translation models without perfectly-paired data.
Our approach addresses inconsistencies across coarsely aligned images in several ways.
First, as two images taken at different times are unlikely to be captured from the exact same camera pose, we propose a misalignment-tolerating L1 loss at the pixel level that first searches for the best local pixel alignment.
Second, as foreground objects (\textit{e.g.}\xspace, vehicles, pedestrians, and cyclists) may vary in terms of their identities and placements, we propose to mask them and apply reconstruction loss only to unmasked regions.
Furthermore, to accommodate stochastic variations (\textit{e.g.}\xspace, due to tree swaying or shadow patterns), we propose a novel usage of the noise contrastive estimation (NCE) loss~\cite{park2020contrastive}.
This type of loss has been used in different contexts (\textit{e.g.}\xspace, image retrieval) to overcome significant misalignment effectively.
We validate \method{CAPIT}\xspace on the Ithaca365 dataset~\cite{diaz2022Ithaca365} with driving scenes through the same routes under various weather conditions (\textit{e.g.}\xspace, sunny, snowy, and night).
We show that \method{CAPIT}\xspace results in a more faithful image translation model compared to existing paired or unpaired algorithms, not only in the quality of generated images but also in the performance of downstream tasks such as semantic segmentation~\cite{zhao2017pyramid}, depth estimation~\cite{lee2019big}, and visual localization~\cite{naseer2018robust}.
While we focus on autonomous driving applications, our proposed approach is general and applicable to other applications where coarsely-aligned data are easily obtainable. For example, in remote sensing~\cite{campbell2011introduction,dubayah2000lidar}, different sensors (\textit{e.g.}\xspace, airborne LiDAR and satellite cameras) capture different channels (\textit{e.g.}\xspace, height, infrared, and RGB images) with different characteristics for downstream tasks. While these sensors mainly operate alone, they may have surveyed some common areas at different times. Therefore, we can pair their data by GPS locations to learn to transfer one information channel to the others.
\section{Proposed Approach: \method{CAPIT}\xspace}
\subsection{Problem Setup}
Our goal is to learn an image mapping function $G$ from a source domain $\mathcal{X}$ (\textit{e.g.}\xspace, adverse weather) to a target domain $\mathcal{Y}$ (\textit{e.g.}\xspace, sunny). We assume two set of training images $X = \{x^{n}\}_{n=1}^N \in \mathcal{X}$ and $Y=\{y^n\}_{n=1}^N \in \mathcal{Y}$ from both domains. In many scenarios, it is infeasible to obtain paired images. However, in autonomous driving application, we can still obtain \textit{coarse} pairs $\{(x^n\in\mathcal{X}, y^n\in\mathcal{Y})\}_{n=1}^N$ if they are collected from the same route/ area.
For example, the Ithaca365 dataset~\cite{diaz2022Ithaca365} collects images and GPS on the same route under different weather conditions (\textit{e.g.}\xspace, sunny, night, and snowy). Therefore, we pair images from different traversals
based on the closest GPS poses (see \autoref{fig:intro}-mid for a training data sample). We note that these paired images are \emph{not perfectly aligned} due to potential GPS differences/errors, changes in semantic layouts at different collection times, and shifts in camera pose.
We train $G$ using a GAN-style loss~\cite{isola2017image,CycleGAN2017}. However, instead of directly employing the paired translation algorithms~\cite{isola2017image} or entirely ignoring the paired information by applying an unpaired algorithm~\cite{CycleGAN2017}, we propose \method{CAPIT}\xspace, which is dedicated to learning image translation using \textit{coarsely-aligned} image pairs.
\subsection{Baseline Algorithm}
\label{ss_baseline}
To begin with, let us introduce a baseline paired translation algorithm, inspired by~\cite{isola2017image}. \cite{isola2017image} assumes that the training data are perfectly paired images, and introduces an L1 loss comparing paired images $x^{n}$ and $y^{n}$ in pixel space.
\begin{equation}
\mathcal{L}_{L1}(G) = \mathbb{E}_{(x,y)\sim(X,Y)}\lVert G(x) - y \rVert_1
\end{equation}
Here, $G(x^{n})_{i,j}$ and $y^{n}_{i,j}$ are the pixel values at the $(i,j)$ location and $K^{n}$ is the number of pixels in image $y^{n}$.
Along with the L1 loss, \cite{isola2017image} further employs a GAN loss~\cite{goodfellow2014generative}. Let $D$ denotes a discriminator (\textit{i.e.}\xspace, binary classifier) that tells a real image $y\in\mathcal{Y}$ from a fake, generated one $G(x)$, where $x\in\mathcal{X}$.
The GAN loss~\cite{goodfellow2014generative} can be defined as follows:
\begin{multline}
\mathcal{L}_\text{GAN}(G,D) = \\ \mathbb{E}_{y\sim Y}\left[\log D(y)\right] + \mathbb{E}_{x\sim X}\left[\log \left(1-D(G(x))\right)\right].\label{e_baseGAN}
\end{multline}
Through minimizing both losses with respect to $G$, we encourage the generated images to look real and similar to the target images\footnote{\autoref{s_exp} provides a study to compare different designs of $D$.}.
\begin{figure}[t
\centering
\includegraphics[width=1\linewidth, trim={0cm 4cm 0 3cm},clip]{figures/Approach-Figure.png}
\caption{Illustration of our proposed approach: \method{CAPIT}\xspace. We aim to train an image-to-image translation model $G$ using coarsely-aligned image pairs.\label{fig:diagram}}
\vskip -8pt
\end{figure}
\subsection{Learning with Coarsely-Paired Images}
\label{ss_loss}
Learning with the above two loss functions does not explicitly account for the misalignment in the coarsely-aligned image pairs and may adversely degrade the resulting generative model. One obvious solution is to apply unpaired translation algorithms~\cite{CycleGAN2017,huang2018munit}, which, however, entirely ignore any useful paired information in the training data. To address these challenges and drawbacks, we propose novel loss functions tailored to coarsely-aligned image pairs.
\subsubsection{Foreground-Masked Loss}
Our first insight about the coarsely-aligned data is that, unlike stationary backgrounds, foreground objects (\textit{e.g.}\xspace, cars and pedestrians) change in presence and location over time. Thus, two images taken at different timestamps mostly have different foreground contents and layouts, even if they are taken from exactly the same camera pose. To address this issue, we propose to mask out pixels of the foreground objects in calculating any paired loss (\textit{e.g.}\xspace, the L1 loss). For example, we only calculate the L1 loss on pixels that belong to the backgrounds in both the source and the corresponding target images. Let us denote the background pixel indices as $M(x,y)$:
\begin{align}\label{eq:fgmask}
M(x,y)=\{(i,j) | c(x_{i,j}) \not\in \text{fg} \wedge c(y_{i,j})\not\in \text{fg}\},
\end{align}
where $c(\cdot)$ denotes the label of a pixel; fg means the set of foreground classes. We can segment these objects using annotated masks or via off-the-shelf segmentation models~\cite{he2017mask}.
\subsubsection{Misalignment-Tolerating L1 Loss}
The original L1 loss penalizes differences in pixel values between the generated image and the target image at each pixel location.
However, with coarsely-aligned data, the source and target images are unlikely from the exact same camera poses. Thus, a conventional L1 loss would exaggerate even with just a few pixel shifts between the two images (\textit{e.g.}\xspace, along the edges), adversely forcing the generator to fit noises.
To address this problem, we propose to compare $G(x)_{i,j}$ not just to $y_{i,j}$ as in \autoref{e_L1}, but to a set of pixels within a small rectangle window around $i,j$. Specifically, we refer to those pixel indices as
\begin{equation}\label{eq:k}
r(i,j)=\{(a,b): |a-i|\leq k_H \wedge |b-j|\leq k_W \}
\end{equation}
,where $k_H$ and $k_W$ are the size of the rectangular window. With this window, given $x_{i,j}$ we then first find the best corresponding pixel in $y$ to overcome the slight pixel-to-pixel misalignment. The misalignment-tolerating L1 loss can be formulated as:
\begin{multline}
\mahthcal{L}_1^{\dagger}(G) = \frac{1}{N}\sum_n\frac{1}{K^{n}}\sum_{i,j} \min_{(i',j')\in r(i,j)} \lVert G(x^{n})_{i,j}-y^{n}_{i',j'}\rVert_1. \label{e_new_L1}
\end{multline}
\autoref{e_new_L1} can be further combined with the foreground masking defined in \autoref{eq:fgmask} into the final L1 loss:
\begin{align}
& \mathcal{L}_1^{\star}(G) = \nonumber \\
& \mathbb{E}_{(x,y)\sim(X,Y)} \mathbb{E}_{(i,j)\sim M(x,y)} \min_{(i',j')\in \{r(i,j)\cap M(x,y)\}} \lVert G(x)_{i,j}-y_{i',j'}\rVert_1.
\end{align}
\subsubsection{Ranking-based Loss to Overcome Large, Stochastic Variations}
The misalignment-tolerating L1 loss can effectively compensate for small pixel shifts due to different camera poses. However, even if two images are from the same camera pose and the foreground objects are masked out, the fact that the two images are captured at different time implies that there are inevitable \emph{stochastic} factors that could affect the image contents and cause large misalignment. For example, the same tree might be posed differently due to the wind, which in turn can cause different shadow patterns on the sidewalks. These stochastic variations could significantly distort the pixel values, which cannot be alleviated by the misalignment-tolerating L1 loss.
To address these stochastic variations, we propose to employ a paired loss term beyond minimizing the differences at same pixel locations. Concretely, we build upon a key insight --- while there exist stochastic variations, patches of the generated and target images at the same position should capture similar structures and contents, such that their similarity are larger than patches at different positions. We realize this idea using a loss similar to the patchNCE loss \cite{park2020contrastive}\footnote{\cite{park2020contrastive}, being an unpaired translation algorithm, computes the patchNCE loss between the source and translated images. However, we compute the patchNCE loss between the translated and coarsely-aligned target images.}.
Let $H$ be a feature extractor and $H_s$ be the feature at the spatial location $s$ on the feature map, our usage of the patchNCE loss is defined as follows:
\begin{equation}
\begin{split}
\sum_s \ell\left(H_s(G(x)), H_s(y), \{H_{s'}(y) | s'\neq s\}\right), \\
\text{where } \ell(v,v^{+}, \{v^{-}_{s'}\})=\\
-\log\left[\frac{\exp(v\cdot v^{+}/\tau)}{\exp(v\cdot v^{+}/\tau)+\sum_{s'} \exp(v\cdot v^{-}_{s'}/\tau)}\right].
\end{split}
\end{equation}
Here, $\tau$ is a temperature constant scalar. Through minimizing this loss with respect to $G$, we encourage $H_s(G(x))$ and $H_s(y)$ to have higher inner product similarity than $H_s(G(x))$ and other patches of $y$. By applying this loss to all the training pairs, we have
\begin{multline}
\mathcal{L}_\text{patchNCE}(G,H)=\\ \frac{1}{N}\sum_s \ell\left(H_s(G(x^{n})), H_s(y^{n}), \{H_{s'}(y^{n}) | s'\neq s\}\right).
\end{multline}
Our implementation follows~\cite{park2020contrastive} to extract features from multiple layers. Furthermore, we apply $M(x,y)$ to mask out the foreground objects.
\subsection{Final \method{CAPIT}\xspace Training Objective}
Combining the GAN loss with the loss terms introduced in \autoref{ss_loss}, our \method{CAPIT}\xspace training objective is:
\begin{equation}
\mathcal{L}_\text{GAN}(G,D) + \lambda_1 \mathcal{L}_\text{patchNCE}(G,H)+ \lambda_2 \mathcal{L}_1^\star(G).
\end{equation},
where $\lambda_1$ and $\lambda_2$ are weighting coefficients.
Specifically, we learn $G$ and $H$ to minimize the objective, while $D$ to maximize the objective.
\subsection{Implementation and Training Detail}
\method{CAPIT}\xspace is in theory agnostic to different model architectures of the generator $G$ and the discriminator $D$. In this paper, we adopt a coarse-to-fine generator $G$ and a multi-scale discriminator $D$ from the Pix2pixHD \cite{wang2018high}.
For training, we use mini-batch stochastic gradient descent with a learning rate of $0.0002$ for the first $50$ epochs and linearly decrease the learning rate to $0$ in the next $50$ epochs.
We use Adam optimizer~\cite{kingma2014adam} with $\beta_{1}=0.5$ and $\beta_{2}=0.999$.
\iffalse
First, unlike stationary backgrounds and structures, foreground objects (e.g., cars and pedestrians) change in terms of locations and presence over time. Thus, two images taken at different time instances will have different foreground content even if they are taken from exactly the same global pose. To address this issue, we propose to calculate L1 only on pixel locations whose semantic in both input $x$ and coarse target image $y$ are not foreground. Specifically, we refer to the pixel indices as $M(x,y)$:
\begin{equation}
M(x,y)=\{(i,j): 0<i<H \wedge 0<j<W \wedge c(x_{ij}) \not\in \text{fg} \wedge c(y_{ij})\not\in \text{fg}\}
\end{equation}, where $c(\cdot)$ denotes semantic label of a pixel, fg means foreground, and $H,W$ are the height and width dimension of the images.
\fi
\iffalse
Second, while image pairs are matched based on closest GPS poses, they do not come from the same exact pose. Mainly, this is because it is practically impossible to drive on the same exact poses under two different collection drives. Furthermore, GPS also has some inherent error. Thus, image misalignment between the \emph{coarsely}-paired images is inevitable. While the misalignment may be small for some cases, but the impact to L1 may be significant. Concretely, even several pixels shift in image may cause very different change in pixel values, especially on object boundaries. Hence, comparing pixel $G(x)_{ij}$ in the output image to \emph{strictly} $y_{ij}$ in the \emph{coarse} target image may not be appropriate, as pixel $x_{ij}$ may correspond to other nearby pixel $y_{i'j'}$ instead of $y_{ij}$. To this extent, we propose to compare $G(x)_{ij}$ to a set of pixels within a small rectangle window around $y_{ij}$. Specifically, we refer to those pixel indices as:
\begin{equation}
n(ij)=\{(a,b): |a-i|<k_H \wedge |b-j|<k_W \wedge 0<i<H \wedge 0<j<W\}
\end{equation}
where $k_H$ and $k_W$ defines the size of the rectangular window. This is to address the slight pixel-to-pixel correspondence misalignment.
Overall, the final L1 loss can be formally expressed as:
\begin{equation}
\mathcal{L}_1^*(G,X,Y) = \mathbb{E}_{x\sim X}\frac{1}{|M(x,y)|}\sum_{(i,j)\in M(x,y)} \min_{(i',j')\in n(ij)\cap M(x,y)} \lVert G(x)_{ij}-y_{i'j'}\rVert_1
\end{equation}
\fi
\iffalse
\subsubsection{Adapting patchNCE Loss}
patchNCE loss \cite{park2020contrastive}, originally proposed for unpaired image translation, aims to regulate output image to preserve basic structure and content from the input image in a patch-wise way. Specifically, it encourages a patch in the translated image to be more similar (in content) to the corresponding patch in the input image, relative to other patches from the input image. It uses contrastive learning framework, formulated as a $N+1$-way classification problem, where the probability that a query $v$ matches with positive sample $v^{+}$ over negatives \{$v^{-}_{n}\}_{n=1}^N$ is represented by cross-entropy loss:
\begin{equation}
\ell(v,v^{+},v^{-})=-\log\left[\frac{\exp(v\cdot v^{+}/\tau)}{\exp(v\cdot v^{-}/\tau)+\sum_{n=1}^{N}\exp(v\cdot v^{-}_{n}/\tau)}\right]
\end{equation}
where $\tau$ is a temperature constant scalar.
Original patchNCE loss uses features extracted from patches in input image as negative and positive samples. However, since we have \emph{coarsely}-paired images, we can directly use the patchNCE loss between the output image $G(x)$ and its \emph{coarse} coarse correspondence $y$ from the target domain. Unlike L1, small misalignment in image should not degrade the optimality of patchNCE, because contrastive learning does not force strict equality. However, we still need to account for foreground objects difference. Thus $\neg M(x,y)$ pixels are masked in both output and target images for patchNCE loss computation.
Following \cite{park2020contrastive}, features are extracted from $L$ selected layers of interest using the encoder part of our generator $G_\text{enc}$ and two-layer MLP networks.
Specifically, MLP network $H_l$ is used for the output of the $l$-th chosen output layer from $G^l_\text{enc}$, resulting in stacks of features $q_l=H_l(G_\text{enc}^l(G(x)))$ and $w_l = H_l(G_\text{enc}^l(y))$ from output and target images respectively. These features are used for pathNCE loss computation:
\begin{equation}
\mathcal{L}_\text{patchNCE}(G,H,X,Y)=\mathbb{E}_{x\sim X}\sum_{l=1}^{L}\sum_{s=1}^{S_l}\ell(q_{l}^{s},w_{l}^{s},w_{l}^{S\setminus s})
\end{equation}
, where $s \in \{1, \dots, S_l\}$ denotes the spatial location, marking $q_l^s$ and $w_l^s$ as corresponding features, and $q_l^{S\s}$ as features from other non-corresponding locations. This encourages maximum mutual information between the generated image and the reference image. $ q_l^s$ and $w_l^s$ are corresponding features at the $s$-th spatial location and $l$-th layer
\color{red}talk about s and $S/s$\color{black}
\fi
\iffalse
\subsection{Generative Adversarial Loss}
Generative adversarial loss \cite{goodfellow2014generative} is used to encouraged generated images to look similar to images in target domain:
\begin{equation}
\mathcal{L}_\text{GAN}(G,D,X,Y) = \mathbb{E}_{y\sim Y}\left[\log D(y)\right] + \mathbb{E}_{x\sim X}\left[\log \left(1-D(G(x))\right)\right]
\end{equation}
As we employ Pix2pixHD \cite{wang2018high} architecture which contains multi-scale discriminator, the GAN loss above is applied at every scale of the discriminator.
\fi
\section{Proposed Approach: \method{CAPIT}\xspace}
\input{figurestext/pipelinefigure}
\subsection{Problem Setup and Baseline Algorithm}
Our goal is to learn an image mapping function $G$ from a source domain $\mathcal{X}$ (\textit{e.g.}\xspace, adverse weather) to a target domain $\mathcal{Y}$ (\textit{e.g.}\xspace, sunny).
To begin, we introduce a \emph{paired} translation baseline, inspired by pix2pix~\cite{isola2017image}.
\cite{isola2017image} assumes that the training data are perfectly paired images $\{(x^n,y^n)\}_{n=1}^N$, $x^n,y^n\in \mathbb{R}^{h\times w}$, where $x^n \in \mathcal{X}$ corresponds to $y^n \in \mathcal{Y}$. It introduces an L1 loss comparing translated image $G(x^n)$ to its corresponding target image $y^n$ in the pixel space:
\begin{equation}\label{eq:L1original}
\mathcal{L}_{L1}(G) = \mathbb{E}_{x^n,y^n}\lVert G(x^n) - y^n \rVert_1.
\end{equation}
\cite{isola2017image} also employs a conditional GAN loss~\cite{goodfellow2014generative}:
\begin{multline}
\mathcal{L}_\text{cGAN}(G,D) = \mathbb{E}_{x^n,y^n}\left[\log D(y^n;x^n)\right] + \\ \mathbb{E}_{x^n}\left[\log \left(1-D(G(x^n);x^n)\right)\right],\label{e_baseGAN}
\end{multline}
where $D$ denotes a discriminator (\textit{i.e.}\xspace, binary classifier) that tells a real image $y$ from a fake/generated one $G(x)$; the discriminator $D$ is conditioned on $x$.
Through minimizing both losses with respect to $G$, we encourage the generated images to look real and similar to the target images.
\subsection{Learning with Coarsely-Paired Images}
\label{ss_loss}
Obtaining paired images is often infeasible in many scenarios, including autonomous driving, thus prior works typically drop the notion of paired training and utilize the more challenging unpaired training way.
Our method leverages \textit{coarsely-aligned} images, which are much easier to obtain.
Particularly, in autonomous driving, we can obtain \emph{coarse} pairs between $x$ and $y$ if they are collected from the same area, which is common (\textit{e.g.}\xspace, driving to work daily).
The Ithaca365 dataset~\cite{diaz2022Ithaca365} collects images and GPS on the same route under different weather conditions (\textit{e.g.}\xspace, sunny, night, and snowy).
We pair images from different traversals based on the closest GPS pose (see \autoref{fig:intro}-mid for a training data sample). These image pairs are \emph{not perfectly aligned} due to potential GPS differences/errors, changes in semantic layouts at different collection times, and shifts in camera pose.
Learning with the above two loss functions (\autoref{eq:L1original} and \autoref{e_baseGAN}) does not explicitly account for the misalignment in the coarsely-aligned image pairs and may adversely degrade the resulting generative model.
Therefore, instead of directly employing the paired translation algorithms~\cite{isola2017image} or entirely ignoring any useful paired information in the training data by applying an unpaired algorithm~\cite{CycleGAN2017,huang2018munit}, we propose \method{CAPIT}\xspace, Coarsely-Aligned Paired Image Translation, which is dedicated to learning image translation using \textit{coarsely-aligned} image pairs. Specifically, we propose novel loss functions tailored to coarsely-aligned image pairs to address the aforementioned challenges. \autoref{fig:diagram} shows an overview of our approach.
\subsubsection{Foreground-Masked Loss}
Our first insight about the coarsely-aligned data is that foreground objects (\textit{e.g.}\xspace, cars and pedestrians) change in presence and location over time. Thus, two images taken at different timestamps will have different foreground contents and layouts, even if they are taken from the same camera pose. To address this issue, we propose to mask out pixels of the foreground objects in calculating any paired loss. For instance, we calculate the L1 loss only on pixels that belong to the backgrounds in both source and corresponding target images. We denote these background pixel indices as:
\begin{equation}\label{eq:fgmask}
M(x,y)=\{(i,j) | (c(x_{i,j}) \not\in \text{fg}) \wedge (c(y_{i,j})\not\in \text{fg})\},
\end{equation}
where $c(\cdot)$ denotes the class label and fg a set of foreground classes. We can segment these objects using annotated/ground-truth masks or off-the-shelf segmentation models~\cite{he2017mask}.
\subsubsection{Misalignment-Tolerating L1 Loss}
The original L1 loss penalizes differences in pixel values between the generated and target images at each pixel location, assuming that the two images are perfectly aligned.
However, with coarsely-aligned data, the source and target images are unlikely from the same camera poses. Thus, a conventional L1 loss would exaggerate even with a few pixel shifts between the two images (\textit{e.g.}\xspace, along the edges), adversely pushing the generator to fit noises.
To address this problem, we propose to compare pixel $G(x)_{i,j}$ in the translated image
to a set of pixels within a small rectangle window around $i,j$. Specifically, we refer to those pixel indices as
\begin{equation}\label{eq:k}
r(i,j)=\{(a,b): |a-i|\leq k_h \wedge |b-j|\leq k_w \},
\end{equation}
where $k_h$ and $k_w$ are the size of the rectangular window. Given pixel location $i,j$ in the translated image, we first find (within the search window $r(i,j)$) the \emph{best} pixel location $i',j'$ in the target image to overcome the slight pixel-to-pixel misalignment.
The misalignment-tolerating L1 loss can be formulated as:
\begin{equation}
\mathcal{L}_{L1}^{\dagger}(G) =
\mathbb{E}_{x^n,y^n}\frac{1}{hw}\sum_{i,j}\min_{i',j'\in r(i,j)} | G(x^{n})_{i,j}-y^{n}_{i',j'}|. \label{e_new_L1}
\end{equation}
Combining \autoref{e_new_L1} with the foreground masking defined in \autoref{eq:fgmask}, we obtain the final L1 loss:
\begin{multline}
\mathcal{L}_{L1}^{\star}(G) = \mathbb{E}_{x^n,y^n} \frac{1}{\text{card}({M(x,y)})} \\
\sum_{(i,j)\in M(x,y)} \min_{(i',j')\in \{r(i,j)\cap M(x,y)\}} | G(x)_{i,j}-y_{i',j'}|,
\end{multline}
where $\text{card}(\cdot)$ is the cardinality/number of pixels.
\subsubsection{Ranking-based Loss to Overcome Large, Stochastic Variations}
The misalignment-tolerating L1 loss can effectively compensate for small pixel shifts due to slightly different camera poses.
However, two images captured at different times contain inevitable \emph{stochastic} factors that could affect the image contents and cause significant misalignment, even if they are from the same camera pose with masked-out foreground objects.
For example, a tree might be posed differently due to wind, which changes the leaves' position and shadow patterns on sidewalks.
These stochastic variations are not accounted for by the previous two strategies and could significantly distort the pixel values.
Thus, we propose a paired loss term that is more lenient to these stochastic variations.
Concretely, we build upon a key insight: although stochastic variations exist, patches of translated and corresponding target images from the same position should still be more similar (\textit{e.g.}\xspace, similar structures and contents) than patches from a different position.
We realize this idea by \emph{adapting} the patchNCE loss from \cite{park2020contrastive}. Being an unpaired translation method,
\cite{park2020contrastive} computes the patchNCE loss between the source and translated images. Unlike \cite{park2020contrastive}, we compute the patchNCE loss between the translated and coarsely-aligned target images, leveraging our coarsely-aligned data.
Let $H$ be a feature extractor that reuses the encoder of the translation model $G$, followed by a two-layer MLP; and $H_s^l$ be the extracted feature at a spatial location $s$ and layer $l$. Our patchNCE loss is:
\begin{multline}
\mathcal{L}^*_\text{patchNCE}(G,H) =
\mathbb{E}_{x^n,y^n} \\ \sum_{l=1}^{L}\sum_s \ell\left(H_s^l(G(x^{n})), H_s^l(y^{n}), \{H_{s'}^l(y^{n}) | s'\neq s\}\right),
\end{multline}
where $\ell$ is a cross-entropy loss to encourage the query $v=H_s^l(G(x^n))$ to match (more similar) with the positive/correct sample $v^\text{+}=H_s^l(y^n)$ over negative samples $\{v^\text{-}\}=\{H_{s'}^l(y^{n}) | s'\neq s\}$:
\begin{equation}
\ell(v,v^\text{+}, \{v^\text{-}\})=\\
-\log \frac{\exp(v\cdot v^\text{+}/\tau)}{\exp(v\cdot v^\text{+}/\tau)+\sum_{\{v^\text{-}\}} \exp(v\cdot v^\text{-}/\tau)}.
\end{equation}
Here, $\tau$ is a temperature constant scalar.
Furthermore, we apply $M(x,y)$ to mask out the foreground objects on the patchNCE computation.
\subsection{Final \method{CAPIT}\xspace Training Objective}
Our final \method{CAPIT}\xspace training objective combines the loss terms introduced in \autoref{ss_loss} with an unpaired-GAN (uGAN) loss~\cite{CycleGAN2017}:
\begin{equation}
\mathcal{L}_\text{CAPIT} = \mathcal{L}_\text{uGAN}(G,D) + \lambda_1 \mathcal{L}_{L1}^\star(G) + \lambda_2 \mathcal{L}_\text{patchNCE}^*(G,H),
\end{equation}
where $\lambda_1,\lambda_2\in \mathbb{R}^+$ are weighting coefficients.
While it is natural to follow \cite{isola2017image} to use conditional-GAN (cGAN) loss, we \emph{empirically} find that unpaired-GAN (uGAN) loss \cite{CycleGAN2017} results in better performance (see \autoref{sec:ablation} and \autoref{sec:exp-GANloss}); thus we employ uGAN loss.
We learn $G$ and $H$ to minimize the objective, while $D$ to maximize the objective.
\method{CAPIT}\xspace is in theory \emph{agnostic} to different model architectures (\textit{i.e.}\xspace, $G$ and $D$). In this paper, we adopt a coarse-to-fine generator and a multi-scale discriminator from \cite{wang2018high}.
\iffalse
First, unlike stationary backgrounds and structures, foreground objects (e.g., cars and pedestrians) change in terms of locations and presence over time. Thus, two images taken at different time instances will have different foreground content even if they are taken from exactly the same global pose. To address this issue, we propose to calculate L1 only on pixel locations whose semantic in both input $x$ and coarse target image $y$ are not foreground. Specifically, we refer to the pixel indices as $M(x,y)$:
\begin{equation}
M(x,y)=\{(i,j): 0<i<H \wedge 0<j<W \wedge c(x_{ij}) \not\in \text{fg} \wedge c(y_{ij})\not\in \text{fg}\}
\end{equation}, where $c(\cdot)$ denotes semantic label of a pixel, fg means foreground, and $H,W$ are the height and width dimension of the images.
\fi
\iffalse
Second, while image pairs are matched based on closest GPS poses, they do not come from the same exact pose. Mainly, this is because it is practically impossible to drive on the same exact poses under two different collection drives. Furthermore, GPS also has some inherent error. Thus, image misalignment between the \emph{coarsely}-paired images is inevitable. While the misalignment may be small for some cases, but the impact to L1 may be significant. Concretely, even several pixels shift in image may cause very different change in pixel values, especially on object boundaries. Hence, comparing pixel $G(x)_{ij}$ in the output image to \emph{strictly} $y_{ij}$ in the \emph{coarse} target image may not be appropriate, as pixel $x_{ij}$ may correspond to other nearby pixel $y_{i'j'}$ instead of $y_{ij}$. To this extent, we propose to compare $G(x)_{ij}$ to a set of pixels within a small rectangle window around $y_{ij}$. Specifically, we refer to those pixel indices as:
\begin{equation}
n(ij)=\{(a,b): |a-i|<k_H \wedge |b-j|<k_W \wedge 0<i<H \wedge 0<j<W\}
\end{equation}
where $k_H$ and $k_W$ defines the size of the rectangular window. This is to address the slight pixel-to-pixel correspondence misalignment.
Overall, the final L1 loss can be formally expressed as:
\begin{equation}
\mathcal{L}_1^*(G,X,Y) = \mathbb{E}_{x\sim X}\frac{1}{|M(x,y)|}\sum_{(i,j)\in M(x,y)} \min_{(i',j')\in n(ij)\cap M(x,y)} \lVert G(x)_{ij}-y_{i'j'}\rVert_1
\end{equation}
\fi
\iffalse
\subsubsection{Adapting patchNCE Loss}
patchNCE loss \cite{park2020contrastive}, originally proposed for unpaired image translation, aims to regulate output image to preserve basic structure and content from the input image in a patch-wise way. Specifically, it encourages a patch in the translated image to be more similar (in content) to the corresponding patch in the input image, relative to other patches from the input image. It uses contrastive learning framework, formulated as a $N+1$-way classification problem, where the probability that a query $v$ matches with positive sample $v^{+}$ over negatives \{$v^{-}_{n}\}_{n=1}^N$ is represented by cross-entropy loss:
\begin{equation}
\ell(v,v^{+},v^{-})=-\log\left[\frac{\exp(v\cdot v^{+}/\tau)}{\exp(v\cdot v^{-}/\tau)+\sum_{n=1}^{N}\exp(v\cdot v^{-}_{n}/\tau)}\right]
\end{equation}
where $\tau$ is a temperature constant scalar.
Original patchNCE loss uses features extracted from patches in input image as negative and positive samples. However, since we have \emph{coarsely}-paired images, we can directly use the patchNCE loss between the output image $G(x)$ and its \emph{coarse} coarse correspondence $y$ from the target domain. Unlike L1, small misalignment in image should not degrade the optimality of patchNCE, because contrastive learning does not force strict equality. However, we still need to account for foreground objects difference. Thus $\neg M(x,y)$ pixels are masked in both output and target images for patchNCE loss computation.
Following \cite{park2020contrastive}, features are extracted from $L$ selected layers of interest using the encoder part of our generator $G_\text{enc}$ and two-layer MLP networks.
Specifically, MLP network $H_l$ is used for the output of the $l$-th chosen output layer from $G^l_\text{enc}$, resulting in stacks of features $q_l=H_l(G_\text{enc}^l(G(x)))$ and $w_l = H_l(G_\text{enc}^l(y))$ from output and target images respectively. These features are used for pathNCE loss computation:
\begin{equation}
\mathcal{L}_\text{patchNCE}(G,H,X,Y)=\mathbb{E}_{x\sim X}\sum_{l=1}^{L}\sum_{s=1}^{S_l}\ell(q_{l}^{s},w_{l}^{s},w_{l}^{S\setminus s})
\end{equation}
, where $s \in \{1, \dots, S_l\}$ denotes the spatial location, marking $q_l^s$ and $w_l^s$ as corresponding features, and $q_l^{S\s}$ as features from other non-corresponding locations.
This encourages maximum mutual information between the generated image and the reference image. $ q_l^s$ and $w_l^s$ are corresponding features at the $s$-th spatial location and $l$-th layer
\color{red}talk about s and $S/s$\color{black}
\fi
\section{Related Work}
Early image-to-image translation works focus on paired translation methods, where they apply losses that leverage the strong one-to-one correspondence between input images from the source domain and corresponding images in the target domain to guide the image translation.
For example, \cite{isola2017image} applies GAN \cite{goodfellow2014generative} in a conditional setting as well as L1 loss between the input and paired target images.
\cite{wang2018high} further adds feature matching and perceptual loss and employs a multi-scale generator and discriminator.
\cite{Park_2019_CVPR} proposes a spatially-adaptive normalization layer to improve the fidelity and alignment of translated images with the input layout.
\cite{PixelNN} combines GAN with a classic nearest-neighbor approach to generate high resolution image.
However, obtaining perfectly-aligned paired images is often challenging, if not impossible, in many cases (such as autonomous driving). Thus, the loss functions and training framework designed for paired image translation cannot be applied.
Another line of work focuses on unpaired image translation, where image correspondence is not required. For example, CycleGAN~\cite{CycleGAN2017}, ToDayGAN~\cite{anoosheh2019night}, and DiscoGAN~\cite{kim2017learning} use cycle-consistency loss to enforce consistency between the original image and the reconstructed image after one translation cycle. Council-GAN~\cite{nizan2020council} uses a council loss to achieve similar goal.
While cycle-consistency loss is a widely used constraint, ACL-GAN~\cite{zhao2020aclgan} proposes to use adversarial-consistency loss to encourage the translated image to retain important information from the source image without the need of cycle translation (\textit{i.e.}\xspace, translated image translated back to source domain).
CUT~\cite{park2020contrastive} proposes a patchNCE loss that employs a contrastive learning technique~\cite{gutmann2010noise} to maximize mutual information between input and output (translated) images.
Other methods, such as MUNIT~\cite{huang2018munit}, DRIT~\cite{DRIT_plus}, UNIT~\cite{liu2017unsupervised}, and ForkGAN~\cite{zheng2020forkgan} focus on learning domain agnostic features to extract shared features between input and output domains.
Although unpaired image translation methods allow training without paired images, the unpaired setup leads to an under-constrained problem and inferior results compared to paired methods.
Our proposed work \method{CAPIT}\xspace, Coarsely-Aligned Paired Image Translation, offers an intermediate solution by leveraging \emph{coarsely-aligned} image pairs that are easily obtainable for autonomous driving. To our knowledge, the closest work to our proposed \method{CAPIT}\xspace is \cite{porav2018adversarial}.
Our work \method{CAPIT}\xspace significantly differs from \cite{porav2018adversarial} in that, we propose a framework dedicated to learning image translation using coarsely-aligned image pairs, while the main training in \cite{porav2018adversarial} comes from the unpaired cycle-consistency training.
\cite{porav2018adversarial} requires \emph{manually} selecting \emph{well-aligned} images to fine-tune the image translation model after the main unpaired training stage. During fine-tuning, \cite{porav2018adversarial} immediately treats those images with \emph{paired} translation loss and does not address possible misalignment factors, which we show in \autoref{ss_loss} and \autoref{sec:ablation} can adversely degrade the resulting generative model.
While some downstream tasks may require removing only the \emph{scattering medium} (\textit{e.g.}\xspace snowfall and raindrop)~\cite{shi2021zeroscatter} from adverse images, adapting whole street scenes (\textit{e.g.}\xspace, removing snow from both the street and cars) can generally produce images that are closer to the target domain and benefit a variety downstream adaptation tasks that may include backgrounds, such as semantic segmentation and depth estimation.
We reiterate that our contribution in leveraging coarsely-aligned image pair translation is \textit{orthogonal} and \textit{complimentary} to existing novel image translation methods and architectures, and can even improve many of them.
Another line of work is improving downstream tasks (\textit{e.g.}\xspace semantic segmentation) in
adverse weather by directly adapting task networks~\cite{pfeuffer2019robust,naseer2018robust,sharma2020nighttime}; this allows task-specific information to be used to tailor the approach to the task.
However, this often requires task labels (\textit{e.g.}\xspace, semantic mask in the ACDC dataset~\cite{sakaridis2021acdc}), which is often challenging or infeasible to annotate under adverse conditions. Adapting through image translation, on the other hand, does not need additional labels. Moreover, image translation immediately bridges the domain gap for \emph{all} image-based downstream tasks (\textit{e.g.}\xspace, semantic segmentation, visual localization).
|
\section{Introduction}
In recent years, artificial intelligence (AI) has become a powerful tool for bioimage analysis.
A major roadblock for the application of AI is obtaining manually annotated training data, which requires human experts to label image features on a computer screen, a tedious process prone to a high error rate.
Standard approaches for labelling 3D objects use 2D cross-sections to view and annotate the volumes,
which, because single 3D objects can appear disjointed in 2D sections, can yield low accuracy while requiring excessive input from the annotator.
We aim to address this issue with our software, MiCellAnnGELo (\textbf{Mi}croscopy and \textbf{Cell} \textbf{Ann}otation \textbf{G}raphical \textbf{E}xperience and \textbf{L}abelling to\textbf{o}l), an easy-to-use virtual reality (VR) interface for annotating dynamic biological surfaces.
The use of VR for data visualization dates back over 25 years \citep[for example][]{fruhauf1996}.
Recent advances in graphics processing unit (GPU) and VR headset technology have greatly expanded the range of applications of this technology.
ConfocalVR~\citep{stefani2018},
Arivis VisionVR~\citep{conrad2020}, and syGlass~\citep{pidhorskyi2018} are examples of VR software for annotating 3D biological image volumes.
SlicerVR~\citep{pinter2020} is a VR visualization plugin for the open-source software 3D~Slicer~\citep{fedorov2012}.
TeraVR~\citep{wang2019} provides a specialized VR application for neuron tracing in 3D image volumes, with functionality for placing markers and surface visualization, as part of the open-source software Vaa3D~\citep{peng2010}.
ChimeraX~\citep{pettersen2021} is primarily focused on molecular visualization and analysis, but has also been applied to biological image analysis and surface visualization~\citep{driscoll2019,quinn2021}.
Our software is differentiated from the above in a number of ways.
MiCellAnnGELo allows users to interact with time series of dual-channel colour surfaces, a functionality
shared only in the above list with ChimeraX.
While most have some form of marker placement tool, ``mesh painting'' for fast multi-vertex labelling of time series is not offered by any of the above.
We are releasing our software as an open-source tool, which further distinguishes it from most of the above list.
Finally, MiCellAnnGELo is streamlined for VR visualization and annotation, making it easier to use than the tools listed above.
\section{Software Features}
\label{sec:software}
MiCellAnnGELo is a cross-platform software program that provides an immersive environment for the rapid annotation of series of triangulated surface meshes.
Meshes with single- or dual-channel colour mappings can be displayed and annotated in the environment.
The software provides both VR and desktop interfaces, with easy interchange between these modes.
\begin{figure*}
\centering
\begin{subfigure}{0.24\linewidth}
\includegraphics[width=\linewidth]{figureSubFigs/sceneLabeledbw.png}
\end{subfigure}
\begin{subfigure}{0.24\linewidth}
\includegraphics[width=\linewidth]{figureSubFigs/uiLabeledbw.png}
\end{subfigure}
\begin{subfigure}{0.24\linewidth}
\includegraphics[width=\linewidth]{figureSubFigs/markerZoomedLabeledbw.png}
\end{subfigure}
\begin{subfigure}{0.24\linewidth}
\includegraphics[width=\linewidth]{figureSubFigs/paintOutlinesLabeledbw.png}
\end{subfigure}
\\
\begin{subfigure}{0.24\linewidth}
\includegraphics[width=\linewidth]{figureSubFigs/scalesmallLabeledbw.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\linewidth]{figureSubFigs/insideLabeledbw.png}
\end{subfigure}
\begin{subfigure}{0.24\linewidth}
\includegraphics[width=\linewidth]{figureSubFigs/transparencyLabeledbw.png}
\end{subfigure}
\begin{subfigure}{0.24\linewidth}
\includegraphics[width=\linewidth]{figureSubFigs/colourLabeledbw.png}
\end{subfigure}
\caption{
Main features of the MiCellAnnGELo software interface.
\textbf{a}:~The streamlined environment, containing a wall-mounted UI and dual-color surface mesh.
\textbf{b}:~Close up of the wall-mounted UI for loading and saving data and adjusting surface color and opacity.
\textbf{c}:~Placement of a marker (red arrow) using the laser pointer.
\textbf{d}:~Mesh painting in the two-tone annotation mode using the size-adjustable paint tool (green outline).
\textbf{e}:~Adjustments to the position and size of the surface can be made using the VR controllers.
\textbf{f}:~Surfaces can also be viewed and annotated from inside the surface.
\textbf{g}:~Transparency of part of the surface can be adjusted using the laser pointer.
\textbf{h}:~Color threshold can be used to highlight intensity-dependent features.
}
\label{fig1}
\end{figure*}
The user interface is designed with simplicity in mind, allowing the user to load, explore, and annotate data with ease.
As can be seen in Fig~\ref{fig1}\textbf{a}, the environment consists of the surface mesh itself, and a wall-mounted user interface (Fig~\ref{fig1}\textbf{b}), which provides controls for loading and saving data, and for adjusting surface colour and opacity.
Additionally, controller layouts for both VR and desktop modes are displayed on the wall for ease of use.
In both VR and desktop environments, functions including changing annotation modes or surface representations, moving forwards/backwards in time, and playing/pausing the series are mapped to single buttons, enabling rapid multi-modal annotation of a sequence.
Annotations can be made by placing markers (Fig~\ref{fig1}\textbf{c}) and by mesh painting (Fig~\ref{fig1}\textbf{d}).
Marker placement allows features
to be marked with a single shot per frame,
allowing rapid manual tracking of surface features across multiple frames.
Painting allows rapid labelling of larger structures on the surface, with an easily adjustable brush size and eraser mode.
The painted surface can be visualized either as a two-tone image (Fig~\ref{fig1}\textbf{d}) or as a cut-out image with mesh colours being blocked out in unlabelled areas.
A number of view controls have been added to enable easier labelling of data.
In VR mode, the user can freely adjust the surface mesh position and size (Fig~\ref{fig1}\textbf{e}), allowing the user to rapidly change perspectives.
For annotating more complex surfaces,
the user can
annotate from inside the surface (Fig~\ref{fig1}\textbf{f}), and
also reduce the opacity of parts of the surface (Fig~\ref{fig1}\textbf{g}).
Finally,
the mesh colours
can be adjusted using controls on the wall UI (Fig~\ref{fig1}\textbf{h}), allowing increased visibility of colour-dependent features.
Surface meshes can be loaded into the environment from a sequence of .ply files.
This format stores colour data, and can be generated from surfaces in many image analysis software applications.
MiCellAnnGELo takes the first two channels of the .ply files as colour data.
Marker annotations are exported as .csv files encoding the frame number, spatial coordinates, and vertex index in the surface mesh of each marker.
Painted surfaces are exported as .ply files, keeping the original colours in the first two channels and placing the paint labels in the third channel, and can
be retrieved in future sessions by selecting this sequence in the load menu.
\section{Conclusion}
MiCellAnnGELo is a cross-platform open-source software application to allow rapid annotation of surface series using VR technology.
The software is streamlined and easy to use, with a range of viewing and annotation options.
In future developments we aim to introduce functionality for 3D volumetric data and multi-object series.
We aim to develop this software as a community resource project, allowing further development to be geared towards providing a VR software solution for a wide range of microscopy use-cases and easy integration into existing pipelines.
\section*{Acknowledgements}
\textbf{Funding:}
This work was supported by Biotechnology and Biological Sciences Research Council [BB/R004579/1 to T.B.] and Engineering and Physical Sciences Research Council [EP/V062522/1 to E.J.L, T.B.].
|
\section{Introduction}
In recent years, Deep Neural Networks (DNNs) have become the preferred method for solving many computer vision (CV) problems~\cite{CHAI2021100134}.
At the same time,
the complexity in both memory and floating point operations of the aforementioned models has grown tremendously~\cite{growth}. That growth, alongside the increasing variety of CV models and their uses, poses a challenge for deployment to low-resource edge devices such as mobile phones, smart speakers, etc. For this reason, in many CV tasks, the most common approach today~\cite{kang2017neurosurgeon} is to simply avoid deploying resource-intensive models to edge devices. Instead, the bulk of the computation is performed on remote servers (the cloud) equipped with powerful graphics processing units (GPU) or tensor processing units (TPU).
Transmitting an image for inference on the cloud using traditional coding methods, which have been designed to accommodate human perception, has been shown to be sub-optimal both theoretically~\cite{hyomin} and empirically~\cite{hyomin,kang2017neurosurgeon}. Collaborative intelligence (CI)~\cite{kang2017neurosurgeon} suggests dividing a given CV model to a frontend, deployed on the edge, and backend, which remains on the cloud. Inference can then be started on the edge, creating intermediate representations at some layer $l$ of the CV model, which we denote $\mathcal{F}^{(l)}$.
These are then sent to the cloud backend, where inference is finished. Of course, this approach requires defining an efficient method for the communication of said representations (generally floating point tensors) to the cloud, which has been explored in several subsequent works~\cite{saeed_multi_task_learning,Choi2018NearLosslessDF,choi2018deep}.
Inspired by CI, the work of \cite{hyomin} proposes replacing the CV frontend with an encoder from a learnable image coding backbone~\cite{cheng2020learned}. This replaces the floating point tensors with highly compressible latent features, which are named the \emph{base} representation. On the decoder side, these can be used to match the intermediate features for the CV backend, which can then perform inference. The matching of intermediate CV features from the compressible latent space is performed by a learned transform named the \emph{Latent Space Transform} (LST). An important aspect of~\cite{hyomin}, is that it allows for human vision of the analysed image in a scalable manner. The \emph{base} representation is combined with a second representation, named \emph{enhancement}, to reconstruct the original image on the cloud side. In terms of rate-distortion, this method outperforms traditional as well as learnable coding schemes for the CV task, without significant degradation with regards to reconstruction for humans.
In this work we further investigate the behaviour of ~\cite{hyomin} in the context of both object detection and input reconstruction. We provide theoretical proof that performing feature matching on deeper layers of a CV model is beneficial
in the sense of rate-distortion. Based on our proofs, we propose changes to the training loss of~\cite{hyomin} to improve rate-distortion behaviour for object detection. Using the proposed method, we achieve superior rate-distortion performance for the object detection task, which we consider to be empirical evidence of the theory presented in our proofs. Finally, we explore the trade-off in rate-distortion performance between computer and human vision in scalable compression, and discuss the consequences of our approach in that regard.
\section{Background}
\label{sec:background}
\subsection{Feature Matching in Collaborative Intelligence}
\label{subsec:man-machine}
When splitting a model for use in CI, it is important to distinguish between models (known as single-stream) where layer operations are performed purely sequentially, and ones (known as multi-stream) where layers are combined in complex architectures including skip-connections and multi-resolution computations~\cite{Redmon2018_yolov3, ren2015faster, he2017mask}.
Often, a CV model will be comprised of some single-stream layers followed by a multi-stream section. We consider the output of a certain layer, $\mathcal{F}^{(l)}$, to be a single stream feature if we can perform inference by obtaining $\mathcal{F}^{(l)}$ from the input, and then passing it on to the rest of the model. Formally, we require $T = G_l(\mathcal{F}^{(l)})$, where $T$ is the CV task and $G_l(\cdot)$ is the operation described by all subsequent layers after $l$. Conversely, if the output of the task model cannot be inferred solely from $\mathcal{F}^{(l)}$, we consider it a multi-stream feature.
In collaborative intelligence, the intermediate features produced on the edge device must be transmitted (or recreated from a latent representation) to be used downstream by the CV backend. Formally, we need to match, on the server side, the features at some small subset $K \ll L$, of the model's $L$ layers
$\mathcal{F}^{(l_k)}$ where $k = 1,2,..,K$. Following our definitions above, we see that in the case $K=1$, we require $\mathcal{F}^{(l_1)}$ to be a single-stream feature. In most multi-stream models, this is often only achievable by selecting a relatively early layer. Unfortunately, as we will prove below, matching earlier features comes at a cost in terms of rate-distortion performance.
\subsection{Compression and Rate-Distortion Analysis}
Compression, in general, can be divided into lossless compression, where a source random variable (RV) $\mathbf{X}$ is reconstructed perfectly after decompressing; and lossy compression, where after decoding we are left with an approximation $\widehat{\mathbf{X}}$ (approximations or quantizations are denoted with a hat operator). For a single pair of observations $(\mathbf{x},\hat{\mathbf{x}})$, we measure the amount of inaccuracy introduced by the approximation using some distortion metric $d(\mathbf{x},\hat{\mathbf{x}})$. This leads to an important concept in lossy compression known as the \emph{rate-distortion function} $R(D)$~\cite{Cover_Thomas_2006}, given by the following:
\begin{equation}
R(D) = \min_{p(\hat{\mathbf{x}}|\mathbf{x})~:~\mathbb{E} \left[d(\mathbf{X},\widehat{\mathbf{X}})\right]~\leq~D} I(\mathbf{X};\widehat{\mathbf{X}}).
\label{eq:RD_function}
\end{equation}
Here, $p(\hat{\mathbf{x}}|\mathbf{x})$ is the conditional distribution of the approximation given the source; $\mathbb{E}[d(\mathbf{X},\widehat{\mathbf{X}})]$ is the expected distortion with respect to the joint distribution $p(\mathbf{x},\hat{\mathbf{x}})$); $I(\mathbf{X};\widehat{\mathbf{X}})$ is the mutual information between $\mathbf{X}$; and $\widehat{\mathbf{X}}$; and $D$ is some value of distortion.
Because the marginal distribution of the source, $p(\mathbf{x})$ is fixed, the joint distribution, $p(\mathbf{x},\hat{\mathbf{x}})$ only changes through the conditional distribution of the approximation $p(\hat{\mathbf{x}}|\mathbf{x})$.
Using that, we can understand the minimization in the rate distortion function as finding an approximation $\widehat{\mathbf{X}}$ that gives the lowest mutual information with the source $\mathbf{X}$ while allowing an expected distortion no greater than $D$. Then, using the source coding theorem~\cite{Cover_Thomas_2006} and its converse, it can be shown that the resulting bit-rate\footnote{This assumes we calculate mutual information using $\log_2(\cdot)$.} is the lowest achievable rate (per source symbol) giving an expected distortion bound by $D$. In this sense, $R(D)$ is a fundamental bound on the performance of lossy compression, similarly to entropy being a bound on lossless compression.
\section{Theoretical Discussion}
As explained in Section~\ref{subsec:man-machine}, a useful approach for communicating an image to a remote-server for some downstream CV task it to perform feature-matching on some intermediate layer of a DNN model. In fact, \cite{hyomin} provided theoretical and experimental proof that feature matching is preferable, in terms of rate distortion, to transmitting an entire image to the cloud. We build on this proof and show that the rate-distortion function when matching deeper layers is lower (better) compared to matching earlier ones. To provide this proof we first need to define some notation. Let our CV model, with a task output previously denoted $T$ be defined as $f(\cdot)$, so that for a given input $\mathbf{X}$ we have $T=f(\mathbf{X})$. Next, let the mapping from the input $\mathbf{X}$ to a set of intermediate features $\mathbfcal{Y}_1\equiv\mathcal{F}^{(l_1)}$ be $g_1(\cdot)$ so that $\mathbfcal{Y}_1 = g_1(\mathbf{X})$, and the mapping from $\mathbfcal{Y}_1$ to the output $T$ be $h_1(\cdot)$ so that $T=h_1(\mathbfcal{Y}_1)$. Next, define a second, deeper set of intermediate features $\mathbfcal{Y}_2\equiv\mathcal{F}^{(l_2)}$ (with $l_2>l_1$) and mappings $g_2(\cdot)$ and $h_2(\cdot)$ such that $\mathbfcal{Y}_2 = g_2(\mathbfcal{Y}_1)$ and $T = h_2(\mathbfcal{Y}_2)$. Note that under this notation, $f = h_1\circ g_1$ and $h_1 = h_2\circ g_2$. This notation is illustrated graphically below.
\begin{singlespace}
\begin{equation*}
\begin{tikzpicture}[
inner sep=0pt,
outer sep=0pt,
baseline=(x.base),
]
\tikzstyle{arrow} = [thin,->,>=stealth]
\path[every node/.append style={anchor=base west}]
(0, 0)
\foreach \name/\code in {
x/ \mathbf{X}~,
tmp/\,\qquad,
y1/ ~\mathbfcal{Y}_1~,
tmp/\,\qquad,
y2/ ~\mathbfcal{Y}_2~,
tmp/\,\qquad,
t/ ~
} {
node (\name) {$\code$}
(\name.base east)
}
;
\path[
every node/.append style={
anchor=base,
font=\scriptsize,
},
]
(x.base) -- node[above=3.0\baselineskip] (f) {\small $f$} (t)
(x.base) -- node[below=0.2\baselineskip] (g1) {\small $g_1$} (y1)
(y1.base) -- node[above=1.8\baselineskip] (h1) {\small $h_1$} (t)
(y1.base) -- node[below=0.2\baselineskip] (g2) {\small $g_2$} (y2)
(y2.base) -- node[below=0.2\baselineskip] (h2) {\small $h_2$} (t)
;
\draw [arrow] (x) -- (y1);
\draw [arrow] (y1) -- (y2);
\draw [arrow] (y2) -- (t);
\begin{scope}[
>={Stealth[length=5pt]},
thin,
rounded corners=2pt,
shorten <=.3em,
shorten >=.3em,
]
\def\GebArrow#1#2#3{
\draw[->]
(#2.south) ++(0, -.3em) coordinate (tmp)
(#1) |- (tmp) -| (#3)
;%
}
\GebArrow{x}{f}{t}
\GebArrow{y1}{h1}{t}
\end{scope}
\end{tikzpicture}
\label{eq:mappings}
\end{equation*}
\end{singlespace}
In the case of compression for machines, we are mainly interested in the output of our CV task, and thus we measure the distortion at $T$. By denoting this as $d_T$, we have:
\begin{equation}
\begin{aligned}
\tilde{d}(\mathbf{x},\hat{\mathbf{x}}) = & d_T\left(f(\mathbf{x});f(\hat{\mathbf{x}})\right), \\
\tilde{d}(\mathbf{y}_1,\hat{\mathbf{y}}_1) = & d_T\left(h_1(\mathbf{y}_1),h_1(\hat{\mathbf{y}}_1)\right), \\
\tilde{d}(\mathbf{y}_2,\hat{\mathbf{y}}_2) = & d_T\left(h_2(\mathbf{y}_2),h_2(\hat{\mathbf{y}}_2)]\right).
\end{aligned}
\label{eq:distortion}
\end{equation}
It is important to note that the choice of distortion metric $d_T(\cdot,\cdot)$ depends on the CV task, and might generally differ from the metrics used for human vision. Next, we define the set of all approximations and their respective conditional distributions, which achieve a task distortion of at most $D$ as:
\begin{equation}
\mathcal{P}_{\mathbf{X}}(D)=\left\{p\left(\hat{\mathbf{x}}|\mathbf{x}\right)~:~\mathbb{E}\left[\tilde{d}\big(\mathbf{X},\widehat{\mathbf{X}}\big)\right]\leq D\right\}.
\label{eq:quant_X}
\end{equation}
Using~(\ref{eq:distortion}), we can obtain equivalent formulations for $\mathcal{P}_{\mathbfcal{Y}_1}(D)$ and $\mathcal{P}_{\mathbfcal{Y}_2}(D)$. This notation allows us to rewrite~(\ref{eq:RD_function}) as:
\begin{equation}
R_{\mathbf{X}}(D) = \min_{p(\hat{\mathbf{x}}|\mathbf{x})~\in~\mathcal{P}_{\mathbf{X}}(D)} I(\mathbf{X};\widehat{\mathbf{X}}).
\end{equation}
Similarly, we can write equivalent formulations for $R_{\mathbfcal{Y}_1}(D)$ and $R_{\mathbfcal{Y}_2}(D)$.
We are now ready to state our first result.
\begin{theorem}
\label{thm:lossy}
For any distortion level $D\geq 0$, the minimum achievable rate for compressing $\mathbfcal{Y}_1$ is an upper bound to the minimum achievable rate for compressing $\mathbfcal{Y}_2$, that is,
\begin{equation}
R_{\mathbfcal{Y}_2}(D) \leq R_{\mathbfcal{Y}_1}(D).
\label{eq:RD_bound}
\end{equation}
Equality occurs when $\mathbfcal{Y}_1$ can be recreated exactly from $\mathbfcal{Y}_2$.
\end{theorem}
\begin{proof}
In~\cite{hyomin}, the authors prove that for any one intermediate layer $\mathbfcal{Y}_1 = g_1(\mathbf{X})$, we have $R_{\mathbfcal{Y}_1}(D) \leq R_\mathbf{X}(D)$.
To prove our theorem, we simply need to show that we can replace $\mathbf{X}$ with $\mathbfcal{Y}_1$, as well as replace $\mathbfcal{Y}_1$ with $\mathbfcal{Y}_2$, all the while maintaining the conditions of the proof.
First, we recall that the only condition on the source $\mathbf{X}$ is that it must have a fixed distribution $p(\mathbf{x})$ in the sense that it does not depend on the approximation (or quantization) process. However, since $\mathbfcal{Y}_1$ is directly computed from $\mathbf{X}$, its own distribution, which we denote $p_1(\mathbf{y}_1)$, induced by $\mathbfcal{Y}_1 = g_1(\mathbf{X})$ is also fixed (because $g_1$ is fixed).
To finish, we want to show that the structure relating $\mathbf{X}$, $\mathbfcal{Y}_1$, and $T$, is maintained between $\mathbfcal{Y}_1$, $\mathbfcal{Y}_2$, and $T$. However, because there are no special requirements on the functions $f,g_1,h_1$ in the original formulation (not even that they are deterministic), this is trivial. To do this, we can simply replace the function $f$ with $h_1$, $g_1$ with $g_2$, and $h_1$ with $h_2$, thus concluding our proof.
\end{proof}
In~\cite{hyomin}, the authors define a two-layer\footnote{The word layer here refers to the base and enhancement portion of the model, and not to DNN layers.} network where the \emph{base} representation is used to compress an input for object detection by a YOLOv3 vision backend \cite{Redmon2018_yolov3}, and the \emph{enhancement} is used (together with \emph{base}) for input reconstruction. To compress the image for object detection, the authors perform feature matching on the 13\textsuperscript{th} layer of YOLOv3,
because $\mathcal{F}^{(13)}$ is the deepest single-stream feature in YOLOv3.
In their three-layer model, the authors in~\cite{hyomin} utilize the \emph{base} and first \emph{enhancement} layers to perform feature matching on more complex architectures of Faster R-CNN~\cite{ren2015faster} and Mask R-CNN~\cite{he2017mask}, which are used for object detection and segmentation, respectively. There, because of the multi-stream structure of a feature-pyramid network (which is part of both R-CNN architectures), they use an alternative approach to feature matching. In this method, the target feature is reconstructed in two steps. First, using an LST, a single-stream feature $\mathcal{F}^{(l_0)}$ is reconstructed for some small $l_0$. Next, using pre-trained portions of the CV model (denoted $H_k$, and collectively named the CV mid-model), several multi-stream features $\mathcal{F}^{(l_k)} = H_k(\mathcal{F}^{(l_0)}), k=1,2,..., K$ are recreated. For clarity, when using this approach we refer to $l_0$ as the partition point~\cite{kang2017neurosurgeon}, and $\mathcal{F}^{(l_k)}, k=1,2,..., K$ as downstream features.
Evaluating distortion in the manner used in Theorem~\ref{thm:lossy}, is equivalent to using the two-step approach to reconstructing the task output $T$, from the intermediate features $\mathbfcal{Y}_1$ or $\mathbfcal{Y}_2$. Using this distortion during training is impractical because it would require a labeled dataset of uncompressed images for each of the desired tasks. Unfortunately, datasets containing uncompressed images with CV task labels are rare,
making supervised training of a scalable model impractical. Instead, both models in~\cite{hyomin} use mean squared error (MSE) of the reconstructed features to evaluate distortion during training. We will prove next, that the use of deeper layers in this practical setting is still beneficial.
To evaluate the effect of measuring distortion on intermediate features, we need to extend the previously defined notation. First, we define the distortion measured at some variable $\mathbfcal{Z}$ as $d_\mathbfcal{Z}(\mathbf{z},\hat{\mathbf{z}})$. Next, the set of all (including two-step) reconstructions of \emph{$\mathbfcal{Y}_2$ from $\mathbfcal{Y}_1$} that achieve a distortion of no more than $D$ (as measured at $\mathbfcal{Y}_2$) is defined as:
\begin{equation}
\resizebox{.9\hsize}{!}{$\mathcal{P}_{\mathbfcal{Y}_{21}}(D)=\left\{p_{21}\left(\hat{\mathbf{y}}_2|\mathbf{y}_1\right):\mathbb{E}\left[d_{\mathbfcal{Y}_2}\left(g_2(\mathbfcal{Y}_1),\widehat{\mathbfcal{Y}}_2)\right)\right]\leq D\right\}$}
\label{eq:cross_quant}
\end{equation}
Analogously, the rate-distortion function of $\mathbfcal{Y}_2$ from $\mathbfcal{Y}_1$ is:
\begin{equation}
R_{\mathbfcal{Y}_{21}}(D) = \min_{p_{21}(\widehat{\mathbf{y}}_2|\mathbf{y}_1)~\in~\mathcal{P}_{\mathbfcal{Y}_{21}}(D)} I(\mathbfcal{Y}_1;\widehat{\mathbfcal{Y}}_2)
\label{eq:cross_RD}
\end{equation}
Note that $\mathcal{P}_{\mathbfcal{Y}_{11}}(D)$ is simply the conventional rate-distortion function of $\mathbfcal{Y}_1$. Finally, we say that a function $\phi$ has a distortion magnitude of $\delta$ if, for any conditional distribution $q(\hat{\mathbf{z}}|\mathbf{z})$, we have
$\mathbb{E}\left[d_{\phi(\mathbfcal{Z})}(\phi(\mathbfcal{Z}),\phi(\widehat{\mathbfcal{Z}}))\right] = \delta\cdot\mathbb{E}\left[d_\mathbfcal{Z}(\mathbf{z},\hat{\mathbf{z}})\right]$.\footnote{Note that
commonly used operations, such as batch normalization, are designed to maintain relatively uniform feature magnitude in each layer, which in turn promotes distortion magnitude close to $\delta\!=\!1$. Alternatively, the distortion function $d_{\phi(\mathbfcal{Z})}$ can be scaled to achieve this behaviour in practice.} We are now ready to state and prove our next result.
\begin{figure*}[htbp]
\centerline{\includegraphics[width = 0.95\textwidth]{images/Block2.png}}
\caption{Block diagram of two-step feature matching.}
\label{fig:block}
\end{figure*}
\begin{theorem}
\label{thm:distortion}
Given a function $g_2$ with a distortion magnitude of $1$, the minimum achievable rate for compressing $\mathbfcal{Y}_2$ from $\mathbfcal{Y}_1$ is upper bounded by the minimum achievable rate for compressing $\mathbfcal{Y}_1$ from itself. That is:
\begin{equation}
R_{\mathbfcal{Y}_{21}}(D) \leq R_{\mathbfcal{Y}_{11}}(D).
\label{eq:RD_bound_distortion}
\end{equation}
\end{theorem}
\begin{proof}
We begin by assuming we have an approximation $\widehat{\mathbfcal{Y}}_1$, represented by $p_1^{*}(\hat{\mathbf{y}}_1|\mathbf{y}_1)$, which achieves the conventional rate-distortion function of $\mathbfcal{Y}_1$. This means that $\mathbb{E}\left[d_{\mathbfcal{Y}_1}\left(\mathbfcal{Y}_1,\widehat{\mathbfcal{Y}}_1\right)\right]\leq D$ and also $I(\mathbfcal{Y}_1;\widehat{\mathbfcal{Y}}_1)= R_{\mathbfcal{Y}_{11}}(D)$. We then define $\tilde{\mathbfcal{Y}}_2 = g_2(\widehat{\mathbfcal{Y}}_1)$, which induces a conditional distribution $p_{21}(\tilde{\mathbf{y}}_2|\mathbf{y}_1)$. Because the distortion magnitude of $g_2$ is $\delta = 1$ , we know that $p_{21}(\tilde{\mathbf{y}}_2|\mathbf{y}_1)\in \mathcal{P}_{\mathbfcal{Y}_{21}}(D)$ as:
\begin{equation}
\begin{aligned}
\mathbb{E}\left[d_{\mathbfcal{Y}_2}\left(g_2(\mathbfcal{Y}_1),\tilde{\mathbfcal{Y}}_2\right)\right] = & \: \mathbb{E}\left[d_{g_2(\mathbfcal{Y}_1})\left(g_2(\mathbfcal{Y}_1),g_2(\widehat{\mathbfcal{Y}}_1)\right)\right]\\
= & \: \mathbb{E}\left[d_{\mathbfcal{Y}_1}\left(\mathbfcal{Y}_1,\widehat{\mathbfcal{Y}}_1\right)\right] \leq D.
\end{aligned}
\end{equation}
Next, we note that the conditional distribution $p_{21}(\tilde{\mathbf{y}}_2|\mathbf{y}_1)$ is induced by the the following Markov chain:
\begin{equation*}
\begin{tikzpicture}[
inner sep=0pt,
outer sep=0pt,
baseline=(x.base),
]
\tikzstyle{arrow} = [thin,->,>=stealth]
\path[every node/.append style={anchor=base west}]
(0, 0)
\foreach \name/\code in {
x/ \textcolor{white}{\widehat{\mathbf{X}}}\mathbfcal{Y}_1~,
tmp/\,\qquad \qquad,
x_hat/ ~\widehat{\mathbfcal{Y}}_1~,
tmp/\,\qquad,
t_tilde/ ~\widetilde{\mathbfcal{Y}}_2%
} {
node (\name) {$\code$}
(\name.base east)
}
;
\path[
every node/.append style={
anchor=base,
font=\scriptsize,
},
]
(x.base) -- node[below=0.2\baselineskip] (p) {~~~$p_1^{*}(\hat{\mathbf{y}}_1|\mathbf{y}_1)$} (x_hat)
(x_hat.base) -- node[below=0.2\baselineskip] (g) {$g_2$} (t_tilde)
;
\draw [arrow] (x) -- (x_hat);
\draw [arrow] (x_hat) -- (t_tilde);
\end{tikzpicture}
\label{eq:producing_T_tilde}
\end{equation*}
We apply the data processing inequality~\cite{Cover_Thomas_2006} to this chain to get $I(\mathbfcal{Y}_1;\tilde{\mathbfcal{Y}}_2) = I\left(\mathbfcal{Y}_1;g_2(\widehat{\mathbfcal{Y}}_1)\right) \leq I(\mathbfcal{Y}_1;\widehat{\mathbfcal{Y}}_1) = R_{\mathbfcal{Y}_{11}}(D)$.
Because we have already shown $p_{21}(\tilde{\mathbf{y}}_2|\mathbf{y}_1)\in \mathcal{P}_{\mathbfcal{Y}_{21}}(D)$ we can use the definition of the rate-distortion function of $\mathbfcal{Y}_2$ from $\mathbfcal{Y}_1$. As defined in~(\ref{eq:cross_RD}), $R_{\mathbfcal{Y}_{21}}(D)$ is the minimum of the mutual information across all conditional distributions in $\mathcal{P}_{\mathbfcal{Y}_{21}}(D)$ and thus we see that $R_{\mathbfcal{Y}_{21}}(D) \leq I(\mathbfcal{Y}_1;\tilde{\mathbfcal{Y}}_2)\leq R_{\mathbfcal{Y}_{11}}(D)$, thereby concluding our proof.
\end{proof}
\section{Experiments}
\subsection{Experimental setup}
To verify that our results hold in a practical setting, we modify the two-layer model presented in~\cite{hyomin}. Utilizing the original training algorithm, we use the following toss formulation to ensure good performance in terms of both task performance as well as bit-rate:
\begin{equation}
\mathcal{L} = R + \lambda(D_{enh} + w \cdot D_{base}).
\label{eq:loss}
\end{equation}
Here $\mathcal{L}$ is the training loss, $R$ is the bit-rate as estimated by a learned entropy model inspired by~\cite{cheng2020learned}, $D_{enh}$ and $D_{base}$ are the distortions of the image reconstruction and feature matching, respectively (both use mean square error, MSE, as their metric), and $\lambda$ and $w$ are Lagrange multipliers. By training the model using various values of $\lambda$, we produce rate-distortion curves for object detection and input reconstruction.
Similarly to the approach taken in the three-layer case, we train the two-layer model using two-step feature matching, which can be seen in Fig.~\ref{fig:block}. In practice, using this two-step approach is equivalent to changing the distortion $D_{base}$ from being measured at the partition point to measuring it at some downstream layer features. Based on the multi-resolution structure of YOLOv3, we explore several combinations of layers, seen in Table~\ref{tbl:mid_models}, to use for feature matching during training. When using more than one layer we combine the distortions by first flattening the tensors in question, then concatenating them, before calculating the MSE. Note that layer 108 is the final layer of YOLOv3, and it's output consists of multiple tensors, corresponding to different resolutions.
\begin{table}[h]
\centering
\caption{Experimental Setup and Results}
\label{tbl:mid_models}
\smallskip\noindent
\resizebox{1.0\linewidth}{!}{%
\begin{tabular}{@{}cccc@{}}
\toprule
Method & Feature & Base & Enhancement \\
& Layers & BD-Rate[\%]& BD-PSNR [dB] \\ \midrule
HEVC & N/A & $89.27$ & $-1.04$ \\
VVC & N/A & $66.40$ & $\mathbf{0}$ \\
Cheng \textit{et al.}\cite{cheng2020learned} & N/A & $54.46$ & $-0.22$ \\ \midrule
Choi \textit{et al.}\cite{hyomin}, $w =0.06$ & $13$ & $0$ & $-0.92$ \\
Choi \textit{et al.}\cite{hyomin}, $w =0.12$ & $13$ & $-13.89$ & $-1.37$ \\ \midrule
Proposed - Mid-Model 1 & $\{13, 38\}$ & $-12.11$ & $-0.88$ \\
Proposed - Mid-Model 2 & $\{13, 38, 63\}$ & $-13.87$ & $-0.83$ \\
Proposed - Mid-Model 3 & $108$ & $\mathbf{-50.55}$ & $-2.56$ \\
\end{tabular}}
\end{table}
We train our model following a similar scheme to the one presented in~\cite{hyomin}, using a batch size of 16, and training in two stages. At first, we train for 400 epochs using a dataset comprised of randomly cropped patches from both the CLIC~\cite{clic_dataset} and JPEG-AI~\cite{jpeg_ai_dataset} datasets. Then we replace the dataset to a subset of VIMEO-90K~\cite{xue2019video_vimeo} and proceed to train for another 350 epochs (the use of a subset is due to the large amount of models to be trained). The model is trained using an Adam optimizer with a fixed learning rate of $10^{-4}$ for the first stage which is then reduced with a polynomial decay every 10 epochs during the second stage.
After training, we evaluate our models in terms of rate-distortion performance for both object detection (base) and reconstruction (enhancement) tasks. For the base task we use the validation set of the COCO-2014 dataset~\cite{COCO}, which we resize to $512\times 512$ following the procedure in~\cite{hyomin}. In accordance with common practice for object detection, we use the mean average precision (mAP) as our metric. For the enhancement task we use the Kodak dataset~\cite{kodak_dataset}, and PSNR as our evaluation metric. As we have proven in our theoretical discussion, we expect our modifications to lead to improved rate-distortion performance in the computer vision task, perhaps at some cost to the reconstruction task.
We compare our model against the original two-layer model, which we retrain using the same data for a fair comparison. Since our method favors the base task slightly, we also train a second version of the two-layer model, using the original layers for distortion but with a larger Lagrange multiplier for $D_{base}$, $w=0.012$. Lastly, we also compare our model against three baselines: HEVC~\cite{hevc_std_2019}, VVC~\cite{vvc_std}, and a learned compression model~\cite{cheng2020learned}, where an image is fully decoded on the server-side before being passed to YOLOv3 for detection.
\subsection{Results}
We begin by reviewing the results for the base task, as seen in Fig.~\ref{fig:results_base}. We can see that, as expected, the use of deeper layers, has led to improved rate-distortion performance, with the largest improvement obtained using the final layer of YOLOv3. To quantify the differences in rate-distortion performance, we compute the BD-Rate metric~\cite{Bjontegaard}, which measures the average difference in bits required to achieve equal performance (in terms of mAP in this case) to a baseline (we use~\cite{hyomin}). Once again, using the final layer achieves the best performance. Interestingly, increasing the Lagrange multiplier for the base task ($w$ in~\ref{eq:loss}) to 0.012, had a comparable effect to that of our method using mid-model 1 or 2.
\begin{figure}[htbp]
\centering
\includegraphics[width=\linewidth]{images/two_layer_multi_task_base_on_coco2014_new.png}
\begin{footnotesize}
\caption{Mean Average Precision (mAP) vs. Bits per Pixel (bpp) for object detection using YOLOv3.}
\label{fig:results_base}
\end{footnotesize}
\end{figure}
Next, we observe our results on the enhancement task, shown in Fig.~\ref{fig:results_enh}. We summarise the performance using the BD-PSNR metric which measures the average difference in PSNR compared to a baseline (we use VVC~\cite{vvc_std}) at an equal bit-rate, seen in Table~\ref{tbl:mid_models}. Comparing the performance here to the base task we clearly see the trade-off between the two. This is most notable in the performance of our method using mid-model 3, which was the best method for object detection, and is the weakest here on image reconstruction. Importantly, our method using mid-model 1 or 2 performs slightly better than the modified version of \cite{hyomin} using $w=0.012$, which had comparable performance on the base task.
\begin{figure}[htbp]
\centering
\includegraphics[width=\linewidth]{images/two_layer_multi_task_enhancement_on_kodak_new.png}
\begin{footnotesize}
\caption{PSNR vs. Bits per Pixel (bpp) for image reconstruction.}
\label{fig:results_enh}
\end{footnotesize}
\end{figure}
\section{Summary and Discussion}
The growing prevalence of DNNs in the field of computer vision have lead to an increasing volume of images seen predominantly by machines. This, in turn, requires efficient coding methods for machines and a good understanding of their rate-distortion behaviour. In this paper we provided important theoretical background for understanding the implication of layer selection when encoding an image for feature matching in a downstream DNN model. We have proved that using deeper layers achieves superior rate-distortion performance compared to earlier ones. This is true both when evaluating distortion at the task level, as well as when the distortion is measured at some intermediate point.
To demonstrate the usefulness of our theoretical results, we used them to modify the training loss of an image coding method for humans and machines~\cite{hyomin}. The results clearly show that using our approach yields improvement in the rate-distortion performance of the machine side of the model, as expected. Importantly, the modifications do not require any change to the encoder side architecture, or to the decoding process, making our approach easy to implement in practice. This is critical as it does not require more edge-device resources than simply using the learned image coding backbone~\cite{cheng2020learned}, on which it is based.
As might have been expected, the improved performance on the machine performance comes at some cost on the human side. Using a hyper-parameter ($w$), the framework in~\cite{hyomin} allows a designer to balance between performance on each side during training. When considering either model for a specific application, one might estimate the frequency at which images will need to be fully reconstructed and adjust the training procedure accordingly. Notably, when both models are adjusted to give equal rate-distortion performance on object detection, our model resulted in superior performance on image reconstruction compared with~\cite{hyomin}.
\bibliographystyle{IEEEtran}
\small
|
\section{Introduction}
Quantum chemistry and physics deal with the description of many interacting particles. Often we limit ourselves to a single particle species. In quantum chemistry these are usually electrons, but in physics also bosonic particles are of interest. Though the many-body Schrödinger equation involves only linear operators, the daunting dimensionality of the many-body wave function renders a direct solution intractable, but for a few particles. This is one of the prime reasons to aim directly for reduced quantities.
In 1964 Hohenberg and Kohn presented their revolutionary work about density functional theory (DFT)~\cite{hohenberg1964inhomogeneous}. They showed that any observable can be regarded as a functional of the density. Especially the Kohnn--Sham (KS) formulation~\cite{kohn-sham1965} has been important to the success of DFT. Their idea was to approximate the true kinetic energy by the kinetic energy of the KS system: a non-interacting system with the same density as the interacting system.
The KS kinetic energy turns out to be a decent approximation to the true kinetic energy of the interacting system. The difference in the kinetic energy is then lumped together with the interaction beyond Hartree (classical Coulomb) in the exchange-correlation energy functional. Though formally exact, in practice KS-DFT has some weaknesses, since the exchange-correlation functional needs to be approximated. A famous example is the stretching of H\textsubscript{2} bond~\cite{cohen2012challenges, becke2014j, vuckovic2015hydrogen}.
One way to bypass some of these problems in constructing an approximate exchange-correlation energy functional is (one-body) reduced density matrix (1RDM) functional theory. One advantage over DFT is that we have also an explicit expression for the kinetic energy while still having the total energy as a functional of the 1RDM~\cite{gilbert1975hohenberg}. However, in the zero temperature setting, mapping back from 1RDMs to (non-local) potentials is problematic as already noted by Gilbert~\cite{gilbert1975hohenberg} and others~\cite{pernal2005, Leeuwen2007, PhD-Baldsiefen2012, giesbertz2019one}. This is most clear in the case of non-interacting particles, since typically ground state 1RDMs are idempotent. It therefore seems that non-idempotent 1RDMs cannot be $v$-representable in the absence of interactions. There is the possibility for orbital energies to be degenerate, however, which allows fractionally occupied orbitals and hence non-idempotent 1RDMs~\cite{RequistPankratov2008, GiesbertzBaerends2010}. But one quickly realizes that the scaled identity operator is the one-body Hamiltonian which has all 1RDMs as ground state 1RDM. It is clear that the interaction should play a crucial role in the back mapping, but there has been no progress in this direction.
An alternative to regularize the theory is to introduce entropy, i.e.\ work at finite temperature as proposed more than a decade ago~\cite{Leeuwen2007, PhD-Baldsiefen2012, baldsiefen2013minimization, baldsiefen2015reduced}
Though this is a theoretical motivation to introduce temperature, also physically this is a well justified choice, since most experiments are conducted at $T > 0$. Important examples where temperature plays an important role are metal-insulator transitions in transition metal oxides~\cite{YooMaddoxKlepeis2005}, high $T_c$ super conductors~\cite{NagamatsuNakagawaMuranaka2001}, hot plasmas~\cite{Dharma-wardanaPerrot1982}, etc.
In Ref.~\cite{giesbertz2019one} 1RDM functional theory (1RDMFT) was presented for the grand canonical ensemble within a finite basis set. However, the use of a grand canonical ensemble is inappropriate if the number of particles is relatively low as in ultra cold atom experiments~\cite{EwaldFeldkerHirzler2019}, but also in the low temperature limit the grand canonical ensemble can lead to unphysical results~\cite{Bedingham2003, MullinFernandez2003}. A canonical formulation of 1RDMFT is therefore desirable and will be the goal of this article.
In classical thermodynamics the grand potential can be reached by a Legendre transformation of the Helmholtz free energy with respect to the number of particles. In the quantum mechanical setting we can not do that. The reason is that the grand potential and the Helmholtz functional act on different spaces: the Fock space and the $N$-particle Hilbert spaces respectively. Thus, we can not simply transform it back to obtain the canonical case.
Recently some progress was done in this area. In~\cite{barghathi2020theory} the occupation numbers for a non-interacting system were recursively calculated.
Another major difference w.r.t.\ the grand canonical ensemble is that in the non-interacting case the occupation numbers are not explicitly given by either the Fermi/Bose function for fermions/bosons. Instead, they need to be calculated recursively, using auxiliary partition functions~\cite{barghathi2020theory}.
In Ref.~\cite{baldsiefen2015reduced} it therefore remained an open question whether every thermal 1RDM (only fractional occupation numbers) would be non-interacting $v$-representable and not much progress could be made.
However, we do not rely on such an explicit relation and are able to prove one-to-one correspondence between thermal 1RDMs (all occupation numbers fractional) and (non-local) potentials for any interaction. The non-interacting system is just a particular case.
This result justifies the existence of an algorithm which finds for any thermal 1RDM the corresponding non-interacting Hamiltonian as published recently by Kooi~\cite{kooi2022canonical}.
In this work we present 1RDMFT in a rigorous way for fixed number of particles, finite basis set and elevated temperature. We show that the universal functional is differentiable and it holds
\begin{align}
\label{differentiability of universal functional}
\frac{\partial F_{N,\pm}}{\partial \g}=-v
\end{align}
where $\g$ is the ground state 1RDM for the potential $v$. If we have a handy expression for $F_{N,\pm}[\g]$ then we can circumvent the handling of the density-matrix operator to compute the free energy and instead we only need to deal with the reduced quantity $\g$. The minimizer for the Helmholtz functional can then be determined through the above relation \eqref{differentiability of universal functional}.
This work is built up in the following way. In Section~\ref{section setting} we introduce all the relevant spaces, then, in Section~\ref{section general approach}, we present the Helmholtz functional, its minimizing density-matrix operator and the general approach for 1RDMFT. For this task we make use of the universal functional $F_{N,\pm}$. To show differentiability of $F_{N,\pm}$ we utilize results from convex analysis. In Section~\ref{section general properties of the helmholtz functional and implications on the universal functional} we show that all the relevant functionals are convex. Additionally, we show that two potential differing by more than a constant can not generate the same density-matrix operator. The proof of differentiability of $F_{N,\pm}$ is finalized in Section~\ref{section final result}.
\section{Setting}
\label{section setting}
We build our $N$-particle space from a finite number of single particle states $\ket{i}$, for $i \in \{1, \dotsc , N_b\}$ and $N_b < \infty$. We require the states to be orthonormal. The one-particle Hilbert space $\oneH$ is now the ${\mathbb C}$-vector space generated by the states $\ket{i}$, i.e.\ $\oneH\coloneqq \mathrm{span}\{\ket{1}, \dotsc , \ket{N_b}\} \cong {\mathbb C}^{N_b}$. To build the $N$-particle space we need to distinguish between bosons and fermions.
\begin{description}
\item[Bosons] A system with $N$ bosons is described by a symmetric wave function. Therefore, the bosonic $N$-particle Hilbert state, denoted by $\oneH^N_+$, consists of all symmetric tensors of order $N$, i.e.\ $\oneH_+^N\coloneqq\mathrm{Sym}^N(\oneH)$. The dimension of $\oneH_+^N$ is $\binom{N_b+N-1}{N}$.
\item[Fermions] Fermionic systems are described by anti-symmetric wave functions. Thus, we consider the space of anti-symmetric tensors of order $N$, i.e.\ $\oneH_-^N\coloneqq\wedge ^N \oneH$. The dimension is given by $\binom{N_b}{N}$. Note that we need to have $N_b \geq N$. The case $N_b=N$ is trivial since we have only one possible state. Thus, we will only consider $N_b>N$.
\end{description}
The set of density-matrix operators on the $N$-particle space $\oneH_{\pm}^N$ is defined as
\begin{equation}
\label{definition set of density-matrix operators}
\NdensMat \coloneqq \Set{\hat{\r}:\oneH_\pm^N \to \oneH_\pm^N | \hat{\r}=\hat{\r}^\dagger, \hat{\r}\geq 0, \Trace\{\hat{\r}\}=1}
\end{equation}
which we endow with the norm
\begin{equation}
\label{defintion norm for density-matrix operators}
\norm{ \hat{\r} }_2= \big( \Trace\{\abs{\hat{\r}}^2\} \big)^{1/2}.
\end{equation}
A density-matrix operator $\hat{\r} \in \NdensMat$ has a spectral decomposition
\begin{equation}
\label{density operator in spectral decomposition}
\hat{\r}= \sum \limits_l \l_l \ket{\psi_l}\bra{\psi_l}
\end{equation}
and its kernel is given by
\begin{multline}
\label{kernel of density-matrix operator}
\r(x_1, \ldots, x_N; y_1, \ldots, y_N) \\
= \sum \limits_l \l_l \psi_l(x_1, \ldots , x_N) \overline{\psi_l(y_1, \ldots, y_N)}.
\end{multline}
Here and in the following we use the notation $\overline{\psi_l}$ for the complex conjugate of $\psi_l$.
We can define the 1RDM $\g$ by \footnote{Creation and annihilation operator are only defined pairwise for a $N$ particle space.}
\begin{align*}
\g_{ij}[\hat{\r}]=\Trace\{\hat{\r}\,\crea{a}_j \anni{a}_i\}.
\end{align*}
It turns out (see Appendix) that the relevant spaces for the 1RDMs are subsets of the space of all hermitian $N_b \times N_b$ matrices denoted by $\mathbb{H}(N_b)$
\begin{subequations}
\begin{align}
\label{definition spaces for 1RDM (bosons)}
\overline{\mathscr{N}}_{N,+}&\coloneqq \Set{ \g \in \mathbb{H}(N_b) | \g \geq 0, \trace\{\g\}=N}, \\
\label{definition spaces for 1RDM (fermions)}
\overline{\mathscr{N}}_{N,-}&\coloneqq \Set{ \g \in \mathbb{H}(N_b) | \g \geq 0, \g^2 \leq \g, \trace\{\g\}=N }.
\end{align}
\end{subequations}
We have used $\trace\{\cdot \}$ to emphasize that the trace is over the one-particle Hilbert space $\oneH$ as opposed to the the $\Trace\{\cdot\}$ which is over a $\oneH^N_\pm$ Hilbert space.
By convention, the eigenvalues and eigenstates of the 1RDM $\g$ are called natural occupation numbers and natural orbitals (NO) respectively. Coleman has shown that all elements of $\overline{\mathscr{N}}_{N,\pm}$ can be obtained from a density-matrix operator in $\NdensMat$, so is a true 1RDM~\cite{coleman1963structure}.
\begin{thm}[Coleman]
\label{thm N representability of 1RDMs}
For any $\g \in \overline{\mathscr{N}}_{N,\pm}$ there is a density matrix $\hat{\r} \in \NdensMat$ which generates $\g$.
\end{thm}
The proof can be found in the Appendix.
\section{General Approach}
\label{section general approach}
The Helmholtz functional for the canonical ensemble is defined as
\begin{equation}
\label{definition Helmholtz functional}
\Omega_v[\hat{\rho}]\coloneqq E_v[\hat{\r}]-\b^{-1}S[\hat{\r}],
\end{equation}
where
\begin{equation}
\label{energy functional}
E_v[\hat{\r}]\coloneqq\Trace[\hat{\r}\hat H_v]
\end{equation}
is the energy of a system with Hamiltonian $\hat{H}_v\coloneqq\hat{H}_0+\hat{V}_v$ ($\hat H_0$ contains the kinetic and interaction part and $\hat V_v$ is the potential with kernel $v(x,x')$). The second term contains the entropy
\begin{equation}
\label{definition entropy}
S[\hat{\r}]\coloneqq -\Trace\{\hat{\r}\log (\hat{\r})\}
\end{equation}
and the inverse temperature $\b=1/T$. With $\log$ we mean the natural logarithm.
The minimizer $\hat{\r}_v$ of the Helmholtz functional can be found by variations in the density-matrix operator which yields the equation
\begin{equation}
\label{functional derivative of Helmholtz functional}
\Trace\{\d \hat{\r}(\hat{H}_v+\b^{-1}\log(\hat{\r}_v)\}+\b^{-1}\Trace\big\{\d \hat{\r}\big\}=0.
\end{equation}
From the unit trace condition and \eqref{functional derivative of Helmholtz functional} it follows that
\begin{align}
\label{minimizer of the Helmholtz functional}
\hat{\r}_v &= e^{-\b\hat{H}_v}/Z[v], &
&\text{where}&
Z[v] &\coloneqq \Trace\bigl\{e^{-\b \hat{H}_v}\bigr\}.
\end{align}
The minimizer $\hat{\r}_v$ is called Gibbs state.
Note that we only have a proper solution for $0<Z[v]<\infty$. This is always the case since we work in a finite basis setting with fixed number of particles, so the trace only runs over a finite number of elements.
One aim is to show that the map from the potential $v$ to the density-matrix operator $\hat{\r}_v$ is invertible. However this is only doable up to a constant since adding a constant to the potential does not change the density-matrix operator. To achieve a one-to-one correspondence we allow only potentials from the following set
\begin{equation}
\label{set of all potentials}
\ensuremath{\mathcal V}\coloneqq \Set{v \in \mathbb{H}(N_b) | \trace\{v\}=0}.
\end{equation}
We can also think of $v\in \ensuremath{\mathcal V}$ being a representative of the equivalence class containing potentials differing by a constant.
In Theorem~\ref{thm N representability of 1RDMs} we have seen that all $\g \in \overline{\mathscr{N}}_{N,\pm}$ are $N$-representable. However, physically relevant are only the 1RDMs that are associated with a Gibbs state $\hat{\r}_v$. Thus, we denote the set of all $v$-representable 1RDM by
\begin{equation}
\label{definition set of v-representable 1RDMs}
\VoneMat \coloneqq \Set{ \g \in \overline{\mathscr{N}}_{N,\pm} | \exists \; v \in \ensuremath{\mathcal V} \mapsto \g}.
\end{equation}
The approach is to partition the minimisation in the Helmholtz functional as
\begin{equation}
\label{infimum of the Helmholtz funcitonal over 1RDM }
\Omega_{N,\pm}[v] \coloneqq \inf_{\mathclap{\hat{\r} \in \NdensMat}} \Omega_v[\hat{\r}]
= \inf_{\mathclap{\g \in \overline{\mathscr{N}}_{N,\pm}}} \big(F_{N,\pm}[\g]+\trace\{v\g\}\big)
\end{equation}
where
\begin{align}
\label{defintion universal functional}
F_{N,\pm}[\g] \coloneqq{}& \inf_{\hat{\r}\to \g}\Omega_0[\hat{\r}] \notag \\*
{}={}& \inf_{\hat \r \to \g}\Trace\big\{\hat \r\big(\hat H_0 +\b^{-1}\log(\hat{\r})\big)\big\}
\end{align}
is called the universal functional which takes the value $\infty$ in case no $\hat{\r} \to \g$ exists. The aim is to show that $F_{N,\pm}$ is differentiable. Then the minimizer can be found through the relation
\begin{equation*}
\frac{\partial F_{N,\pm}}{\partial \g}=-v
\end{equation*}
and we know that $\g$ is a canonical eq-1RDM which was an open question in \cite{baldsiefen2015reduced}.
\section{General Properties of the Helmholtz Functional and Implications on the Universal Functional}
\label{section general properties of the helmholtz functional and implications on the universal functional}
\begin{thm}
\label{thm invertibility hamiltonian to density matrix}
The mapping $\hat{H}_v \mapsto \hat{\r}_v$ with $v\in \ensuremath{\mathcal V}$ is invertible up to a constant in the Hamiltonian.
\end{thm}
\begin{proof}
Assume that two Hamiltonians $\hat{H}_v$ and $\hat{H}_v'$ differing in their potential yield the same density-matrix operator $\hat{\r}_v$. From~\eqref{functional derivative of Helmholtz functional} it follows that $\hat{\r}_v$ fulfills
\begin{align*}
\frac{1}{\b}\log(\hat{\r}_v)+\hat{H}_v&=C\\
\frac{1}{\b}\log(\hat{\r}_v)+\hat{H}_v'&=C'.
\end{align*}
Subtracting these equations gives \(\hat{H}_v-\hat{H}_v' = C-C' \).
\end{proof}
\begin{remark}
Since we have fixed number of particles in the Hilbert space, the constant in Theorem~\ref{thm invertibility hamiltonian to density matrix} can be of the form $ f(\hat{N})$ where $f: {\mathbb R} \to {\mathbb R}$, so this includes the arbitrary constant shift in the potential.
\end{remark}
\begin{Cor}
\label{cor v to density matrix is invertible}
The map $v \mapsto \hat{\r}_v$ with $v \in \ensuremath{\mathcal V}$ is invertible.
\end{Cor}
Note that we only have a one-to-one correspondence because we require $\trace\{v\}=0$. Otherwise a constant shift in the potential would lead to the same density-matrix operator.
At this point we want to mention that the density-matrix operator $\hat{\r}_v$ is positive definite, $\hat{\r}_v >0$, and lies in the following subspace of $\NdensMat$
\begin{equation}
\label{space where statistical density operators lie}
\NTdensMat\coloneqq \Set{ \hat{\r}: \oneH^N_\pm \to \oneH_\pm^N | \hat{\r}=\hat{\r}^\dagger, \hat{\r}>0, \Trace\{\hat{\r}\}=1}.
\end{equation}
It follows that the natural occupation numbers $n_i$ are positive and in the fermionic case additionally $n_i<1$. To see this let $\phi_1, \dotsc , \phi_{N_b}$ be the NO basis and $\hat{\r}_v=\sum_j \l_j \ket{\psi_j}\bra{\psi_j}$ be the spectral decomposition of the density-matrix operator. Then, as the $\psi_j$'s build a basis of $\oneH_\pm^N$, each NO $\phi_i$ contribute to at least one of the eigenstates. So,
\begin{multline}
\label{natural occupation number for statistical ensemble positiv}
n_i= \sum \limits_j \l_j \int \mathrm{d} x \mathrm{d} y \mathrm{d} x_2 \dotsi \mathrm{d} x_N \overline{\phi_i(x)}\phi_i(y) \\
\psi_j(x, x_2, \dotsc, x_N) \overline{\psi_j(y,x_2, \dotsc, x_N)}>0
\end{multline}
where we used the fact that all weights $\l_j=e^{-\b E_j}/Z$ are positive. In case of fermions we have already showed that $n_i \leq 1$. The $i$th NO can not be present in all $\psi_j$'s (in case $N \neq N_b)$, so
\begin{multline}
\label{fermionic natural occupation number for statistical ensemble smaller than 1}
n_i=\sum \limits_j \l_j \int \mathrm{d} x \mathrm{d} y \mathrm{x_2} \dotsi \mathrm{d} x_N \overline{\phi_i(x)}\phi_i(y) \\
\psi_j(x, x_2, \dotsc, x_N) \overline{\psi_j(y,x_2, \dotsc, x_N)}<1
\end{multline}
because the integral is at least for one $j$ not equal to 1.
Summarized, the 1RDMs produced by a potential are contained in
\begin{align}
\mathscr{N}_{N,+}&\coloneqq \Set{ \g \in \mathbb{H}(N_b) | \g>0 }, \\
\mathscr{N}_{N,-}&\coloneqq \Set{ \g \in \mathbb{H}(N_b) | \g>0, \g^2<\g }.
\end{align}
We want to show that the most important functionals are either convex or concave to be able to use results from convex analysis. We start with the functional $\Omega_{N,\pm}[v]$ which is achieved through a minimization and thus turns out to be concave \cite{eschrig2010t}.
\begin{thm}
The functional $\Omega_{N,\pm}[v]= \min \limits_{\hat{\r} \in \NdensMat} \Omega_v[\hat{\r}]=\min \limits_{\hat{\r} \in \NTdensMat} \Omega_v[\hat{\r}]= -\b^{-1}\log\big(Z[v]\big)$ is strictly concave in $v$.
\end{thm}
\begin{widetext}
\begin{proof}
Let $v_1 \neq v_2$ be two potentials in $\ensuremath{\mathcal V}$ and let $0<t<1$. Then we have
\begin{align*}
\Omega_{N,\pm}[tv_1+(1-t)v_2
&=\min_{\hat{\r}\in \NTdensMat}\Trace\Big\{\hat{\r}\big(t\hat{H}_{v_1} + t \frac{1}{\b}\log(\hat{\r})\big)\Big\}+\Trace\Big\{\hat{\r}\big((1-t)\hat{H}_{v_2}+(1-t)\frac{1}{\b}\log(\hat{\r})\big)\Big\}\\
&>t \min_{\hat{\r}_1 \in \NTdensMat} \Trace\Big\{\hat{\r}_1\big(\hat{H}_{v_1}+\frac{1}{\b}\log(\hat{\r}_1)\big)\Big\} +
(1-t) \min_{\hat{\r}_2 \in \NTdensMat}\Trace\Big\{\hat{\r}_2\big(\hat{H}_{v_2}+\frac{1}{\b}\log(\hat{\r}_2)\big)\Big\}\\
&=t \Omega_{N,\pm}[v_1]+(1-t)\Omega_{N,\pm}[v_2]
\end{align*}
where the strict inequality follows from Corollary~\ref{cor v to density matrix is invertible}.
\end{proof}
\end{widetext}
With Corollary~\ref{cor v to density matrix is invertible} it is possible to show a generalization of the Hohenberg--Kohn theorem for 1RDMs and non-local potentials~\cite{mermin1965thermal}.
\begin{thm}
\label{Mermin's theorem for canonical ensemble}
The map $v \mapsto \g_v$ for $v \in \ensuremath{\mathcal V}$ is invertible.
\end{thm}
\begin{proof}
Assume there are two potential $v_1\neq v_2 \in \ensuremath{\mathcal V}$ yielding different density matrix operators $\hat{\r}_1 \neq \hat{\r}_2$ but the same 1RDM $\g$. Then we get
\begin{align*}
\Omega_{N,\pm}[v_1]&=\Omega_{v_1}[\hat{\r}_1]
=\Omega_{v_2}[\hat{\r}_1]+\trace\big\{\g(v_1-v_2)\big\} \\
& > \Omega_{v_2}[\hat{\r}_2]+\trace\big\{\g(v_1-v_2)\big\}\\
&= \Omega[v_2]+\trace\big\{\g(v_1-v_2)\big\}.
\end{align*}
Changing the role of $v_1$ and $v_2$ and adding the two inequalities gives
\begin{align*}
\Omega_{N,\pm}[v_1]+\Omega_{N,\pm}[v_2]<\Omega_{N,\pm}[v_2]+\Omega_{N,\pm}[v_1]
\end{align*}
which is a contradiction.
\end{proof}
One aim is to show that the universal functional is convex. For this purpose we first show that the entropy is strictly concave (\cite{ruelle1969statistical}, \cite{lieb1975convexity}, \cite{wehrl1978general}).
\begin{thm}
The entropy is strictly concave, i.e.\ for any $\hat{\r}_0, \hat{\r}_1 \in \NdensMat$ and $\l \in (0,1)$ we have $S[\l \hat{\r}_0+(1-\l)\hat{\r}_1]> \l S[\hat{\r}_0]+(1-\l)S[\hat{\r}_1]$.
\end{thm}
\begin{proof}
Let $\hat{\r}_\l= \l \hat{\r}_0+(1-\l)\hat{\r}_1 = \sum_k w_k \ket{\psi_k}\bra{\psi_k}$. We use strict concavity of the function $s(x)=-x\log(x)$ and we get
\begin{align*}
S[\hat{\r}_\l]&=-\sum_k w_k \log(w_k) = \sum_k s\big(\braket{\psi_k|\hat{\r}_\l|\psi_k } \big) \\
&= \sum_k s\big(\l \braket{ \psi_k|\hat{\r}_0|\psi_k } +
(1-\l) \braket{ \psi_k | \hat{\r}_1|\psi_k}\big)\\
&> \l \sum_k s\big(\braket{ \psi_k| \hat{\r}_0|\psi_k } \big) +
(1-\l) \sum_k s\big(\braket{ \psi_k|\hat{\r}_1|\psi_k} \big)\\
&\geq \l \sum_k \braket{ \psi_k|s(\hat{\r}_0)|\psi_k } +
(1- \l) \sum_k \braket{ \psi_k |s(\hat{\r}_1)|\psi_k } \\
&= \l S[\hat{\r}_0] + (1- \l) S[\hat{\r}_1]
\end{align*}
where we used Jensen's inequality for the last inequality.
\end{proof}
\begin{Cor}
The Helmholtz functional $\Omega_v[\hat{\r}]$ is strictly convex in the density operator $\hat{\r}$.
\end{Cor}
\begin{proof}
It follows directly from the fact that the Helmholtz functional is the sum of a linear and a strictly convex functional.
\end{proof}
\begin{thm}
\label{thm convexity of universal functional}
The universal functional $F_{N,\pm}[\g]$ is convex on $\overline{\mathscr{N}}_{N,\pm}$.
\end{thm}
\begin{proof}
Let $\g_0, \g_1 \in \overline{\mathscr{N}}_{N,\pm}$, $\l \in [0,1]$ and $\g_\l= \l \g_0+(1-\l)\g_1$. Then we get,
\begin{multline*}
\l F_{N,\pm}[\g_0] + (1-\l)F_{N,\pm}[\g_1] \\
\begin{aligned}
&= \l \inf_{\hat{\r}_0 \to \g_0} \Omega_0[\hat{\r}_0] +
(1-\l) \inf_{\hat{\r}_1 \to \g_1} \Omega_0[\hat{\r}_1] \\
&= \inf_{\hat{\r}_0 \to \g_0} \inf_{\hat{\r}_1 \to \g_1} \l \Omega_0[\hat{\r}_0] +
(1-\l) \Omega_0[\hat{\r}_1]\\
&\geq \inf_{\hat{\r}_0 \to \g_1}\inf_{\hat{\r}_1 \to \g_1} \Omega_0[\l \hat{\r}_0 +
(1-\l)\hat{\r}_1] \\
&= \inf_{\hat{\r}\to \g_\l} \Omega_0[\hat{\r}] = F_{N,\pm}[\g_\l]. &\qedhere
\end{aligned}
\end{multline*}
\end{proof}
\section{Final Result}
\label{section final result}
Now we want to show that the universal functional $F_{N,\pm}$ is differentiable. Differentiability is only defined on an open set. However, the set $\overline{\mathscr{N}}_{N,\pm}$ has empty interior in $\mathbb{H}(N_b)$. Thus, we need to embed $\overline{\mathscr{N}}_{N,\pm}$ in a topological space where $\mathscr{N}_{N,\pm}$ is the interior of $\overline{\mathscr{N}}_{N,\pm}$. The idea is to use the following result about sub\-gradient and sub\-differential.
\begin{thm}
\label{properties of the subgradient/subdifferential}
Let $X$ be a finite dimensional vector space, let $f:X \to {\mathbb R} \cup \{\infty\}$ be a convex function with domain $M$. Assume $M$ is contained in $a+\mathfrak{L}$ such that $\mathfrak{L}$ is a subspace with lowest dimension such that there exists $a \in X$ with $M \subset a+ \mathfrak{L}$ \footnote{This means $M$ is contained in an affine space}. Let $\partial _\mathfrak{L}f(x)\coloneqq\partial f(x) \cap \mathfrak{L}$ where $\partial f(x)$ is the subdifferential of $f$ at a point $x$ in the set $M$. Then the following properties hold for $\partial_\mathfrak{L} f(x)$
\begin{enumerate}[label=(\roman*)]
\item the set $\partial_\mathfrak{L} f(x)$ is nonempty,
\item $f$ is differentiable at $x$ if and only if $\partial_\mathfrak{L} f(x)$ contains only one element. In that case this element equals the usual gradient. (With differentiable we mean that there is a linear map $J:\mathfrak{L}\to {\mathbb R}$ such that for all $h \in \mathfrak{L}$ we have
$\lim_{h \to 0}\frac{1}{\norm{h}_\mathfrak{L}}\abs{f(x+h)-f(x)-J(h)}=0$.
\end{enumerate}
\end{thm}
\begin{figure}[t]
\begin{center}
\begin{tikzpicture}[scale=1.7]
\draw[->] (0,0)--(0,2) node[pos=1.1]{$n_3$};
\draw[->] (0,0)--(1.85,0.6) node[pos=1.1]{$n_2$};
\draw[->] (0,0)--(1.2,-1.4) node[pos=1.1]{$n_1$};
\draw[->] (0,0)--(0.5,0.55) node[midway, above left]{$a$};
\filldraw[color=teal!20!white, opacity=0.5] (0,1.4)--(1.3,0.422)--(0.7,-0.81667)--(0,1.4);
\filldraw [teal] (0,1.4) circle (0.5pt) node[left, color=black]{\small{$2$}};
\filldraw [teal] (1.3,0.422) circle (0.5pt) node[above, color=black]{\small{$2$}};
\filldraw [teal] (0.7,-0.81667) circle (0.5pt) node[left, color=black]{\small{$2$}};
\draw[color=teal] (0,1.4)--(1.3,0.422)--(0.7,-0.81667)--(0,1.4);
\end{tikzpicture}
\end{center}
\caption{Representation of $\overline{\mathscr{N}}_{N,+}$ in terms of the occupation numbers for $N_b=3$ and $N=2$ (See proof of Theorem~\ref{thm N representability of 1RDMs}). The vector $a$ shifts a $2$ dimensional subspace to an affine space which contains $\overline{\mathscr{N}}_{N,+}$.}
\label{affine space of 1RDMs}
\end{figure}
The universal functional has domain $\overline{\mathscr{N}}_{N,\pm}$ which is contained in $\mathfrak{L}+a$ with $\mathfrak{L}=\Set{x \in \mathbb{H}(N_b) | \trace\{x\}=0}$ and $a=N/N_b \cdot \mathds{1}$~\footnote{We could have taken a different point in $\mathscr{N}_{N, \pm}$, but the current choice is the most natural in this setting.} and its (relative) interior is $\mathscr{N}_{N, \pm}$. Fig. \ref{affine space of 1RDMs} shows $\overline{\mathscr{N}}_{N,+}$ for $2$ particles and $3$ basis functions. This also justifies the choice of the potential gauge. The aim is to get the relation $\partial F_{N,\pm}/\partial \g=-v$. But as mentioned in Theorem \eqref{properties of the subgradient/subdifferential}, the differential is a map $J:\mathfrak{L}\to {\mathbb R}$, i.e.\ it is contained in $\ensuremath{\mathcal V}$. We can now apply the above theorem for all 1RDMs contained in $\mathscr{N}_{N,\pm}$.
\begin{thm}
\label{main thm differentiation of F and v representability}
If the infimum in \eqref{defintion universal functional} is attained, then
\begin{enumerate}[label=(\roman*)]
\item \( \mathscr{N}_{N,\pm}=\VoneMat \)\\
\item the universal functional $F_{N,\pm}[\g]$ is differentiable on $\mathscr{N}_{N,\pm}$.
\end{enumerate}
\end{thm}
\begin{proof}
Convexity of $F_{N,\pm}$ implies that for any $\g \in \mathscr{N}_{N,\pm}$ there exists at least one subgradient $h \in \ensuremath{\mathcal V}$. So for all $\tilde{\g} \in \overline{\mathscr{N}}_{N,\pm}$ it holds \(F_{N,\pm}[\Tilde{\g}]+ \braket{ -h| \Tilde{\g} } \geq F_{N,\pm}[\g]+\braket{ -h| \g } \) which implies
\begin{align*}
F_{N,\pm}[\g] + \braket{ -h| \g }
\leq \min_{\mathclap{\tilde{\g}\in \overline{\mathscr{N}}_{N,\pm}}} F_{N,\pm}[\Tilde{\g}]+ \braket{ -h| \tilde{\g}}
= \Omega_{N,\pm}[-h].
\end{align*}
Thus, the negative of the subgradient, $-h$, yields a potential generating $\g$ and hence \(\mathscr{N}_{N,\pm}=\VoneMat \). By Theorem~\ref{Mermin's theorem for canonical ensemble}, we get that there is only one such potential. Hence the subgradient is unique and $F_{N,\pm}[\g]$ is differentiable for all $\g \in \mathscr{N}_{N,\pm}$ by Theorem~\ref{properties of the subgradient/subdifferential}.
\end{proof}
We proved $v$-representability under the assumption that the minimum in \eqref{defintion universal functional} is attained. To finish the proof we still need to show that this is indeed the case.
In the following we would like to have that $\NTdensMat$ is the interior of $\NdensMat$. However this is not the case in the topology on $\mathbb{H}(d_\pm)$ where $d_\pm$ is the dimension of $\oneH_\pm^N$.
We make use of the topological subspace $b+\mathfrak{P}$ where $\mathfrak{P}$ is a vector space with lowest dimension such that $\NdensMat \subset b+\mathfrak{P}$. Then $\NTdensMat$ is the relative interior of $\NdensMat$. In the following we restrict ourself to the subspace $b+ \mathfrak{P}$ with the induced metric on it.
The idea is to show that the relevant functions are continuous and then use the fact that continuous functions attain their minima (and maxima) over compact sets.
\begin{proposition}
The energy $E_v[\hat{\r}]$ is Lipschitz continuous on $\NdensMat$.
\end{proposition}
\begin{proof}
The Hamiltonian acts on a finite dimensional space and thus it has a maximum eigenvalue, $\norm{\hat H_v}_\infty < \infty$. For two density-matrix operators $\hat{\r}_0, \hat{\r}_1$ we have
\begin{align*}
\abs{E_v[\hat{\r}_0]-E_v[\hat{\r}_1]}
&\leq \norm{\hat H_v}_\infty \norm{\hat{\r}_0- \hat{\r}_1}. \qedhere
\end{align*}
\end{proof}
\begin{proposition}
The entropy is continuous on $\NdensMat$.
\end{proposition}
\begin{proof}
The function $\hat{\r} \mapsto (a_1, \dotsc , a_m)$ where $(a_1, \dotsc, a_m)$ contains the eigenvalues of $\hat{\r}$ in an ordered fashion and with multiplicity ($a_i \leq a_{i+1}$ for all $i=1, \dotsc, m-1$) is continuous. The statement follows from continuity of the function $(a_1, \dotsc , a_m) \mapsto \sum_{j=1}^m a_j \log(a_j)$.
\end{proof}
To finish the proof we make use of the following theorem.
\begin{thm}
Let $X$ be a compact metric space and let $f:X \to {\mathbb R}$ be a continuous function. Then $f$ is bounded and it attains its maximum and minimum.
\end{thm}
Compact sets in finite dimensional affine spaces (with the usual metric) are fully characterized by closedness and boundedness. Thus, $\NdensMat$ and $\set{\hat{\r} \in \NdensMat | \hat{\r} \to \g }$ are compact and we get the following corollary.
\begin{Cor}
The infima in the the Helmholtz functional $\Omega_{N,\pm}[v]$ and the universal functional $F_{N,\pm}[\g]$ are achieved in the fermionic and bosonic case, so the infima in \eqref{definition Helmholtz functional} and~\eqref{infimum of the Helmholtz funcitonal over 1RDM } can be replaced by minima.
\end{Cor}
\section{Conclusion}
For fixed number of particles, finite numbers of basis functions and elevated temperature the universal functional in 1RDM functional theory is differentiable with $\partial F_{N,\pm}/ \partial \g=-v$ for all 1RDMs $\g$ in $\mathscr{N}_{N,\pm}$. This relation holds for potential with $\trace\{v\}=0$. However, all potentials differing from $v$ by only a constant lead to the same 1RDM. From this relation it follows directly that the map $v \mapsto \g$ is bijective up to a constant in the potential and it gives a characterization of the set of $v$-representable 1RDMs. Additionally, for every potential the Helmholtz functional and the universal functional attain a minimum.
|
Subsets and Splits